Short-Term Photovoltaic Power Prediction Method Based on Fusion of Cloud Map Features and Historical-Power Data

Existing data-fusion-driven photovoltaic (PV) power-forecasting methods do not fully exploit cloud-image information. Simple feature concatenation cannot explicitly describe high-order interactions between cloud imagery and historical power, whereas full high-order fusion tensors introduce substantial parameter and computational overhead. This paper proposes a short-term PV power-forecasting method that fuses ground-based cloud-image features with historical-power data. MAGAN is used to enhance the brightness and texture representation of ground-based cloud images. An improved CNN-Transformer with a multi-scale separable convolution (MSC) module extracts multi-scale spatial cloud features. Low-rank tensor fusion (LTF) models cross-modal interactions between cloud-image and power features under a low-rank constraint and the fused sequence is fed into an improved Pyraformer with an i-Pyra module to forecast four 15 min steps over the next hour. Experiments using 2021 measurements from a 100 kW PV station in Jiangxi Province show that MAGAN-enhanced cloud images reach a PSNR of 7.25 dB and an SSIM of 0.75. On the independent test set, the proposed method obtains MAE, MAPE, and RMSE values of 21.8%, 17.6%, and 32.4%, respectively, reducing the three errors by 13.1%, 15.8%, and 15.2% compared with a cloud-feature and sample-weighted Pyraformer, while reducing fusion-layer parameters by 46.7% relative to full bilinear fusion.