Search

Electricity theft has emerged as a notable concern for smart grids, as fraudulent users illicitly access electricity from utilities without a contractual agreement or manipulate meter readings to evade bill payments. Consequently, the importance of electricity theft detection (ETD) in preserving the cost-effectiveness of smart grids has increased. In recent years, the utilization of deep learning models in electricity theft detection has surged in popularity, owing to their capacity to capture the periodic patterns within time-series data on electricity consumption. Nonetheless, deep learning techniques for electricity theft detection encounter several challenges, such as data imbalance, higher false positive rates, and the determination of periodic cycle timings. In this paper, we present an innovative approach to tackle these challenges: the Multi-Objective Evolutionary Hybrid Deep Learning (MOE-HDL) architecture, specifically designed to identify electricity theft. In MOE-HDL techniques, first, a novel multi-objective evolutionary algorithm based on ranking scheme (RMOE) is proposed to optimize architectural model parameters and the detection window size, with precision and recall as conflicting objectives. Next, an optimized dynamic detection window is generated to effectively capture the periodic patterns in consumption data and convert the one-dimensional data into a grid structure. Finally, a hybrid deep learning model is designed with the help of optimized model parameters to identify abnormal patterns in electricity consumption data. We conducted a comprehensive set of experiments, comparing our approach with eight baseline methods for electricity theft detection. The experimental results demonstrate that the proposed MOE-HDL approach outperforms these alternatives in terms of performance. • In this paper, a novel multi-objective evolutionary hybrid deep learning approach was developed for theft detection in smart grids. • Precision and recall are two conflicting yet effective objective functions designed to optimize the model parameters and detection window. • A dynamic detection window is obtained through optimization to transform electricity consumption data. • The hybrid deep learning model is designed according to the model parameters obtained through evolutionary multi-objective optimization. [ABSTRACT FROM AUTHOR]

The objective of this paper is to optimize the energy consumption and performance of fuel cell electric bicycles (FCEBs) under specific key input parameters. The paper applied an integrated method that includes an artificial neural network (ANN) and genetic algorithm (GA) to forecast and identify an optimal performance and energy consumption of FCEBs. The simulation model of FCEBs is established and simulated in MATLAB-Simulink environment to generate 1000 data points, that are used for training, validating, testing artificial neural network, the ANN architecture containing five input neurons, two hidden neurons and two output neurons, respectively. Furthermore, the GA is integrated to find the maximum performance and energy consumption once the ANN has exactly been trained. The study found that the FCEB configuration can achieve an effective performance at 30.3 km/h with required power of 210.4 W under speed level_5, radius of wheel 0.39 m, frontal area 0.423 m2, slope grade 0%. In order to validate and verify the simulated results, the experimental approach method was conducted in the same condition. The experimental results fit well with the simulated results in the same initial input parameters. • A method integrating the artificial neural network with genetic algorithm is applied. • The consumed power and fuel cell voltage are optimized by combining artificial neural network and genetic algorithm (ANN-GA). • The effective performance area of fuel cell electric bicycle under various a structure and operating parameters is identified. • Deep-learning results reflect major physical mechanism of fuel cell electric bicycle operation. [ABSTRACT FROM AUTHOR]

Timely monitoring of abnormal vibration of machinery in highly harsh environments is essential to ensure the safe operation of mechanical systems. This paper uses a self-charging power module to harvest energy to provide sustainable power for monitoring multidirectional ultra-low frequency mechanical vibrations. The power generation unit consists mainly of a spherical electromagnetic triboelectric hybrid nanogenerator (SETE-HNG). It is equipped with advanced sensing functions to support deep learning to identify signals with different directions of vibration, different frequencies, and different amplitudes of vibration, thereby enhancing the high-precision perception function. The accuracy of the prediction results is as high as 98.6622%, 100%, and 99.3333%, respectively. A Power Management Circuitry (PMC) has been meticulously crafted to maximize the utilization of vibrational energy. It efficiently charges a 40 mAh lithium polymer battery to 3.3 V in just 26 min, all without the requirement of an external power source. This advancement facilitates self-powered Global Positioning System (GPS) tracking of vibrational signals. Moreover, the stored energy is harnessed to energize a microcontroller and a low-power Bluetooth module. This enables real-time monitoring of mechanical vibrations via a mobile phone. The design presented in this paper is a testament to the potential of self-powered multidirectional mechanical vibration monitoring, contributing significantly to the safety and efficiency of mechanical systems. The design in this paper enables self-powered monitoring of abnormal vibrations in mechanical systems in extremely harsh environments, thus ensuring the safe operation of equipment. [Display omitted] • Large-scale vibration energy harvesting provides an innovative and efficient method. • SETE-HNG has multidirectional ultra-low frequency mechanical vibration monitoring and energy harvesting functions. • A power management circuit is designed to store energy efficiently. [ABSTRACT FROM AUTHOR]

To mitigate the impact of market uncertainty on trading investments, this paper proposes a forecasting and investing framework for crude oil market by integrating interval models and machine learning models. Firstly, natural language processing technique is employed to analyze text information from social and news media, enabling the capture of market and societal sentiment. Subsequently, deep learning models are integrated to combine sentiment data with other economic variables for more accurate predictions of crude oil prices. Furthermore, this paper introduces a trading strategy with interval constraints based on interval prediction models to reduce trading risk arising from the uncertainty of point forecasts in investments. Through trading simulations, it is discovered that employing the interval constrained strategy is more effective in reducing trading risk and enhancing investment returns compared to point-based trading strategies. This interval-based strategy offers a novel approach to mitigating investment risk in the crude oil market. • A novel prediction framework is proposed to integrate text data and interval data. • Combining point-valued and interval-valued deep learning models. • An interval-based trading strategy is developed. • The proposed trading strategy can reduce risk while improving investment returns. [ABSTRACT FROM AUTHOR]

• This paper focuses on daily natural gas consumption forecasting. • A novel hybrid model is developed for daily natural gas consumption forecasting. • The most recent datasets of four cities in three climate zones are collected. • An improved singular spectrum analysis is first proposed. In daily natural gas consumption forecasting, the accuracy of forecasting models is vulnerably affected by the noise data in historical time series. Singular spectrum analysis (SSA) is often introduced into hybrid models for denoising. However, as a deterministic-based algorithm, SSA does not give good results when a time series is contaminated with a high noise level. Considering this fact, this paper proposes an improved SSA (ISSA) that modifies the determination method of subseries selection in the reconstruction stage of SSA. Combining ISSA with long short-term memory (LSTM), a novel hybrid model, ISSA-LSTM, is thus developed. Additionally, for validating the robustness and superiority of ISSA-LSTM, the historical datasets of four representative cities located in three climate zones are collected as the training and testing datasets, and a comparison of ISSA-LSTM with five advanced models is performed. The results reveal that SSA would generate negative values when time series close to zero and the contribution of SSA in improving the forecasting accuracy of LSTM is insignificant. In contrast, ISSA avoids generating negative values and reduces the mean absolute range normalized error (MARNE) of LSTM by a range of 0.86–11.86%. Among the models, ISSA-LSTM achieves the best performance and its MARNEs for London (temperate zone), Melbourne (subtropical zone), Karditsa (subtropical zone), and Hong Kong (tropical zone) are 4.68%, 5.72%, 5.76%, and 14.10%, respectively. The MARNE of the tropical city is higher than that of others, which is caused by the complex natural gas consumption pattern of itself. [ABSTRACT FROM AUTHOR]

District heat load forecasting is a challenging task that involves predicting future heat demand based on historical data and various influencing factors. Accurate forecasting is essential for optimizing energy production and distribution in district heating systems. However, most existing forecasting models lack transparency and interpretability and fail to capture the spatial–temporal dependencies in the data. Moreover, they often require a large amount of annotated data for training, which can be costly and time-consuming to obtain. In this paper, we present a novel approach to district heat load forecasting, which involves predicting future heat demand based on historical data and various influencing factors. The proposed approach is based on an Active Graph Recurrent Network (Ac-GRN), which leverages the strengths of active deep learning and graph neural networks to capture the complex spatial–temporal dependencies in the data. The approach also provides explainability for its predictions by using correlation-based attribution methods. The active deep learning component can effectively select the most informative and representative samples from a large pool of data, reducing the frequency and cost of data collection and human effort. The graph neural network component can model both linear and nonlinear relationships among heat meters using bidirectional recurrent connections, enhancing the accuracy and robustness of the predictions. We conduct extensive experiments and compare our approach with eleven state-of-the-art models on a real-world dataset of district heating consumption data from Danish residential buildings. Our results show that our approach outperforms other models in terms of accuracy, robustness, reliability, and computational efficiency for multi-horizon multi-step district heat load forecasting. Our approach also provides meaningful explanations for its predictions by highlighting the most influential factors and heat meters for each prediction. This paper makes a novel contribution to district heat load forecasting with explainability. • Novel approach for district heat load forecasting with active deep learning and graph neural networks. • Explainable predictions using gradient-based attribution methods. • Outperforms state-of-the-art models in accuracy and robustness. • Evaluates on real-world district heating data and conducts ablation studies. [ABSTRACT FROM AUTHOR]

