Leveraging explainable Machine Learning to discover trade-offs between water supply and demand management strategies in California

As water scarcity becomes the new norm in the Western United States, states such as California have increased their efforts to improve water resilience. Achieving water security under climate change and population growth requires an integrated multi-sectoral approach, where adaptation strategies combine water supply and demand management interventions. Yet, most studies consider supply-side and demand-side water management strategies separately. Further, publicly available data to assess the effectiveness of these strategies and their dependency on individual and collective human behavior is often hard to find and unstructured. Water conservation efforts are driven by water scarcity and policy requirements, with conservation targets and water use restrictions often designed assuming a degree of rationality of human behavior and based on cost-effective options and ease of implementation.

In this work, we develop a data-driven analysis aimed at evaluating historical synergies and possible trade-offs between water supply and demand management strategies in California. Our analysis is based on CaRDS – the statewide California Residential water Demand and Supply open dataset, which contains monthly values of water supply and residential water demand for 404 water suppliers in California from 2013 to 2021. In this time span, Californian water agencies had to adapt and mitigate the effects of two droughts (in 2012-2016 and 2020-2022) through residential water demand reductions, as well as address rapid changes in demand associated with the global COVID-19 pandemic (2020). Our trade-off analysis integrates the following three sequential steps: (i) trend analysis – we use Random Forest regression to control for seasonal factors (i.e., temperature and precipitation) that affect water supply and demand at the utility scale; (ii) multi-criteria trade-off analysis – we examine the temporal relationship between water supply and demand by utilizing Dynamic Time Warping to identify trade-offs and management patterns. Next, we cluster water suppliers in 6 groups based on their combined management patterns; (iii) and driver analysis – we utilize explainable Machine Learning by combining SHAP (Shapley values) with LGBM (Light Gradient Boosting Method) to identify the drivers of each cluster. Potential drivers include climatic region, water supply portfolio, indoor vs. outdoor water use, local and state policies,  population, supplier size, and income. We finally validate the results of our analysis by comparing our findings with responses from water supplier interviews carried out in 2017 and reveal differences between intended and actual water management outcomes. This research contributes insights into the combined effects of policies on water supply and demand at a statewide level. Further it facilitates the formulation of adaptive resilience strategies for human actors in water management and decision makers alike to address vulnerability of small and large water systems to a rapidly changing climate and a society with non-linear changes in human behavior.