Developing an evolving multi-layer perceptron network by genetic algorithm to predict full-scale ‎municipal ‎wastewater treatment plant effluent

Abstract

Developing an efficient model for Wastewater Treatment Plant (WWTP) non-linear process simulation is a vital tool for managing wastewater reuse and recycling. A novel evolving Multilayer perceptron (MLP) by the Genetic Algorithm (GA) was applied to a full-scale municipal ‎WWTP providing high-precision models to predict effluent concentration under various conditions and highlight possible bottlenecks’ potential improvements. Besides the MLP architecture optimization by GA, three different scenarios were considered to study the effect of input variable space pre-processing. In the first scenario twenty-one, original input variables were fed into GA-MLP, while in the second scenario, the basic GA-MLP was extended to assign a weight to each input variable simultaneously by architecture optimization, and in the third scenario, input variables were first pre-processed by principle component analysis. Each scenario was repeated 30 times, and a multi-criteria ranking system based on four performance indices was designed to select the best scenario and MLP. Results show that the second scenario and its best MLPs provided the highest level of precision and generalization. The best MLP of the second scenarios’ Nash-Sutcliffe efficiency and root mean square errors for train and test data were (‎1.000‎, 0.004) and (‎0.914, ‎2.741) ‎ for chemical oxygen demand, (‎0.975,‎ 0.329)‎ and (‎0.915, ‎0.689)‎ for biochemical oxygen demand and (‎0.911, ‎1.186) and ‎(0.875, ‎1.415) for total suspended solids. The proposed hybrid intelligent model was able to efficiently develop near-optimal MLPs to simulate the non-linear WWTP processes and provide an accurate monitoring tool for managers to increase the resilience of WWTPs.