Reinforcement Learning-Based Dynamic Probability Distribution Selection for Wind Power Prediction

Wind power prediction is of great significance for wind energy utilization due to its stochastic nature. This paper proposes a novel short-term wind power prediction method, DQHMDN, for accurately predicting short-term future wind power while effectively assessing uncertainty. In this model, a Heterogeneous Mixture Density Network (HMDN) is introduced, which mixes multiple types of probability density functions. To better capture temporal dependencies, a Transformer module is implemented to process time-series features, ensuring that the generated conditional density function closely aligns with real data. Secondly, to achieve real-time selection of the most suitable Mixture Density Network (MDN) model for prediction based on current data features, a Markov Decision Process (MDP) model is constructed and solved by the Deep Q-Network (DQN) algorithm. Computational experiments based on SCADA data from nine wind turbines in a wind farm demonstrate that the DQHMDN model outperforms three benchmark models in both accuracy and prediction reliability.