Professor Dept. of Civil Engineering, University of North Dakota
A Historical Supply Analysis (HSA) is a crucial tool in managing water resources and aiding policy makers in decisions regarding water rights. Especially in areas where water scarcity is becoming a more pressing issue such as the American Southwest. To perform an HSA, two things need to be known. The supply, which is the available flow in a river, and the demand, which is the total permitted water rights adjusted for conveyance and evaporative losses. For the supply side data, this typically requires a stream gauge upstream of the point of diversion. Unfortunately, many times there is no gauge in the area of interest or gauge data does not exist over a sufficient period. In addition, there is a growing interest in incorporating the effects of climate change in these analyses. This paper aims to address this gap in information by investigating machine learning as a method of recreating supply and demand data for a given basin. This was done using a deep learning neural network, specifically a long short-term memory network (LSTM) with forward feature selection (FFS) hyperparameter tuning. The model was trained using the Catchment Attributes and Meteorology for Large-sample Studies (CAMELS) dataset and calibrated using known historic records. The output of this model is currently being analyzed for accuracy and being compared to other, more established methods of reconstructing historic supply data including tree-ring reconstructions and the Sacramento Soil Moisture Accounting (SAC-SMA) model to determine the efficacy.
