An Empirical Study of Environmental Data Prediction in the United States Energy-Water Nexus

Abstract

Electricity generation systems are dependent on water availability and planning for future water scarcity is currently hindered by limited data and predictive models. The Energy-Water-Emissions Dashboard (EWED) is a novel environmental data management system that integrates multiple heterogeneous data sources and provides information for nearly 10,000 individual power plants across the United States. This article describes our empirical research of using machine learning models for electricity prediction and water usage in the context of water availability constraints. We evaluate the use of linear regression, decision tree regression, random forest regression, eXtreme Gradient Boosting (XGBoost), and Artificial Neural Network (ANN). Based on the performance evaluation of each model, we use ANN for generation and water consumption and XGBoost for water withdrawal prediction in the production environment. Model performance evaluation is based on statistical measures including Root Mean Square Error (RMSE), Mean Absolute Error (MAE), coefficient of determination (R2), Willmott's Index of Agreement (WIA), RMSE-observations to Standard deviation Ratio (RSR), Nash-Sutcliffe model Efficiency Coefficient (NSEC), and Percent Bias (PBIAS). This article presents performance improvements of our machine learning approach compared to the conventional coefficient method used by EWED, for example, RMSE decreased 8.1% in generation, 59% in water consumption, and 53% in water withdrawal prediction. The significance of this research is that it covers a wide variety of power plant types, it uses consistent methods across energy and water systems, and provides predictions at multiple management scales across the United States to assist with future planning at the energy-water nexus.