Urban Microclimate Forecasting System
This is a research project I led as part of the Georgia Tech VIP-SMUR (Vertically Integrated Projects — Sustainable Microclimate Urban Research) group. The goal: predict temperature and relative humidity at hyper-local resolution — not city-scale or even block-scale, but at roughly 6-meter spacing across the entire Georgia Tech campus (~3.5 km²), using nothing more than 16 weather stations and publicly available geospatial data.
Model Architecture
The architecture combines four components chosen specifically for the microclimate forecasting problem: adaptive feature gating, learned input weighting, parallel temporal processing at two scales, and multi-head attention for long-range dependencies.
Core Components
Key Design Decisions (with Rationale)
1. Multi-scale parallel LSTM branches. Rather than a single sequential LSTM, we use two parallel branches: a short-term branch (1 LSTM layer, hidden_dim/2) and a long-term branch (2 LSTM layers, hidden_dim/2), concatenated back to the full dimension. Rationale: microclimate evolution has two overlapping timescales — rapid fluctuations (a cloud passing, a building shadow) on minutes, and diurnal cycles (sunrise heating, nighttime cooling) on hours. A single LSTM collapses both; parallel branches preserve them separately.
2. Single-step point prediction. Given the past 24 hours (144 timesteps × 10 min), predict one timestep ahead (+10 min). For longer forecasts, iteratively roll — feed prediction back as input. This avoids distribution mismatch beyond the training horizon and produces cleaner inputs for downstream Regression Kriging (no multi-horizon uncertainty propagation).
3. No station identifiers. No station IDs anywhere in the input. This forces the model to learn location-independent patterns (physics-based features), enabling generalization to unseen stations — validated by our held-out station protocol.
Input Features (12 dimensions, physics-aware)
| Category | Features | Encoding |
|---|---|---|
| Weather observations | Temperature, Relative Humidity | Raw (autoregressive signal) |
| Solar geometry | Altitude angle, Azimuth angle | Sin/cos trigonometric |
| Hour of day | 24-hour diurnal cycle | Sin/cos (period = 24h) |
| Day of year | 365-day seasonal cycle | Sin/cos (period = 365d) |
| Minute | Sub-hourly resolution | Sin/cos (period = 60 min) |
All temporal features encoded as sin/cos pairs to avoid discontinuities (hour 23 → hour 0). Solar angles encoded trigonometrically so the model receives physically meaningful sun-position signals rather than raw degree values.
Data Engineering
~947,000 warm-season observations (April–September 2015–2019) from 16 stations at 10-minute intervals. The preprocessing pipeline:
Scaling: Features use RobustScaler (median-centered, IQR-normalized — resistant to sporadic outliers). Targets use MinMaxScaler (range [0,1]). Both fit on training data only.
Training: AdamW (β₁=0.9, β₂=0.999) + weight decay 1e-3, initial LR 1e-3, ReduceLROnPlateau (factor=0.5, patience=7), gradient clipping max_norm=1.0, early stopping patience=20. Xavier uniform init for linear layers; orthogonal init for LSTM (critical for preventing vanishing gradients in the 2-layer long-term branch). Trained in ~23 minutes on A100 GPU (Phoenix cluster).
Results
Held-Out Station Evaluation
| Variable | RMSE | MAE | MAPE | R² | Max Error |
|---|---|---|---|---|---|
| Temperature | 0.43°C | 0.29°C | 1.15% | 0.9928 | 6.98°C |
| Relative Humidity | 1.28% | 0.90% | 1.51% | 0.9952 | 22.23% |
Overfitting Analysis
| Variable | Train RMSE | Test RMSE | Test/Train Ratio |
|---|---|---|---|
| Temperature | 0.41°C | 0.43°C | 1.057 ✓ |
| Relative Humidity | 1.19% | 1.28% | 1.074 ✓ |
Both ratios below 1.10 — excellent generalization. Ratios below 1.10 = excellent; 1.10–1.20 = good; above 1.20 = potential overfitting.
Feature Importance (Permutation-Based)
| Rank | Feature | RMSE Increase When Removed |
|---|---|---|
| 1 | Relative Humidity | +2789% |
| 2 | Temperature | +372% |
| 3 | Hour (sin) | +7.0% |
| 4 | Solar altitude (sin) | +6.5% |
| 5 | Hour (cos) | +2.4% |
| 6 | Solar azimuth (sin) | +2.0% |
Removing solar angles alone degrades RMSE by 2–6% — confirming the model uses sun position to predict temperature evolution, not just correlations with time-of-day.
Geospatial Inference: Stations → Campus Maps
The model outputs predictions at 16 discrete station locations. To predict at every campus point, I use Regression Kriging (PyKrige) — combining Random Forest regression on spatial covariates with kriging of the residuals.
9 real GIS covariates from OpenStreetMap + elevation data:
- Distance features: Distance to nearest building, park, library, parking lot, footway
- Elevation features: Ground elevation, building height, total elevation
- Shadow features: Monthly shadow ratio (selected by target timestamp's month — computed from building geometry and solar angles)
After kriging, Random Forest feature importances are extracted to analyze which spatial characteristics most influence microclimate variation — connecting predictions back to actionable urban planning insights.
Extreme Scenario Mapping
An automated scenario generator scans the full warm-season dataset, identifies extreme-condition timestamps (hot/cold/dry/humid, filtered for ≥10 stations reporting), and runs the full inference → kriging → visualization pipeline for each:
Each comparison map shows side-by-side temperature and humidity predictions with station observations overlaid as square markers. All output maps are 300 DPI (~1.7–2.1 MB PNG) at 4800×3600 px resolution.
HPC Integration
| Stage | Time | Resources |
|---|---|---|
| Training (16 stations) | ~26 min | A100 GPU, 34 GB RAM |
| Inference (100,283 points) | ~2 min 10 sec | A100 GPU |
| Visualization (3 maps, 300 DPI) | ~1 min 31 sec | CPU |
| All 4 extreme scenarios | ~3 min 4 sec | A100 GPU |
Full SLURM job scripts for each stage, plus a comprehensive HPC tutorial I wrote for onboarding other VIP-SMUR team members (VPN setup, SSH, job submission, file transfer).
What I Learned
The most valuable lesson was about model design as scientific hypothesis testing. Every architectural decision encodes an assumption: parallel LSTM branches assume multiple temporal scales matter; removing station IDs assumes the model learns location-independent physics; using solar angles assumes that sun position causally drives temperature evolution. The performance metrics then confirm or reject each assumption. When removing solar angles degraded RMSE by 6.5%, that confirmed the assumption. This feedback loop — where the model is simultaneously a prediction tool and a scientific instrument — fundamentally changed how I think about model engineering.