Advancing Intelligent Water Systems through New Methods in Model Discovery

Abstract

Flooding, water quality degradation, and ecological decline are persistent challenges in drainage systems, worsened by increasing urbanization and climate change. Existing infrastructure solutions are expensive and cannot be tailored to address the fundamental tradeoffs between water quality and flood prevention. Besides this, the piecemeal, site-scale design paradigm of drainage assets makes these infrastructure networks far less optimized at the system scale than utilities such as transportation and power.

Smart water systems promise to eliminate the fundamental tradeoff of drainage design while enabling system-wide coordination. These next-generation infrastructure systems use Internet-of-Things networks to dynamically redesign entire drainage systems in response to every storm. However, how best to achieve this redesign is an open question. The drawbacks of existing heuristic and process-based methods for generating control algorithms suggest application of tools from state-space control theory. However, the delayed and diffused causation typical of drainage systems makes the system identification necessary for such approaches difficult.

While the use of discrete time-lags has a very successful history in system identification, improvements may be possible. The central hypothesis of this dissertation is that understanding the system as partially observed and directly estimating the unobserved states and processes will generate dynamical systems models well suited to practical prediction and control tasks. This hypothesis will be developed and applied by addressing critical knowledge gaps in drainage systems:

1 We do not know how to automatically create interpretable rainfall-runoff models from short and noisy data records. 2 We do not know how to cheaply and automatically generate water level and flow predictions across entire sensor networks. 3 We do not know how to automatically generate the system approximations of drainage systems necessary for state space control. 4 We do not have methods to ensure dynamically controlled drainage systems are resilient to communications disruptions.

The chapters of this dissertation address these gaps via the following fundamental contributions:

Chapter 2 provides a new conceptual model of runoff generation: the “effective subcatchment.” Performance is validated on nearly 400 real catchments. This chapter also demonstrates the ability of Model Discovery in Partially Observable Dynamical Systems (modpods) to identify nonlinear single-input, single-output (SISO) systems.

Chapter 3 demonstrates an approach which dramatically lowers barriers in deploying predictive sensor networks vital to bridge safety and flood warning systems. This demonstrates how modpods can identify nonlinear multiple-input, single-output (MISO) systems while also detailing enhancements to an existing wireless water level sensor network.

Chapter 4 demonstrates the ability of modpods to identify linear multiple-input, multiple-output (MIMO) systems approximating drainage networks. This expands the application space of modern control approaches to partially observed systems and allows the automatic generation of safe and tunable feedback controllers for drainage systems.

Chapter 5 shows that centralized control of a drainage system can be well approximated for communications outages lasting up to a week. This chapter also provides evidence that distributed agents in a cooperative control problem can infer total system conditions from local measurements when a suitable inference model is available. end{enumerate}

In addition to the fundamental scientific contributions, the primary practical impacts of this dissertation are (1) accessible and automated water level prediction networks and (2) resilient state-space controllers for drainage systems. The modpods method appears to have general applicability beyond water systems and could greatly expand the possibilities for application of tools from dynamical systems theory.