Leakage Detection Of Central Heating-network Based On Machine Learning

At present,heating-network operation faults have been gradually increasing with the rapid development of central-heating in China.The leakage may lead to huge economic losses and damagements of environment.Therefore,detecting the heating-network leakage timely and accurately is extremely important for maintaining a safe and stable operation of the heating system.On the basis of analysing the research situation of heating-network leakage detection methods,the leakage detection method of the central heating-network based on machine learning is proposed in this thesis.This method combines the system simulation and the experiments,in which the leakage detection is considered as the pattern recognition problem.A total of four typical classification algorithms including BP neural networks are researched.The specific contents are as follows:Firstly,hydraulic working-condition simulation model of four experimental heating-networks(branch network with single heat source,branch network with double heat sources,single-ring network with single heat source,double-ring network with double heat sources)are established based on the graph theory and Kirchhoff’s law.Heating-network simulations and experiments are conducted towards 4 kinds of working-conditions,2 types of leakage positions(user nodes leakage and pipes leakage)and 4 kinds of leakage degrees.The results show that for normal working-conditions and leakage working-conditions of the four experimental heating-networks,the absolute value of the maximum relative error for the simulation pressure and the corresponding actual pressure of heat sources and users are 5.01%,5.52%,5.17% and 5.32% respectively;the corresponding absolute value of the maximum relative error for user flow are 5.02%,5.01%,5.51% and 5.77% respectively,which proves the effectiveness of the simulation model.Second,the labels are set,which could represent working-condition types,leakage locations and leakage degrees.After that,the model datasets and the experimental datasets are built via simulations and experiments of heating-network.14 groups of the cross datasets are built according to different cross-data ratios(the ratio of model data and experimental data).For the four heating-networks,the feature number of datasets are 54,36,54 and 36 respectively.Based on the simulation model and the principal component analysis(PCA)algorithm,it could be concluded that the correlation and the feature expression force of the model data are better than that of the experimental data.In order to explore the spatial distribution characteristics of datasets,the stochastic gradient descent,the Gaussian Bayes and the K nearest neighbor are adopted.The results show that the fluctuation of datasets in high-dimension space is large and the distance separation of datasets is great.The experimental datasets and the cross datasets are processed by the data normalization,and the BP neural network is adopted to detect the leakage of heating-networks.The poor detection effectiveness shows that the feature distribution of the experimental data is extremely unstable and different from that of the model data.Then,PCA is adopted as the data pretreatment algorithm,BP neural network,decision tree,random forest and support vector machine are used to detect leakage of heating-networks respectively.All the procedures of leakage detection are compiled with Python.The results show that the PCA-BP neural network trained via the cross datasets have the best prediction effectiveness.Under the cross-data ratio as 100:1,the leakage detection accuracies for four heating-networks are 92.21%,92.02%,89.74% and 92.50% respectively;the prediction effectiveness of the PCA-extreme random forest is the second and the corresponding prediction accuracies are 86.65%,81.68%,89.31% and 92.62% respectively.The prediction results show that the model datasets among the training set could replace the experimental datasets effectively thereby solve the lack problem of actual heating-network leakage data.In addition,the order of algorithms transfer-learning performance towards datasets features is as follows: BP neural network> extreme random forest> general random forest> decision tree> support vector machine.Finally,the city heating-network is taken as the research target and the leakage detection method of this thesis is verified.BP neural network,extreme random forest and support vector machine are selected as the core prediction algorithms.The results show that the prediction effect of PCA-BP neural network is the best one.Under the training set / test set ratio as 9:1,the PCA-BP prediction accuracy are 85.71% with pressure as the datasets,100% with flow as the datasets and 100% with pressure+flow as the datasets respectively.Followed by PCA-support vector machine,and the corresponding prediction accuracy are 85.71%,85.71% and 100%respectively.