Data-driven hybrid SARIMAX-MLP framework for energy consumption prediction in residential micro-grid
Ibrahim Ali Kachalla, Christian Ghiaus, Adeniran Kolade Ademuwagun, Olufemi Babajide Odeyinde, Majid Baseer
Abstract
With electric water boilers (EWBs) energy consumption increasing by over 30% due to global warming, an accurate load demand prediction model incorporating exogenous factors, such as hot water consumption, seasonal variations, and time of day, can mitigate energy efficiency, energy resilience, and optimisation challenges in micro-grid management. This, in turn, reduces dependence on carbon-intensive grids and enhances model predictive control (MPC) strategies for EWBs scheduling in residential applications. This study proposes a hybrid SARIMAX-MLP model that integrates exogenous variables to improve energy consumption forecasting. While SARIMAX captures linear dependencies and seasonal trends, MLP model's complex nonlinear relationships that SARIMAX cannot estimate. A case study of two residential blocks, with one year six months (18 months) of energy consumption data, was utilised to evaluate the model's prediction accuracy using performance metrics (RMSE, MAE, R 2 ) and computational cost. The proposed hybrid model was compared with statistical learning models (ARIMA, ARIMAX, SARIMAX), machine learning models (MLP, SVR), and other hybrid approaches (ARIMA-MLP, ARIMA-SVR, SVR-MLP). Results show that SARIMAX-MLP outperforms both single and hybrid models, offering higher accuracy and robustness across different time horizons. SHAP analysis depicts hot water consumption as the most influential variable, subsequently hour of the day. Regardless of its higher computational cost, the SARIMAX-MLP model achieves a balanced trade-off between accuracy and computation efficiency, making it suitable for real-time energy management and MPC applications residential micro-grid. This study highlights the importance of exogenous variables in predictive modelling and suggests future exploration of deep learning architectures, such as 1D CNNs and Transformers, to further enhance forecasting performance and computational efficiency.