A Hybrid Decomposition–Deep Learning Framework for Accurate Forecasting in Hydro–Wind–Solar Systems

Accurate forecasting in hydro–wind–solar systems is vital for multienergy coordination. However, traditional methods struggle to capture critical temporal patterns when relying solely on raw input data. This study proposes a hybrid decomposition–deep learning framework to extract periodic, trend, and multiscale features. Firstly, forecasting factors are screened using the maximal information coefficient (MIC) from teleconnection indices, meteorological elements, and historical series. Secondly, an improved sparrow search algorithm (ISSA) is integrated with variational mode decomposition (VMD) to adaptively decompose raw sequences into trend, periodic, fluctuation components, and high-frequency noise. Thirdly, a novel model combining convolutional neural network (CNN), improved Transformer (iTransformer), and long short-term memory network (LSTM) is developed to process decomposed sequences and conduct forecasting. Applied to the Upper Yellow River Hybrid Energy Base, the results demonstrate that (1) the ISSA-VMD effectively separates intrinsic mode functions containing high-frequency fluctuations, periodic components, and long-term trends; (2) the proposed forecasting model achieved an average coefficient of determination (R2) of 0.948 (versus 0.744 for comparative models), with 49.51% lower mean absolute error (MAE) and 49.9% lower RMS error (RMSE); and (3) enhanced forecast accuracy benefited from the capability of CNN in local feature extraction, LSTM in temporal dynamic modeling, and iTransformer in global correlation mechanism.