CROSS-CLIMATE ANALYSIS OF FEATURE IMPORTANCEFOR SHORT-TERM BUILDING ENERGY FORECASTING

Abstract

The paper tests the hypothesis that the feature importance structure for short-term building energy forecasting is independent of climate zone. A unified computational pipeline covering load lags, weather forecasts, cyclical, calendar, and other feature groups was applied to two open datasets with contrasting climatic conditions: the Norwegian COFACTOR Drammen dataset, comprising 45 public buildings with heating-dominated loads, and the Californian CityLearn dataset, comprising 15 residential buildings with cooling-dominated loads.Three models of increasing complexity were compared, ranging from a naive baseline to gradient boosting on decision trees. Interpretation was performed using SHAP, which quantifies each feature's contribution to the forecast. Statistical significance of model differences was verified using the Diebold–Mariano test. Additional experiments on cold start and feature ablation were conducted to assess the minimum required training data and optimal number of features.The feature group ranking proved stable regardless of climate: load lags dominate in both cases, while weather forecasts play a comparatively secondary role. At the same time, forecast accuracy differs substantially: heating-dominated buildings are predicted much more accurately than cooling-dominated ones. Unexpectedly, a simple linear model outperformed gradient boosting for cooling loads, an outcome attributed to the more complex model overfitting the stochastic component of cooling demand. Practical recommendations are provided: three to six months of data suffice for heating-dominated buildings, and a moderate reduction in the feature set based on importance rankings preserves most of the model's accuracy. The results are directly applicable to the Ukrainian context, where the climate is similar to Norway's.