A hybrid deep learning model with an optimal strategy based on improved VMD and transformer for short-term photovoltaic power forecasting

The precise forecasting of photovoltaic (PV) power is important for efficient grid management. To enhance the analysis and processing capability of PV characteristics, address the feature extraction challenges for long sequences, and improve forecasting accuracy, this study presents a robust hybrid deep learning model for PV power forecasting. First, a dynamic mean pre-processing algorithm is applied for data cleaning. Subsequently, an improved whale variational mode decomposition (IWVMD) algorithm is proposed for data decomposition in multichannel multi-scale modeling. Furthermore, a novel context-embedded causal convolutional Transformer (CCTrans) structure is used to predict each subsequence, and an optimal strategy is formulated for both input and output under the combined dynamic contextual information and single target variable forecasting (CDCTF) pattern. Finally, the forecasting results are reconstructed. Experiments are conducted to evaluate the performance of the model across different seasons, using publicly available datasets from the Desert Knowledge Australia Solar Center (DKASC). Ablation studies, validation with diverse datasets, and comparisons with other models confirm the effectiveness, accuracy, robustness, and generalizability of the model. In addition, recommendations for optimal forecasting ranges for different seasons are provided.