A Similar Day Clustering and Temporal Convolutional Network Framework for Photovoltaic Power Forecasting
TL;DR
Results demonstrate that the proposed weather-specific TCN framework substantially improves both accuracy and robustness across different weather regimes, offering a practical solution for distributed photovoltaic power forecasting in real-world operation.
Abstract
To improve photovoltaic power forecasting under changing weather conditions, this study proposes a hybrid framework that combines correlation-based feature selection, particle swarm optimization for fuzzy C-means clustering (PSO-FCM), and temporal convolutional network (TCN) modeling. Existing methods often rely on a single model for all weather types, which fails to capture scenario-specific temporal patterns and leads to large errors under cloudy or rainy conditions. The proposed approach first uses Pearson correlation to select key meteorological factors, then clusters historical data into sunny, cloudy and rainy scenarios via PSO-FCM, and finally trains a dedicated TCN model for each scenario to capture its distinct temporal dependencies. Experimental results show that under sunny conditions the TCN model achieves an R2 of 0.9965 with Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) of 6.06 kW and 8.35 kW, respectively; under cloudy and rainy conditions the R2 values are 0.9921 and 0.9297. In contrast, the Gated Recurrent Unit (GRU) model degrades significantly under non-sunny weather—its MAE and RMSE on cloudy days are 3.6 and 3.9 times higher than those of TCN, with an R2 of only 0.8798. These results demonstrate that the proposed weather-specific TCN framework substantially improves both accuracy and robustness across different weather regimes, offering a practical solution for distributed photovoltaic power forecasting in real-world operation.