Accurate short-term forecasting of building cooling load (CLF) in an integrated energy system (IES) is essential for effective building energy management. However, the existing CLF models for IES often treat each building as an independent entity and neglect the spatiotemporal correlation among buildings. To address this research gap and achieve accurate CLF, this paper proposes a new hybrid deep learning model that considers spatiotemporal coupling. First, the coupled spatial–temporal features among different buildings were analyzed, and the meteorological factors were screened based on the Spearman's rank order correlation coefficient (SROCC). Second, synchrosqueezing wavelet denoising (SWT) was adopted to denoise the historical cooling load (CL) data, remove high-frequency noise, and improve data quality. Third, the TTGAT-GTC model was constructed for the CLF of an IES. A temporal trend-aware graph attention network (TTGAT) captured the spatial correlation of CL between buildings. A gated temporal convolution layer (GTC) was constructed to extract the trend in the dynamic temporal variation in historical load. Residual and skip connections were applied to avoid gradient disappearance and increase the computational efficiency of the model. To validate the effectiveness of the proposed SWT-TTGAT-GTC model, this paper compared the proposed model with four benchmark models using MAPE, RMSE, MAE, and R2. The experimental results showed that the performance of the proposed CL forecasting model is superior and that the proposed model appropriately introduces the spatio-temporal coupling information between buildings. • The spatiotemporal information between related buildings is appropriately considered in CL. • TTGAT captures the spatial correlation of cooling load between buildings. • GTC extracts the trend in dynamic temporal variation between buildings. • Residual and skip connections are used to avoid gradient disappearance. [ABSTRACT FROM AUTHOR]

This paper introduces and formalizes the concept of T-shape data, which arises in several engineering and natural contexts, where the initial data are richer and cover a wider range of operations than the data acquired in the following time. The specific context considered is the Li-ion battery experimental testing before the actual operation, where the T-shape data is cast as right-censored data in the application of battery remaining useful life (RUL) prediction. Existing RUL prediction techniques focused on RUL point estimation and provided rough calculations concerning the RUL distribution, thus are insufficient for battery degradation information. Based on the T-shape structure, this paper investigates a time-varying construction of the RUL probability density function. The proposed computationally efficient method, the multiple non-crossing quantiles Long Short-Term Memory (MNQ-LSTM), learns long-term dependencies among battery RUL and operation process quantities, including operating conditions. With several predicted quantile levels, the construction of the RUL conditional probability density function can capture the underlying survival distribution and statistical inferences of battery RUL with richer and more accurate information. Numerical results verify the performance of the proposed MNQ-LSTM with T-shape data. Compared to the conventional LSTM, Gaussian process regression, and a hybrid deep learning model of convolutional neural network and the bi-directional gated recurrent unit, the proposed model outperforms and can achieve the coefficient of determination up to 95.74% regarding point predictions. 100% testing results of battery RUL are not outside the range of 90% prediction intervals even in the worst case of 80% T-shape data. • Formalize a novel T-shape data concept that arises in several practical contexts. • Develop first time multiple quantile LSTM for probabilistic battery RUL prediction. • Introduce Kaplan–Meier weight for quantile LSTM model training with T-shape data. • Resolve crossing quantile problem without imposing sophisticated constraints. [ABSTRACT FROM AUTHOR]

• A method integrating the artificial neural network with genetic algorithm is applied. • The MATLAB-Simulink model of electric assisted bicycle is employed for data generation and validation. • The required power and electric consumption are optimized by combining artificial neural network and genetic algorithm (ANN-GA). • The effective performance area of electric assisted bicycle under various a structure and operating parameters is identified. • Deep-learning results reflect major physical mechanism of electric assisted bicycle operation. The purpose of this research is to study how the operating and structure parameter affect the dynamic, require power and electric consumption of electric assisted bicycle (EAB). The paper applied a combined an artificial neural network (ANN) with genetic algorithm (GA) method to predict the required power, electric consumption of EAB and find an effective performance area. The MATLAB-Simulink simulation model is established to create 1000 data point, that is applied in an ANN for training, testing and verifying. The ANN-GA method is applied to identify the optimized required power and electric consumption under four typical slope grades in the area of (0–0.65%), four typical wheel radius in the area of (0.3–0.0.39 m), four typical frontal areas in the range of (0.423–1.323 m2), four typical speed levels from speed level 1 to speed level 4. After the ANN is trained, it is applied into the genetic algorithm to identify the effective performance. The prediction results reflect the major physical mechanism that governs ES performance and are validated against the MATLAB-Simulink simulation model. Additionally, the electric assisted bicycle can reach an effective performance area at 27.16 km/h with the optimized required power of 545.5 W at slope grade of 0%, wheel radius of 0.39 m, speed level 4, frontal area of 0.423 m2. The results show that the ANN-GA method is suitable to identifying structure and operating parameters including various slope grades, frontal areas, wheel radius, speed level for optimizing effective performance of EAB, which contributes a helpful method for EAB design and control. Beside that, the experimental test was conducted on real road test at Taehwa river to verify the simulated results. The experiment results and simulation results have the same trend at the same conditions. (The short version of the paper was presented at ICAE2022, Bochum, Germany, Aug 8–11, 2022. This paper is a substantial extension of the short version of the conference paper.) [ABSTRACT FROM AUTHOR]

Power grid damage and blackouts are increasing with climate change. Load forecasting methods that integrate climate resilience are therefore essential to facilitate timely and accurate network reconfiguration during periods of extreme stress. Our paper proposes a generalised Wildfire Resilient Load Forecasting Model (WRLFM) to predict electricity load based on operational data of a Distribution Network (DN) in Australia during wildfire seasons in 2015–2020. We demonstrate that load forecasting during wildfire seasons is more challenging than during non-wildfire seasons, motivating an imperative need to improve forecast performance during wildfire seasons. To develop the robust WRLFM, comprehensive comparative analyses were conducted to determine proper Machine Learning (ML) forecast structures and methods for incorporating multiple factors. Bi-directional Gated Recurrent Unit (Bi-GRU) and Vision Transformer (ViT) were selected as they performed the best among all 13 recently trending ML methods. Multi-factors were incorporated to contribute to forecast performance, including input sequence structures, calendar information, flexible correlation-based temperature conditions, and categorical Fire Weather Index (FWI). High-resolution categorical FWI was used to build a forecasting model with climate resilience for the first time, significantly enhancing the average stability of forecast performances by 42%. A sensitivity analysis compared data set patterns and model performances during wildfire and non-wildfire seasons. The improvement rate of load forecasting performance during wildfire seasons was more than two times greater than in non-wildfire seasons. This indicates the significance and effectiveness of applying the WRLFM to improve forecast accuracy under extreme weather risks. Overall, the WRLFM reduces the Mean Absolute Percentage Error (MAPE) of the forecast by 14.37% and 20.86% for Bi-GRU and ViT-based models, respectively, achieving an average forecast MAPE of around 3%. • Novel climate-resilient load forecast model for use during extreme wildfire seasons. • Bi-GRU and ViT outperform 11 recently trending machine learning models. • Categorical FWI with high resolution enhances forecast stability by 42%. • Optimised calendar, temperature and FWI input reduce errors by 17.6%. • Forecast improvement during wildfires is twice that in non-wildfire seasons. [ABSTRACT FROM AUTHOR]

Currently, building energy consumption prediction models are usually based on a large amount of historical operational data in high demands of building operating hours and monitoring systems. However, many buildings may lack operational data due to relatively limited monitoring systems, causing the failure to use data-driven methods to characterize the energy profile. In this context, transfer learning is a promising method to establish the knowledge transfer between many high-quality building operation datasets and a small amount of target building data, and to help predict energy consumption in the target building. This paper studies the possibility of employing transfer learning to achieve both short and long-term building energy consumption prediction. Firstly, a similarity analysis method, based on variable modal decomposition and dynamic time warping, is proposed for identifying the source buildings with the most similar energy features to the target building. Then, transfer learning for long-term prediction air-conditioning energy consumption is developed with weather parameters generated by the Morphing method as inputs. For the short-term single-step prediction, the proposed model CV-RMSE improves 81.3% (AEC) and 77.4% (EEC), respectively, compared to the prediction model that does not implement the transfer learning strategy and directly uses the target BEC data. As for the short-term multi-step prediction, the proposed model CV-RMSE improves 62.0% (AEC) and 65.5% (EEC), respectively. For the long-term prediction, the average CV-RMSE for the whole year is 6.62% and 11.15% for the proposed and directly target domain-based model, respectively. The proposed method explores the practicality of transfer learning in building energy forecasting, contributing to the use of existing building operation data for energy management at different timespan. • A similarity analysis method for building energy consumption data based on VMD-DTW is proposed. • The Seq2Seq-Transformer model is applied to short-term multi-step prediction of building energy consumption. • Transfer learning idea is first introduced to the long-term building energy consumption prediction. [ABSTRACT FROM AUTHOR]

Accurate solar irradiance forecasts can make solar power forecasts more reliable, which can help the power grid dispatch reasonably. In practice, Numerical Weather Prediction (NWP) is widely applied in solar irradiance forecasts. In this paper, the statistical NWP Global Horizontal Irradiance (GHI) error analysis shows that the characteristics of NWP GHI error vary obviously under different weather conditions. However, existing correction methods are not designed contrapuntally for different weather conditions, resulting in poor correction performance. To solve this problem, a hybrid method is proposed to get day-ahead correction results for NWP GHI. Specifically, the hybrid method consists of three key parts, including Deep Clustering (DC), Variational Mode Decomposition (VMD), and an Encoder–Decoder based Correction model (EDC). DC is used to categorize the historical samples into three clusters, the input of which is the multi-dimensional information series containing observed data and NWP data. After a feature selection by VMD, the correction models are trained on each cluster respectively. The performance of the proposed model is evaluated with a public dataset and an actual field dataset, and the results demonstrate that the accuracy has been effectively improved by the proposed method compared with other models. In addition, we find that adopting VMD is effective in improving the correction accuracy, and the root mean square error is reduced by 5.70% and 9.32% compared with models without it. • Propose a novel hybrid model for day-ahead NWP solar irradiance correction. • NWP GHI error analysis shows it varies obviously under different weather conditions. • Jointly optimize the feature learning and clustering assignment by DC. • CLSTM is applied to learn effective features of multi-dimensional input data. • Remove redundant information of solar irradiance by VMD to reduce correction error. [Display omitted] [ABSTRACT FROM AUTHOR]

