TCM-Former: a transformer with temporal convolution for photovoltaic power forecasting

Abstract

To tackle the problem of modeling long-term trends and short-term and high-frequency variations in PV time series, a strong and efficient forecasting model of photovoltaic (PV) power generation is advanced. This paper presents STL-TCM-Former, a hybrid model that breaks down the raw PV signal with seasonal-trend decomposition with LOESS (STL) into trend and seasonal components. These elements are after that processed via a dual path encoder decoder transformer architecture that is improved with a temporal convolutional module (TCM). The period (transformer-based) path is used for capturing the global, long-range dependencies in seasonal component and temporal (TCM) path is used to extract the localized, short-term dynamics. Trend component is processed with a dedicated TCM branch, minimizing component interferences and providing a multiscale temporal representation, specific to PV generation patterns. The proposed model was evaluated on the Yulara (Uluru) solar dataset under short-term (60-hour) and long-term (300-hour) forecasting horizons. Compared with benchmark models including LSTM, WOA-LSTM, VMD-LSTM, WOA-VMD-LSTM, and hybrid WOA/VMD/LSTM configurations, STL-TCM-former achieved superior performance with R² = 99.76%, MAPE ≈ 2.24%, RMSE = 16.63, and MAE = 10.36. In addition, the PJM interconnection dataset was also employed to evaluate the generalization capability of the proposed method. The results demonstrate high accuracy, stability, and strong generalization capability under varying environmental conditions.