Search

Background: Deep learning has the potential to assist the medical diagnostic process. We aimed to identify facial anomalies associated with endocrinal disorders using a deep-learning approach to facilitate the process of diagnosis and follow-up., Methods: We collected facial images of patients with hypercortisolism and acromegaly, and we augmented these images with additional negative samples from public databases. A model with a pretrained deep-learning network was constructed to automatically identify these hypersecretion statuses based on characteristic facial changes. We compared its performance to that of endocrine experts and further investigated key factors upon which the best performing model focused., Findings: The model achieved areas under the receiver operating characteristic curve of 0.9647 (Cushing's syndrome) and 0.9556 (acromegaly), accuracies of 0.9593 (Cushing's syndrome) and 0.9479 (acromegaly), and recalls of 0.7593 (Cushing's syndrome) and 0.8089 (acromegaly). It performed better than any level of our endocrine experts. Furthermore, the regions of interest on the part of the machine were primarily the same as those upon which the humans focused., Interpretation: Our findings suggest that the deep-learning model learned the facial characters based merely on labeled data without learning prerequisite medical knowledge, and its performance was comparable with professional medical practitioners. The model has the potential to assist in the diagnosis and follow-up of these hypersecretion statuses., (© 2019 S. Karger AG, Basel.)

The stochastic cloud cover on photovoltaic (PV) panels affects the solar power outputs, producing high instability in the integrated power systems. It is an effective approach to track the cloud motion during short-term PV power forecasting based on data sources of satellite images. However, since temporal variations of these images are noisy and non-stationary, pixel-sensitive prediction methods are critically needed in order to seek a balance between the forecast precision and the huge computation burden due to a large image size. Hence, a graphical learning framework is proposed in this study for intra-hour PV power prediction. By simulating the cloud motion using bi-directional extrapolation, a directed graph is generated representing the pixel values from multiple frames of historical images. The nodes and edges in the graph denote the shapes and motion directions of the regions of interest (ROIs) in satellite images. A spatial-temporal graph neural network (GNN) is then proposed to deal with the graph. Comparing with conventional deep-learning-based models, GNN is more flexible for varying sizes of input, in order to be able to handle dynamic ROIs. Referring to the comparative studies, the proposed method greatly reduces the redundancy of image inputs without sacrificing the visual scope, and slightly improves the prediction accuracy. [ABSTRACT FROM AUTHOR]

Developing solar power generation technology is an efficient approach to relieving the global environmental crisis. However, solar energy is an energy source with strong uncertainty, which restricts large-scale photovoltaic (PV) applications until accurate solar energy predictions can be achieved. PV power forecasting methods have been widely researched based on existing predictions of satellite-derived solar irradiance, whereas modeling cloud motion directly from satellite images is still a tough task. In this study, an end-to-end short-term forecasting model is proposed to take satellite images as inputs, and it can learn the cloud motion characteristics from stacked optical flow maps. In order to reduce the huge size of measurements, static regions of interest (ROIs) are scoped based on historical cloud velocities. With its well-designed deep learning architecture, the proposed model can output multi-step-ahead prediction results sequentially by shifting receptive attention to dynamic ROIs. According to comparisons with related studies, the proposed model outperforms persistence and derived methods, and enhances its learning capability relative to conventional learning models via the novel architecture. The model can be applied to PV plants or arrays in different areas, suitable for forecast horizons within three hours. [ABSTRACT FROM AUTHOR]

Due to the stochastic and non-stationary characteristics of wind speed, the wind power generation is highly uncertain and fluctuating, which significantly challenges the operation of the power system and the associated electricity market. In this article, a new spatial-temporal method is proposed for short-term wind power prediction based on image inputs and augmented convolutional network. First, the geographical locations of various wind farms and the relevant wind vectors are processed into a series of multiframe spatial-temporal wind images, which can be handled by the convolutional networks. Then, wind power conversion and prediction models are developed based on those networks, where recurrent paths and attention mechanism are introduced to enhance the model architecture. The testing results have validated the high performance of the proposed method within a forecast horizon of up to seven hours. In particular, even when the terrain information is not available, the implicit wind flow field within the original inputs can still be approximately learned by the proposed convolutional networks. [ABSTRACT FROM AUTHOR]

