In order to improve the wind power consumption level of the new power system with new energy as the main body, it is necessary to accurately predict the wind speed. The key is to refine the dynamic trend of the wind power system and the potential physical structure in the wind speed sequence. Firstly, on the basis of the library Koopman dynamics theory and the encoder structures, a physically constrained spatio-temporal neural network is built, which generates the linear evolution moment of the nonlinear variables of the wind farm. Secondly, the system trend is approximated by the linear evolution matrix, and the forward and backward dynamics are fully considered in the prediction process. Then, the bidirectional correlation prediction mechanism and the cost function adapted to different objects are set to reduce the requirements for the reversibility and stability of the prediction sequence. Meanwhile, the hidden vector of feature space is visualized to show the feature interval dependency in the system. Finally, the effectiveness of the proposed method is verified by the wind speed measurement data of Prince Mountain in Beipiao. The results show that the proposed method has high prediction accuracy, strong generalization ability, and high interpretability for strong random and strong fluctuation wind speed series.

No abstract

No abstract

No abstract