Evaluation of different artificial intelligent methods for predicting dam piezometric water level

• Published in: Modeling earth systems and environment 2022-06, Vol.8 (2), p.2715-2731
• Main Authors: Ziggah, Yao Yevenyo, Issaka, Yakubu, Laari, Prosper Basommi
• Format: Article
• Language: English
• Subjects: Artificial neural networks; Back propagation; Back propagation networks; Chemistry and Earth Sciences; Computer applications; Computer Science; Correlation coefficient; Correlation coefficients; Dam safety; Dam stability; Dams; Deformation; Earth and Environmental Science; Earth Sciences; Earth System Sciences; Ecosystems; Environment; Gaussian process; Group method of data handling; Hydraulic structures; Math. Appl. in Environmental Science; Mathematical Applications in the Physical Sciences; Methods; Neural networks; Original Article; Physics; Piezometers; Predictions; Production methods; Radial basis function; Root-mean-square errors; Seepage; Software; Statistics for Engineering; Structural health monitoring; Support vector machines; Water levels
• ISSN: 2363-6203; 2363-6211
• DOI: 10.1007/s40808-021-01263-9

Stability is an important measure to consider when dealing with dam structural health management system. Dams are hydraulic structures built with impenetrable materials which serve as a barrier to the flow of water. Geodetic and geotechnical observables such as seepage clarity and flow, water level, piezometric water level, pressure, temperature variation, deformation and loading conditions are often measured for dam safety control. This study is focussed on piezometric water level which is an important parameter to support seepage analysis of dams. The efficiency of least squares support vector machine (LSSVM), group method of data handling (GMDH), M5 prime and Gaussian process regression (GPR) were explored for the first time in piezometric water level prediction. These methods were then compared with the widely used backpropagation neural network (BPNN), support vector machine (SVM) and radial basis function neural network (RBFNN). The seven methods were tested on experimental data collected at four different piezometers located at different positions of the dam for a period of 2 years 4 months in Ghana. It was generally observed from the prediction outputs that all the methods applied could produce very reasonable and applicable results. However, ranking the results according to root mean square error (RMSE), percent mean absolute relative error (PMARE), Correlation Coefficient (R), Loague and Green (LG), and variance accounted for (VAF) revealed the GMDH as the best prediction approach for all the piezometers. It was concluded that the implemented artificial intelligent techniques constitute reliable computational tools for dam piezometric water level prediction.