Increasing shares of renewable generation are leading to more volatile electricity prices, presenting an opportunity for Energy Storage Systems (ESS) participating in short-term electricity markets. Model Predictive Control (MPC) has been shown to be a powerful tool to leverage the latest information at the time of optimization, yet its efficacy depends on the quality of the employed price forecasts. So far, these forecasts have been developed with traditional forecasting methods instead of value-oriented approaches, which consider the downstream decision problem during the forecaster training phase. Existing value-oriented methods, however, often rely on a specific downstream problem structure. This paper addresses these shortcomings by introducing a universally applicable, value-oriented forecasting methodology that employs a generalized loss function designed to account for inter-temporal price variability, using the downstream value (i.e., profit from ESS market participation) as the selection criterion in the hyperparameter tuning step. The proposed methodology is tested on a case study considering different types of ESS participating in the Belgian balancing market through MPC. The method is benchmarked against other forecasting techniques including a neural network trained in traditional, accuracy-oriented fashion. Using real-life data over a test set of two months, we show that the methodology outperforms those traditional techniques in terms of ex-post out-of-sample profit.

Energy Storage Systems (ESS) play a crucial role in managing renewable energy variability. Forecast-informed optimization is typically used to maximize ESS profit in electricity markets. Whereas traditional forecaster training methods use accuracy-based loss functions, Decision-Focused Learning uses a task-aware loss function with the aim of improving ESS profits. This can be achieved by integrating the downstream optimization in the forecaster training procedure. A common task-aware loss function is the Smart Predict-then-Optimize (SPO+) loss. However, its current implementation is prone to overfitting and is limited to linear forecasting models. Here, we extend the SPO+ framework to neural network forecasters with non-linear activation functions while introducing an interior-point training method to mitigate overfitting risks. When applied to an ESS participating in the day-ahead market, our approach outperforms both traditional and other decision-focused benchmarks in terms of obtained ESS profits.