Ensemble Decision Tree Models Using RUSBoost for Estimating Risk of Iron Failure in Drinking Water Distribution Systems

Abstract

Safe, trusted drinking water is fundamental to society. Discolouration is a key aesthetic indicator visible to customers. Investigations to understand discolouration and iron failures in water supply systems require assessment of large quantities of disparate, inconsistent, multidimensional data from multiple corporate systems. A comprehensive data matrix was assembled for a seven year period across the whole of a UK water company (serving three million people). From this a novel data driven tool for assessment of iron risk was developed based on a yearly update and ranking procedure, for a subset of the best quality data. To avoid a ‘black box’ output, and provide an element of explanatory (human readable) interpretation, classification decision trees were utilised. Due to the very limited number of iron failures, results from many weak learners were melded into one high-quality ensemble predictor using the RUSBoost algorithm which is designed for class imbalance. Results, exploring simplicity vs predictive power, indicate enough discrimination between variable relationships in the matrix to produce ensemble decision tree classification models with good accuracy for iron failure estimation at District Management Area (DMA) scale. Two model variants were explored: ‘Nowcast’ (situation at end of calendar year) and ‘Futurecast’ (predict end of next year situation from this year’s data). The Nowcast 2014 model achieved 100% True Positive Rate (TPR) and 95.3% True Negative Rate (TNR), with 3.3% of DMAs classified High Risk for un-sampled instances. The Futurecast 2014 achieved 60.5% TPR and 75.9% TNR, with 25.7% of DMAs classified High Risk for un-sampled instances. The output can be used to focus preventive measures to improve iron compliance.