Hybrid Prophet-NAR Model for Short-Term Electricity Load Forecasting

Electricity load forecasting is crucial for effective energy management, particularly in minimizing energy production and distribution costs. Traditional models like SARIMA and Singular Spectrum Analysis (SSA) have been widely used but often need to capture complex nonlinear patterns and deal with data uncertainties. This study aims to develop and evaluate a hybrid forecasting model that combines the Prophet model with a Neural Network Autoregressive (Prophet-NAR) model, referred to as PropNAR. The objective is to enhance the accuracy of hourly electricity load forecasting in Malaysia by addressing the limitations of existing models. The proposed PropNAR integrated the strengths of the Prophet model in capturing deterministic structures, such as trends, seasonality, and holiday effects with the NAR model's ability to handle nonlinear stochastic relationships. Additionally, SSA-based bagging is employed to manage data uncertainties, and ensemble techniques are applied to further refine the forecasting accuracy. The hybrid PropNAR model demonstrated significant improvements in forecasting accuracy. Specifically, it reduced Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE) by approximately 21%-86% compared to the standalone Prophet model. On the Malaysian electricity load datasets, the PropNAR model achieved MAE values ranging from 527.26 to 1023.78, RMSE values from 752.70 to 1498.54, and MAPE values from 1.12% to 2.04%. These results indicate a substantial enhancement over SARIMA and SSA-NAR in handling outliers and data variability. The proposed hybrid PropNAR model offers a robust solution for Malaysian short-term electricity load forecasting, outperforming conventional models in accuracy and reliability.