Architecture Overview · Park Whisperer Platform

ML-Based Weather Nowcasting System

An end-to-end ML pipeline ingesting METAR reports from 17 FAA stations, computing atmospheric indices, training 6 Random Forest models, and serving sub-30-second nowcast predictions for park operations weather risk management.

● Production · GCP + BigQuery scikit-learn · Random Forest 44K+ Training Observations

Click any component to explore it in the full case study.

Source A
FAA METAR Observations
17 Central Florida aviation weather stations polled via aviationweather.gov. Raw METAR strings decoded: thunderstorm groups (TS/VCTS/LTG), fog codes (BR/FG), dewpoint depression, rolling precip windows
17 stations · 5-min polling
Source B
Upper-Air Radiosonde Data
University of Wyoming soundings archive, Tampa Bay station 72210. Provides 7 atmospheric thermodynamic indices: CAPE, CIN, Lifted Index, K-Index, Total Totals, Showalter Index, PWAT — the most predictive features for thunderstorm onset
UWY soundings · Station 72210
Data Pipeline — weather-updater-v6 · Continuous ingestion
Step 1
METAR Parse
Decode raw strings → structured fields + adverse event flags
›
Step 2
Sounding Fetch
Parse CAPE/CIN/LI/K-Index/TT/SI/PWAT from UWY archive
›
Step 3
Feature Assembly
ml_training_features BigQuery view joins METAR + sounding into hourly vectors
›
Step 4
BigQuery Upsert
weather_updater_v3 table — 8+ years of history, growing daily
🏋️
Model Training — weather-ml-trainer · On-demand after data milestones
Reads from ml_training_features view. Trains 6 independent Random Forest classifiers with class-weight balancing. Serializes each model to GCS as a pickled .pkl file.
44,000+ training observations Random Forest · class-weight balanced BigQuery → scikit-learn → GCS pkl 8+ years of historical data Drift-triggered retraining
Six Trained Models — All Random Forest · GCS pickle storage
Thunderstorm Nowcast
RF n=150 depth=12
F1 0.87
All 7 atmos indices
Precipitation
RF · balanced
F1 0.82
METAR precip groups
High Wind
RF · balanced
F1 0.78
Wind speed trends
Fog
RF n=100 depth=8
F1 0.84
Surface only (no sounding)
Venue Impact
RF · KISM primary
F1 0.79
Nearest WDW station
High-Impact Venue
RF · high threshold
Precision 0.91
Severe event filter
Inference Engine
weather-ml-predictor
Loads pickled models from GCS, queries current conditions + spatial storm context via BigQuery. Applies spatial boost: when nearby/upwind stations report active storms, thunderstorm probability is adjusted. Total execution: <30 seconds across all 17 stations
<30s total runtime
Published Output
Nowcast JSON → GCS (Public)
Per-station nowcast payload: thunderstorm/precip/wind/fog probabilities + confidence scores. Venue-impact JSON consumed by park operations risk engine, ride closure recommendations, and content generation pipelines
GCS public bucket · JSON API
17
FAA stations monitored
6
Trained ML models
44K+
Training observations
<30s
Inference runtime
Stack GCP BigQuery scikit-learn Random Forest Google Cloud Storage Python aviationweather.gov API UWY Soundings Archive