Keywords
water resources; sustainability; graphical models; latent variables; convex optimization
Start Date
27-6-2018 3:40 PM
End Date
27-6-2018 5:00 PM
Abstract
Water is a precious commodity, especially in the state of California. Our state frequently experiences cycles of major state-wide precipitation deficits – most notably the 2012–2015 drought which was the worst to occur in the past 1200 years. The focus of this work is to develop a state-wide model of the California reservoir network to address the following scientific questions: 1) what are the dependencies among reservoirs? 2) Are there unmodeled phenomena (denoted as latent variables) that are influencing the network globally? Could these latent variables cause a system-wide catastrophe (e.g. exhaustion of multiple large reservoirs)?
Since reservoirs are strongly impacted by human activities (i.e. latent variables) that cannot be directly incorporated into physics-based models, traditional approaches have been unsuccessful at modeling the reservoir variability at the scale of the California reservoir network. For the first time, we develop a statistical model over 55 large representative California reservoirs to overcome these challenges and address the preceding questions. This model was built from reservoir level data gathered over the 13-year time period (2003-2016) that includes the latest drought. Our approach is based on substantial advancements of the classical technique of Gaussian graphical modeling; this technique characterizes interdependencies among reservoirs, identifies latent variables influencing the network, and links these latent variables to physical processes. One of the appealing property of our methodology is that the model can be fitted to data via a convex optimization program, which is guaranteed to obtain optimal solution in polynomial time complexity in terms of the number of variables. In particular, the execution time for learning our statistical model over the 55 reservoirs is approximately 10 seconds.
With this model, we precisely characterize the system-wide behavior of the network to hypothetical drought conditions, and provide risks of catastrophe for each individual reservoir in the network. We further use our findings to propose guidelines for more sustainable water management policies in the future.
These results and their implications will be illustrated in a user friendly output based on collaboration with SEED LA: http://seedcg.org/
From Data Science to Hydrology: California Reservoirs During Drought
Water is a precious commodity, especially in the state of California. Our state frequently experiences cycles of major state-wide precipitation deficits – most notably the 2012–2015 drought which was the worst to occur in the past 1200 years. The focus of this work is to develop a state-wide model of the California reservoir network to address the following scientific questions: 1) what are the dependencies among reservoirs? 2) Are there unmodeled phenomena (denoted as latent variables) that are influencing the network globally? Could these latent variables cause a system-wide catastrophe (e.g. exhaustion of multiple large reservoirs)?
Since reservoirs are strongly impacted by human activities (i.e. latent variables) that cannot be directly incorporated into physics-based models, traditional approaches have been unsuccessful at modeling the reservoir variability at the scale of the California reservoir network. For the first time, we develop a statistical model over 55 large representative California reservoirs to overcome these challenges and address the preceding questions. This model was built from reservoir level data gathered over the 13-year time period (2003-2016) that includes the latest drought. Our approach is based on substantial advancements of the classical technique of Gaussian graphical modeling; this technique characterizes interdependencies among reservoirs, identifies latent variables influencing the network, and links these latent variables to physical processes. One of the appealing property of our methodology is that the model can be fitted to data via a convex optimization program, which is guaranteed to obtain optimal solution in polynomial time complexity in terms of the number of variables. In particular, the execution time for learning our statistical model over the 55 reservoirs is approximately 10 seconds.
With this model, we precisely characterize the system-wide behavior of the network to hypothetical drought conditions, and provide risks of catastrophe for each individual reservoir in the network. We further use our findings to propose guidelines for more sustainable water management policies in the future.
These results and their implications will be illustrated in a user friendly output based on collaboration with SEED LA: http://seedcg.org/
Stream and Session
B3 (DMTES)