Abstract: Owing to the bay’s limited hydrodynamic exchange and the substantial influx of land-derived pollutants from adjacent rivers, Shenzhen Bay experiences frequent harmful algal blooms and pronounced water-quality fluctuations. Accurately and promptly forecasting these dynamics is therefore essential. This study systematically evaluates the predictive skill of deep-learning ensemble models that integrate Variational Mode Decomposition (VMD) against conventional, non-decomposed deep-learning models for water-quality forecasting in Shenzhen Bay. Within the proposed “decomposition-prediction-reconstruction” framework, the water-quality and nutrient time-series are first decomposed by VMD into five stationary intrinsic mode functions (IMFs). These IMFs are then input into Temporal Convolutional Network (TCN) and Long Short-Term Memory (LSTM) models to predict each individual component function. Finally, the predicted results are aggregated to generate the ultimate prediction.Key findings are as follows.(1). The hybrid TCN-VMD and LSTM-VMD models demonstrated significantly improved prediction performance compared to the baseline TCN and LSTM models, with the maximum improvement rate reaching 91.69%.(2). TCN-VMD performs better than the LSTM-VMD model in terms of the Nash-Sutcliffe efficiency coefficient, achieving a maximum improvement of 95.56%. The corresponding Nash-Sutcliffe efficiencies are 0.998 for water temperature, 0.971 for pH, 0.988 for salinity, 0.960 for dissolved oxygen, 0.909 for chlorophyll-a, 0.911 for ammonium, 0.909 for phosphate, 0.969 for nitrate, and 0.968 for nitrite; the TCN-VMD framework faithfully captures observed peaks and troughs.

Key words: VMD, TCN, LSTM, seawater quality forecasting, Shenzhen Bay