Utilizing reserve market price prediction in multi-market optimization of energy storage

Abstract

The ongoing transition towards a carbon-neutral energy system is fundamentally reshaping electricity markets, increasing both the share of variable renewable generation and the demand for system flexibility. In Finland, the rapid expansion of wind power, coupled with electrification of multiple sectors, is expected to significantly increase electricity consumption while simultaneously introducing greater variability and uncertainty into power system operation. These developments have led to growing importance of reserve markets, where flexible resources such as battery energy storage systems (BESS) play a central role in maintaining system balance. Efficient participation in these markets relies heavily on accurate short-term price forecasting, which enables optimal allocation of asset capacity across multiple market products. This thesis investigates the creation of reserve market price forecasting models and their integration into a multi-market optimization framework for energy storage operation. The study focuses on Finnish reserve markets, specifically aFRR and mFRR products in both upward and downward directions. Existing forecasting approaches, based on traditional machine learning models and extensive feature sets, are assessed and used as a baseline for further development. To address issues related to overfitting and high-dimensional input data, a structured feature selection methodology is applied. Different forecasting models are implemented and compared, including gradient-boosted decision trees (XGBoost) and transformer-based time series models (Chronos-2), the latter both in zero-shot and fine-tuned configurations. The models are trained using a rolling time series validation approach to ensure robustness under changing market conditions. Performance is evaluated using standard error metrics, and the results are benchmarked against a naïve baseline model based on historical price persistence. The predictive performance of each model is further assessed in terms of its impact on the economic optimization of battery operation, with a focus on revenue generation in a multi-market setting. The results demonstrate that models based on machine learning can achieve meaningful improvements in forecasting accuracy compared to baseline methods, although the magnitude of improvement varies across market products and time periods. In particular, the comparison between XGBoost and Chronos-2 highlights differences in model behavior when applied to noisy and regime-dependent prices. The findings also indicate that even moderate improvements in forecast accuracy can translate into measurable gains in economic performance when integrated into an optimization framework. Fine-tuned Chronos-2 model performed the best measured both by prediction errors and economical returns. In the bidding simulation, the best model achieved 95.0 % of possible revenues. In total, the BESS would have generated 195 561 € during the examined test period when allocating 15 MW of capacity for every hour. It beat the baseline of naïve prediction by 3.3 %. The error metric MAE reached values between 0.35 and 1.63 €/MW depending on the market. Overall, this thesis contributes to the understanding of price forecasting in reserve markets and provides practical insights into feature selection and domain-based analysis under realistic market conditions. The results support the role of advanced machine learning methods as a key enabler for efficient utilization of battery energy storage in multi-market optimization.