Adaptive Weighted Combination Approach for Wind Power Forecast Based on Deep Deterministic Policy Gradient Method

Accurate wind power forecast (WPF) is critical for ensuring secure and economic operation of a power system, and combination forecasting approaches for WPF have been proved effective on attaining accurate forecasting results. However, the weights in a combination forecasting model are usually predetermined based on the global performance on the training set, limiting the adaptability of the model to different scenarios. To further enhance the accuracy and effectiveness of the combination forecasting approaches, this article proposes an adaptive weighted combination forecasting approach based on the deep deterministic policy gradient (DDPG) so as to consider the local behavior accompanied by the change of external environment. Three sub-models are first selected considering the equal-likelihood and dispersion indices to construct a combination model. Then, the DDPG agent is trained to act as a weight generator by interacting with the environment and to maximize the reward. Thus, the DDPG agent can perceive the environmental changes online and dynamically weight the sub-models to attain accurate forecasting results. Case studies demonstrate that the forecasting accuracy of the proposed approach is better than that of all sub-models and several benchmark combination forecasting approaches.