Optimization of deep learning models for forecasting performance in the water industry using genetic algorithms

Abstract

The objective of this study is to develop an optimization approach for performance prediction in the water industry by integrating feature selection (FS), machine learning and genetic algorithms (GAs). To select the best set of input variables, FS methods such as Pearson correlation coefficient (PCC), principal component analysis (PCA), and GAs were employed. Next, a GA method was used to optimize the hyperparameters and the architecture of a deep learning network. The methodology was tested using bi-directional long-short term memory networks for univariate prediction of effluent quality and biogas production. The results demonstrate that FS methods are not universal and depend on the desired goal. Additionally, optimized machine learning applications to wastewater process data may require a shallow hidden architecture consisting of two hidden layers with an average of 25 total neurons, resulting in models that are more computationally efficient and interpretable. This proposed methodology has a wide range of applications, including predictive big data analytics, optimal data-driven development of soft-sensors, multi-objective optimization, and model predictive controllers.