Interpretable Machine Learning for Municipal Residential Gas Consumption: Integrating Climate, Building Stock, and Socioeconomic Drivers

Dutch municipalities lead the phase-out of residential natural gas, and need to know why gas use differs between them. We merge four Dutch open datasets on climate, building stock, and socioeconomic conditions into a panel of 2905 municipality–years (408 municipalities, 2016 to 2023). A global Random Forest estimated annual gas use per dwelling from 14 predictors, without past consumption, and SHapley Additive exPlanations (SHAP) was used to rank them. On the held-out 2020 to 2021 test set, the Random Forest model reached an R2 of 0.78 (95% CI 0.73 to 0.81) and a mean absolute percentage error of 7.6%. A persistence baseline that repeats each municipality’s last known value was more accurate (0.86 and 5.9%), so the model’s value is explanation, not prediction. For municipalities left out of training, R2 fell to 0.62. The strongest predictors were average floor area, heating degree days, population density, and the pre-1975 dwelling share, holding 73.6% of the SHAP importance. Building stock variables carried most of the accuracy; climate did not improve it. In the 2022 to 2023 energy crisis, the predicted level broke down (R2 = −5.83) but the municipal ranking held (Spearman 0.92 in 2022 and 0.85 in 2023). Gas use above the municipal median amounts to 0.48 Mt CO2-eq per year, linking the results to SDG targets 7.3, 11.3, and 13.2.