Comparative study on vulnerability assessment for urban buried gas pipeline network based on SVM and ANN methods

Abstract

Vulnerability assessment is an effective way to identify systemic weaknesses, thus to provide scientific guidance to make management decisions and avoid potential accidents. Support Vector Machine (SVM) and Artificial Neural Network (ANN) have shown great superiority on vulnerability assessment in many fields. However, their applications on the relationship analysis between the state features of urban buried gas pipeline network and the corresponding vulnerability level have not yet been reported. This paper proposed a uniform framework of the novel application of SVM and ANN methods to vulnerability assessment of urban buried gas pipeline network, and performed a comparison between these two methods on various aspects. The analysis methodology follows a four-stage procedure. First, various indexes influencing vulnerability are selected and quantified for further data generation process, by which enough training and validation samples are obtained. The architecture, essential algorithms and optimized parameters for the application of these two methods are determined by the next process of model selection. The model selection phase for ANN is far more complex than that for SVM, which makes it easy for ANN to come to over-fitting. The third stage is to train SVM and ANN on the training set and to evaluate the performance on the executions of the validation set. The results show that the training outputs of SVM (MSE = 2.74E-4) are better fitted with the desired outputs than that of ANN (MSE = 1.92E-2). The SVM model (SMAPE = 0.79%) is able to output satisfying values when applied to unknown samples and is more accurate than the ANN model (SMAPE = 8.64%) in prediction. Finally, a sample gas pipeline network was used to demonstrate the feasibility and practicability of these two models, where the results are similar and consistent with the reality. The proposed methods can be used in practical applications to support better safety management.