Abstract:

Due to weak water exchange capacity and significant input of land-based pollutants from surrounding rivers, Shenzhen Bay experiences frequent red tide occurrences and marked seawater quality fluctuations. Timely and accurate prediction of dynamic changes in Shenzhen Bay's seawater quality is crucial for early warning of marine disasters in the region. Traditional water quality prediction models or single machine learning models face limitations such as low computational efficiency and insufficient capture of long-term dependencies. To address these challenges, this study proposes a hybrid deep learning model based on variational mode decomposition (VMD) and compares its performance with basic deep learning models in seawater quality prediction for Shenzhen Bay. The "decomposition-prediction-reconstruction" hybrid model first applies VMD to decompose seawater quality and nutrient data into five stable intrinsic mode functions (IMFs). These IMFs are then fed into temporal convolutional network (TCN) and long short-term memory (LSTM) models for individual prediction, with the final prediction value generated by summing all IMF predictions. The results demonstrate that: (1) The TCN-VMD and LSTM-VMD hybrid models significantly outperform the basic TCN and LSTM models, with the highest improvement in Nash-Sutcliffe efficiency coefficient (NSE) reaching 91.69%; (2) TCN-VMD performs better than LSTM-VMD in terms of NSE, achieving a maximum improvement rate of 95.56%. For water temperature (0.998), pH (0.971), salinity (0.988), dissolved oxygen (0.960), chlorophyll-a (0.909), ammonia nitrogen (0.911), phosphate (0.909), nitrate (0.969), and nitrite (0.968), TCN-VMD accurately captures the peaks and troughs of the true values; (3) TCN-VMD provides a more accurate and efficient deep learning framework for seawater quality prediction in Shenzhen Bay, effectively supporting early warning of coastal marine disasters.

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