Search

International audience; Spatially aggregating renewable power plants is beneficial when participating in electricity markets. In this context, a substantial number of features is available from various data sources. In machine learning, feature selection is common so as to relieve the curse of dimensionality and avoid overfitting. However, there is no guarantee that the selected features result in reliable forecasts and post-processing can therefore be valuable. In this study, we combine model agnostic feature selection with linear and nonlinear probabilistic forecast combination techniques. Moreover, the filters automatically compute the weights for our analog ensemble (AnEn) forecast model. We verify our model chain by generating intra-day forecasts of the aggregated output of 20 photovoltaic power plants using 831 input features in total. We show that the collection of filters selects a heterogeneous feature set but that each individual AnEn-filter combination results in underdispersed forecasts, which is efficiently remedied by the forecast combination techniques.

International audience; High temporal resolution intra-day and day-ahead photovoltaic (PV) power forecasts are important to maximize the value of PV systems because they enable stakeholders to participate in both the energy and ancillary services markets. Whereas most day-ahead electricity markets feature an hourly temporal resolution, intra-day markets may require forecasts at 5-minute resolution. In addition, battery integration can improve power system management in isolated grids with high PV power penetration, but battery control requires high temporal resolution forecasts. We propose an efficient method based on pattern matching to generate multivariate probabilistic forecasts, approximated by trajectories, at high temporal resolution and without the need to separately forecast the marginals and estimate the covariance matrix. We compare the proposed method against quantile regression forests in combination with copula theory and show that our method reduces the forecast time by approximately 98% and simplifies the modeling chain while incurring a minor performance penalty.

International audience; Standard practices of decision-making in energy systems are dynamic, non-linear, complex, and chaotic processes in nature. Trading the power produced by solar photovoltaic (PV) plants in electricity markets is an important decisionmaking problem which receives increasing attention in the past few decades. The main objective of this paper is to build an interpretable data-driven decision aid model for the case study of a solar power plant with the objective to minimize imbalance costs and thus maximise the revenue, using Symbolic Regression (SR) through Genetic Programming. The use of SR in the experiments and analysis developed in this paper show numerous advantages. SR evolves linear combinations of nonlinear functions of the input variables. Three penalty metrics are introduced to enhance the interpretability of the final solutions. SR shows robust results, especially in the case study.

International audience; Energy forecasting models deployed in industrial applications face uncertainty w.r.t. data availability, due to network latency, equipment malfunctions or data-integrity attacks. In particular, the case when a subset of features that has been used for model training becomes unavailable when the model is used operationally, poses a major challenge to forecasting performance. In this work, we present a principled approach to introducing resilience against missing features in energy forecasting applications via robust optimization. Specifically, we formulate a robust regression model that is optimally resilient against missing features at test time, considering both point and probabilistic forecasting. We develop three solution methods for the proposed robust formulation, all leading to Linear Programming problems, with varying degrees of tractability and conservativeness. We provide an extensive empirical validation of the proposed methods in prevalent applications, namely, electricity price, load, wind production, and solar production, forecasting, and we further compare against well-established benchmark models and methods of dealing with missing features, i.e., imputation and retraining. Our results demonstrate that the proposed robust optimization approach outperforms imputationbased models and exhibits similar performance to retraining without the missing features, while also maintaining practicality. To the best of our knowledge, this is the first work that introduces resilience against missing features into energy forecasting.

International audience; The participation of renewable generators in electricity markets involves employing a number of forecasting and decision-making tools. The standard approach consists in forecasting power output and market quantities, and then inputting the results into an optimization problem to derive optimal decisions. Typically, forecasting models are trained to optimize accuracy without considering the subsequent decision-making process. In this paper, we consider training forecasting models with a value-oriented approach that aims to minimize the suboptimality of decisions induced by a set of predicted inputs. We consider a risk-aware renewable generator participating in a day-ahead market subject to imbalance costs, and train ensembles of decision trees to forecast the imbalance penalty by directly minimizing trading costs for the provided strategy. The results indicate that our innovative approach leads to improved trading performance, compared to the standard method in which forecasting models are trained to minimize prediction errors.

