Interpreting transboundary water resources system operation using machine learning

Water reservoirs serve vital roles in energy and food security through storage for hydropower generation and food production. Multiple reservoirs are typically used in a river basin to meet growing demands. To promote cooperative reservoir operation and mitigate water use conflicts, it is important to understand how operational decisions develop in time and interact, especially when considering multiple operating objectives. Interpreting multi-purpose, multi-reservoir dynamics is particularly challenging in transboundary contexts, where interdependencies develop across multiple riparian countries. To address this problem, we develop a model-agnostic interpretation framework to explain water systems globally and locally using machine learning methods. Specifically, we use the Gradient Boosting Regression Tree to approximate reservoir operation policies and combine it with recursive feature elimination to select the most important features or input variables in reservoir operation decision-making. We also attribute specific optimized reservoir decisions to the selected input variables using the Shapley additive explanation method and further explain their dependencies by examining the partial dependence plots of reservoir release decisions on water availability and water demand. Using a case study of cascade reservoir operation in the Nile River Basin, we elucidate trade-offs between power generation and irrigation water supply in Ethiopia, Sudan, and Egypt and interpret the relationship among reservoir operation decisions, water availability, and water demand in the Nile River Basin system. Our results show that the trade-offs between power generation and water supply in the Nile Basin are strongly related to the decision-making criteria of the different objectives (the criteria here represent the riparian countries' requirements for the performance of the Nile operation). For example, the conflict between power generation from the High Aswan Dam (HAD) and irrigation water supply in Sudan is evident regardless of the power generation output, whereas the conflict between power generation from the HAD and irrigation water supply downstream of the HAD becomes evident and difficult to mitigate only when considering a certain criterion of HAD power generation (i.e. HAD power generation is required higher than 860 GWh/month). In addition, the most important features or input variables selected for the operation of each reservoir in the Nile system vary significantly with operational performance and decision preferences. By incorporating the proposed framework, we can select the most important features in multi-objective water system operations and infer the relationship between operational decisions, water availability, and water demand at each time step.