Multi-Index Water Quality Prediction Based on Dual-Branch Temporal Convolutional and Gated Recurrent Units

Abstract

Wastewater treatment processes generate highly dynamic time-series data characterized by strong nonlinearity and complex spatiotemporal dependencies, making the accurate prediction of effluent quality and energy consumption significantly challenging. To address these limitations in conventional single-structure networks, this paper proposes a novel hybrid prediction architecture integrating a Temporal Convolutional Network (TCN) and a Gated Recurrent Unit (GRU). The front-end TCN utilizes dilated causal convolutions to efficiently extract multi-scale abrupt features from multivariable water quality data in parallel, strictly preventing future information leakage. Subsequently, the back-end GRU processes these high-dimensional features to robustly construct long-range temporal memory, adaptively capturing both longterm evolutionary biochemical trends and short-term severe fluctuations caused by sudden influent disturbances. Comparative experiments demonstrate that the proposed TCN-GRU model significantly outperforms baseline approaches, including traditional backpropagation and standard GRU networks. The hybrid model achieves the lowest prediction errors and the highest fitting accuracy for critical operational indicators. Ultimately, this architecture exhibits exceptional generalization capability, providing a robust and highly accurate data foundation for the subsequent multi-objective optimization and intelligent control of wastewater treatment processes.