With the rapid growth of wind power penetration, its inherent stochasticity and uncertainty will seriously affect the stable operation of power systems. How to effectively characterize the uncertainty of wind power is a great challenge for day-ahead power system dispatching, scenario generation is an important method to describe the uncertainty of wind power. Currently, most of the wind power scenarios are generated using a generative adversarial network with two-dimensional convolution as the main structure, which may make it difficult to adequately characterize the temporal features, the day-ahead mode properties, and seasonality of wind power. In this paper, we first establish an auxiliary classification time-series generation adversarial network based on error stratification, construct the numerical characteristic conditional labels that can reflect the fluctuation characteristics of day-ahead wind power and power output level, and design the temporal embedding function that captures the seasonal characteristics of wind power. On this basis, to fully extract the dynamic variation characteristics and global effective information of wind power prediction error sequences, Informer is combined with a time-series generative adversarial network, and a joint loss function incorporating supervised learning and unsupervised learning is constructed. Subsequently, the generated set of prediction error sequences is superimposed with the day-ahead predicted value of wind power to obtain the day-ahead wind power scenario set. Finally, to verify the effectiveness of the proposed method, two datasets from different geographic locations are used to comprehensively evaluate the generated day-ahead wind power scenario set in terms of three aspects: temporal correlation characteristics, fluctuation characteristics, and accuracy. The experimental results indicate that the scenario generation method proposed can improve the quality of the day-ahead wind power scenario set and has an excellent performance in describing wind power uncertainty compared with other methods. • A novel generation model for wind power based on Probspare self-attention and TimeGANs is developed. • A method for generating day-ahead wind power scenarios considering numerical characteristics of prediction errors is proposed. • A joint loss function is designed by combining auxiliary classification supervised training and unsupervised adversarial training. [ABSTRACT FROM AUTHOR]

A large part of French electricity consumption variation is due to temperature fluctuations. While HVAC (heating, ventilation and air-conditioning) systems consumption are directly affected by the temperature, other systems (refrigerator, freezer, water heater) can also be driven by weather changes making thermal contribution to overall consumption difficult to extract. This paper presents a "by-design" unsupervised data-driven method to separate the consumptions due to the weather in the overall electricity consumption. The proposed deep-learning model is based on the separation of meteorological parameters from calendar ones within the model architecture. The performances of this model, in particular its ability to split consumption mechanisms, is tested on a synthetic dataset and on the french consumption dataset. Being relatively simple and interpretable, this approach can be generalized to other countries whereasenergy sobriety represents an important challenge we are facing. • The hourly power load of a country due to temperature is unsupervisingly identified. • Deep learning model provides Large Scale Non Intrusive Load Separation. • Engineering-driven neural network design separates weather and calendar consumption. • Model's performance is assessed on country's actual electricity consumption. [ABSTRACT FROM AUTHOR]

Electric Vehicles (EVs) and their charging infrastructure have witnessed rapid adoption into the smart grid owing to their ability to reduce the transportation sector's emissions. However, this hasty deployment has left the grid prone to attacks initiated through the charging ecosystem. This paper develops a family of robust dynamic load-altering attacks that manipulate the EV load to destabilize the power grid. The attack strategy is formulated based on feedback control theory and Linear Matrix Inequalities (LMIs) and incorporates mathematical modeling of the power grid uncertainties. The effectiveness of our proposed robust EV attack strategy is demonstrated through extensive simulations of attack scenarios against the New England (NE) 39-bus grid. The results demonstrate that the proposed attacks were able to destabilize the grid frequency inducing a deviation of 1.5 Hz and forcing the generators to trip after compromising as little as 6.6% of the total grid load. We also present a two-tier Deep Learning (DL) attack detection mechanism. We optimize this mechanism to achieve optimal results against dynamic and switching attacks. The proposed mechanism was evaluated against several other models and proved to be superior with an accuracy of 99.9%. • We propose a robust EV Dynamic Attack Design that takes advantage of compromised EVs to cause power grid instability. • The controller design accounts for several uncertainties that can impact the attack success. • We perform extensive time domain simulations to demonstrate the devastating impact of such attacks even in the presence of uncertainties. • The proposed scheme demonstrates its ability to cause generator tripping even with just 6.7% of the total grid load. • We then propose a two-tier detection mechanism that achieved a detection accuracy above 99.9%. [ABSTRACT FROM AUTHOR]

Highlights • Multiple optimization targets are considered for a generalized energy Internet. • Power supply-demand balance is realized in the entire energy Internet system. • A model-free approach is applied with real-world power data. • Advanced deep reinforcement learning approach is applied. Abstract This paper investigates the energy management problem in the field of energy Internet (EI) with interdisciplinary techniques. The concept of EI has been proposed for a while. However, there still exist many fundamental and technical issues that have not been fully investigated. In this paper, a new energy regulation issue is considered based on the operational principles of EI. Multiple targets are considered along with constraints. Then, the practical energy management problem is formulated as a constrained optimal control problem. Notably, no explicit mathematical model for power of renewable power generation devices and loads is utilized. Due to the complexity of this problem, conventional methods appear to be inapplicable. To obtain the desired control scheme, a model-free deep reinforcement learning algorithm is applied. A practical solution is obtained, and the feasibility as well as the performance of the proposed method are evaluated with numerical simulations. [ABSTRACT FROM AUTHOR]

Highlights • Review of the application of reinforcement learning (RL) for demand response (DR) • DR is relevant for integrating renewable energy source into the smart grid. • Considering RL/DR from energy generation and storage, to supply and user satisfaction. • Typical algorithms and open research directions are discussed. • Most articles focus on single-agent systems in stationary environments. Abstract Buildings account for about 40% of the global energy consumption. Renewable energy resources are one possibility to mitigate the dependence of residential buildings on the electrical grid. However, their integration into the existing grid infrastructure must be done carefully to avoid instability, and guarantee availability and security of supply. Demand response, or demand-side management, improves grid stability by increasing demand flexibility, and shifts peak demand towards periods of peak renewable energy generation by providing consumers with economic incentives. This paper reviews the use of reinforcement learning, a machine learning algorithm, for demand response applications in the smart grid. Reinforcement learning has been utilized to control diverse energy systems such as electric vehicles, heating ventilation and air conditioning (HVAC) systems, smart appliances, or batteries. The future of demand response greatly depends on its ability to prevent consumer discomfort and integrate human feedback into the control loop. Reinforcement learning is a potentially model-free algorithm that can adapt to its environment, as well as to human preferences by directly integrating user feedback into its control logic. Our review shows that, although many papers consider human comfort and satisfaction, most of them focus on single-agent systems with demand-independent electricity prices and a stationary environment. However, when electricity prices are modelled as demand-dependent variables, there is a risk of shifting the peak demand rather than shaving it. We identify a need to further explore reinforcement learning to coordinate multi-agent systems that can participate in demand response programs under demand-dependent electricity prices. Finally, we discuss directions for future research, e.g., quantifying how RL could adapt to changing urban conditions such as building refurbishment and urban or population growth. [ABSTRACT FROM AUTHOR]

Highlights • A deep recurrent neural network (RNN) model is used for medium-term thermal load prediction. • The deep RNN model outperforms a simple multilayer perceptron for time series forecasting. • An optimization framework is proposed for sizing a thermal storage tank to meet thermal loads. • The optimization method can be used in selecting a thermal storage tank for use with a building. • The combined predictions and optimization allow for estimating performance of the tank. Abstract This paper evaluates the performance of deep recurrent neural networks in predicting heating demand for a commercial building over a medium-to-long term time horizon (⩾ 1 week), and proposes a modeling framework to demonstrate how these longer-term predictions can be used to aid design of a stratified thermal storage tank. The building sector contributes significantly to primary energy consumption in the US, and as such, there is a need to predict heating demand in buildings over longer time horizons, and to develop methods that can facilitate installation, planning and management of distributed generation and thermal storage to meet these heating demands. Key objectives of this paper are: (a) Investigate how a deep recurrent neural network model performs in predicting heating demand in campus buildings at University of Utah over multiple weeks, and (b) develop an optimization framework that which can provide definitive guidelines on sizing a stratified thermal storage tank without requiring high performance computing resources. The results showed that the predictions by the deep RNN are comparatively more accurate than those by a 3-layer MLP, and that these deep RNN predictions can adequately serve as proxy for future demand while considering sizing in the design of a complementary stratified thermal storage tank. [ABSTRACT FROM AUTHOR]