A prior knowledge of residential load demand is critical for power system operations at the distribution level, such as economic dispatch, demand response and energy storage schedule. However, as residential customers perform more casual and active consumption behaviors, prediction of such highly volatile loads can be much harder. Owing to the development of sensor technology, micrometeorological data can be sampled with a high geographic resolution. Those data that represent the weather condition on the land surface show a strong relationship to the residential load evidently, whereas it remains unsolved on how to fully utilize those great number of datasets. This paper proposes a day-ahead probabilistic residential load forecasting method based on a novel deep learning model, named convolutional neural network with squeeze-and-excitation modules (CNN-SE), and micrometeorological data. The model can employ multi-channel input data with dissimilar weights, suitable for analyzing massive relevant input factors. A feature extraction method is adopted for customer consumption pattern based on sparse auto-encoder (SAE), which can help correct probabilistic forecasting results. A case study that covers 8 residential communities and 18 micrometeorological sites is conducted to validate the feasibility and accuracy of the proposed hybrid method. [ABSTRACT FROM AUTHOR]

Solar energy is a strongly intermittent renewable energy source, which is affected by varied meteorological conditions, and thus produces arbitrary power outputs in photovoltaic (PV) power generation. Complex weather variations make it challenging to develop an efficient PV power forecasting method. In this study, a graph modeling method is proposed for short-term PV power prediction. Unlike many conventional machine-learning models, the proposed model is capable of evaluating interconnections among various meteorological input factors. This study details the design and operation of graph modeling, including graph construction, node feature construction, message transfer, and readout. An entire model is established consisting of spectral graph convolution, multiple graphical edges and a hierarchical output manner. The testing results suggest that the proposed multi-graph model outperforms other benchmark models in terms of accuracy under day-ahead forecasting cases. Besides, the single-graph model achieves a reduced cost of training time comparing to deep-learning benchmark models. [ABSTRACT FROM AUTHOR]

Power dispatching systems currently receive massive, complicated, and irregular monitoring alarms during their operation, which prevents the controllers from making accurate judgments on the alarm events that occur within a short period of time. In view of the current situation with the low efficiency of monitoring alarm information, this paper proposes a method based on natural language processing (NLP) and a hybrid model that combines long short-term memory (LSTM) and convolutional neural network (CNN) for the identification of grid monitoring alarm events. Firstly, the characteristics of the alarm information text were analyzed and induced and then preprocessed. Then, the monitoring alarm information was vectorized based on the Word2vec model. Finally, a monitoring alarm event identification model based on a combination of LSTM and CNN was established for the characteristics of the alarm information. The feasibility and effectiveness of the method in this paper were verified by comparison with multiple identification models. [ABSTRACT FROM AUTHOR]

Wind energy is a commonly utilized renewable energy source, due to its merits of extensive distribution and rich reserves. However, as wind speed fluctuates violently and uncertainly at all times, wind power integration may affect the security and stability of power system. In this study, we propose an ensemble model for probabilistic wind speed forecasting. It consists of wavelet threshold denoising (WTD), recurrent neural network (RNN) and adaptive neuro fuzzy inference system (ANFIS). Firstly, WTD smooths the wind speed series in order to better capture its variation trend. Secondly, RNNs with different architectures are trained on the denoising datasets, operating as sub-models for point wind speed forecasting. Thirdly, ANFIS is innovatively established as the top layer of the entire ensemble model to compute the final point prediction result, in order to take full advantages of a limited number of deep-learning-based sub-models. Lastly, variances are obtained from sub-models and then prediction intervals of probabilistic forecasting can be calculated, where the variances inventively consist of modeling and forecasting uncertainties. The proposed ensemble model is established and verified on less than one-hour-ahead ultra-short-term wind speed forecasting. We compare it with other soft computing models. The results indicate the feasibility and superiority of the proposed model in both point and probabilistic wind speed forecasting. [ABSTRACT FROM AUTHOR]

