Multi-source fusion-based cross-dimensional temporal feature-enhanced photovoltaic forecasting model

Accurate photovoltaic power forecasting is essential for the secure and efficient operation of power systems. However, existing methods often focus on single data sources or multi-source data combinations, overlooking cross-dimensional feature interactions. Therefore, this paper proposes a photovoltaic power forecasting framework that combines a hybrid neural network with data cleansing and reconstruction. The framework fuses photovoltaic operational data, on-site meteorological observations, and numerical weather prediction data. The hybrid network model, combining Temporal Convolutional Network and Crossformer, enhances cross-dimensional and cross-temporal feature extraction. In the cross-temporal feature extraction module, Temporal Convolutional Network is employed to extract temporal features from heterogeneous data streams. These features are then fed into a Multilayer Perceptron to learn nonlinear correlations and are integrated with the native Multi-head Self-Attention mechanism of Crossformer to capture interdependent feature weighting. In the cross-dimensional feature extraction module, the router mechanism of Crossformer is preserved to selectively reinforce critical feature information. For data processing, a data cleansing and reconstruction strategy based on the Horned Lizard-Optimized Variational Mode Decomposition algorithm is adopted, and Spearman-based combinatorial evaluation is employed for data reconstruction, helping mitigate data acquisition noise. Experimental results show that, for a 6-h forecasting horizon, the proposed model reduces mean absolute error by 4.24% to 19.71% compared with the original Crossformer and by 4.37% to 7.93% compared with the best single-source baseline models. In addition, the model achieves a coefficient of determination greater than 0.90 in all seasons and maintains robust trend-tracking performance under extreme weather conditions.