Real-Time Guidance for Hydrant Flushing Using Sensor-Hydrant Decision Trees

A utility may detect contaminant in a water distribution network through water quality sensor information, which indicates that a biological pathogen or chemical contaminant is present in the network. A utility manager should identify actions that can be taken to protect public health, and flushing a contaminant by opening a set of hydrants can be an effective response action. Hydrants should be selected and timed to flush the contaminant; however, accurately ascertaining the characteristics of the contaminant source may be impossible, which creates difficulties in developing a hydrant flushing strategy. This research develops a decision-making approach that is designed to select hydrant flushing strategies in response to sensor activations and does not require information about the characteristics of the contaminant source. A sensor-hydrant decision tree is introduced to provide a library of rules for opening and closing hydrants based on the order of activated sensors. Sensor-hydrant decision trees are developed for a wide range of contaminant events using a simulation-optimization methodology. Potential contamination events are generated using Monte Carlo simulation and are simulated using a water distribution system model. Events are classified based on the order of the activation of water quality sensors in the network, and a noisy genetic algorithm is used to identify hydrant strategies for each class of events. Three sensor-hydrant decision trees are developed to represent risky, risk-averse, and adaptive management strategies. A risk-averse strategy specifies immediate actions to achieve average performance over many events. A risky strategy specifies specialized actions based on the prediction of the plume movement or a decision to wait to receive more information. An adaptive strategy specifies the actions for opening hydrants as each sensor is activated. An adaptive approach does not require predictions of the plume movement, but may result in lower performance due to delays in taking actions. The methodology is demonstrated to develop sensor-hydrant decision trees for a virtual midsized municipality, Mesopolis.