This paper presents a model for the realistic planning of multi-microgrids in the context of Active Distribution Networks with the assistance of Geographic Information Systems. The model considers the distribution system grid as well as the geographic features of the Region of Interest. It also includes long-term purchase decisions and short-term operational constraints, and considers uncertainties associated with electricity demand and Renewable Energy Resources using an existing Two-Stage Stochastic Programming approach. Geographic Information Systems along with Deep Learning are used to estimate the areas of rooftops within the Region of Interest and model the Low Voltage grid. The planning model is used to study the feasibility of implementing a multi-microgrid system consisting of 4 individual microgrids at an Active Distribution Network in a municipality in the state of São Paulo, Brazil. The results of the model presented in this paper are compared with the results obtained using Monte Carlo Simulations and an existing, less detailed, Two Stage Stochastic model. It is demonstrated that the stochastic solutions are close to those obtained with Monte Carlo at a lower computational cost, and that the use of Geographic Information allows to determine both the capacity and location of the PV panels, batteries, and distribution transformers on the microgrids grid, thus providing more precise and useful planning results. • Geographic Information (GI) -based model for the planning of Multi-microgrids (MMGs). • An existing Two-Stage Stochastic Programming (TSSP) model is used. • Uncertainties in demand and solar generation are considered for study and analysis. • GI allows the modeling of the Low Voltage (LV) distribution network. • Results show where rooftop solar panels and batteries can be physically deployed. [ABSTRACT FROM AUTHOR]

• This paper proposes a point classification method for PQDs based on DAOM. • Hilbert transform (HT) based local feature attention mechanism (LFAM) is proposed. • The classification model optimized-CAM is realized by squeeze-and-excitation network. • The proposed method can identify the type of each sampling point. The rapid development of the power grid makes power quality disturbances (PQDs) more complex. Accurately classifying PQDs and measuring the duration of each disturbance element are both crucial for studying the causes of PQDs and subsequent countermeasures. To satisfy the requirements, a novel dual-attention optimization model (DAOM) is presented in this paper for points classification. First, the local feature attention mechanism (LFAM) based on Hilbert transform (HT) is proposed in the input layer to enhance the local features of PQDs. Subsequently, on the basis of convolutional neural network (CNN), a channel attention mechanism (CAM) based on squeeze-and-excitation network (SENet) is introduced to each convolutional layer to achieve the purpose of automatically learning the importance of each channel feature. Finally, each sampling point is classified in the form of multiclass-multioutput. The proposed model is built through the Keras framework, and a synthetic database containing 49 types of PQDs is established to test the model. The proposal achieves an average classification accuracy of 98.95% in a 30 dB white noise environment, which is more precise and robust than other deep learning-based models. At the same time, the proposed model consistently performs best when evaluated through imbalanced data classification metrics and nonparametric test results. Unlike traditional methods, the proposal can accurately identify the occurrence and end time of each element in complex PQDs. Moreover, through a hardware experiment based on the AC source, the proposed model achieves an average accuracy of 98.30%, which is ahead of other comparison models as well. [ABSTRACT FROM AUTHOR]

This paper proposes a novel method namely WaveletKernelNet-Convolutional Block Attention Module-BiLSTM for intelligent fault diagnosis of drilling pumps. Initially, the random forest method is applied to determine the target signals that can reflect the fault characteristics of drilling pumps. Accordingly, the WaveletKernelNet-Convolutional Block Attention Module Net is constructed for noise reduction and fault feature extraction based on signals. The Convolutional Block Attention Module embedded in WaveletKernelNet-CBAM adjusts the weight and enhances the feature representation of channel and spatial dimension. Finally, the Bidirectional Long-Short Term Memory concept is introduced to enhance the ability of the model to process time series data. Upon constructing the network, a Bayesian optimization algorithm is utilized to ascertain and fine-tune the ideal hyperparameters, thereby ensuring the network reaches its optimal performance level. With the hybrid deep learning model presented, an accurate fault diagnosis of a real five-cylinder drilling pump is carried out and the results confirmed its applicability and reliability. Two sets of comparative experiments validated the superiority of the proposed method. Additionally, the generalizability of the model is verified through domain adaptation experiments. The proposed method contributes to the safe production of the oil and gas sector by providing accurate and robust fault diagnosis of industrial equipment. • A novel intelligent drilling pumps fault diagnosis method, WCCN-BiLSTM, is proposed. • The WCCN is constructed for noise reduction and feature extraction of signals • The CBAM module embedded in WCCN enhances the feature representation of channel and spatial dimension. • The diagnostic ability and generalization ability of the model are verified by drilling pump data under different working conditions. [ABSTRACT FROM AUTHOR]

The fast proliferation of electric vehicles (EVs) profoundly affects the stable operation of the power system. Since EV charging demand is affected by users volatile behaviors and many other environmental and social factors, their uncertainty is quite pronounced. Traditional deterministic forecasting methods which provide little information on the uncertainty may not be informative to the decision-making of power system with massive EVs. This paper proposes a novel feature-enhanced deep learning method (FEDM) for EV charging demand probabilistic forecasting of individual station, in which a two-stage feature selection model is constructed to introduce the Pearson correlation coefficient between charging demand and external features as a prior knowledge to reweight the initial weights. The performance is improved compared to which is without feature-enhanced. In the comparison with eight benchmarks, FEDM shows superiority in both deterministic and probabilistic evaluation metrics. With high computational efficiency, FEDM has low dependence on the historical data stock, which can still achieve good performance in the case of low data volume. It provides a fast and accurate forecasting scheme for the demand forecasting of individual stations under the limitation of low data stock. • The proposed approach improves the accuracy rate, qualified rate, sharpness and computational efficiency. • A novel feature enhanced mechanism is proposed to introduce the correlation coefficient as the prior knowledge of the neural network. • The proposed approach can provide an excellent forecasting performance in the constraints of low data volume. [ABSTRACT FROM AUTHOR]

Energy storage systems enable balancing supply and demand and facilitate the integration of intermittent renewable energy sources. In particular, latent heat thermal energy storage units are attractive for deployment in thermal systems due to their high energy density. However, knowledge of the state-of-charge of a thermal store is crucial to effectively regulate its charging and discharging cycles. To achieve this, continuous-time non-linear observers may be employed to estimate the state-of-charge at the expense of a high computational cost. The emergence of artificial intelligence solutions may be helpful to reduce computational burden, but their adoption for thermal stores has been limited. This paper bridges this research gap by presenting a novel approach to predict the state-of-charge of a latent heat unit. It employs a discrete-time estimator based on a recurrent neural network, which is based on a long short-term memory structure and the regression method for estimation. The estimator offers a reduced computation time for state-of-charge estimation and allows flexible sampling adjustments without sacrificing accuracy. Additionally, the presented approach simplifies data collection by independently handling charging and discharging processes through internal state resets. The estimator, trained using MATLAB's deep learning toolbox, uses a dataset comprising various operating conditions obtained from simulations of a physics-based model. When compared against a more traditional state-of-charge estimation method using a discrete-time non-linear observer, the advantages of a recurrent neural network-based estimator are evidenced, highlighting its potential for practical applications. The presented method exhibited high accuracy with a maximum root mean square error under 0.73% and a mean absolute error below 0.41% with respect to direct state-of-charge calculation. Although the discrete-time non-linear observer exhibited a marginal higher accuracy, the recurrent neural network-based estimator achieved a significant improvement in computational efficiency. These findings make the proposed approach a robust tool facilitating enhanced control strategies, optimised energy management, and increased overall thermal system performance. • Highly accurate AI-based SoC estimator for thermal stores. • RNN training framed within a regression problem to predict SoC. • Implementation of the SoC estimator through simple matrix operations. • Significantly reduced computational resource compared to other methods. [ABSTRACT FROM AUTHOR]

This paper presents a recurrent neural network model to make medium-to-long term predictions, i.e. time horizon of ⩾ 1 week, of electricity consumption profiles in commercial and residential buildings at one-hour resolution. Residential and commercial buildings are responsible for a significant fraction of the overall energy consumption in the U.S. With advances in sensors and smart technologies, there is a need for medium to long-term prediction of electricity consumption in residential and commercial buildings at hourly intervals to support decision making pertaining to operations, demand response strategies, and installation of distributed generation systems. The modeler may have limited access to information about building’s schedules and equipment, making data-driven machine learning models attractive. The energy consumption data that is available may also contain blocks of missing data, making time-series predictions difficult. Thus, the main objectives of this paper are: (a) Develop and optimize novel deep recurrent neural network (RNN) models aimed at medium to long term electric load prediction at one-hour resolution; (b) Analyze the relative performance of the model for different types of electricity consumption patterns; and (c) Use the deep NN to perform imputation on an electricity consumption dataset containing segments of missing values. The proposed models were used to predict hourly electricity consumption for the Public Safety Building in Salt Lake City, Utah, and for aggregated hourly electricity consumption in residential buildings in Austin, Texas. For predicting the commercial building’s load profiles, the proposed RNN sequence-to-sequence models generally correspond to lower relative error when compared with the conventional multi-layered perceptron neural network. For predicting aggregate electricity consumption in residential buildings, the proposed model generally does not provide gains in accuracy compared to the multi-layered perceptron model. [ABSTRACT FROM AUTHOR]

