Emulation of monthly water allocations using LSTM models: a case study of the Colorado River Basin in Texas

Abstract As concerns about water scarcity grow due to increasing aridity and extreme weather, there is a need for computational test beds to help explore water management challenges and elucidate interdependences within coupled human-natural systems. Prior approaches for water management analyses have ranged from process-based simulations to data-driven modeling. However, development of data-driven models for water availability can be challenging due to poor data availability and software limitations. To overcome these challenges, this work aims to develop a parsimonious, deep learning, long short-term memory (LSTM)-based surrogate of a process-based water management model, the Water Rights Analysis Package (WRAP). Synthetic streamflows were generated to create an ensemble of scenarios across a range of hydrologic conditions. These data were inputted into WRAP to obtain associated water allocation data for LSTM training, testing, and validation. We applied this computational framework to a case study of the Colorado River Basin in Texas. The trained LSTM emulates the water allocation process with low overall error and demonstrates minimal performance impact when out-of-distribution drought scenarios are used. Exploratory analysis of error patterns shows that the LSTM effectively captures the spatio-temporal water allocation patterns well and has no significant bias for underlying system attributes. These findings suggest that LSTMs have the potential to serve as surrogates for water management models. The computational framework developed could be used to explore water management scenarios for other basins to understand potential impacts from increasingly extreme hydrologic conditions.