A comparative framework for evaluating machine learning models in forecasting electricity demand for port microgrids

This study presents a framework for forecasting electricity demand in port microgrids using advanced machine learning models, including Random Forest, Least Squares Boosting Ensemble, and Gaussian Process Regression. These models were evaluated under different forecasting setups (fixed origin, expanding windows, and rolling windows) and compared against simpler baseline methods, such as Linear Regression and Naive models. The study assessed the effectiveness of machine learning models in handling dynamic electricity demand patterns in port environments and highlighted the advantages of data-driven models. Results indicate that the Random Forest (expanding window) model outperforms the other models, achieving a root mean square error of 1.1848 MW and a mean average percentage error of 7.2483 %. Gaussian Process Regression with Exponential kernel follows closely with a root mean square error of 1.1904 MW and a mean average percentage error of 7.5017 %. In contrast, the Naive Method (previous day) shows the poorest performance with a root mean square error of 4.5357 MW and a mean average percentage error of 18.1485 %. Partial Dependence Plots reveal that features such as weighted port calls play a significant role in improving prediction accuracy. These findings highlight the effectiveness of machine learning models in accurately forecasting port microgrid demand and optimizing energy management. • Developed a machine learning framework to forecast port electricity demand. • Framework evaluated using Random Forest, Least-Squares Boosting, and Gaussian Process Regression. • Random Forest achieved best accuracy (RMSE: 1.18 MW, MAPE: 7.25%) over baseline models. • Improved model interpretability by analyzing key features with Partial Dependence Plots. • Validated the framework’s practical use with real-world port data.