Gradient-Descent based Short-Term Electricity Demand Forecasting: A Comparison with SARIMAX
TL;DR
A comparative analysis of two independent forecasting approaches: a statistical Seasonal AutoRegressive Integrated Moving Average with eXogenous variables (SARIMAX) model and a feature-enhanced gradient descent (GD) based learning model demonstrates that the gradient-based learning approach captures short-term load variations more effectively than the traditional statistical model.
Abstract
Accurate short-term load forecasting (STLF) is essential for efficient energy management and grid stability. This study presents a comparative analysis of two independent forecasting approaches: a statistical Seasonal AutoRegressive Integrated Moving Average with eXogenous variables (SARIMAX) model and a feature-enhanced gradient descent (GD) based learning model. The objective is to evaluate their predictive performance for short-term electricity load forecasting. Both models were trained and tested on the same dataset, and their forecasting accuracy was assessed using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). A simple mean-based baseline model was also implemented to provide a reference for performance comparison. The results indicate that the gradient descent model achieves superior forecasting accuracy outperforming and the baseline model. The results demonstrate that the gradient-based learning approach captures short-term load variations more effectively than the traditional statistical model. Overall, this study highlights the effectiveness of data-driven optimization-based learning models for short-term load forecasting applications and provides a structured performance comparison between statistical and machine learning approaches.