Bridging deterministic and stochastic scheduling: Learning risk-aware forecast adjustments for PV-BESS operation

Abstract

The participation of hybrid photovoltaic (PV) and battery energy storage system (BESS) installations in electricity markets requires modeling multi-temporal forecast uncertainty via temporal trajectories and integrating this information into bidding optimization frameworks. Most state-of-the-art approaches rely on stochastic or robust optimization methods, which are computationally intensive and offer limited interpretability to decision-makers. Consequently, deterministic optimization approaches remain the industry standard in practical market applications. To bridge this gap, this paper proposes a novel learning-to-optimize framework that combines a deterministic mixed-integer linear programming formulation with a context-aware artificial neural network policy trained through reinforcement learning. The policy operates as a learned cost function approximation , translating probabilistic forecast metrics into time-dependent multiplicative forecast adjustment parameters. This approach preserves the simplicity and industry acceptance of deterministic optimization while effectively embedding information related to forecast uncertainty and risk-awareness into the day-ahead bidding and scheduling process for hybrid PV-BESS systems. For a real-world electric vehicle charging station case study in Portugal, equipped with 375 kWp of PV generation and a 1.258 MWh BESS, the proposed learning-to-optimize framework substantially outperforms the deterministic baseline. Specifically, it captures 39.27% and 27.26% of the maximum theoretical cost-saving potential achievable under perfect forecasts during autumn and summer peak periods, respectively. At the intraday level, the learned policy effectively mitigates systematic forecast errors, reducing the cumulative daily negative imbalance volume by 48.22% under standard operating conditions. Moreover, the framework maintains a robust operational safety margin, reducing imbalance exposure by 10.02% during large unforeseen PV generation drops caused by clouds.