HydroForecast takes a physics-informed neural network approach to forecasting inflows. At the core of our model is a deep neural network specifically designed to process time series data in a dynamic system.
The data used to train the model as well as the model structure, training processes and model validation all draw heavily from physical hydrology theory. By combining physics and machine learning we allow our models to produce highly accurate forecasts which are consistent with hydrologic theory, and allow for advanced model interpretability which builds trust with users and allows for location-specific context to be integrated into the forecasting process.
HydroForecast currently provides forecasts at four horizons:
- Short-term: 10-day, hourly time-step, updated twice daily
- Seasonal: 3-month, 10-day time-step, updated daily
- Extended seasonal: 1-year, 10-day time-step, updated daily
- Long-term: daily streamflow distributions aggregated by month from 2020-2100