Occupancy-Aware Spatio-Temporal Building Energy Forecasting with a Hybrid Long ShortTerm Memory and Graph Neural Network Benchmark Using Public Datasets

Accurate short-term forecasting of building energy demand is complicated by coupled temporal dynamics, cross-meter spatial effects, and occupancy-driven variability. We present an occupancy-aware spatiotemporal framework that uses a Long Short-Term Memory (LSTM) branch and a Graph Neural Network (GNN) branch, augmented with calibrated occupancy probabilities transferred from labeled sources to public corpora lacking occupancy labels. Using BDG2 and ASHRAE GEPIII, we construct physical, correlation kNN, and learned kNN graphs; engineer calendar– weather–lag/rolling features; and evaluate with forward-chaining splits across horizons t+1…t+24. Primary (MAE, RMSE, MAPE) and domain metrics (CVRMSE, NMBE) follow ASHRAE Guideline 14. The hybrid attains RMSE 2.766 kWh (BDG2) and 2.740 kWh (ASHRAE GEPIII), yielding 33.44% and 33.52% reductions versus a ridge/XGBoost baseline, and statistical parity with LSTM-only (ΔRMSE −0.23% on BDG2; +0.02% on ASHRAE GEPIII; paired tests p<0.05). Horizon-wise curves show stable gains—especially during business hours—and learned kNN typically provides the lowest average error. Per-meter distributions indicate 100% of meters satisfy CVRMSE ≤ 30% and ∣NMBE∣ ≤ 10%, supporting calibration and retro-commissioning use. These findings demonstrate that using temporal and graph-based spatial cues with transferable occupancy signals delivers robust, label-efficient multi-meter forecasting, with units standardized (kWh, °C) and |NMBE| consistently denoted for clarity.