Solar photovoltaic power plays a vital role in global renewable energy power generation, and an accurate solar power forecast can further promote applications in integrated power systems. Due to advanced artificial intelligence technologies, various deep-learning models have been developed with the benefits of improved prediction precision, but these models inevitably sacrifice their interpretability compared to linear methods. Since a 100% accurate forecast is impossible to achieve, an opaque black-box model will always raise doubts for the operators of renewable power-grids, especially when the prediction deviation may produce higher economic costs and even a system turbulence. Motivated by this, the present study summarizes the requirements of deep-learning solar power forecast models from the power-grid application perspective. Post-hoc evaluation and discussion are conducted to analyze the performances of a typical deep-learning benchmark model based on open-access dataset for solar forecasting. Based on the results, the aim of this study is to increase confidence of deep-learning-based intelligent models into the practical engineering utilization of solar power forecasting. The case studies indicate that some simple evaluation procedures can aid a better understanding of the factors that influence the performances of opaque models, and these procedures can help in the design methods for model modifications. [ABSTRACT FROM AUTHOR]

Accurate photovoltaic (PV) power forecasting is crucial to ensure the safety and stability of power systems, given the penetration of solar energy. Extracting spatial-temporal features from ground-based sky images can greatly improve ultra-short-term PV power forecasting. Previous studies have primarily focused on extracting holistic spatial-temporal features from sky images without considering their interaction, leading to a loss of partial critical features that restricts the improvement of forecasting performance. Hence, this study proposes a novel framework considering the interaction of spatial-temporal features for ultra-short-term PV power forecasting. First, a two-stream network is used to extract spatial and temporal features separately from sky images, aiming to eliminate the negative impact of spatial-temporal feature interaction. Then, a gate unit is employed to fuse the extracted features adaptively. Subsequently, a PV-guided attention mechanism is proposed to enhance forecasting performance by identifying dominant regions within the fused feature map. Last, a time series inference model based on progressive architecture is proposed to forecast future PV power. Comparative results demonstrate that the proposed framework outperforms benchmark frameworks and exhibits higher generalization and robustness in ultra-short-term PV power forecasting. • An end-to-end image-based framework is developed for photovoltaic power forecasting. • Spatial-temporal features are extracted based on considering spatial-temporal feature interaction. • The correlation between image features and photovoltaic power is investigated. • A time series inference model based on progressive architecture is introduced. • The proposed framework outperforms benchmark frameworks for all case studies. [ABSTRACT FROM AUTHOR]

Accurate solar irradiance forecasting with fine spatiotemporal correlations is essential for photovoltaic power generation. However, current spatiotemporal methods are unable to extract spatial features appropriately, leading to rough prediction results. Consequently, a deep graph neural network is proposed in this paper for multi-site solar irradiance forecasting. Self-adaptive adjacency matrices were developed to better capture the intrinsic influence between sites under different climate types at various time steps. Based on the adaptive graph convolution and long short-term memory network (AGCLSTM), the spatiotemporal features of each site were extracted to obtain prediction results for all the sites in the research area simultaneously. In the experiments conducted on the datasets of three climate types, the average root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient (R) of the proposed model are observed to be in a range of 47.519–54.854 W/m2, 18.968–24.426 W/m2, and 0.9737–0.9889, respectively. Compared to the benchmark models, the average RMSE and MAE reduce by 1.31%–10.86% and 3.75%–18.57%, respectively. Moreover, the proposed method proves to be effective for scenes of solar irradiance fluctuations with the help of self-adaptive adjacency matrices, which enhances its generalization ability and prediction accuracy for one-hour-ahead solar irradiance forecasting. [ABSTRACT FROM AUTHOR]

