Forecasting Natural Gas Consumption by User Type Using Machine Learning: A Comparative Study

This study aims to develop user-type-specific machine learning models to forecast natural gas consumption for residential and commercial user groups in İzmir, Turkey. Multiple Linear Regression, Random Forest, LightGBM, and XGBoost algorithms were implemented, and model performance was enhanced through hyperparameter optimization. The models were evaluated using MAE and RMSE metrics. Results indicate that LightGBM and Random Forest provided the most accurate forecasts overall, with LightGBM performing best in the residential group and Random Forest slightly outperforming others in the commercial group. In contrast, MLR underperformed due to the non-linear nature of the data. Residential consumption patterns were found to be more predictable, leading to lower error rates, whereas the commercial group exhibited higher variability and forecast challenges. The study highlights the importance of distinguishing user types and employing well-tuned machine learning algorithms for improved energy demand forecasting.