Predicting Urban Reservoir Levels Using Statistical Learning Techniques

Abstract

Abstract Urban water supplies are critical to the growth of the city and the wellbeing of its citizens. However, these supplies can be vulnerable to hydrological extremes, such as droughts and floods, especially if they are the main source of water for the city. Maintaining these supplies and preparing for future conditions is a crucial task for water managers, but predicting hydrological extremes is a challenge. This study tested the abilities of eight statistical learning techniques to predict reservoir levels, given the current hydroclimatic conditions, and provide inferences on the key predictors of reservoir levels. The results showed that random forest, an ensemble, tree-based method, was the best algorithm for predicting reservoir levels. We initially developed the models using Lake Sidney Lanier (Atlanta, Georgia) as the test site; however, further analysis demonstrated that the model based on the random forest algorithm was transferable to other reservoirs, specifically Eagle Creek (Indianapolis, Indiana) and Lake Travis (Austin, Texas). Additionally, we found that although each reservoir was impacted differently, streamflow, city population, and El Niño/Southern Oscillation (ENSO) index were repeatedly among the most important predictors. These are critical variables which can be used by water managers to recognize the potential for reservoir level changes.