A double deep reinforcement learning-based adaptive framework for decision-optimal wind power interval prediction

Chenghan Li, Yinliang Xu

Abstract: Prediction intervals (PI) effectively quantify forecasting uncertainty and serve as inputs for subsequent decision-making tasks. While it is traditionally assumed that reducing prediction errors will correspondingly reduce decision errors, this assumption is not invariably valid. This paper introduces an adaptive decision-optimal framework for optimal interval forecasting in wind power, designed to alleviate the economic dispatch challenges posed by wind power uncertainty within power systems. Methodologically, this framework employs a closed-loop method based on the Double Deep Q-Network algorithm, where the forecasting module leverages a pre-trained model with Bi-Directional Long Short-Term Attention to enhance feature extraction from historical data and improve quantile forecast precision. The Double Deep Q-Network then selects decision-optimal quantile proportions. The validity of the framework is demonstrated through experiments utilizing real-world wind power data from the Belgian Elia company, validated on IEEE 6-bus and IEEE 30-bus test systems. The proposed method reduces average operational cost and risk by 0.36%/4.38% on the IEEE 6-bus system and by 0.76%/6.29% on the IEEE 30-bus system compared with benchmark methods. This framework provides a robust solution for probabilistic wind power forecasting and decision-making in power systems, thereby enhancing the stability and economic efficiency of power system operations.