Abstract

Background: Most NN (neural network) research only conducted qualitative analysis, analyzing its accuracy, with certain limitations, without studying its NN model, error convergence process, and pressure ratio. There is relatively limited research on the application of NN optimized by GA (genetic algorithm) to oil and gas pipelines; Moreover, the residual strength evaluation of GA-BP NN (genetic algorithm backpropagation neural network) has the advantages of high global search ability, efficiency not limited by constant differences, and the use of probability search instead of path search, which has a wide application prospect.

Objective: Using MATLAB software, establish GA-BP NN models under five residual strength evaluation criteria and introduce the relative error of the parameters and the pressure ratio to comprehensively analyze the accuracy and applicability of GA-BP NN.

Methods: Firstly, using MATLAB software, a GA-BP NN model was established based on five residual strength evaluation criteria: ASME B31G Modified, BS7910, PCORRC, DNV RP F101, and SHELL92, by changing five factors that affect the residual strength of oil and gas pipelines: diameter, wall thickness, yield strength, corrosion length, and corrosion depth; Second, the trained GA-BP NN model is used to predict the residual strength of the same set of evaluation criteria test data and compared with the calculation results of five residual strength evaluation criteria. The relative error of the parameters and pressure ratio are introduced to comprehensively analyze the accuracy and applicability of the GA-BP NN.

Results: The error convergence time of the BP NN is longer, and the optimized GA-BP NN has a shorter convergence time. By comparing the convergence training times of different models, it can be obtained that for the five sets of residual strength evaluation criteria of ASME B31G Modified, BS7910, PCORRC, DNV RP F101, and SHELL92, the optimized GA-BP NN model significantly reduces convergence training times, significantly improves convergence speed, and further evolves the system performance. From the relative error and local magnification, it can be seen that for the ASME B31G Modified evaluation criteria, the maximum relative error of the BP NN model is 1.4008%, and the maximum relative error of the GA-BP NN model is 0.7304%. For the evaluation criterion BS7910, the maximum relative error of the BP NN model is 0.7239%, and the maximum relative error of the GA-BP NN model is 0.5242%; for the evaluation criteria of DNV RP F101, the maximum relative error of the BP NN model is 1.1260%, and the maximum relative error of the GA-BP NN model is 0.4810%; for the PCORRC evaluation criteria, the maximum relative, error and the maximum relative error of the GA-BP NN model is 0.8004%; for the SHELL92 evaluation criterion, the maximum relative error of the BP NN model is 1.2292%, and the maximum relative error of the GA-BP NN model is 0.8346%. The results of the GA-BP NN prediction are closer to the results of the calculation of the five residual strength evaluation criteria, and the prediction effect is better, which can more accurately predict the residual strength of the oil and gas pipelines. Based on the pressure ratio, the average pressure ratio A of the BP NN model under the five residual strength criteria is 1.0004, and the average pressure ratio A of the GA-BP NN model is 0.9998. The results predicted by the GA-BP NN model are more accurate.

Conclusion: These findings have crucial implications for the forecast of the residual strength of corrosive oil and gas pipelines.

Keywords: MATLAB, residual strength, GA-BP NN, genetic algorithm, SHELL92.

Graphical Abstract