Data pre-post processing methods in AI-based modeling of seepage through earthen dams. (December 2019)

• Record Type: Journal Article
• Authors: Sharghi, Elnaz; Nourani, Vahid; Behfar, Nazanin; Tayfur, Gokmen
• Is Part Of: Measurement. Volume 147(2019)
• Publication Date: 2019-12
• Subjects: AI artificial intelligence -- ANN artificial neural network -- ANFIS adaptive neural fuzzy inference system -- ARIMA autoregressive integrated moving average -- BP back propagation -- CDF cumulative distribution function -- R2 coefficient of determination -- FFNN feed forward neural network -- MI mutual information -- MF membership function -- PDF probability density function -- RMSE root mean square error -- SOM self-organizing map -- GA genetic algorithm -- SVM support vector machine -- SVR support vector regression; Artificial intelligence -- Seepage -- Ensemble method -- Jittering -- Mutual information; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2019.07.048
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: Data pre-post processing evaluated for AI-based modeling of seepage in earthfill dam. Jittering data pre-process was used to enlarge the training sample space. Ensemble of outputs from AI models was computed in data post-processing stage. There ensemble techniques applied to enhance overall predictions of piezometric heads. Results show superiority of nonlinear neural ensemble technic over linear methods. Abstract: In this paper, seepage of Sattarkhan earthen dam in northwest Iran was simulated using various artificial intelligence (AI) models (e.g., Feed forward neural network, Adaptive neural fuzzy inference system and Support vector regression) and linear ARIMA model based on different input combinations. Both jittering pre-processing and ensembling post-processing methods were also used in order to enhance the performance of the used AI-based data driven methods. For this purpose, various jittered datasets were produced by imposing noises (at different levels) to the original time series to enlarge the training data sample space. Further, three techniques of simple linear, weighted linear and nonlinear neural averaging were considered for pre-post processing purpose. The obtained results indicated that using both jittering and ensembling (especially neural ensemble) enhanced the modeling performance by almost 30% in the testing phase.