Particle Swarm Optimization for Hyperparameter Tuning of One-Dimensional Convolutional Neural Networks in Wastewater Classification: An Open Problem in Feature-Dependent Generalization

Abstract

Wastewater treatment plants generate large, heterogeneous, and often imbalanced operational records, and classifying the resulting environmental and process indicators remains a persistent obstacle to automating water-quality management. This paper condenses a broader investigation into a hybrid model that couples a one-dimensional Convolutional Neural Network (1D-CNN) with Particle Swarm Optimization (PSO) for hyperparameter and feature selection, evaluated on a six-year operational record (1,383 samples, 18 variables) from Melbourne’s Eastern Wastewater Treatment Plant. Feature selection driven by PSO raised classification accuracy to 99% for Total Rainfall and 96% for Average Inflow, but the same procedure left Energy Consumption classification at only 62% accuracy, with several minority classes returning zero precision, recall, and F1-score, and a subset of ROC computations returning non-numeric (NaN) values due to insufficient class support. Rather than presenting these results as a settled solution, this paper argues that they expose an unresolved and label-dependent limitation of PSO-CNN pipelines: feature selection that is optimal for one operational target can be actively unsuitable for another drawn from the same dataset, and severe class imbalance undermines both the reliability of the reported metrics and the trustworthiness of the resulting model for operational deployment. We characterize the boundary conditions under which the PSO-CNN approach succeeds and fails, and we frame the imbalance-sensitive, target-dependent nature of feature selection as an open problem that current single-fold, single-objective PSO formulations do not resolve. The paper closes by outlining concrete directions (class-balancing prior to optimization, multi-objective PSO formulations, cross-validated fitness functions, and target-specific feature subsets) that future work is invited to pursue.