A fully co-optimized hybrid approach for short-term PV power forecasting that addresses the challenges of diurnal trends and residuals

In recent years, intra-day forecasts of PV power generation are becoming increasingly crucial for the optimal operation of modern power systems. In the absence of satellite images and Numerical Weather Predictions, data-driven PV power forecasting models solely rely on on-site numerical measurements, which have been proven sufficient for forecasting horizons up to two hours. This paper introduces a novel, data-driven, hybrid PV power forecasting model for the generation of one- hour ahead forecasts at a minute-scale resolution. The original PV power time-series is decomposed into a diurnal trend and residual components, which are separately handled by Long Short-Term Memory networks and Random Forests, respectively. The proposed model is fully co-optimized using Ant-Lion Optimization, to improve the forecasting accuracy while maintaining the scalability of the overall approach. An ablation study is conducted to assess the performance of the proposed model on real historical data under different weather conditions. The proposed model outperforms all benchmark models under more challenging cloudy conditions, reducing the forecasting error by up to 5.71% and 7.94% regarding the RMSE and the absolute percentage error, respectively, while maintaining the computational efficiency at low levels.