Statistically Optimized Back-Propagation Neural-Network Model and Its Application for Deformation Monitoring and Prediction of Concrete-Face Rockfill Dams

Abstract

Concrete-face rockfill dams (CFRDs) are widely used in hydropower engineering. Deformation monitoring and safety operation of CFRDs is of great significance in ensuring the safety of human life and property in downstream areas. A new prediction model for the horizontal displacement of CFRDs, the statistically optimized back-propagation neural network model, was proposed by combining a statistical model and a back-propagation neural network (BPNN) model, which was applied to Deze Dam, Yunnan Province, Southwest China. First, the thermometer selection method is improved based on the correlation coefficients between the measured values of thermometers and the horizontal displacement. Further, combined with water level and time factors, three thermometers with large correlation coefficients were selected and applied to Deze Dam’s model training. On this basis, an improved statistical model for the horizontal displacement of CFRDs is proposed. Subsequently, prediction results of the improved statistical model are taken as an input vector of the traditional BPNN model. Then the statistically optimized BPNN model, a combination of the improved statistical model and BPNN model, is proposed to predict the horizontal displacement of CFRDs. Compared with the improved statistical model and the BPNN model, the statistically optimized BPNN model has a higher prediction accuracy and a strong nonlinear prediction capability, which can compensate for the errors of statistical models and overcome the defects of overfitting and local minima. In addition, the statistically optimized BPNN model proved to have a strong generalization capability by changing the training sample sizes.