Few-Shot Short-Term Wind Power Forecasting (FS-STWPF) is designed to develop accurate short-term wind power forecasting models with limited training data, reducing the losses suffered by wind farms and power systems due to the data scarcity. Based on the idea of extracting valuable knowledge from the source wind farms and then applying it to the target wind farm, a novel Meta-Learning approach (WG-Reptile) has been proposed in this paper. Building on the existing Reptile algorithm, two specific designs have been made in WG-Reptile for FS-STWPF: (1) Within-Task Samples Assignment method based on Operational Scenario (WTSAOS) has been proposed to improve the adaptability of the models to changing conditions. (2) Gradients Conflict Attenuation method based on Cosine Similarity (GCACS) has been proposed to enhance the effect of knowledge fusion from different source wind farms. Two open wind power forecasting datasets and three deep learning models have been used to implement 24-h-ahead FS-STWPF experiments with different amounts of training data. The results illustrate that the proposed WG-Reptile is able to outperform the other few-shot learning approaches. Intuitively, with only 30-day training data, the accuracy of the proposed WG-Reptile can be equivalent to the conventional supervised learning approaches trained on 6-month. • A novel Meta-Learning approach, WG-Reptile, has been proposed to perform FS-STWPF, which is able to obtain accurate STWPF models with a limited amount of training data. • WTSAOS (Within-Task Samples Assignment method based on Operational Scenario) has been devised to reasonably assign within-task training samples to different updating steps in Inner Loop , which is able to improve the applicability of the models to the changing conditions. • GCACS (Gradients Conflict Attenuation method based on Cosine Similarity) has been devised to alleviate gradient conflicts by scaling the outer learning rates for different layers of the models, which is able to filter valuable gradients to enhance the effectiveness of the knowledge extraction. [ABSTRACT FROM AUTHOR]

As the significance of participation in the Day-Ahead Market (DAM) for stakeholders managing the charging of Electric Vehicle (EV) fleets increases, the necessity for precise EV Load Curve (EVLC) forecasting emerges as crucial. This paper presents an extensive investigation of nine diverse EVLC forecasting methodologies, encompassing statistical, machine learning, and deep learning techniques. These methodologies are evaluated on four public, real-world EV datasets, keeping in line with the specific forecasting horizon required for DAM. The study incorporates models with and without online historical data, ensuring broad applicability across varied data availability scenarios. An exploration of seasonal variations in forecasting performance is conducted via one year-long rolling simulations, providing deep insight on seasonal patterns. Furthermore, a detailed methodology for constructing EVLCs from session-based, tabular EV datasets is presented. The conducted research establishes a first-of-its-kind comprehensive comparison of EVLC forecasting methodologies for the real-world challenge of DAM participation. The research results provide indispensable guidance to wholesale electricity market participants, namely suppliers and electric vehicle aggregators, advancing the understanding in the field and significantly contributing to optimization of DAM participation for EV fleets. • Evaluation of forecasting in specific horizon required for DAM participation. • Nine diverse EV Load Curves forecasting methods. • Forecasts are evaluated on four public, real-world EV datasets. • Evaluation for a year-long period and diverse characteristics of EV charging data. • Methodology for constructing EV Load Curve from session-based tabular EV dataset. [ABSTRACT FROM AUTHOR]

Deep learning (DL) is effective in identifying the dominant instability mode (DIM) of power systems. However, regular supervised learning for training DL models requires a large number of training samples with accurate labels provided by power system experts, which is prohibitively costly in practice. To address this issue, this paper proposes a weakly supervised learning framework to train DL models for DIM identification based on cheap but potentially inaccurate (noisy) labels from non-expert engineers. The framework comprises two stages to mitigate the detrimental effects of label noise. At Stage I, an auxiliary model is proposed to intelligently detect detrimental noisy samples while preserving truly-labeled informative hard samples based on the entire training loss dynamics of base DL models. Then Stage II incorporates virtual adversarial training to utilize all samples, including noisy ones with labels removed, to train a smooth DIM identification model in a semi-supervised learning way. It can help further mitigate the effects of undetected label noise. Case studies are conducted on CEPRI 36-bus system and Northeast China Power System (2131 buses). The results verify that the proposed framework can tolerate high-intensity feature-(in)dependent label noise and build reliable DL models for DIM identification with significantly less reliance on experts. • Robust deep learning for power system dominant instability mode identification. • Two-stage weakly supervised learning to mitigate label noise from non-experts. • Noisy samples are detected while hard ones are preserved. • Virtual adversarial training is proposed to use noisy samples as unlabeled ones. • Intelligent dominant instability mode identification becomes more affordable. [Display omitted] [ABSTRACT FROM AUTHOR]

Lithium-ion batteries age continuously during usage due to their characteristics and the influence of various external factors, but as degradation deepens, it can lead to an apparent decrease in battery safety and reliability. Therefore, predicting remaining useful life from the current to end state and preventing possible dangerous incidents are essential for battery health management. A novel method from the perspective of providing aging features reference and balance prognostics speed and precision is proposed. Firstly, the used dataset is preprocessed in many ways. Then, through feature engineering, 79 features are extracted in the dataset from three perspectives: direct, evolution, statistics, and most representative 16 features are filtered to form the final feature set based on the method of combining discretization and importance. Feature engineering is established to find crucial information within the dataset that correlates highly with remaining life and represents most battery recession paths. Finally, the sparse autoencoder and Transformer integrated approach is proposed to build the life prediction model, and it can learn the temporal relationship between the feature set and remaining cycles quickly and accurately. According to the forecasting results of 62 batteries with different aging conditions, the error can reach 7.43% by only using the feature sequence under 30 cycle lengths. For early life prediction, the error of using the first 100 cycles data is at most 2.6%. • A novel aging feature filtering approach is proposed based on feature engineering. • A prediction model is established by using a sparse autoencoder and Transformer. • More accurate RUL and early life of a battery can be predicted by using fewer cycle data. • After the feature engineering presented in this paper, the final aging feature set selected has strong applicability. [ABSTRACT FROM AUTHOR]

The virtual power plant (VPP) that aggregates demand-side resources, is a new type of entity to participate in the electricity market and demand response (DR) program. Accurate customer baseline load (CBL) estimation is critical for DR implementation, especially the financial settlement in incentive-based DR. However, this is a challenging task as CBLs cannot be measured and are not equal to actual loads when DR events occur. Moreover, VPPs with different aggregation scales form heterogeneous electricity customers, which increases the difficulty of CBL estimation. In order to address this challenge, this paper proposes a novel deep learning-based CBL estimation method for varied types of electricity customers with different load levels. Specifically, we first transform the CBL estimation problem into a time-series missing data imputation issue, by regarding actual load sequences as CBL sequences with missing data, during DR periods. Then, we propose an attention mechanism-based neural network model to learn load patterns and characteristics of various CBLs, and also create the DR mask to avoid the disturbance of actual loads of DR periods on CBL's normal pattern. Further, we develop the generative adversarial networks (GAN)-based data imputation framework to produce the corresponding complete CBL sequence according to the actual load sequence, and then recover the missing values accordingly. Finally, comprehensive case studies are conducted based on public datasets, and our proposed method outperforms all benchmarks, where the mean and standard deviation of its estimation percentage error are 5.85% and 1.74%, respectively. This validates the effectiveness and superiority of the proposed method. • The customer baseline load estimation for virtual power plants in demand response. • Proposed baseline load estimation method is generalized for heterogeneous customers. • An attention mechanism-based model is proposed to learn customer load characteristic. • A generative adversarial networks framework is proposed to improve the robustness. [ABSTRACT FROM AUTHOR]

Wind power serves as a clean and sustainable form of energy. However, its generation is fraught with variability and uncertainty, owing to the stochastic and dynamic characteristics of wind. Accurate forecasting of wind power is indispensable for the efficient planning, operation, and grid integration of wind energy systems. In this paper, we introduce a novel forecasting method termed Temporal Collaborative Attention (TCOAT). This data-driven approach is designed to capture both temporal and spatial dependencies in wind power generation data, as well as discern long-term and short-term patterns. Utilizing attention mechanisms, TCOAT dynamically adjusts the weights of each input variable and time step based on their contextual relevance for forecasting. Furthermore, the method employs collaborative attention units to assimilate directional and global information from the input data. It also explicitly models the interactions and correlations among different variables or time steps through the use of self-attention and cross-attention mechanisms. To integrate long-term and short-term information effectively, TCOAT incorporates a temporal fusion layer that employs concatenation and mapping operations, along with hierarchical feature extraction and aggregation. We validate the efficacy of TCOAT through extensive experiments on a real-world wind power generation dataset from Greece and compare its performance against twenty-two state-of-the-art methods. Experimental results demonstrate that TCOAT outperforms existing methods in terms of both accuracy and robustness in wind power forecasting. Moreover, we conduct a generality study on an additional real-world dataset from a different climate condition and wind power characteristics. The results show that TCOAT can achieve comparable or better performance than the state-of-the-art methods, confirming the generalization ability of TCOAT. • Novel deep learning model for wind power forecasting • Temporal collaborative attention and fusion for wind data • Superior performance on real-world wind power dataset • Attention maps for interpretability and end-to-end training [ABSTRACT FROM AUTHOR]

