ENHANCING SHORT-TERM SOLAR GENERATION FORECASTING ACCURACY USING A CNN-TRANSFORMER MODEL

Abstract

The increasing penetration of solar energy into power systems poses significant challenges to grid dispatch and operation due to its inherent variability and strong dependence on meteorological conditions. Recently, Transformer-based deep learning models have demonstrated great potential in time-series analysis owing to their self-attention mechanism. However, the vanilla Transformer architecture tends to prioritize long-term dependencies, which may result in an insufficient capture of local, short-term fluctuations in meteorological data. This study proposes a hybrid CNN-Transformer model for short-term solar power forecasting. A one-dimensional convolutional neural network (1D-CNN) is employed to extract local spatio-temporal features, which are subsequently fed into the Transformer block to learn long-term dependencies. Experimental results conducted on a real-world dataset from a 50MW solar power plant indicate that the CNN-Transformer model achieves superior performance. Specifically, the model significantly reduces the error metrics, with an RMSE of 2,919.42 kW and an NMAPE of 3.45%, consistently outperforming the vanilla Transformer model.