Recurrent Neural Network Based Model Development for Energy Consumption Forecasting

Abstract

The world population is increasing day by day. As a result, limited resources are decreasing day by day. On the other hand, the amount of energy needed is constantly increasing. In this sense, decision makers must accurately estimate the amount of energy that society will require in the coming years and make plans accordingly. These plans are of critical importance for the peace and welfare of society. Based on the energy consumption values of Germany, it is aimed at estimating the energy consumption values with the GRU, LSTM, and proposed hybrid LSTM-GRU methods, which are among the popular RNN algorithms in the literature. The estimation performances of LSTM and GRU algorithms were obtained for MSE, RMSE, MAPE, MAE, and R2 values as 0.0014, 0.0369, 6.35, 0.0292, 0.9703 and 0.0017, 0.0375, 6.60, 0.0298, 0.9650, respectively. The performance of the proposed hybrid LSTM-GRU method, which is another RNN-based algorithm used in the study, was obtained as 0.0013, 0.0358, 5.89, 0.0275, and 0.9720 for MSE, RMSE, MAPE, MAE and R2 values, respectively. Although all three methods gave similar results, the training times of the proposed hybrid LSTM-GRU and LSTM algorithms took 7.50 and 6.58 minutes, respectively, but it took 4.87 minutes for the GRU algorithm. As can be understood from this value, it has been determined that it is possible to obtain similar values by sacrificing a very small amount of prediction performance in cases with time limitations.