• A co-simulation platform for Hybrid Electric Vehicles is developed for virtual assessment of Energy Management, performance, CO 2 & pollutant emissions. • CO 2 & pollutant emissions models are based on convolutional neural network for accurate transient state prediction. • A continuous deep reinforcement learning based EMS of HEV is proposed. • DDQL is trained with random generated driving cycles based on Markov Chain approach from real recordings. • Cumulative d error for predicted pollutants emissions is below 8.5% while HEV simulation is running up to 10 times faster than real time in a standard laptop. The Energy Management Strategy (EMS) in an HEV is the key for improving fuel economy and simultaneously reducing pollutant emissions. This paper presents a methodology for developing hybrid models that enable EMS testing as well as the evaluation of fuel consumption, CO 2 and pollutant emissions (CO, NO x and THC). In this context, pollutant emissions are hard to quantify with static models such as the well-known map-based approach which is mainly due to the pronounced impact of transient effects. The novelty of this paper primarily comes from the characterization of pollutant emissions through Convolutional Neural Networks (CNN), providing high accuracy for both, instantaneous and cumulative values. The input parameters are classical Internal Combustion Engine (ICE) measurements such as engine speed, air mass flow, torque and exhaust temperature. The proposed CNNs are reduced to a minimum for low complexity and fast computability. These models are developed with experimental data from chassis dyno testing of a conventional turbo-charged gasoline engine vehicle. The pollutant emission models are used in conjunction with physical models of the remaining powertrain allowing for real time simulations of the complete HEV vehicle. The Double Deep-Q learning algorithm is proposed for the EMS and compared to the Dynamic programming (DP) solution. The introduced methodology is developed in a co-simulation framework between MATLAB-Simulink and AMESIM. The resulting model runs between 8 and 10 times faster than real time in an off-the-shelf PC. This provides the capability for developing models suitable for HIL (hardware-in-the-loop) and SIL (software-in-the-loop) applications. The final error in predicted CO 2 remains below 2.5% while the final cumulative error for pollutants is below 8.5% in the case of CO and HC emissions. [ABSTRACT FROM AUTHOR]

The State-of-Health (SOH) estimation of Lithium-ion (Li-ion) batteries is critical for the safe and reliable operation of the batteries. Deep learning technologies are currently the popular methods for SOH estimation due to the advantages of no modeling and automatic feature extraction. However, existing methods require a large amount of annotated data to ensure model fitting, and the collection and labeling of battery aging data are time-consuming and laborious. Therefore, a self-supervised framework incorporating weak labels (SSF-WL) is proposed in this paper to obtain excellent estimation results on a small amount of annotated data. First, a novel data processing method based on the Gramian angular field, difference calculation, and raw data is proposed to enrich information and enhance features. Then, a five-layer Transformer encoder is constructed in SSF-WL for feature extraction. Finally, the model is pre-trained and fine-tuned on the proposed SSF-WL to obtain the estimated results of SOH. The proposed method is validated on the 124 commercial battery and Oxford databases. Experiments indicate that when using only 30% of the annotated training data, the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) obtained by SSF-WL are 0.5219%/0.6085% lower than traditional supervised learning on the 124 commercial battery database, respectively. Moreover, the SSF-WL pre-trained model on a large unannotated database can be transferred to different types of batteries with a small annotated database and obtains on-par or better estimation results than the model trained from scratch. • A data processing method is proposed for feature enhancement. • A five-layer ViT encoder is constructed to improve the estimation performance. • A self-supervised framework is proposed to reduce the model's reliance on data. • The framework enables performance transfer between different types of batteries. [ABSTRACT FROM AUTHOR]

With the rise in household computing power and the increasing number of smart devices, more and more residents are able to participate in demand response (DR) management through the home energy management system (HEMS). However, HEMS has encountered challenges in developing the most effective energy management strategies, including the complexity of modeling user comfort, uncertainty in electricity price and photovoltaic (PV) output, and the challenge of solving high-dimensional time-coupled decision problems. To address these challenges, a novel data-driven deep reinforcement learning (DRL)-base HEMS optimization framework considering uncertain household parameters is proposed. Firstly, a thermal comfort evaluation model based on integrated learning is proposed. Then, a prediction model based on the bidirectional gated recurrent unit neural network (BiGRU-NN) algorithm is proposed to mine the time series PV output and electricity price data. Finally, combining the PV output and electricity price forecasting, along with the thermal comfort evaluation, an optimal decision-making method based on soft actor-critic (SAC) algorithm for the HEMS is established. The results of numerical experiments show that the proposed method can effectively solve the high-dimensional integrated decision-making problem with uncertainty. By participating in DR, the household electricity cost can be reduced by 17.7% and the total cost can be reduced by 8.4%. Furthermore, the comparison result shows that the method proposed in this paper performs better than the existing optimization models based on proximal policy optimization (PPO) algorithm and twin-delayed depth deterministic policy gradient (TD3) algorithm. [Display omitted] • Mathematical models for different household appliances are established. • A thermal comfort evaluation model based on the XGBoost algorithm is designed. • A prediction model based on BiGRU-NN algorithm with bidirectional information. • An optimal decision-making method based on soft actor-critic algorithm is proposed. [ABSTRACT FROM AUTHOR]

This paper reports a new knowledge-driven engineered feature learning approach in response to the Global AI Challenge for Building E&M Facilities held by the Electrical and Mechanical Service Department (EMSD) of the Hong Kong SAR. The results were awarded with a Grand Prize by the competition organizer. A dynamically engineered multi-modal feature learning (DEMMFL) method is proposed for predicting the cooling load of two office buildings. The DEMMFL model is estimated with the Lasso-ridge regression and compared with other well-known methods such as the Lasso. The novel approach applies control system knowledge to engineer useful features and explore load patterns for multi-mode modeling. Deep learning methods including LSTM, GRU, and AutoGluon are implemented for automated machine learning and tested in parallel to compare the performance of the proposed model with existing methods. The proposed model is demonstrated to predict long-term cooling load most accurately using engineered features from weather information only. • A new dynamic feature engineering method for office building cooling load prediction. • Slow and fast dynamic features engineered from available weather data only. • Control-knowledge based feature to learn on-hours dynamic responses. • Multi-modal approach with eight day-type features to learn occupancy loads. • Deep recurrent learning and AutoML methods compared to the proposed method. [ABSTRACT FROM AUTHOR]

Integrated energy system (IES) is an important way for energy structure transition and development. Considering the characteristics of large data volume, strong randomness, and multi-energy coupling of IES, this paper proposes a novel short-term multi-energy load forecasting method for IES based on feature separation-fusion technology and improved CNN. Firstly, based on the distribution pattern of pixels in static images, the irregular multi-energy load is reconstructed into 3D load pixel matrix, giving them certain correlation features in both horizontal and vertical directions respectively. Secondly, the feature separation-fusion technology is employed to differentially process distinct features based on their information value differences. Finally, the extracted features are combined and input into a multi-task learning framework with BiLSTM as the shared layer. The hard parameter sharing mechanism is employed to learn the IES multi-coupling information and extract temporal characteristics of the load sequence through BiLSTM. In particular, three different structures of fully connected neural network are designed as feature interpretation modules to accommodate the different prediction requirements of various loads. The simulation results show that the proposed model achieves a weighted mean accuracy of 98.01% during winter days, with an average standard deviation of relative error as low as 0.0242. Among all the contrast models, it exhibits better prediction accuracy and stable error distribution. • Proposing a feature separation-fusion technology based on the difference in information value. • Proposing an improved CNN based on multi-scale fusion and reconstructing the original load into a 3D pixel matrix. • Designing a multi-task hard sharing learning framework and using feature interpretation modules with different structures. • The simulation results show that the proposed model achieves a WMA of 98.01% during winter days, with an RESD as low as 0.0242. [ABSTRACT FROM AUTHOR]

• This paper addresses the challenge of health-conscious vehicle battery modeling method. • A cloud-based health-conscious battery modeling and management framework is designed. • A healthy perception model is built to analyze aging state of different battery packs. • The deep transfer learning method is used for boosting the training of battery model. • Electric buses operation data in cloud platform is used to validate the proposed method. Establishing an accurate mathematical model is fundamental to managing, monitoring, and protecting the battery pack in electric vehicles (EVs). The application of the deep learning algorithm-based state estimation method can significantly improve the accuracy and stability of the battery model but is hindered by the great demand for training data. This paper addresses the challenge of health-conscious battery modeling by utilizing multi-source data based on a novel deep transfer learning method. Firstly, a cloud-based battery management framework is designed, which is able to collect and process battery operation data from various EVs and provide a foundation for deploying the transfer learning method. Battery healthy state information in the collected dataset is labeled by a generic perception model, which can be commonly used to quantify the aging state of different battery packs and facilitate the knowledge transfer process. Additionally, a deep transfer learning method is developed to boost the training process of the battery model, where the operation data from different types of EVs can be used for establishing state estimators. The method is verified by the battery operation data collected from two types of electric buses. With the developed healthy state perception model and transfer learning method, battery model error can be limited to 2.43% and 1.27% in the whole life cycle. [ABSTRACT FROM AUTHOR]

• The convolution and pooling operations are used to extract the wind power feature. • A double Gaussian activation function is proposed for Radial Basis Function network. • Coefficients of determination (r-square) of the results exceed 0.88 in 4 seasons. Wind power generation is always associated with uncertainties as a result of fluctuations of wind speed. Accurate predictions of wind power generation are important for the efficient operation of power systems. This paper presents a hybrid deep learning neural network for 24 h-ahead wind power generation forecasting. This novel method is based on a Convolutional Neural Network (CNN) that is cascaded with a Radial Basis Function Neural Network (RBFNN) with a double Gaussian function (DGF) as its activation function. The CNN is utilized to extract wind power characteristics by convolution, kernel and pooling operations. The supervised RBFNN, incorporating a DGF, deals with uncertain characteristics. Realistic wind power generations, measured on a wind farm, were used in simulations. The proposed method is implemented using TensorFlow and Keras Library. Comparative studies of different approaches are shown. Simulation results reveal that the proposed method is more accurate than traditional methods for 24 h-ahead wind power forecasting. [ABSTRACT FROM AUTHOR]

