基于模型融合的供水管网渗漏预测研究

Research on leakage prediction of water supply pipe network based on model fusion

供水管网泄漏问题对水资源造成了严重浪费和污染。通过利用华中某企业的管网GIS数据库建立数据集,在传统分类模型中分析学习曲线和可靠性曲线,筛选出随机森林(RF)、逻辑回归(LR)和BP神经网络(BP)模型,并对其进行优化。最后,为克服单一预测模型的不足,引入了基于堆叠法(Stacking)和投票法(Voting)的融合模型。在融合模型中,将这三种模型作为基础模型,LR作为元模型。通过扩大模型的宽度,成功提高了模型的性能。实验结果表明,融合模型在测试集中预测供水管网漏水的准确率达到94.9%,AUC值为0.970。融合模型的预测能力明显优于任何仅使用单一特征构建的分类器,并具备良好的泛化能力和鲁棒性。

The leakage problem of water supply pipe network poses a threat of serious waste of water resources and water pol⁃lution.By using the pipe network GIS database of an enterprise in central China to build a dataset,the learning curve and reliability curve are analyzed in the traditional classification model,and the random forest(RF),logistic regression(LR)and BP neural net⁃work(BP)models are screened out and optimized.Finally,a fusion model based on Stacking(Stacking)and Voting(Voting)is in⁃troduced to overcome the shortcomings of a single prediction model.In the fusion model,these three models are used as the base model and LR as the meta⁃model.By expanding the width of the model,the performance of the model was successfully improved.The experimental results show that the fusion model predicts water leakage in the water supply network with an accuracy of 94.9%and an AUC value of 0.970 in the test set.The prediction ability of the fusion model is significantly better than any classifier con⁃structed using only a single feature with good generalization ability and robustness.