Water supply network pollution source identification by random forest algorithm

Abstract

A novel approach for identifying the source of contamination in a water supply network based on the random forest classifying algorithm is presented in this paper. The proposed method is tested on two different water distribution benchmark networks with different sensor placements. For each considered network, a considerable amount of contamination scenarios with randomly selected contamination parameters were simulated and water quality time series of network sensors were obtained. Pollution scenarios were defined by randomly generated pollution source location, pollution starting time, duration of injection and the chemical intensity of the pollutant. Sensor layout's influence, demand uncertainty and imperfect sensor measurements were also investigated to verify the robustness of the method. The proposed approach shows high accuracy in localizing the potential sources of pollution, thus greatly reducing the complexity of the water supply network contamination detection problem.

Machine Learning and Simulation-Optimization Coupling for Water Distribution Network Contamination Source Detection

“…The injected contaminant is chemically or biologically unspecified and is treated as a mass which is introduced into the water distribution network since the investigated frameworks are independent of the transport model used in the simulation. The sensor positioning was the same as the one introduced in [ 48 ] and which showed in [ 39 ] to include a good number of suspect nodes when used in conjunction with the RF algorithm for classification. The NET3 water distribution network with the selected sensor layout can be seen in Figure 1 .…”

“…Previously, Decision Trees (DT) were utilized for water network contamination source area isolation [ 38 ] and more recently, the RF algorithm has also been successfully utilized for potential water supply network contamination source node identification [ 39 ] and for determining the number of contamination sources in a water distribution network [ 40 ]. The RF algorithm was trained with Monte Carlo (MC) generated input data of sensor water quality readings through a time interval and the true source nodes as the output data.…”

“…Two algorithmic frameworks are constructed which are based on the methodology. Both frameworks utilize a machine learning approach based on the RF algorithm (implemented in the Python machine learning module scikit-learn 0.21.3 [ 43 ]) for potential contamination source search space reduction (as presented in our previous work [ 39 ]). The first investigated framework couples the simulation-optimization procedure directly with the RF classifier in order to determine the contamination start time, end time and injected contaminant concentration for each RF model predicted node separately and for this framework, three different stochastic optimization algorithms were investigated for one water distribution network benchmark.…”

“…However, for the purposes of examining the algorithm proposed in this study the simplified EPANET2 complete mixing model is good enough as the whole procedure is not dependent on the accuracy of the mixing processes occurring in the water distribution network. Monte Carlo simulations are made to train the RF model for classification (as described in [ 39 ]) with the sensor water quality measurements being the input features and the true source node being the output. The RF model classifier then predicts the most probable contamination source nodes which are then submitted either to the stochastic simulation-optimization procedure (for the first framework) or to RF regression (trained with previously generated MC EPANET2 data) which predicts their start and end times and injected contaminant concentration.…”

Machine-Learning Classification of a Number of Contaminant Sources in an Urban Water Network

“…Machine learning tools have been increasingly used in contamination detection, where Random Forest has been used for groundwater source of contamination detection [20] and source detection in a river [21]. In Grbčić et al [22] Random Forest algorithm was used to predict contamination event parameters in water distribution networks and in Grbčić et al [23] new machine learning-based algorithm was proposed. A great advantage of prediction models is that they can be constructed before an accident occurs, so when a contamination event is detected prediction can be made even for large networks in a computationally efficient way.…”