Electrical Load Prediction Using Statistical, Deep Learning, and Hybrid Time Series Models

Abstract

The efficient operation, planning and reliability of modern power systems rely on accurate electrical load forecasting. Accurate demand forecasting allows for optimal use of resources by cutting unnecessary expenditures and improving grid stability. Since electricity usage is particularly volatile and nonlinear in time, conventional forecasting methods frequently have trouble identifying the complex temporal patterns. Finally, this study examines the performance of models based on time-series and deep learning for electrical load forecasting over several time horizons. Several statistical models (Autoregressive Integrated Moving Average — ARIMA and Seasonal ARIMA — SARIMA) are compared to a Long Short-Term Memory (LSTM) neural network to evaluate their forecasting performance. In addition, a hybrid ARIMA–LSTM model is presented to jointly reflect linear, seasonal, and nonlinear features of load series. The temporal and climatic-enhanced historical electricity consumption data are preprocessed and subject to stationarity tests and data consistency checks. Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE) are used to assess the performance of model in short-, medium- and long-term forecasting, respectively. From the experimental results, we show that the hybrid ARIMA–LSTM model enables to always outperform individual models, producing the lowest prediction errors over all forecasting horizons. The results show that incorporating classical statistical method with deep learning methods can lead to better accuracy and robustness in this electrical load prediction problem.