Forecasting Summer Water Demand Using Optimized Random Forests: A Case Study of 12 US Utilities
Abstract
Forecasting municipal water demands is an essential requirement of long-term water supply planning, seasonal operations, and daily management. Long-term water demand forecasting is challenged by changes in water pricing, conservation, overall shifts in per capita consumption and changes in climate. Traditional parametric approaches have proven useful although these have often resulted in significant overestimations of demand. This paper presents an approach that applies a Random Forest (RF) machine learning technique to a variety of physical, climate, and economic variables to improve the water demand forecasts. This method, denoted as the “Random Forest Water Demand Forecasting (RF-WDF) framework,” is applied to twelve utilities throughout the US over the period of 2004-2018. For the RF optimization, a randomized search cross validation (RandomizedSearchCV) process is employed to obtain optimal hyperparameters and avoid overfitting. The RF-WDF is evaluated using repeated random holdout, and leave-one-year-out experiments, separately in each region. The water demand forecasts generated by the RF-WDF show improvements by producing high linear correlation (0.71-0.91) and Kling-Gupta Efficiency values (0.72-0.94) for the twelve cities. In applying this technique to cities with water demands that vary significantly, the magnitude of systematic and random error for the RF-WDF generated water demand was noticeably decreased (by 2-10%) compared to more traditional approaches. Finally, a k-fold cross-validation experiment was conducted to examine the transferability of the RF-WDF framework for 12 US Utilities. We conclude that the proposed RF-WDF ensemble framework has the potential to improve the skill in demand forecasts which could directly benefit water resources planning and management.