Incorrect predictions or underestimation of a city's solar potential can result from neglecting common features of photovoltaic (PV) panels from remote sensing images. This paper proposes an improved approach to address the challenge of accurately segmenting PV panels from remote sensing images using deep learning methods. The proposed method incorporates common features of PV panels and a constraint refinement module (CRM) to perform the localization of PV panel regions and shape regularization more accurately. Specifically, the method uses a color loss function based on prior knowledge of color to refine the predicted region with correct color information among confusing objectives, and a shape loss function based on multi-layer shape targets calculation to refine the initial segments and constrain the edge information of predicted regions. Different CRMs are embedded into the four refined initial segment modules, respectively, to improve the detection IoU of PV panels by up to 7.44%. The best CRM-integrated model performs the best IoU of 74.66% when segmenting PV panels. The proposed method has important implications for urban PV panel segmentation at the city level and provides a promising solution for remote sensing image-based PV plate segmentation tasks in challenging scenarios. • We proposed an improved feature-oriented segmentation method to enhance the accuracy and contextual information of PV panels. • We incorporated common features of PV panels and CRM to perform PV panel region localization and shape regularization. • We designed a color constraint refinement module via prior knowledge to refine the predicted region among confusing targets. • We designed a shape constraint refinement module based on multi-layer shape targets to constrain the edge of predicted areas. [ABSTRACT FROM AUTHOR]

Accurate ultra-short-term wind power forecasting is a prerequisite for decision making related to the management of power systems. Existing approaches used to forecast wind power ignored the correlation between wind power outputs under similar wind power, which provides important information for wind power forecasting. A spatiotemporal approach based on an indirect strategy and a multi-factor extraction model is proposed in this study to achieve more accurate power prediction. Specifically, wind power and other variables are classified into four categories according to wind direction, and multivariate feature extraction and wind power forecasting are performed separately for each category. In the experimental study, taking two real-world datasets of wind farms in Northeast China as examples, the proposed approach is compared with eight benchmarking approaches, and the results demonstrate the effectiveness and robustness of our approach. In addition, we further conduct the DM test, and the sensitivity analysis to discuss the effect of hyperparameters on forecasting performance. The results of DM verify the superiority of the proposed approach in statistic. This study provides a new high-precision approach and a new forecasting strategy for future wind power forecasting. • An indirect strategy is adopted in the wind power forecasting. • The dependence between data under similar wind direction is considered. • The approach characterizes the spatiotemporal effects of wind power. • This paper uses various information extraction and filtering methods. • Comprehensive experimental studies are conducted. [ABSTRACT FROM AUTHOR]

The increasing effects of global warming and energy depletion have raised concerns about the pollution caused by traditional oil and fossil energy usage. Distributed energy resources (DERs) have emerged as promising techniques to address these issues. However, the growing proportion of power injection from DERs presents technical challenges such as power quality decline and power flow reversal, which may destabilize the grid. Non-intrusive load monitoring (NILM) is a promising and cost-effective approach that can provide residential power information to improve grid scheduling, dispatching, and optimize residential power consumption behavior. This paper proposes a novel NILM method that can decompose the power of both household appliances and DERs integrated with residential houses. The proposed approach employs a data segment method that utilizes adaptive window lengths to identify the behavior of DERs based on the operating periods and characteristics of different equipment. A non-intrusive load monitoring method is proposed based on the ResNet-seq2seq framework to address the problem of model degradation and gradient vanishing. Post-processing techniques are applied to the output results using a feature list of the target equipment. Experiments are conducted on a dataset organized based on the REDD and Pecan Street datasets to demonstrate the accuracy and effectiveness of the proposed method. The results show that compared to the start-of-the-art CNN-seq2seq model, the proposed method achieved 54.95%, 69.18%, 56.95%, and 84.98% reduction in MAE for PV, EV, refrigerators, and microwaves, respectively. • A novel ResNet-seq2seq-based NILM approach is proposed for load disaggregation of residential houses integrating DERs. • Wavelet denoising and adaptive window length data segments are used for data pre-processing and accurate disaggregation. • Residual connections are employed to solve model degradation and vanishing gradient issues as network depth increases. • A post-processing technique is applied to modify the output of the network and improve disaggregation accuracy. [ABSTRACT FROM AUTHOR]

The increasing penetration of distributed renewable energy resources brings a great challenge for real-time voltage security of distribution grids. The paper proposes a safe multi-agent deep reinforcement learning (MADRL) algorithm for real-time control of inverter-based Volt-Var control (VVC) in distribution grids considering communication delay to minimize the network power loss, while maintaining the nodal voltages in a safe range. The multi-agent VVC is modeled as a constrained Markov game, which is solved by the MADRL algorithm. In the training stage, the safety projection is added to the combined policy to analytically solve an action correction formulation to promote more efficient and safe exploration. In the real-time decision-making stage, a state synchronization block is designed to impute the data under the latest timestamp as the input of the agents deployed in a distributed manner, to avoid instability caused by communication delay. The simulation results show that the proposed algorithm performs well in safe exploration, and also achieves better performance under communication delay. • A MADRL based decentralized control strategy is proposed to achieve real-time voltage regulation under fast fluctuations. • A safe MADRL algorithm is proposed to achieve efficient exploration with guaranteed safety. • A data synchronization strategy is proposed to achieve guaranteed performance under communication delay. [ABSTRACT FROM AUTHOR]

The increase of electric vehicles (EVs) in the power system, in addition to environmental benefits, also has economic advantages. Charging parking lots can perform better economically through aggregators the benefits of participating in energy markets and developing different charging programs. The main challenge for participating in the electricity market for parking lots is the uncertainties in the behavior of EV owners and the day-ahead market (DAM) price. Therefore, this paper proposes a deep learning-based framework for day-ahead optimal charging scheduling, which is in three stages. In the first stage, the parameters of charging sessions are modeled using the proposed Copula generative adversarial network (CopulaGAN). In the second stage, the DAM price is forecasted using the time series multi-input nonlinear autoregressive neural network for 24 h ahead. In addition, in the proposed framework, the balancing energy market (BEM) is used to keep the charging power balance on the implementation day of the optimal charging program. In the third stage, optimal day-ahead EV scheduling is performed. The simulation is carried out using Grey Wolf Optimization (GWO) based on the Caltech charging sessions dataset and the Germany energy market dataset. The simulation results show that using the proposed deep learning framework reduces the charging cost for the parking lots compared to the fixed price approaches, the use of the BEM, the time of use (TOU) program, and using normal distribution modeling about 19.94 €, 159.97 €, 14.27 €, and 114.03 € respectively. • Presenting a deep learning framework for modeling day-ahead scheduling. • The proposed framework is applied to Caltech and the German energy market dataset. • Results show a 15% reduction in charging costs compared to conventional methods. [ABSTRACT FROM AUTHOR]

Accurate building energy consumption forecasting is crucial for developing efficient building energy management systems, improving energy efficiency, and local building energy supervision and management. However, short-term building energy consumption forecasting is challenging due to highly non-smooth and volatile trends. In this paper, we present a novel methodology that combines interpretable decomposition methods with an interpretable forecasting model. We first illustrate a daily energy consumption pattern recognition (DECPR) method, which decomposes daily energy consumption patterns into interpretable energy consumption subsequences. To achieve satisfactory forecasting performance, we design the vector representation of each subsequence as a static input to the temporal fusion transformers (TFT) model. This vector representation integrates the DECPR method into the TFT model. The TFT model produces interpretable outputs, such as the attention analysis of different step lengths and the visualization of the importance ranking of exogenous variables, including meteorological data, calendar information, and the vector representation. Empirical studies demonstrate that our proposed DECPR-TFT system outperforms comparable models with a mean absolute percentage error (MAPE) of 6.11%, which is significantly lower than other models. These interpretable outputs provide valuable insights for researchers seeking to develop energy-saving operation strategies in buildings. Overall, our methodology offers a promising solution for short-term building energy consumption forecasting that can contribute to more efficient building energy management and energy-saving operation strategies. • This study proposes DECPR, an interpretable decomposition method, explaining decomposed subsequences. • TFT, a deep learning-based approach with interpretability, predicts hourly energy consumption for the next day. • The interpretability of TFT model are used to analyze the effects of exogenous variables on energy consumption • The proposed DECPR-TFT model outperforms others, with 6.11% MAPE, significantly lower than competitors. [ABSTRACT FROM AUTHOR]

Recent proliferation of renewable energy has increased the installation of residential energy sources (e.g., roof-top photovoltaic (PV) panel and residential wind turbine) in households. To manage electricity usage incurred by renewable energy and residential load, home energy management systems (HEMSs) provide intelligence to home by real-time monitoring and controlling appliances. In this paper, we propose a novel HEMS framework considering multiple uncertainties from renewable generation and load profiles. First, we generate scenarios of each uncertainty through deep learning. Then, we propose an algorithm called clustered quantile scenario reduction (CQSR) to reduce computation time while preserving the stochastic properties of generated scenarios based on the Wasserstein-1 distance. We prove that solution of CQSR is determined by the number of clustered scenarios. Also, we show provable upper bound of performance degradation incurred by the scenario reduction. Simulation results show that the optimality gap and computation time of the proposed framework is substantially reduced compared to other HEMS algorithms, e.g., by up to 81.4% and 93.7%, respectively. Furthermore, although the original scenarios are generated through different scenario generation algorithms, HEMS using CQSR is less vulnerable to performance degradation incurred by scenario reduction. • Propose HEMS framework under multiple uncertainties. • Propose a scenario reduction algorithm that leverages the cardinality of the cluster. • Provide upper bound of performance degradation incurred by scenario reduction. • Achieve less optimality gap of cost and computation time compared to baselines. • Reduce vulnerability to performance degradation incurred by scenario reduction. [ABSTRACT FROM AUTHOR]

