Short-term PV plant power forecasting model using CNN-GRU with XGBoost algorithm

Abstract

Accurate short-term forecasting of photovoltaic (PV) power output is crucial for the efficient integration and operation of PV systems in smart grids. However, the intrinsic unpredictability and hierarchical discontinuity of PV data complicate the attainment of good predictive accuracy through conventional regression procedures. This research introduces a combined deep learning and gradient boosting model for short-term PV power forecasting. A convolutional neural network (CNN) is initially utilized to get important spatial features from input variables. Then, these features are coupled with historical PV power measurements in a gated recurrent unit (GRU) to capture temporal dependencies. The XGBoost technique is also used to represent complex nonlinear relationships in the data, which makes predictions even more accurate. The model is tested using actual PV plant data and gets a R² of 0.79, an RMSE of 23.30 MW, and a MAPE of 16.68%. These improvements show that the hybrid model works well to make predictions more accurate and reliable for real-world use.