Accurate short-term solar irradiance forecasting is crucial for ensuring the optimum utilization of photovoltaic power generation sources. This study addresses this issue by proposing a spatiotemporal correlation model based on deep learning. The proposed model first applies a convolutional neural network (CNN) to extract spatial features from a two-dimensional matrix composed of meteorological parameters associated with a target site and its neighboring sites. Then, a long short-term memory (LSTM) network is applied to extract temporal features from historical solar irradiance time-series data associated with the target site. Finally, the spatiotemporal correlations are merged to predict global horizontal irradiance one hour in advance. The prediction performance and generalization ability of the proposed CNN-LSTM model are evaluated within a whole year, under diverse seasons and sky conditions. Three datasets are involved for case studies, which are collected from 34 locations spread across three different climate zones in Texas, USA. Moreover, the performance of the CNN-LSTM model is compared with those obtained using the CNN, LSTM, and other benchmark models based on five evaluation metrics. The results indicate that the proposed model has advantages over the other models considered and provides a good alternative for short-term solar radiation prediction. • Meteorological data and GHI are reconstructed to extract spatiotemporal features. • The CNN-LSTM is proposed to improve the accuracy of solar radiation prediction. • Analyzed data are collected from 34 locations spread across three climate zones. • Testing cases are within a whole year, under diverse seasons and sky conditions. [ABSTRACT FROM AUTHOR]

Day-ahead residential load forecasting is crucial for electricity dispatch and demand response in power systems. Electrical loads are characterized by volatility and uncertainty caused by external factors, especially for individual residential loads. With the deployment of advanced metering infrastructure, the acquisition of electricity consumptions of multiple residential customers is available. This paper proposes a novel day-ahead residential load forecasting method based on feature engineering, pooling, and a hybrid deep learning model. Feature engineering is performed using two-stage preprocessing on data from each user, i.e., decomposition and multi-source input dimension reconstruction. Pooling is then adopted to merge data from both the target user and its interconnected users, in a descending order based on mutual information. Finally, a hybrid model with two input channels is developed by combining long short-term memory (LSTM) with self-attention mechanism (SAM). The case studies are conducted on a practical dataset containing multiple residential users. Performance of the proposed load forecasting method using data pools of different groups of users as well as different input forms is compared. The effectiveness of input dimension reconstruction and hybrid model is also validated. The overall results demonstrate the superiority of the proposed load forecasting method through comparison with other benchmark methods. • The multi-source input is reconstructed to promote the feature extraction for model. • LSTM-SAM model is proposed to improve the accuracy of residential load forecast. • Pooling strategy is adopted to construct training samples for forecasting models. • Optimal data pool and input form are investigated for better prediction performance. [ABSTRACT FROM AUTHOR]

• Two novel deep convolutional networks, i.e. ResNet and DenseNet, are introduced to photovoltaic power forecasting. • A meta learning strategy is proposed to enhance the reliability of extracted features and forecasting results. • Historical PV power series, historical weather elements and weather type prediction are processed into 2D input features. • Probabilistic forecasting models are built based on already-trained point forecasting models using transfer learning. The outputs of photovoltaic (PV) power are random and uncertain due to the variations of meteorological elements, which may disturb the safety and stability of power system operation. Hence, precise day-ahead PV power forecasting is crucial in renewable energy utilization, as it is beneficial to power generation schedule and short-term dispatch of the PV integrated power grid. In this study, a novel day-ahead PV power forecasting approach based on deep learning is proposed and validated. Firstly, two novel deep convolutional neural networks (CNNs), i.e. residual network (ResNet) and dense convolutional network (DenseNet), are introduced as the core models of forecasting. Secondly, a new data preprocessing is proposed to construct input feature maps for the two novel CNNs, which involves historical PV power series, meteorological elements and numerical weather prediction. Thirdly, a meta learning strategy based on multi-loss-function network is proposed to train the two deep networks, which can ensure a high robustness of the extracted convolutional features. Owing to the learning strategy and unique architectures of the two novel CNNs, they are designed into relatively deep architectures with superb nonlinear representation abilities, which consist of more than ten layers. Both point and probabilistic forecasting results are provided in the case study, demonstrating the accuracy and reliability of the proposed forecasting approach. [ABSTRACT FROM AUTHOR]