A reinforcement learning-based ensemble forecasting framework for renewable energy forecasting

Abstract

The randomness and intermittency of wind and photovoltaic power generation can negatively affect the stability of power systems. Therefore, accurate forecasting of these energy outputs is crucial for effective power system management. Among various forecasting methods, ensemble forecasting has gained attention for its superior performance and reliability. However, traditional ensemble methods, such as weighted averaging and stacking, use fixed model combinations that fail to adapt to varying wind and radiation conditions, thereby limiting their accuracy. To overcome this limitation, this study proposes a novel ensemble forecasting framework based on reinforcement learning. The framework uses a deep Q-network to dynamically select the appropriate base model for different wind and radiation conditions. The learning process is supported by a model control module, a basic forecasting module, and a performance evaluation module. Independent experiments conducted across 14 regions in China validate the framework's effectiveness, showing a significant improvement in forecasting accuracy. The framework achieved an average improvement of 12.18 % in mean absolute scaled error over base models and 4.84 % over other ensemble methods. Additionally, this study analyzes the impact of different reinforcement learning models and sample sizes on the framework's performance.