Performance Evaluation of Deep Learning Models for Photovoltaic Power Prediction across PV Systems in China

Accurate prediction of photovoltaic (PV) power generation is essential for incorporating solar energy into the existing power grid system to achieve effective renewable energy operations. The uncertain nature of meteorological data such as solar radiation and temperature makes accurate prediction more difficult. In this paper, five different deep learning frameworks Artificial Neural Network (ANN), Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) are compared against each other. The framework is formulated as a regression and nowcasting problem, mapping meteorological inputs at time t to PV power output at time t. During operational deployment, this framework can be coupled with Numerical Weather Prediction (NWP) data to generate PV power forecasts up to 15 days ahead. The performance of the models is evaluated using actual operational data from monocrystalline and thin-film PV technologies across five different geographical regions. The performance is measured in terms of Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), coefficient of determination (R²), and Mean Bias Error (MBE). The results show that recurrent and fully-connected neural network models, such as RNN and ANN, produce more robust predictions compared to others. Although CNN proves to be successful in many applications, its inability to capture long dependencies limits its use for forecasting PV generation, which is treated as a one-dimensional time series mapping. Despite usually producing smaller MAPE values, GRU can exhibit larger RMSE values under certain conditions, which underlines the necessity of using different metrics when measuring prediction success. In general, all models underestimate PV production during periods of high irradiation, being largely affected by specific weather conditions. The presented paper contributes to the field by providing a dual nowcasting and multi-day forecasting evaluation framework for deep learning models under realistic technological and data constraints.