On the Use of AI Planning for Water Management of the Red River Basin in Vietnam

Abstract

The management of reservoir-based hydrological systems is challenging due to complex dynamics and strong interdependencies. Traditional rule-based policies, derived from hydrological models, expert experience, and regulatory standards, are often too rigid, leading to inefficient use of water resources and limited ability to react to unexpected events such as off-season floods. We investigate how AI planning can support water management through the lens of numeric planning. The study focuses on the Hoa Binh reservoir, a key component of the Red River Basin in Vietnam, under hydrologically challenging dry-season conditions. Building on a previously validated PDDL-based simulator of the basin, we recast simulation inputs as planner-controlled release actions and study the resulting numeric planning task from a search perspective. The proposed formulation captures the complex system dynamics by defining actions that directly control the amount of water released on a daily basis. We first provide a fully domain-independent planning setting that makes no use of problem-specific information and observe that planners struggle to capture storage-dependent constraints, cumulative flows, and downstream interactions. Motivated by these limitations, we then integrate problem-specific structure and control knowledge into the domain-independent formulation to guide the planner toward higher-quality operating policies. Results indicate that planning can efficiently explore the decision space and generate plans of competitive quality compared to historical operation while achieving stricter adherence to operational constraints, particularly when enhanced with control knowledge.