HydroForecast is a software product that models inflow into dams and rivers. It is the only distributed deep-learning machine learning model in operation today.
Streamflow is a measure of flowing water, usually expressed as a rate, e.g. cubic feet per second. It tells us how much water will move through a single point (e.g., a given spot in a river). Sometimes streamflow is expressed as a volume over longer time horizons, e.g. total flow over a day or month, in acre-feet or cubic meters.
Streamflow forecasting models are built on an underlying hydrologic model that takes inputs from influential variables and accounts for complex interactions in the environment to output future flow predictions. These models can range from physically based to statistically based, and each has advantages. Standard hydrological model inputs include meteorological data, such as precipitation and temperature, land surface characteristics, e.g. elevation and slope.
Excellent question! A specific type of machine learning model has recently been growing in usage within the hydrologic community, both in academic literature and in operational applications. Upstream Tech’s research has been a part of that growth, both with co-authored publications and in case studies that demonstrate customer use cases. This research paper laid the groundwork of benchmarking experiments showing the accuracy improvements using a machine learning prediction model over a traditional physically based model.
Read our guide for those interested in learning more about machine learning in hydrology.
Briefly, a few advantages of machine learning models for streamflow prediction:
This is a difficult question, and one that many scientists are trying to untangle and quantify. Generally, extreme storms, drought, and shifting rain patterns are testing the resiliency of our water infrastructure and challenging forecast models built for a more stable period.
Read our post on why streamflow forecasts are so important.
Unlike physical models, which use equations based on empirical studies of how interactions occur in the natural world, statistical models are a blank slate. They learn about relationships by combing through data and identifying signals. This can be problematic if not properly managed because coincidences can often be mistaken for patterns.
The idea behind theory-guided machine learning is to design a model that can take advantage of the benefits from both physical and statistical approaches. HydroForecast is our solution to this challenge for streamflow forecasts.
HydroForecast is designed to adapt to different needs. Different needs often call for a different time frame, or scope, of interest – what we call horizons. We offer 6 horizons that can be tailored fit to your organization. In brief, the 6 horizons:
