Presenters-0725
Abstract
GaugePredict started as a research question rather than a software project: can we predict water level at a downstream diversion days to weeks in advance using only upstream gauge observations? GaugePredict was developed as an open source machine learning framework for multi-lead forecasting (eg. hours or days). It learns how signals move through a river network from upstream gauges to a downstream target location. In our example, we focus on hydrodynamics using U.S. Geological Survey (USGS) gauge data and the way routing and storage shape downstream water levels and discharge.
GaugePredict uses a hybrid CNN–LSTM neural network. The 1D Convolutional Neural Network (CNN) layers capture short term events (eg. rapid water level rises and falls) to learn relationships between gauges and the Long Short-Term Memory (LSTM) layers represent the context of those events, including background conditions. Models are trained for user-defined lead times from hours to weeks, which supports both operational forecasting and longer range planning. GaugePredict also includes interpretability into its framework using SHAP (Shapley Additive Explanations) to quantify how much each upstream gauge at each lead time contributes to each prediction. This allows gauge-network reduction, where a smaller set of predictors can retain skill while decreasing runtime and model size.
We demonstrate how GaugePredict can be used with two applications in the Mississippi River Basin. First, we forecast daily water level at the Bonnet Carré Spillway, a freshwater diversion near New Orleans. We then convert predicted water level to diversion discharge using a broad-crested weir equation, which provides discharge forecasts through the diversion while it is open as well as leakage through the structure during closures. Second, we use GaugePredict to backfill missing data during outages at the Mississippi River at Baton Rouge, which is useful for hydrodynamic and water quality models that require continuous boundary conditions and calibration time series data. GaugePredict is designed for applications beyond these examples. Inputs can include any set of basin-wide USGS gauge parameters, and target sites can be any USGS gauge(s) or user-defined time series (for example, CSV-based variables), providing a flexible framework that can be extended to new gauges, variables, and forecasting objectives.
Please acknowledge the original contributors when you are using this material. If there are any copyright issues, please let us know (CSDMSweb@colorado.edu) and we will respond as soon as possible.
Of interest for:
