Hybrid CEEMDAN–deep learning models for COD prediction in wastewater treatment plants

Abstract

Precise and timely chemical oxygen demand (COD) prediction is key to regulatory compliance and process control in a full-scale wastewater treatment, but laboratory assays and sluggish online measurements hinder responsiveness. Thus, a deployment-conceived hybrid pipeline for near-term COD predictions is proposed in this study. This method introduces a deployment-oriented hybrid framework that integrates complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), gated recurrent units (GRU), and a multi-objective observer-teacher-learner-based optimizer (MOOTLBO) for near-term COD forecasting. Daily measurements from the Gebze municipal wastewater treatment plant (Türkiye) between 2019 and 2023 were collected from regularly available measured data together with total suspended solids, total nitrogen, total phosphorus, and pH with statistically enlightened lags to encode process memory. Intrinsic mode functions were obtained from the COD series with the application of CEEMDAN, and the intrinsic mode functions were concatenated with exogenous features and forwarded to recurrent learners. Long short-term memory (LSTM) and gated recurrent units (GRU) were studied as two potential configurations in the study. Hyperparameters were optimized with a multi-objective observer-teacher-learner optimizer (MOOTLBO) in order to balance fidelity, parsimony, and training cost. Evaluation occurred on a held-out test set (80/20 partition) with RMSE, MAE, R², MAPE, MBE, and KGE metrics. Across alternatives, decomposition-then-learn decreased error compared with non-decomposed baselines. A CEEMDAN–GRU variant turned out to be the most accurate configuration, which reached good performance with RMSE = 2.84 mg L⁻¹, MAE = 2.28 mg L⁻¹, R² = 0.9999, MAPE = 0.46%, MBE = 1.47 mg L⁻¹, and KGE = 0.9974. In an optimized CEEMDAN–GRU–MOOTLBO configuration, similarly strong accuracy (RMSE = 4.19 mg L⁻¹, MAE = 2.66 mg L⁻¹, R² = 0.9999, MAPE = 0.52%, MBE = 0.16 mg L⁻¹, KGE = 0.9997) was obtained with excellent bias control. In view of accuracy, stability, and runtime, the CEEMDAN–GRU family represents the most suitable choice for real-time soft sensing of COD, facilitating proactive control and early warning with the presence of changing influent conditions in challenging plant practice.