A Hybrid machine learning–statistical based method for short-term energy consumption prediction in residential buildings

Abstract

Accurate short-term load forecasting is essential for modern power systems, enabling efficient energy management and supporting grid reliability amid increasing demand and variable weather conditions. This study addresses the challenge of forecasting household electricity consumption by proposing SSRXLR—a novel hybrid method that integrates statistical and machine learning techniques including a sparse, Seasonal Autoregressive Integrated Moving Average Exogenous model, Random Forest, Extreme Gradient Boosting, Long Short-Term Memory, and a Residual Correction step to leverage both linear trends and complex nonlinear relationships. We have analyzed one year of high-resolution (5-minute interval) energy and weather data from a household in Las Vegas, Nevada. Through a rigorous feature selection process, we have identified the four most influential features, i.e., sea level pressure, temperature, feels-like temperature, and dew point. The proposed method has demonstrated strong prediction performance across multiple metrics. Compared to well-known models, the proposed method achieved a root mean square logarithmic error of 0.043, which surpassed the Random Forest method by 0.066 and the Seasonal Autoregressive Integrated Moving Average Exogenous model by 0.106 in reducing the Root Mean Square Logarithmic Error (RMSLE). The coefficient of determination for the proposed method attained a 0.97 value, outperforming Random Forest (0.92) and the Seasonal Autoregressive Integrated Moving Average Exogenous model (0.67). These results highlight the effectiveness of combining advanced statistical modeling, machine learning, and targeted feature selection for precise short-term load forecasting. The proposed framework offers a scalable solution for smart grid operations, resource planning, and integration of renewable energy in diverse environments.