Quality forecasts of urban water demand are critical to effective water resource management. Machine learning techniques have been widely adopted to provide accurate deterministic predictions, but have less emphasis on quantifying prediction uncertainties. This study evaluates an array of machine learning-based uncertainty quantification techniques on a large dataset recorded by Tampa Bay Water, a water wholesaler in Florida. We quantify the uncertainty of monthly water demand forecasts on a house-hold level. The results show that the random forest model provides the best calibration quality and the long short-term memory model yields the best deterministic forecasting accuracy, as judged by multiple metrics and from different perspectives. The resulting models are promising for assisting long and short-term water resource management.
Yi Han is a Ph.D. student, major in machine learning, at the University of Florida (UF) Agricultural and Biological Engineering Department. He received his master degree of quantitative genetics in Fall 2016 and his master degree of statistics in 2019. Yi’s current research projects focus on applications of machine learning and deep learning methods to various agricultural problems. Yi Han areas of expertise include statistics, genetic data analysis, zero-inflated data modeling, time-series data modeling, uncertainty quantification, and anomaly detection with generative models.