A high-frequency time-series prediction model for dissolved oxygen based on multi-source data in estuarine areas

Accurately predicting the intermittent and highly fluctuating hypoxia at the estuary of a small river, particularly in urbanized areas where the upstream water is notably affected by human activities, presents a significant challenge. The dissolved oxygen concentration (DO) in tidal estuaries is affected by multiple driving factors, and the dynamic changes in the impacts of these factors make it difficult to develop an hourly forecasting model. However, accurate prediction of DO is of utmost importance for water ecological security and aquaculture. To tackle this issue, this study combined an "end - to - end" deep learning model and incorporated an encoder - decoder architecture based on the latest time series decomposition forecasting model, TimeMixer. It realized dynamic feature selection via multi - scale time encoding and set up an attention mechanism at both the time step level and across components to emphasize the dynamic contributions of driving factors. In hourly forecasting, TimeMixer outperformed the enhanced long short - term memory (LSTM) model, and the enhanced Fusion Attention - TimeMixer (FAT) model enhanced the overall prediction by an average of 17% across all indicators. Specifically, the Nash - Sutcliffe Efficiency (NSE) increased by 24.74%, and the Mean Absolute Error (MAE) decreased by 14.83%. The F1 Score for hypoxia prediction went up by 3.24%, and the forecast error for peak values dropped by 12.56%. Additionally, this study explored the impact of adjusting the input and output windows and integrating multi - source monitoring data on the prediction accuracy of the FAT model. Specifically, when predicting DO for the next 48 hours, a 48 - hour historical window input is optimal for the model's prediction performance. If the prediction length is extended, a longer historical data window than the future window should be used as input. During the model improvement process, it was discovered that the introduction of feature - level attention mechanisms is not advantageous for DO forecasting during rainfall periods. Nevertheless, conducting attention calculations at the time - step dimension and realizing dynamic feature selection through multi - scale time encoding can respectively improve the root mean square error (RMSE) and mean absolute error (MAE) by 7.87% and 12.15% for DO forecasting during rainfall periods. Zero - shot prediction experiments were carried out on the improved deep learning models. Although the prediction performance of most models declined, the FAT model still maintained satisfactory performance and achieved acceptable prediction results even with a reduced number of features. It was discovered that when only tide level and meteorological indicators were used as inputs, the NSE of the FAT model's prediction could reach 0.82. Finally, the prediction effectiveness of the FAT model for DO in five typical scenarios at the estuary was tested to provide references for different station establishment schemes, and the results were presented via a web interface to allow users to predict and evaluate the model. Overall, this study validated the feasibility of predicting non - stationary DO sequences at estuaries using multi - source monitoring data, offering crucial methodological support for real - time prediction and systematic management of estuaries.