Low-Cost Aerial Wind Sensing with Multirotor Drones
Physics-Informed Neural Networks · Field Deployment · Flow Reconstruction
Problem
Conventional wind measurement (met masts, lidar, weather stations) is sparse, expensive, and stationary. Drone-based sensing could fill gaps at lower cost — but extracting ambient wind from drone telemetry is non-trivial due to rotor downwash, attitude coupling, and calibration drift.
Approach
Low-cost multirotors equipped with IMU, GPS, and reference anemometers. Wind speed and direction are estimated from vehicle attitude response, power draw, and wind-relative direction using calibrated regression, then a PINN reconstructs the local wind field with mass-conservation and smoothness constraints.
Methods
SDG 7 — Clean Energy
SDG 13 — Climate Action
Physics-Guided Wind & Grid Forecasting
Adaptive Feature Selection · Regime-Aware Loss · Edge Deployment
Problem
Short-horizon (0–3h) wind forecasting is critical for renewable integration and grid stability, but standard data-driven models degrade under wind ramps, extreme gusts, and site-dependent variability. Deploying these models on turbine-edge hardware adds strict latency and memory constraints.
Approach
Physics-guided sequence models that fuse adaptive temporal feature selection with physical wind-field structure (boundary-layer behavior, diurnal cycles, wake transport). Regime-adaptive focal losses penalize errors more heavily during high-risk ramps. Models are optimized for Jetson-class edge hardware.
Key Results
→ 6 first-author papers (2024–2025)
→ Shallow hybrid model runs on <2 GB RAM with <200 ms inference
→ Dynamic Bayesian optimization for hyperparameter tuning on-device
→ Hardware-centric design space exploration of Transformer–LSTM architectures
SDG 7 — Clean Energy
SDG 13 — Climate Action
Urban Heat-Risk Early Warning
Ensemble Learning · Spatio-Temporal Correction · Public Health
Problem
Urban heat extremes are rising in frequency and intensity. City-center thermal stress concentrates in dense cores where nighttime cooling is suppressed. Most forecasts operate at city-wide resolution, missing the neighborhood-scale hotspots that drive mortality risk.
Approach
Physics-informed ensemble learning (PIEL-NET) that forecasts grid-cell temperature at ~100 m–1 km resolution, several hours ahead. Residual spatio-temporal correction preserves spatial context and physical plausibility. Outputs can trigger targeted cooling guidance for vulnerable populations.
Methods
SDG 11 — Sustainable Cities
SDG 13 — Climate Action