International audience; This paper presents the solutions on renewable energy forecasting proposed by the Horizon2020 Project Smart4RES. The ambition of the project is twofold: (1) increase substantially the performance of short-term forecasting models of Renewable Energy Sources (RES) production and associated weather forecasting and (2) optimize decisions subject to RES uncertainty in power systems and electricity markets. Developments are based on latest advances in meteorology and original use of data science (combination of multiple data sources, data-driven approaches for trading and grid management). Finally, solutions such as flexibility forecast of distributed resources and data markets are oriented towards value for power system stakeholders. Although the project covers a broad scope, in this paper we focus on a selection of use cases that concern the integration of renewables in distribution grids.

The aggregation of multiple renewable plants located in distinct climate zones, using different energy sources, enables to reduce the production uncertainty when compared to the production of a single plant. Such aggregations, controlled by a Virtual Power Plant (VPP) system, are good candidates for the provision of ancillary services. Stochastic optimization models are available to optimize biddings on ancillary services and energy markets (see for instance [1]). These models require trajectories of the renewable VPP production that anticipate production uncertainty and reproduce correctly the temporal correlations observed in the production signal. This is particularly important in ancillary services markets, where a reserve bid must be guaranteed over a production duration or validity period during which power fluctuations are significant (e.g. lasting currently 24 hours on the European common market for Frequency Containment Reserve, with a foreseen evolution to 4 hours by July 2020 [2]). Production trajectories may be obtained by coupling probabilistic forecasts and a model of temporal dependencies between forecast horizons [3] and possibly spatial dependencies in the case of a multivariate forecast at the scale of a region or a portfolio [4]. In the case of a renewable VPP, the aggregated production is primarily of interest. In this work, we propose a methodology to generate trajectories of aggregated production from probabilistic forecasts obtained with decision-tree based models or neural networks. A copula models the dependency between forecast horizons and the space defined by the plants contained in the aggregation. The model is tested in a day-ahead forecasting configuration on a 54 MW VPP comprising 15 plants with 3 different energy sources (Photovoltaics, Wind, Hydro). The comparison of trajectories generated from a direct forecast of the aggregated production and from forecasts at lower levels of the aggregation shows that the latter solution reproduces with more accuracy the temporal variability of the aggregated production over the whole horizon range, especially when Photovoltaics dominates the production capacities in the aggregation (15 % improvement of the Variogram Score). [1]: Soares, T., & Pinson, P. (2017). Renewable energy sources offering flexibility through electricity markets. Technical University of Denmark.[2]: ENTSO-E. (2018). TSO’s proposal for the establishment of common and harmonised rules and processes for the exchange and procurement of Balancing Capacity for Frequency Containment Reserves (FCR) TSOs’ proposal for the establishment of common and harmonised rules and pro-c, (October), 1–9.[3]: Pinson, P., Madsen, H., Nielsen, H. A., Papaefthymiou, G., & Klöckl, B. (2009). From probabilistic forecasts to statistical scenarios of short-term wind power production. Wind Energy, 12(1), 51–62. [4]: Golestaneh, F., Gooi, H. B., & Pinson, P. (2016). Generation and evaluation of space–time trajectories of photovoltaic power. Applied Energy, 176, 80–91.

