Energy usage prediction with ensemble graph convolutional recurrent neural networks: case study of Greek islands

This study investigates the use of graph convolutional and recurrent neural networks to forecast energy consumption on five Greek islands. The model is trained on historical energy usage data, using graph convolutional networks (GCNs) with graphs constructed from power production and their geographical location which their results are averaged. The evaluation metrics mean absolute error (MAE), mean squared error (MSE), and mean absolute percentage error (MAPE) were utilised to measure the accuracy of the proposed prediction algorithms. The results show that all the suggested EGCRNN models with LSTM, Bi-LSTM, and GRU demonstrated excellent performance, strong model fit, and enhanced predictions of energy consumption. In particular, models which utilise long-short term memory (LSTM) and gated recurrence units (GRU) followed with better prediction in one model each, Thira and Rhodes, respectively.