Control and Information Design for Water Infrastructure Operations

Water systems are increasingly strained by rising demand, climate change, and more frequent extremes. These pressures intensify tradeoffs among municipal supply, agriculture, hydropower, flood control, and ecosystems. In regulated basins, water reservoir operations are the primary tool for managing these competing objectives. Mathematically, reservoir operations are typically framed as a model-based, closed-loop control problem, where dam releases (controls) are determined as a function of reservoir storage (state) under uncertain inflows (disturbances). Classical methods from control theory and operations research have long guided policy design. Yet the problem remains challenging: system dynamics are nonlinear, objectives are multiple, non-convex, and asymmetric, the state space is high-dimensional, and disturbances are highly uncertain and non-Gaussian. This talk presents a control-theoretic perspective on emerging approaches to reservoir operations, which increasingly use reinforcement learning to approximate solutions abandoning optimality, while preserving the complexity of the real system. Within this field, I focus on the critical, often overlooked question of policy representation: what information should a control policy condition on when both the state and disturbances are high-dimensional and uncertain? I introduce a framework that jointly learns the control policy and the information set it relies on, enabling richer representations that incorporate real-time indicators such as inflow forecasts and alternative aggregations of system state. Results show that strategic selection of policy inputs can substantially improve performance, particularly in multi-objective settings with delayed feedbacks.

Marta Zaniolo

Assistant Professor of Civil & Environmental Engineering, Duke University on September 5, 2025 at 10:15 AM in EB2 1231
Join Zoom Webinar

Dr. Marta Zaniolo is an assistant professor at Duke Civil and Environmental Engineering department. Her research addresses timely water sustainability challenges including how to ensure reliable and equitable access to water resources in the face of scarcity, competing demands, and a changing climate. She combines the fields of hydrology with machine learning, and optimal control techniques to enable more informed decision-making about water use at scales large and small. Dr. Zaniolo holds a BS and MS degree in Environmental Engineering and a PhD in Information Technology earned in the Environmental Intelligence lab at Politecnico di Milano. Prior to joining Duke she was a postdoctoral researcher in the CEE department at Stanford University.

This lecture series features exciting and dynamic visiting and virtual speakers from across the range of ECE disciplines. Take some time every Friday morning to be inspired by these great scientists and engineers before heading into the weekend!