International audience; In this presentation we will outline the research directions and innovative solutions developed in the European Horizon 2020 project Smart4RES (http://www.smart4res.eu) aiming at improving all main steps of the model and value chain of renewable energy forecasting. The aim is to optimise integration of renewable energy (RES) production (i.e. wind, solar, run-of-the-river hydro) into power systems and electricity markets. Smart4RES started in November 2019 and gathers experts from several disciplines, from meteorology and renewable generation to market- and grid-integration. It aims to contribute to reach very high RES penetrations in power grids of 2030 and beyond, through thematic objectives including:• Improvement of weather forecasting with focus on variables that are important for RES applications;• Improvement of RES power forecasting and better estimation of uncertainties; • Streamlined extraction of optimal value through new forecasting products, data market places, and novel business models;• New data-driven optimization and decision-aid tools for market and grid management applications;• Validation of new models in living labs and assessment of forecasting value vs costly remedies to hedge uncertainties (i.e. storage). Smart4RES focuses both on improving forecasting models of weather (e.g. physical models, data assimilation, Large Eddy Simulation, enabling weather forecasts seamless) and RES production (e.g. seamless models, highly resolved predictions), and on addressing applications in power grids. Developments in the project have been formalized in Use Cases that cover a large range of time frames, technologies and geographical scales. For example, use-cases on power grids refer to the provision of ancillary services to the upper-level grid (e.g., balancing power) and the local grid (e.g., voltage control and congestion management), where the accurate forecasts of variable generation are key for accurate decision-making. A grid state forecasting will quantify dynamically the flexibility potential of RES in distribution grids. Collaborative forecasting investigates the improvement associated to local data sharing between distributed RES plants. This data sharing paves the way to a data market where agents exchange measurements, predictions or other types of valuable data. Lastly, data-driven approaches will streamline decision-making by simplifying the model chain of bidding RES production, storage dispatch or predictive management electricity grids. They will also provide interpretable hindsight to decision-makers by integrating the decisions of experts (human-in-the-loop) and will be tested in realistic laboratory conditions (software-in-the-loop).In this presentation we focus on the work done for the development of seamless numerical weather prediction (NWP) approaches to be able to couple outputs of NWP models with different resolutions and on the contribution of ultra-high resolution NWPs based on Large Eddy Simulation. A second group of results will focus on the optimal use of probabilisitic RES production and market price forecasts for trading in energy and ancillary service markets. Results from a new paradigm will be presented, where artificial intelligence methods are used to simplify the complex model chain (forecasting/trading) by a single decision making model that derives directly the bidding decisions using the available data. Results are based on data from real world test cases considered in the project.

International audience

International audience

International audience; Standard practice of decision-making in energy systems relies largely on complex modeling chains to address technical constraints and integrate numerous sources of uncertainty. The increased penetration of Renewable Energy Sources (RES) such as wind and photovoltaic plants adds complexity due to the weather dependency of their electricity production. Artificial Intelligence (AI) based tools have proven their efficiency in different applications in the energy sector ranging from forecasting to optimization and decision making. They permit to simplify modeling chains and to improve performance due to higher learning capabilities compared to state-of-the-art methods.However, decision-makers of the energy sector need to understand how decision-aid tools construct their outputs from the data. AI-based tools are often seen as black-box models and this penalizes their acceptability by end-users (traders, power system operators a.o.). The lack of interpretability of AI tools is a major challenge for the wider adoption of AI in the energy sector and a fundamental requirement to better support humans in the decision-aid process. Agents of energy systems expect very high levels of reliability for the various services they provide. As energy systems are impacted by multiple uncertainty sources (e.g. available power of RES plants, climate change, market conditions), developed AI tools should not only be performant on average situations but be able to guarantee robust solutions in the case of an extreme event.Therefore, our research focuses on understandable representations of data-driven decision-aid models for human operators in the energy sector. In order to enhance the interpretability of the AI models, a technique borrowed from the computer science domain is explored and further developed. Genetic programming and more precisely Symbolic Regression is used to derive a symbolic representation for the data-driven model that can take the form of a single equation. This equation results according to a specific reward function. The optimal solutions are selected naturally mimicking the biological theory of survival of the fittest.The main outcome is the production of symbolic representations of the AI models that require minimum changes when applied to different case studies. In this presentation a real-world use-case is considered, to demonstrate the added value of the proposed tools for decision making when trading the production of wind and solar power plants to the day-ahead market. An annual period of data is considered to train and test the proposed model. The typical modeling chain involves as many as 12 models for forecasting RES production and market quantities together with stochastic optimization to derive trading decisions. This complex chain is here replaced by a single AI-based model. Such simplification is a significant enhancement to the modeling chain interpretability and facilitates trust to the human decision-maker.This work is carried out in part in the frame of the European project Smart4RES supported by the H2020 Framework Program and in part in the frame the Marie-Curie COFUND project Ai4theSciences