Urban Microclimate Forecasting with Deep Learning

Overview

Designed a deep learning architecture for urban microclimate forecasting, combining multi-head attention and multi-scale LSTM to process 10 years of high-resolution environmental data for Georgia Tech campus.

Key Technologies

• Deep Learning: Temporal Fusion Transformer (TFT), Multi-head Attention, Multi-scale LSTM
• Geospatial ML: Regression Kriging, Random Forest
• Data Processing: 1.5M+ 10-minute observations, Physics-informed features

Achievements

• Processed 1.5M+ observations (10 years) with TFT-inspired deep learning architecture
• Generated 100K+ high-resolution grid-level predictions across Georgia Tech campus (~3.5 km²)
• Enabled robust 24-hour temporal modeling with strong seasonal generalization
• Preparing research paper for submission to top-tier SCI Q1 journal (IF ≈ 7.1)

Technical Innovations

• Physics-Informed Features: Solar-angle and cyclic-time encodings for temporal modeling
• Multi-Scale Architecture: Combined multi-head attention with multi-scale LSTM
• Spatial Interpolation: Regression Kriging with 9 geospatial covariates (LULC, elevation, shadow ratio)
• High-Resolution Output: Grid-level predictions for urban planning and climate analysis

Applications

• Urban planning and development
• Climate resilience assessment
• Energy optimization for buildings
• Environmental policy decision support