Using Deep Learning Models for Decision Support for Large River-Reservoir/Lake Water Systems

Abstract Submission: Large river-lake/reservoir systems, especially those found in the Western United States, face several daunting challenges precipitated by an uncertain climate and increasing water demands. While various legal frameworks govern long term water sharing and delivery obligations (e.g., Colorado River Compact), there is a need to balance these long-term obligations with shorter term operational objectives including flood control, poor water quality mitigation, ecosystems protection and maintenance, and maximizing hydro power generation. The Lake Powell Glen Canyon Dam and Colorado River system provides an example, where release volumes and mechanisms must be timed appropriately to meet power generation, water delivery, and downstream ecosystem objectives at sub daily time scales. While a model predictive control (MPC) approach could provide a path forward for balancing these often-competing objectives, models available for simulating lake stratification dynamics, releases, and fate downstream in the Colorado River are based on first principles and are computationally expensive. Data-driven modeling approaches could be used to achieve MPC but are typically only viable at locations where measurements exist. Because ecosystems have broad spatial extents, a continuum of water quality predictions is needed for wholistic ecosystem modeling. In this presentation, we will present a physics constrained graph-based deep learning modeling approach for predicting flow and temperature at all locations along the length of the river. This model is coupled with another that predicts release temperatures from Glen Canyon Dam. The coupled model is used to forecast temperature dynamics and is used in a reinforcement learning framework to attempt to inform sub-daily release magnitudes and mechanisms to meet power generation and water temperature objectives for fish ecosystems that exist at critical locations along the river.

Learning Objectives/Expected Outcome (Optional) :