Forecasting Energy Consumption Using Time Series Models : A Comparison of Naïve, ARIMA, and Auto-ARIMA Approaches

Abstract

Accurate forecasting of energy consumption is critical for optimizing resource management and supporting sustainable development. This paper presents a comparative analysis of forecasting models for household energy consumption, evaluating the performance of simple naïve methods (Daily, Weekly, and Year-Over-Year persistence) against established statistical models (ARIMA and Auto-ARIMA). The study utilizes a real-world household power consumption dataset, employing data preprocessing, normalization, and a standardized training-testing split. Performance is rigorously evaluated using the Root Mean Square Error (RMSE) metric. The results demonstrate a clear hierarchy in forecasting accuracy: naïve models exhibit the highest error (RMSE 543.82–638.93), followed by the manually configured ARIMA model (RMSE 160–200). The Auto-ARIMA model, which automates parameter optimization, achieves the lowest error (RMSE 100–120), confirming its superior capability in capturing complex temporal patterns. This study concludes that while naïve models offer simplicity, advanced time series models like Auto-ARIMA are essential for accurate forecasting, providing a reliable tool for informed energy planning and efficiency strategies. Future work should focus on integrating exogenous variables and exploring deep learning architectures to further enhance predictive performance.