CoolShift: Lightweight modeling of building cooling demand with causal machine learning

We present CoolShift, a lightweight CausalML framework for counterfactual cooling-demand prediction under indoor setpoint interventions. CoolShift estimates condition-specific effects (CATE) via double machine learning with compact covariates, then composes effects into building-level counterfactuals—supporting fast “what-if” screening and aggregation to city-scale impacts. To evaluate both levels and effects, we build a quasi-random simulation corpus and the Setpoint–Shift Benchmark (SSB) with seen/unseen splits across heterogeneous buildings. CoolShift outperforms strong non-causal baselines (LightGBM, XGBoost, TabNet), maintaining low error on counterfactual levels (MAE $\approx$ 0.023–0.024) and accurate effects (MAE $\approx$ 0.028) on both splits, while baselines collapse on unseen effects (negative $R^2$). Results show that explicitly estimating causal effects, rather than differencing level predictors, is key for intervention queries and out-of-distribution generalization, enabling rapid portfolio- or city-scale DSM assessments.