An adaptive neural network embedded genetic algorithm approach for inverse water quality modeling

Abstract

[1] This paper proposes a neural network (NN)-embedded genetic algorithm (GA) approach for solving inverse water quality modeling problems to overcome the computational bottleneck of inverse modeling by replacing a water quality model with an efficient NN functional evaluator. An existing one-step, NN-embedded GA approach is found incapable of solving an inverse water quality modeling problem because it tends to fail in guiding the global search process to converge toward the near optima. As a remedy, an adaptive NN-GA approach is proposed to achieve a gradual convergence toward the near optima through an iterative network learning method. The proposed approach is applied to a full-scale, numerical example, and the result shows that the adaptive NN-GA approach is capable of obtaining near-optimal solutions for the inverse problem of a complicated water quality model.

Estimation of Hydraulic Conductivity and Its Uncertainty from Grain-Size Data Using GLUE and Artificial Neural Networks

“…Because of the very low gradients applied, the permeability of the samples is not believed to be altered. Additional 27 hydraulic conductivity measurements were performed using a modified constant-head permeameter adapted to medium-to-low K s -values as described in Wemaere et al (2002). For the latter, cylindrical samples are transferred into a stainless steel cylindrical cell with two sintered stainless steel filters (pore diameter of 10 µm) at both ends.…”

“…Moreover, it has been proven theoretically that multilayer feedforward networks are universal approximators (Hornik et al 1989;Hornik 1991). Environmental applications of this technique include stream flow, flood and rainfall forecasting (Kişi 2007;Jain and Kumar 2007;Tiwari and Chatterjee 2010;Valverde Ramírez et al 2005), processing remote sensing data (Linderman et al 2004), flow and transport simulations (Morshed and Kaluarachchi 1998), groundwater table prediction (Coppola et al 2005;Joorabchi et al 2009) and water quality modelling (Zou et al 2007;Khalil et al 2011). In soil science, ANNs have been used extensively for predicting soil hydraulic properties from more easily measurable parameters.…”

An Indirect Simulation-Optimization Model for Determining Optimal TMDL Allocation under Uncertainty

“…Taken together, the total computational time taken to find out the approximate optimal solutions was only about 33 h. In contrast, running a traditional direct SOM framework using a GA approach might require several days (e.g., if a population size of 50 is used and the model is iterated for 300 generations, it takes over 10 days to obtain a solution) [11]. More time would be required to increase the chances of obtaining globally-approximate optimal solutions, because multiple GA runs would need to be executed for the same problem [21,22]. Moreover, the computation efficiency of the proposed BRRT-EILP model can be significantly higher than that of the traditional direct SOM approach when Equation (3) is used to evaluate solutions for different objective functions, water-quality criteria or system constraints.…”

“…For example, traditional NP algorithms, including both gradient and non-gradient based ones, were limited to local optima upon solving the aforementioned SOM framework [11,21]. Although heuristic global search algorithms, including genetic algorithms, evolutionary algorithms and simulated annealing, are capable of surpassing the local optima limitations, their applications in the optimization of TMDL allocation are still restricted by their extremely high computational cost [18,22].…”

“…Generally, the framework with the approximation function needs to be sufficiently generalized to achieve optimal or approximate-optimal solutions of the original direct SOM system. This means that the regression tree model [25], the response surface [14] or the artificial neural network (ANN) [22] should truly capture the underlying functional form of the simulation model rather than just the training data used to calibrate the models. Although the ANN approach is one powerful platform for developing an approximation function, it has not yet been shown to provide explicit expressions of nonlinear stressor-response relationships and effectively quantify uncertainties in dynamic water body systems.…”