Machine Learning Methods to Predict Failures in Water Distribution Network
In recent years,failures have occurred frequently to water distribution networks throughout China,resulting in huge economic losses and serious social impacts.However,the application of state-of-art algorithms to predict pipe failures is still to be explored and there is little research on systematic comparison of machine learning algorithms.Therefore,it is of great importance to accurately predict pipeline failures and ensure the operational safety of water distribution networks economically and efficiently.To this end,the failure prediction problem is described.The basic theories of four machine learning algorithms,i.e.,logistic regression,random forest,artificial neural network and 1-D convolution neural network,are introduced.The failure prediction performance of the four models is verified and compared with a case network in an industrial park of a city in the south of China.The results show that the 1-D convolutional neural network has the highest accuracy but random forest is the most cost-effective algorithm.The impact of each pipe feature on the failure probability is analyzed as well.It is found that pipe diameter,pipe length,road class and qualification of construction enterprise are the four most important features.The failure probability is negatively correlated with pipe diameter and road class while positively correlated with pipe length and qualification of construction enterprise.
water pipesfailure predictionmachine learningimpact of pipe feature
国家科技期刊平台
NSTL
万方数据
