A Short-Term Photovoltaic Power Forecasting Method Based on Similar Days and WOA-MS-TFformer-BiTCN

Accurate short-term photovoltaic (PV) power forecasting is important for grid dispatch and PV integration. However, PV power under complex weather conditions has strong fluctuation, non-stationarity, and multi-frequency coupling. These features make accurate forecasting difficult. This paper proposes a short-term PV power forecasting model named WOA-MS-TFformer-BiTCN. The model first constructs similar-day samples through daily feature extraction, Gaussian mixture clustering, and physical consistency correction. Then, the whale optimization algorithm (WOA) optimizes the key parameters of variational mode decomposition (VMD) and the forecasting network. VMD decomposes the original power sequence into modes with different frequency features. The multi-scale frequency-domain perception (MS) module extracts multi-scale frequency-domain features from these modes. TFformer captures global temporal relationships, while BiTCN models local dynamic changes. Experiments are conducted using PV data from Gansu, China. The Alice Springs PV dataset is used for cross-regional validation. The results show that the proposed model achieves the lowest MAE, RMSE and the highest R2 in all 16 season-weather cases, corresponding to four seasons and four weather types, for the 15 min-ahead task. Its average MAE, RMSE and the highest R2 are 0.5439, 0.7910, and 0.99898, respectively. The model also performs best on rainy samples from the Alice Springs dataset. In addition, prediction intervals based on validation-set residual quantiles provide uncertainty information for point forecasts. The results show that the proposed method improves the accuracy and stability of short-term PV power forecasting under complex weather conditions.