New Approach for Prediction of Water Distribution Network Pipes Failure Based on a Intelligent Hybrid Model (Case Study: Gorgan Water Distribution Network)

Urban water distribution networks consider as one of the essential infrastructural facilities and equipment in urban areas. The pipes are one of the primary and essential components of a water distribution network break during operation due to various factors. So, developing models for pipes failure rate prediction can be one of the most crucial tools for managers and stakeholders to the optimal operation of the water distribution network. In the last decade, various studies have performed to predict the failure rate of water distribution pipes using statistical and soft models - each of which has strengths and weaknesses. This study aims to present a new approach based on the development of a hybrid prediction model, considering the capabilities of soft and statistical models, to more accurately predict the water distribution network pipes failure rate compared to statistical and soft models used in previous research.

In order to achieve the study goals, 4-year (2015-2018) time duration statistics of Gorgan water distribution network characteristics including diameter, length, age, depth of installation, and the number of pipe failures used to predict future pipes failure rates. To modeling the pipe failure rate of the investigated network, five different models, including three statistical models (linear regression, generalized linear regression, support vector regression) and two soft models (feed-forward neural network, and radial basis function neural network) has studied. Optimal parameters of the models were selected based on appropriate statistical error indicators, including correlation coefficient (R), Mean Square Error (MSE), and Correlation Mean square error Ratio (CMR) for the training and testing data. In order to select the best model from different models to predict the failure rate of network pipes, the values of R and MSE indicators of the above models were calculated in the validation stage and compared with each other. Finally, to predict pipes failure rate more accurately, a new approach is developed based on the hybrid prediction model in which the predicted values of pipe failure rates by statistical and soft computing models considered as independent variables of the best model inputs and the observed values of failure rates as dependent variables of the best model outputs.

Comparing the values of R and MSE indicators of each statistical and soft computing model used in this study in the validation phase show that these models cannot predict the pipes' failure rate with reasonable accuracy. Feed forward neural network model with the highest R = 0.69 and the lowest MSE = 0.062 values has the best estimates. Using the new approach developed based on hybrid soft and statistical models, the R index is equal to 0.96, and the MSE index is equal to 0.046.

A significant increase in the R index (39%) and decrease in the MSE index (25%) through using the proposed hybrid approach compared to the feed-forward neural network model demonstrates that using the new approach provides perfect accuracy prediction of the pipes failure rate of the water distribution network.