Seasonal Forecasting of Global Horizontal Irradiance for Grid-Connected PV Plants: A Combined CNN-BiGRU Approach

Abstract

The quest for environmental sustainability in power systems necessitates the incorporation of renewable energy sources into the grid infrastructure. Among these renewable sources, solar energy has risen to prominence due to its widespread availability. However, the variable nature of solar irradiance poses challenges in operational and control aspects of its integration. A potential solution lies in predictions of global horizontal irradiance (GHI). This study introduces an ensemble deep learning-based forecasting approach, leveraging a Convolutional Neural Network and Bidirectional Gated Recurrent Unit (CNN-BiGRU). The efficacy of this approach is evaluated against three ensemble models: The Convolutional Neural Network Bidirectional Long Short Term Memory (CNN-BiLSTM), Convolutional Neural Network Gated Recurrent Unit (CNN-GRU), the Convolutional Neural Network Long Short Term Memory (CNN-LSTM). The comparative analysis is centered on seasonal GHI forecasting in Alice Springs, Australia, with a 1-hour time horizon. Four metrics are employed to gauge the accuracy of the models: coefficient of determination (R2), mean absolute error (MAE), normalised root mean square error (nRMSE), and root mean square error (RMSE). The findings reveal that the proposed ensemble bidirectional model outperforms its counterparts in all seasons. Specifically, in terms of seasonal forecasting, the CNN-BiGRU model achieves a maximum nRMSE of 0.0955, indicating its superior performance.