Simulation of Regulated Streamflow using Noah‐MP Land Surface Model and Machine Learning

Abstract:

Accurate prediction of streamflow is key for efficient and informed operation and management of water resources systems. Advances in meteorological forecasts coupled with land surface models (LSMs) have provided valuable information for short- and long-term operation of surface reservoirs. However, these forecasts and the models on which they are based are often trained and validated by direct measurement of streamflow, which is affected by human activities and reservoir regulation. While climate-land surface models are able to predict unimpaired steamflows with usable accuracy, they generally neglect these effects and are unable to predict regulated streamflow accurately. Having reliable flow predictions, in particular inflow to reservoirs, is crucial due to the potential damages caused by flood events and water shortages during droughts. Advances in Machine Learning (ML) provide an opportunity to represent human effects on streamflow and to adjust the natural flow predictions produced by LSMs, especially for peak and base flows. Here, we use decomposition techniques to generate natural and regulated flow intrinsic mode functions and train an ML model, e.g. Random Forest, to adjust streamflow forecasts produced by the Noah-MP LSM. We configured Noah-MP for the upper American River Basin, CA, driven by the meteorological forcings from gridded observations at 1/32° resolution, as well as the NOAA/Climate Testbed Subseasonal Experiment (SubX) reforecasts after downscaling and bias correction. We compare the Noah-MP and ML-based adjusted forecasts are compared with various observational data sets, including U.S. Geological Survey streamflow observations. Compared with observations, the adjusted results of Noah‐MP show significant improvement in hydrological modeling of upper American River basin and forecasts of the inflow to Folsom Reservoir.