Long Short-Term Memory Recurrent Neural Network (LSTM-RNN) Power Forecasting

Abstract

The geographical position of the Kingdom of Saudi Arabia has significant potentials for utilizing renewable energy resources, which aligns with the country&#x0027;s vision for 2030. This paper proposes a solution to achieve energy sustainability by forecasting future load demands through adopting three different scenarios. We used the outsourced Individual Household Electric Power Consumption Dataset, University of California-Irvine repository, for testing our proposed system. We utilized the Long Short-term Memory-Recurrent Neural Network (LSTM-RNN) algorithm to estimate the whole house power consumption for different horizons: every 15 minutes, daily, weekly, and monthly. Next, we evaluated the performance of the system by Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Square Error (RMSE), and <tex>$R^{2}$</tex> score metrics. Then, we applied the Mean Absolute Percentage Error (MAPE) to find its accuracy. The results showed that the monthly forecasting interpretation scenario was the best performing model. That scenario used (n-1) months for training and the last month for testing. The scores for that model were 0.034 (MAE), 0.001 (MSE), 0.034 (RMSE), and 97.16&#x0025; (accuracy). The constructed model successfully achieved its goals of predicting the active power of the household and now can be accommodated on energy applications not only in Saudi Arabia but also in any other country.