Collaborative Research: AI-Enabled Decision Support for Rate Setting and Infrastructure Investment in Decentralized Water Utilities under Drought

The objective of this research is to understand and improve how water utilities can set rates and invest in infrastructure under drought by building integrated models that make explicit the feedback among water supply, household demand, and utility finances. Drought preparedness across the United States depends on decentralized decisions made by local water utilities, which plan and finance most infrastructure renewal and expansion through water bills. The consequences of investment and rate setting, however, unfold through coupled feedback among water availability, household demand, and utility finances, which are poorly understood and not often captured in decision support models. In cases where an investment to improve water reliability requires raising rates, for example, the resulting increase in water bills can reduce water demand and erode utility revenue, ultimately constraining future investment capacity and altering the reliability benefits the investment was meant to deliver. Despite extensive work on individual components of this system, such as hydrologic, infrastructure, and demand response models, empirical evidence and tools that trace the feedbacks end-to-end is still lacking. Also lacking is an understanding of how hundreds of local decisions aggregate through shared supplies and conveyance to shape large-scale reliability and performance. Through this project, an integrated team of economists and engineers come together to close these gaps in knowledge, using modern artificial intelligence and machine learning (ML) methods to make the analysis feasible at scale. The researchers will document how utilities set rates and plan investments under drought stress using large language models to extract investment and financial variables from unstructured planning documents. They then apply supervised ML models to screen which socioeconomic, climatic, and decision-making factors are associated with performance, and quasi-experimental methods to test which of those associations are causal. Data and code will be released as open resources that can be updated in future years. The work is conducted in partnership with three utilities with distinct sizes and operating contexts to develop dynamic decision-support models leveraging multi-objective policy optimization techniques derived from Reinforcement Learning (RL). These models represent coupled water–demand–finance feedbacks under stochastic hydrology and demand response. Finally, the observed rate structures will be benchmarked against optimized rates for hundreds of utilities under competing objectives to infer how utilities balance tradeoffs and to quantify the magnitude of rate inefficiency across the sector. Institutional analysis will be applied to diagnose observed inefficiencies. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.