Abstract
Spatially and sectorally disaggregated gas demand projections are essential for infrastructure planning and decarbonisation, yet records of consumption and its drivers often span only a few years, which is insufficient for conventional time-series methods and risks overfitting in complex machine learning models. Pooling observations across regions can provide the sample size and variation in demand and its drivers these models need. We test this approach on residential natural gas demand in Turkey’s 81 provinces, comparing Random Forest, CatBoost, Support Vector Regression and linear regression with fixed effects for projections to 2050 under five Shared Socioeconomic Pathways (SSPs). In cross-validation, the machine learning models roughly halved linear regression’s relative error, but tree-based models stay largely within the range of historical consumption, flattening projections for the highest-consuming provinces. Model choice shapes the 2050 projections more than the socioeconomic pathway. Under SSP2, the four models span 74 TWh, whereas for any one model, the five scenarios, which differ only in their population and GDP trajectories, span 4–17 TWh. Past consumption is the strongest predictor in the machine learning models. Among the other drivers, population and household composition outweigh climatic variables such as degree days. Total residential gas demand in Turkey nearly doubles from 209 TWh in 2025 to 412 TWh in 2050, with models and scenarios ranging from 358 to 463 TWh. Under the no-new-policy baseline, direct emissions reach 67–80 MtCO