• Inputs
  • PSPS Report Processing, Duration of events: Dec 2020 - Jan 2021
    • 2 months of data
    • 3 events
  • CIMIS Weather Station Data
    • 200 Weather Stations
    • 13 readings from the weather station: wind, humidity, temp, etc.
    • Total records: 230k, 1 hour each
• Outputs
  • How can it be used?
    • Ex. Next 5 days, low fire risk at weather stations x,y,z with prob %
    • Forecasting Humidity, Wind
    • Predicting "crossings", extreme values, negative correlations of humidity and wind
  • What metrics do we have?
• Methods

  1. Preprocessing

     • Document reading/table extraction for PSPS data
     • During PSPS event, extreme wind/humidity values were observed

  2. Processing

     • Get weather station data from APIs (CIMIS, MesoWest, etc.)
     • Join/match data sources based on time

  3. Prediction

     • Built a prediction model to anticipate extreme wind/humidity

  4. Model Building

     • Technical details of inputs, wind distribution, simplicity of predicting humidity, temp
     • Technology, Software, Algorithms: Gluon-ts
     • How is the model built? Layers, Tuning
     • Performance, Accuracy: Time to train per config, metrics
     • Potential pitfalls, challenges, failures

  5. Future Enhancements

     • Train model on morning data only
     • Domain knowledge, constraints on humidity, wind
     • Data feed morning/evening. How to handle new data?
     • Input from metereology to improve/tune models

  6. Results

     • Temp, Humid, Wind Prediction Accuracy
     • Relationship between wind/humidity, wind brings humidity down, increasing fire risk
     • 3-day and 7-day forecast (how long do PSPS events last for? what granularity do we need?)
       • 4, 6-hour intervals, find chance of PSPS events in each interval
       • Distill data/predictions into clear visualizations

  7. Reuseability

     • Demand Forecasting
     • Use cases: Energy and Utilities, Load shedding