Short-Term Wind Power Forecast Based on Continuous Conditional Random Field

The randomness and volatility of wind power severely challenge the safety and economy of power grids. Most short-term forecasting models exclusively concentrate on the correlation of numerical weather prediction (NWP) with wind power, while ignoring the temporal autocorrelation of wind power. To take both of them into consideration, this paper proposes a continuous conditional random field (CCRF) model integrated with the bidirectional LSTM (Bi-LSTM) and Gaussian Kernels (GKs). Firstly, through establishing the weather research and forecasting model, NWP data of high temporal-spatial resolution are generated as the input features of forecasting model. Secondly, Bi-LSTM is employed as the unary potential function to construct the non-linear relationship of the feature sequence with the wind power sequence, and the pre-defined two GKs are supplemented as the pairwise potential function to learn the interaction of wind power sequence. Thirdly, with the two potential functions, the CCRF model is constructed and trained by applying the mean-field theory, avoiding the complex gradient derivation in the learning process. Finally, the proposed CCRF model is tested through case studies and the results show that the forecasting accuracy can exceed that of any selected benchmark model.