Flexible Methods to Support Sustainable Energy and Water use

The increase in populations and urban development has driven demands for energy as well as fresh and ground water in agricultural, industrial, commercial and residential sectors. According to United States Government Accountability Office (GAO) Report in 2014 [1], with the current average water usage, 40 out of 50 states would expect water shortages in parts of their states over the next decade. However, the problem of water shortage is not due to its limited supply in the nature, but that of mismanagement of how it is used and distributed [2]. In addition, shortages of water effects energy generation since 40% of water withdrawal in the United States is used for thermoelectric power plant. Consequently, there is a need to understand how water and energy is being used in order identify the stress points. This thesis leveraged the power of flexible statistical methods in order to estimate water, energy and water-energy nexus as a (non-linear) function of various geographic, climatic and socio-economic variables. More specifically, this thesis focused on analyzing the total, per capita water usage at the state-level in all sectors in the U.S., as well as the residential electricity demand and also residential water-energy demand nexus. This thesis shows that for both total water usage and residential electricity demand analysis, the widely used ‘traditional’ statistical models such as Multiple Lin- ear Regression (MLR) cannot adequately capture the complex structures in the data. Hence, there is a need for leveraging flexible method such as Random Forest (RF). In both water and residential electricity analysis, RF outperforms many other statistical models (e.g., MLR) in both fit and predictive accuracy, with RF improvements of al- most double in-terms of R-squared and predictive accuracy. In total per capita water usage, the predictive model based on the method of Random Forest helped identify factors such as percentage of irrigated farmland, coal production and heating-degree-days (HDD) as the key predictors of water usage. In residential electricity consumption, the developed predictive model identified electricity cost and climate variables as the important variables. In addition, due to the lack of analysis in end-user sector for water-energy nexus, this thesis proposed a framework to analyze water-energy nexus in the residential sector using Multivariate Regression Tree Boosting to study the effect of predictors given multiple responses (residential water, electricity and natural gas usage)