EWT-Hilbert-CACLNet Hybrid Forecasting Model for Short-Term Photovoltaic Power

Photovoltaic power prediction is an important component of the comprehensive energy system planning and scheduling for a park. High-precision photovoltaic power prediction helps with the dynamic stability dispatch of the power system. To address issues such as high volatility noise and unstable fitting performance of single models in photovoltaic power prediction, this paper proposes a combined forecasting method based on Empirical Wavelet Transform (EWT), Hilbert Transform (HT), and Cycle-Adaptive Convolutional and Bi-LSTM Network (CACLNet). First, EWT is used to decompose the historical photovoltaic power generation data into several sub-sequences of different frequencies, which are then reconstructed based on the frequency energy of each sequence. The Hilbert transform is used to further improve the resolution of the main components and break down their frequency domain information. Finally, Convolutional Neural Network (CNN) is applied to predict the decomposed main components. For the high-frequency minor components, the CACLNet neural network model is used to fully extract both the explicit and implicit trends within a cycle to obtain the prediction values. The final photovoltaic power prediction is obtained by linearly combining the results. Experimental results show that the proposed combination prediction method effectively improves the accuracy of photovoltaic power prediction.