Integrated forecasting and deep reinforcement learning for price-based self-scheduling of PV-BESS: Utility-scale...

Area of Science:

• Energy Systems Engineering
• Artificial Intelligence
• Renewable Energy Integration

Background:

• Deep Reinforcement Learning (DRL) shows promise for optimizing Battery Energy Storage Systems (BESS) coordinated with Photovoltaic (PV) plants.
• Most existing studies rely on simulations, lacking validation with real-world operational data.

Purpose of the Study:

• To empirically validate an integrated forecast-and-control framework for utility-scale PV-BESS assets using real operational data.
• To bridge the gap between simulated DRL performance and practical application in energy storage optimization.

Main Methods:

• Developed an integrated framework coupling a Sequence-to-Sequence (Seq2Seq) LSTM forecaster with DRL agents (PPO, SAC).
• Trained DRL agents on 1,000 probabilistic scenarios per site using two years of operational data (2022-2023).
• Benchmarked DRL policies against Oracle, predict-then-optimize, Model Predictive Control (MPC), and Dummy policies over 14-day horizons.

Main Results:

• The Seq2Seq forecaster improved price prediction accuracy (34.5% RMSE reduction vs. SARIMAX).
• DRL agents consistently outperformed the predict-then-optimize baseline, achieving average 14-day profits near USD 55k.
• DRL policies demonstrated robust, adaptive contracyclical behavior without excessive battery cycling.

Conclusions:

• The study provides empirical validation for data-driven BESS control using DRL in real-world conditions.
• The developed framework offers a reproducible blueprint for practical PV-BESS optimization.
• DRL-based control demonstrates significant economic benefits and practical viability for energy storage systems.