Regime-Aware Seasonal Physics-Informed Neural Networks for Long-Horizon Building Temperature Prediction

Commercial buildings consume 40–60% of building energy via HVAC Goldfeder et al. [2024], making accurate long-horizon temperature prediction critical for carbon reduction. We present Seasonal-PINN, extending a hybrid XGBoost–PINN framework Saha and Shinde [2026] with data-driven thermal regimes on the Google Smart Building SB1 dataset (123 zones, 93,858 sq. ft., 32.2M sensor readings across 2022–2024). Regime Discovery. k-means clustering (k=3, silhouette = 0.515) identifies Cold (Jan–Mar, Nov–Dec), Transition (Apr–May), and Warm (Jun–Oct) regimes from annual temperature swings of 6.9°F. Inter-zone variance (σCold=3.09°F vs. σWarm=2.14°F) justifies regime-specific physics weights (λCold=0.35, λTrans=0.20, λWarm=0.12) replacing the original fixed λ=0.1 Saha and Shinde [2026]. Winter–Summer inter-zone correlation stability (r=0.777) further motivates regime-specific adjacency matrices (A1 for Cold/Transition, A2 for Warm) capturing seasonal shifts in thermal coupling between zones. Imputation and Weighting. 40% of 32.2M readings are missing; physics-guided LightGBM imputation Arisaka et al. [2025] achieves 0.258°F MAE on high-quality partitions (2022a/b), degrading gracefully to 1.64°F on the low-quality 2024 test set (∼85% imputed). V1 uses regime λ with equal weights; V2 adds imputation-aware sample weights (w=0.3 imputed, w=1.0 real), reducing Cold Stage-4 MAE by 71.4% (0.908→0.260°F) and demonstrating that imputation quality propagates directly into downstream prediction accuracy. Architecture. An 81-feature pipeline (zone-specific lags, setpoint gaps, time encodings, adjacency ∆T, exogenous sensors) feeds a sequential XGBoost–PINN. Our 9-stage pipeline progressively scales within each regime: S1–S2 (1-day, 1-zone), S3 (1-week, 1-zone), S4–S7 (1-week to 1-month, all 123 zones), S8–S9 (1-cycle, full-regime), with S10 (1-year) and S11 (2-year) cross-regime chaining evaluating long-horizon generalisation. The PINN activates only when ≥1,000 real samples exist, otherwise falling back to XGBoost alone, with regime warm-starting (Cold→Transition→Warm) transferring cross-regime knowledge, and an ensemble variant (w · XGB + (1−w) · PINN) stabilising predictions under high imputation rates. Results. Table 1 reports MAE across all stages and regimes. V2 delivers the largest gains in short-horizon and imputation-heavy settings (Cold S4: −71.4%), while destabilising data-scarce single-zone training (S3), empirically motivating the ≥1,000-sample PINN activation threshold. On cross-regime chaining, V1 achieves 95.5%/93.4% improvement on 1-yr/2-yr horizons (0.127/0.188°F vs. 2.826°F baseline); V2 achieves 94.2%/89.5% (0.163/0.296°F). A Cold setpoint gap of −5.83°F quantifies direct climate impact via reduced heating waste across all 123 zones.