Accurate energy load forecasting can not only provide favorable conditions for ensuring energy security but also reduce carbon emissions and thereby slow down the process of global warming. With the continuous construction of smart grids, energy consumption information can be easily obtained, which provides a good foundation for forecasting work. In the household energy consumption forecasting field, the current mainstream methods combine signal processing with artificial intelligence methods to improve forecasting accuracy. However, due to differences in residents' daily habits and sensor errors, the energy consumption data of individual households exhibit high volatility and contain noise. Existing methods' accuracy, robustness, generalization ability, and efficiency are challenged when considering multi-time scales and multi-samples. Based on this, this paper proposes a deep learning framework using multi-feature fusion recurrent neural networks for univariate energy consumption data. The proposed framework employs improved Singular Spectrum Analysis (SSA) and multivariate input recurrent neural network (RNN). To verify the effectiveness of the proposed framework, a series of comparative experiments were conducted on the Moroccan buildings' electricity consumption dataset (MORED). Experiments reveal that the proposed framework achieves adaptive decomposition of raw time series data and multi-feature fusion input. Moreover, under the evaluation of multiple metrics, the proposed framework outperforms the state-of-the-art methods in terms of forecasting accuracy, robustness, generalization ability, and efficiency. • A deep learning framework is proposed for forecasting energy consumption. • The framework achieves adaptive decomposition and multi-feature fusion. • The proposed framework's performance has been validated using MORED. [ABSTRACT FROM AUTHOR]

Automated fault detection and diagnosis (AFDD) plays a crucial role in enhancing the energy efficiency of air conditioning systems. Deep learning has emerged as a promising tool for image classification, and its application in the context of AFDD of HVAC systems is gaining traction due to its exceptional performance. However, the deployment cost of deep models in practical scenarios is increased due to the large number of parameters and the lack of interpretability. This paper focuses on improving the potential of deep learning models for AFDD in real HVAC systems. We use pruning to reduce the number of parameters in the model and use layer-wise relevance propagation (LRP) to improve the interpretability of the model. The case study builds a simulation model and 31 kinds of fault data sets based on the actual HVAC in Japan. Based on the findings, Without loss of accuracy, the pruning method can reduce the model size by more than 99 % and maintain 90% classification accuracy. The LRP score allows model users to find out the input data that most affects the results at each diagnosis, improving interpretability. • The study presents a solution for cutting costs of deep learning deployment in HVAC system AFDD. • The method combines pruning and LRP, reducing model size by over 99 • LRP score improves model interpretability by showing impactful input data on diagnosis results. • The study shows the potential of deep learning in HVAC AFDD, enhancing their practicality and usefulness. [ABSTRACT FROM AUTHOR]

Optimal decision-making compels us to anticipate the future at different horizons. However, in many domains connecting together predictions from multiple time horizons and abstractions levels across their organization becomes all the more important, else decision-makers would be planning using separate and possibly conflicting views of the future. To this end, this work proposes a novel multi-dimensional hierarchical forecasting method built upon structurally-informed machine-learning regressors. A generic formulation of multi-dimensional hierarchies, reconciling spatial and temporal dimensions under a common frame is initially defined. Next, a coherency-informed hierarchical learner is developed built upon a custom loss function leveraging optimal reconciliation methods. The coherency of the produced hierarchical forecasts is then secured using similar reconciliation techniques, granting decision-makers a common view of the future serving aligned decision-making. The method is evaluated on two different case studies to predict building electrical loads across spatial, temporal, and spatio-temporal hierarchies. Although the regressor natively profits from computationally efficient learning, results displayed disparate performances, demonstrating the value of hierarchical-coherent learning in only one setting. Yet, existing obstacles were clearly delineated, presenting distinct pathways for future work. Overall, the paper expands and unites traditionally disjointed hierarchical forecasting methods providing a fertile route toward a novel generation of forecasting regressors. • Multi-dimensional hierarchy definition • Hierarchical forecasting with machine learning • Custom coherency loss function built from optimal reconciliation methods • Deep-learning model design and hyper parameter tuning • Smart-building electrical load forecasting from open dataset [ABSTRACT FROM AUTHOR]

Accurate mid-term gas demand forecasting plays a crucial role for gas companies and policymakers to achieve reliable gas supply plans, supply contracts management, and efficient operation to meet the increasing gas demand. However, mid-term gas demand forecasting faces the problems of data paucity caused by the low frequency of collecting monthly data and heterogeneous consumption patterns of various usage categories. This paper proposes a novel Fed erated Con trastive pretraining - L ocal C lustered F inetuning paradigm (FedCon-LCF) by integrating federated learning, deep contrastive learning, and clustering approaches. The proposed method can utilize data from multiple gas companies to overcome data paucity issues in a privacy-preserving way, and high-performance forecasting can be achieved by local clustered regression considering the heterogeneous patterns. An improved hierarchical contrastive loss and multi-scale regression loss are integrated to develop the F orecasting- O riented C ontrastive L earning model (FOCL), which can effectively extract information and generate fine-grained representations of time series for accurate forecasting. The proposed method is evaluated on a dataset collected from 11 gas companies in 11 different Chinese cities with a total of 17648 clients over 10 usage categories. The proposed method outperforms the benchmark LSTM model with an average improvement of 25.30% in MSE and 16.52% in MAE for 3-month-ahead, 6-month-ahead, 9-month-ahead, and 12-month-ahead gas demand forecasting. • Propose a new paradigm to address data scarcity and diverse demand patterns problems. • Develop a new method to generate fine-grained representations from gas demand data. • Provide new findings for mid-term gas demand forecasting of different ranges. [ABSTRACT FROM AUTHOR]

Towards the low-carbon goal, a smart grid features remote connection, data sharing, and cyber–physical integration to increase the flexibility of energy supplies, to reduce electricity wastage, and to enhance safety under intermittent renewables. In this context, machine learning (ML) is increasingly employed in smart grids to solve tasks that require deep data analytics. However, it is well recognized that ML models are vulnerable to adversarial examples that might be contained in open shared data, which creates opportunities for potential attackers to launch cyberattacks. In particular, system operating states acting as critical inputs to ML-based smart grid applications (MLsgAPPs) can be manipulated by malicious actors to mislead the system operator into making wrong decisions that in turn may cause major blackouts and other damaging cascading events. It is thus imperative to advance vulnerability assessment for MLsgAPPs. In this paper, we propose a physics-constrained robustness evaluation framework for MLsgAPPs, which is a first attempt of using tree ensemble (TE) models. Unlike adversarial attacks against prior studied image-classification tasks, adversarial examples against TE-based smart grid applications (TEsgAPPs) must not only cheat human intuition but also follow laws of physics and bypass error-checking mechanisms for power systems. Thus, formulation of the traditional robustness evaluation problem must be reconsidered to account for domain-specific misclassification, physical limit, power balance, and error-bypass constraints, in order to obtain tight lower bounds of the magnitudes of tolerated adversarial perturbations. We adopt a formal modeling approach to construct a discrete tree structure and to specify the model evasion conditions. We propose an efficient robustness evaluation method by transforming variables considering ℓ 1 and ℓ ∞ norms, followed by a generalization that expands the variable space to admit other types of norms. Using TE-based power system security assessment as an example, comprehensive simulations are conducted for evaluating the physics-constrained robustness under different TE models and model parameters. [ABSTRACT FROM AUTHOR]

• The proposed control framework mitigates assumptions about accurate physical models. • A novel exploration mechanism incorporates historical information representation. • Each MG controller can manage the energy of MG in a decentralized manner. Reliability and cost-effectiveness in the operation of the multiple microgrid (MMG) system depend on the skillful management of its energy resources. Traditional energy management approaches are physics-model-based, which rely on the precise system parameters (e.g., line parameters) of the electricity and heat network. It is difficult to establish the precise system parameters in practice because they depend on a variety of factors. In this context, this paper proposes a physics-model-free control framework for the energy management of MMGs with heat-electricity energy, consisting of proposed surrogate model and multi-agent deep reinforcement learning (MADRL) approach. An important step is to use historical data to train a surrogate model in supervised manner that can imitate the realistic power and thermal flow calculations. Meanwhile, the energy management problem is reformulated as a Markov game. It is solved by the proposed MADRL-based approach by modeling each MG controller as an agent with a specific objective. The historical trajectories representation, parameter space technology, and deep dense architecture in reinforcement learning are introduced in MADRL to overcome the negative impact brought by the time series data from the input state on the decision-making process and construct an efficient exploration mechanism to overcome inefficient optimization of the MMG system in a multi-agent setting. During training period of MADRL, the trained surrogate models integrate into the environment of the MADRL, which can develop optimal energy management strategy based on the continuous interaction with the surrogate models. The proposed surrogate model enabled MADRL approach can reduce the reliance on precise physical systems and prevent having an impact on the real system while being trained involving trial and error process. Simulation results demonstrate the effectiveness of the proposed control framework. [ABSTRACT FROM AUTHOR]