Using Artificial Neural Networks to Determine the Concentration of Sediments at the Limit of Deposition in Sewers

The self-cleansing design concept in drainage sewers is fundamental to avoid sediment deposits inside the drainage sewers which implies in the loss of discharge capacity of the structure. Unfortunately it is known that most of available design equations are based on simple regression models.

The aim of this study is to apply the neural networks before regression in order to improve the design equation. As a result, this study proposes one new regression equation using the results from the artificial neural networks. The study use 508 experimental data obtained from the literature of four authors (Mayerle (1991), Ab Ghani (1993), May et al. (1996) and Ota (1999)). The limit of deposition was analyzed with the proposing the use of feedforward artificial neural networks (Multi-layer perceptron) for the determination of volumetric concentration of sediments at the limit of deposition in drainage sewers.

The study presents the performance of 25 different neural networks architectures, trained using half of the experimental data analyzed. After the training phase, all the architectures was evaluated through one dataset of the other half of the datasets obtained in the literature. The selection of the architecture with the best performance to generalize the problem were made using the correlation coefficient (R²) and the standard deviation, analyzing the results presented by the all the neural networks architectures and comparing then with the results provided in the experiments.

After that, it was presented another regression equation obtained using the results provided by the artificial neural network. The study concluded that the use of artificial neural networks is proper in the determination of the volumetric concentration at the limit of deposition in drainage sewers. When compared with the other regression models and with the experimental data utilized the artificial neural network shown good results.