Secondary Hybrid Decomposition Strategy for Wind Power Prediction Using Long Short-Term Memory With Crisscross Optimization

This paper presents a hybrid forecasting model for a wind power named Secondary Hybrid Decomposition (SHD)-Long Short-Term Memory (LSTM) with Crisscross Optimization (CSO).The model integrates three key techniques to improve prediction accuracy. Firstly, the SHD mitigates the limitations of single-stage decomposition by employing a two-step Empirical Mode Decomposition (EMD) process. This improves the predictability of wind power series by decomposing it into its Intrinsic Mode Functions (IMFs) and residual components. Secondly, LSTM networks are utilized to forecast each subseries and residue generated by SHD, capturing temporal dependencies within wind power data effectively. Finally, the CSO is used to optimize the hyperparameter of individual LSTM networks, enhancing global search performance. Two sets of wind farm data are used to validate the accuracy of the model. A comparative study shows that the proposed SHD-LSTM-CSO model performs better than other hybrid models such as SHD-LSTM, EMD-LSTM, LSTM based on variable mode decomposition (VMD) (VMD-LSTM), Gray Worm Optimization-based backpropagation neural network (GWO-BPANN) and EMD-based Artificial Neural Network (EMD-ANN) methods.