Abstract:
Multi-reservoir systems require adaptive control policies capable of managing evolving hydroclimatic variability and human demands across a wide range of time scales. However, traditional operating rules are static, ignoring the potential for coordinated information sharing to reduce conflicts between multi-sectoral river basin demands. This study shows how recent advances in multi-objective control enable the design of coordinated operating policies that continuously adapt as a function of evolving hydrologic information. The benefits of the proposed control innovations are demonstrated for the Red River basin of Vietnam, where four major reservoirs serve to protect the capital of Hanoi from flooding, while also supplying farmers with irrigable water supply and the surrounding region with electric power. Operating policies recently proposed by the Vietnamese government seek to improve coordination and adaptivity in the Red River using a conditional if/then/else rule system that triggers alternative control actions using information on current storage and recent hydrology. However, these simple, discontinuous rules fail to protect Hanoi to even the 100-yr flood. Our policy diagnostics illustrate how recent advances in multiobjective control make better use of coordinated system information to reduce food-energy-water conflicts in the basin. These findings accentuate the need to explicitly explore critical multisectoral tradeoffs, vulnerabilities, and system dependencies to enable major river basins to better adapt to evolving hydroclimatic variability and socioeconomic change.
Bio:
Dr. Reed’s Decision Analytics for Complex Systems research group has a strong focus on the sustainability of Food-Energy-Water systems given conflicting demands from ecosystem services, expanding populations, and climate change. The tools developed in Dr. Reed’s group bridge complexity science, risk management, economics, multiobjective decision making, artificial intelligence, and high performance computing. Engineering design and decision support software developed by Dr. Reed is being used broadly in academic, governmental, and industrial application areas. More recently, Dr. Reed is facilitating the development of the MultiSector Dynamics Community of Practice to advance complex adaptive human-Earth systems modeling to better address the interdependent challenges of climate change, energy transitions, and sustainability.
Summary:
Climate change is pushing our infrastructure to its design limits
e.g. recent heavy rains put heavy strain on Missouri river flood controls
Need to
Focus on extreme scenarios to which systems are exposed
Systemic failure modes
Focus: Red river basin and its impact on Hanoi, Vietnam
Vietnamese government is building reservoirs on the river to control flooding
But there are multi-sectoral demands on this system
Hydro-power
Water supply for agriculture (76% of Vietnamese agriculture is irrigated, 70% of Vietnamese employed in agriculture, 20% of GDP is agriculture, which is shrinking)
Flood protection (affects population, industry and agriculture)
Goals:
Maximize hydro production
Minimize water supply deficit
Minimize flooding in Hanoi
Want to convert goals into control strategies
Hanoi delta fed by 5 rivers
Control points at reservoirs (on 3 of 5 rivers)
Confluence point among rivers is at Hanoi flows into delta to the ocean
Official guidelines have separate algorithms for
Flood season
Dry season
Between seasons
State-dependent rules based on flow rates, demand, etc., nest of conditional decisions
Improved control algorithm: Evolutionary Multi-Objective Direct Policy Search (EMODPS)
Goal is to optimize policy for making decisions, not directly optimizing a particular set of decisions
Want to find the tradeoff surface between alternative decisions
Evaluate policies using a high-fidelity simulation of the target system
Learn EMODPS policies using radial basis functions of the same state variables as the official policy
Action space: hourly releases of water from reservoirs
Used parallel genetic search to explore the space of policies
Gradient-based search is infeasible since derivatives are rarely available for the high-fidelity models
Bayesian algorithms and fast emulation may be helpful in optimizing the search process
Largest study so far: 500k samples on 40k compute cores
The learned policies discover the Pareto optimization surface
Simulations cover 100k simulated years (sampled from a distribution of weather scenarios)
Surface completely dominates the official guidelines
Examining alternative scenarios
Drivers
Climate
Water demand
Can examine climate/demand conditions that most contribute to
Flood control failures
Hydro power generation failures
… separate dynamics for each scenarios
Can learn a separation between success and failure scenarios
Hydro and flood criteria are at odds: hydro requires lots of elevated water, which is a flood risk
Very important to consider tradeoffs between these goals
Current official river management guidelines are outside the Safe Operating Space across climate scenarios: don’t protect against 100 year flood with 95% probability
Official guidelines are limited because they were made by committees that didn’t have access to comprehensive simulation and optimization resources
General approach related to Bayesian Optimization and Reinforcement Learning, but with a different emphasis:
Detailed simulations to evaluate possible outcomes of decisions
Consider many extreme scenarios and different estimates of their probabilities
Highlight impacts on many competing stakeholders who have different priorities
Applicable to optimizing many coordination problems, both in the top-down control setting and market-based coordination
Related papers: