<p>Accurate prediction of urban surface water quality is essential for effective water-environment monitoring and management. However, the strong nonlinearity and non-stationarity of water-quality time series remain a major challenge for existing data-driven models. To address this issue, a hybrid forecasting model (CEEMDAN-BiTCN model) is proposed. The complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) based signal decomposition was integrated to enhance data quality and mitigate noise-induced instability. Moreover, the bidirectional temporal convolutional network (BiTCN) was developed to simultaneously capture forward and backward temporal dependencies in multivariate sequences. Consequently, a unified framework was constructed, which effectively combines signal processing and deep learning for robust water-quality prediction. The proposed model in this work was evaluated using real monitoring data from the Lincheng River basin in Zhoushan, China, covering six key indicators, including temperature, pH, dissolved oxygen (DO), ammonium nitrogen (AN), total phosphorus (TP), and total nitrogen (TN). Comparative results demonstrate that the proposed approach outperforms several benchmark models, achieving high prediction accuracy with R² values ranging from 0.881 to 0.980. In addition, generalization experiments conducted on data from other regions show that the model maintains strong predictive performance, with R² values close to 0.90, indicating good transferability and scalability. These findings indicate that the CEEMDAN-BiTCN framework provides a reliable and effective solution for forecasting complex urban water quality.</p>