Observability-based sensor placement improves contaminant tracing in
river networks
Abstract
This study presents a new methodology for identifying near-optimal
sensor locations for contaminant source tracing in river networks. To
establish a physical basis for the problem, we first derive a linear
time-invariant (LTI) model for riverine contaminant transport using the
one-dimensional advection-diffusion equation. We then formulate an
optimization problem to find the sensor placement that maximizes the
observability of the modeled system, and identify two heuristics for
efficiently achieving this goal. Evaluating each sensor placement
strategy on its ability to reconstruct initial contaminant loads from
observed outputs, we find that the best sensor placement is obtained by
greedily maximizing the rank of the LTI system’s Observability Gramian.
In addition to providing the best approximate reconstruction of internal
states, this strategy makes it possible to perfectly recover any initial
contaminant load while only monitoring a small subset of river branches
(~14%). Our methodology will enable researchers to
build sensor networks that better interpolate pollutant loads in ungaged
locations, improve contaminant source identification, and inform more
effective pollution control strategies.