Dynamic Weighted Combination Approach for Short-Term Wind Power Forecast Considering Meteorological Similarity

Wind power forecast (WPF) system has been an indispensable component for a wind farm connected to power grid, and combination approach has been proved effectiveness to improve the robustness and accuracy of forecasting results. Most existing combination approaches fix the weights for selected sub-models and ignore the performance variation of sub-models under different meteorological conditions. This paper proposes a dynamical weighted combination approach considering the meteorological similarity to improve short-term WPF. Firstly, three sub-models consisting of gate recurrent unit, light gradient boosting machine and least square support vector machine are determined to be the members of the combination model. Secondly, according to the constructed meteorological features, a certain number of historical similarity days of the target forecasting day are found. Thirdly, the particle swarm optimization algorithm is applied to optimize the weights of sub-models. Finally, the combination weights change dynamically along with the variation of meteorological condition. Case studies demonstrate the validity of proposed model and the forecasting results for the test dataset are better than the sub-models and selected benchmarks according to the normalized mean absolute error and normalized root mean square error.