BRITE Pivot: Learning-based Optimal Control of Streamflow with Potentially Infeasible Time-bound Constraints for Flood Mitigation

点击翻译按钮获取中文摘要

This Boosting Research Ideas for Transformative and Equitable Advances in Engineering (BRITE) Pivot award will fund research that enables the intelligent deployment of optimal strategies for mitigating the damaging effects of rain-induced flooding, thereby promoting the progress of science and advancing the national prosperity, welfare, and health. Climate change is causing more frequent extreme weather events, like heavy rains, with disastrous consequences to infrastructure, public health, and national security. As the population grows and new urban centers develop, flood mitigation becomes a complex task that requires high-level coordination, is time critical, occurs in the presence of uncertainty and lack of full observability, and may be only partially feasible due to infrastructure constraints. This project will build a control framework powered by artificial intelligence to operate reservoirs in an optimal way for regulating streamflow while accounting for incomplete data acquisition, unpredictable effects of extreme weather, and ethical decision-making. The results from this research will benefit the scientific communities of hydrologic systems, control, and robotics, with applications also to intelligent systems with machine ethics. In addition, this project will provide undergraduate research opportunities and outreach activities, including educational materials for K-6 students to learn how climate change affects people’s lives, with emphasis on enhancing diversity, equity, and inclusion.This research aims to make fundamental contributions to methods for combining physics-informed and recurrent neural networks to predict the evolution of dynamic systems while also quantifying the effects of uncertainty, as well as for constructing learning-based control synthesis algorithms for complex high-level tasks that are temporally constrained and potentially infeasible in a partially observable environment. Data collected from US Geological Survey stations will be used to parameterize hillslope-link hydrologic models for streamflow forecasts. Small model simulations will then be combined with machine learning techniques to forecast streamflows with uncertainty quantification. Next, a formal description of the flood mitigation task that also accounts for lack of observability will be used to characterize the cost of violating temporal and economic constraints and ethical preferences. Finally, reinforcement learning techniques will be used to train a control agent to intelligently accomplish infeasible tasks to the greatest possible degree. A case study of the flood of 2008 of the Iowa-Cedar Watershed will be used to demonstrate the model development and control framework.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.