Convolutional Neural Network (CNN) Extended Architectures for Photovoltaic Power Production Forecasting

The incorporation of PV technology in the energy mix of countries helps in bringing economic and environmental benefits. However, the operation of the power system may be fraught with many challenges owing to the randomicity of PV power output which is highly dependent on weather conditions. The development of highly accurate forecasting models for photovoltaic power generation is in great demand to ensure grid stability and high-quality electricity for end consumers. Motivated by recent development in deep learning models and their promising performance in different time series applications, two extended architectures of the convolutional neural network CNN were proposed for short-term predictions of a PV plant in this paper. The proposed models are utilized for one-day and two-days-ahead PV power forecasting. Two datasets were used a univariate one containing only power production records, and a multivariate dataset with weather variables and power output records. The performance of the proposed models is tested with a case study using an actual dataset collected from Rabat, Morocco. The values of the three error metrics, MAE, MAPE, and RMSE, show that the extended versions of the CNN model exhibit upper performance in terms of prediction accuracy and stability.