Short-Term Photovoltaic Power Forecasting Based on a CEEMDAN-PE-ISSA-TCN-Transformer Framework

This study proposes a CEEMDAN-PE-ISSA-TCN-Transformer framework for short-term photovoltaic (PV) power forecasting. Six pre-specified meteorological variables were retained, and their Pearson correlation coefficients with PV power were calculated using the training interval only. The power sequence was decomposed into high-frequency, medium-frequency, and low-frequency reconstructed components using CEEMDAN and permutation entropy (PE). A branch-specific residual TCN-self-attention forecasting network was constructed for each component, and an improved sparrow search algorithm (ISSA) was used to optimize the number of filters, the number of attention heads, and the initial learning rate. Four seasonal datasets, each containing 10 consecutive days of five-minute observations, were divided chronologically into training, validation, and untouched test intervals in a 7:1:2 ratio. For validation and test forecast origins, CEEMDAN used only the most recent 288 available historical power observations. Across the four within-season test intervals, the proposed model obtained RMSE values of 0.5905, 0.9661, 1.0645, and 1.2476; MAE values of 0.4434, 0.4796, 0.9104, and 0.7066; R2 values of 0.9631, 0.8499, 0.8819, and 0.8401; and WAPE values of 16.98%, 25.27%, 34.48%, and 26.78% in spring, summer, autumn, and winter, respectively. The results indicate improved point-forecast accuracy for the evaluated within-season configurations.