Energieopslagsystemen in kortetermijn-elektriciteitsmarkten: Een verkenning van de interactie tussen optimalisatie en machine learning

Since 2020, the dynamic landscape of electricity markets has witnessed significant changes, marked by increased penetration of renewable energy sources and geopolitical upheavals. These factors have introduced volatility and unpredictability into electricity prices, presenting both challenges and opportunities for market participants. Energy Storage Systems (ESS) have emerged as a crucial technology to navigate these challenges, offering the ability to store and release electricity, thus helping to stabilize the power system and facilitate the transition to a more sustainable energy system. The main objective of this dissertation is to enhance the integration of ESS in managing short-term intermittency in electricity markets by developing and implementing innovative, profitable market participation strategies. This research is underpinned by cutting-edge optimization and Machine Learning (ML) methodologies aimed at maximizing the profitability of ESS operations. The dissertation focuses on two aspects of this objective, relating to innovations in (i) optimization modeling approaches and (ii) forecasting techniques based on ML. The first aspect is assessing the viability of the novel market participation strategy of implicit balancing, exposing market participants to the imbalance price without the need to reserve capacity in the day-ahead stage, from both an ESS as well as a system operator perspective. To that end, the dissertation explores a Model Predictive Control (MPC) algorithm, involving iterative bi-level optimization over a rolling horizon. This approach enables the continuous incorporation of the latest information to improve the accuracy of System Imbalance (SI) predictions. To ensure realistic modeling, the optimization program considers the ESS to be a price-maker, incorporates risk aversion, and explicitly accounts for degradation effects, as such assuming a battery-based storage system. The findings of a case study of the Belgian balancing market indicate that implicit balancing is more profitable than day-ahead market participation, and that ESS engaging in this strategy generally improve the balance in the system. Extending the MPC method to a two-stage optimization framework, which combines participation in the day-ahead market with implicit balancing, demonstrates the potential of this extension to further improve profits compared to executing either strategy separately. This enhancement is particularly evident when adopting a risk-averse approach to avoid aggressive gaming behaviour, which arises from anticipating inter-market arbitrage opportunities, in the day-ahead stage. The second aspect explores the effectiveness of ML techniques for price forecasting, tailored to the problem of ESS participation in short-term electricity markets. While traditional ML training techniques focus on minimizing a statistical error metric, Decision-Focused Learning (DFL) aims to maximize the value resulting from the decision problem in which the forecasts are utilized. First, we consider generalizable DFL, i.e., methods that are not limited by the structure of the downstream decision problem. In this context, a loss tuning approach is introduced, defining a parametric family of loss functions, the parameters of which are considered as hyperparameters of the ML model. The trained model yielding the highest downstream value is deployed. Our findings indicate that this method outperforms traditional training methods in terms of obtained ESS profit for both the applications of implicit balancing, as well as participation in the day-ahead market. On the other hand, integrated DFL methods directly train the price forecaster to maximize ESS profits. Methods like implicit differentiation of the Karush-Kuhn Tucker conditions and adopting a surrogate loss function for the ESS profit suffer from long train times as they require optimizing over the entire train dataset with every forward pass of the training procedure. To mitigate this, the novel approach of ML proxies, which mimic the optimal outcome of the ESS profit maximization problem by leveraging duality theory, are shown to significantly increase training speed while yielding similar profit performance as the aformentioned techniques. On the other hand, the reformulation approach to the surrogate loss function method suffers from overfitting to the train data. A novel approach based on interior point methods to solve the forecaster training problem was introduced to address this issue. Case studies of the Belgian electricity system reveal that these integrated DFL methods can outperform traditonal and generalizable DFL methods for downstream problems with low complexity, such as ESS participation in the day-ahead energy market. However, for the downstream problem of implicit balancing though MPC, they were unable to improve upon the performance resulting from implementing the loss tuning approach. This can be attributed to a mismatch between the downstream MPC value and the objective function of the optimization program.