Abstract: Real-time variations in water quality environment and the internal coupling pose challenges to predicting water quality efficiently and accurately. To extract further information from the time series data and enhance the accuracy of prediction models, we propose a novel combined prediction model for Dissolved Oxygen (DO). To begin with, the water quality data is decomposed using the Seasonal and Trend decomposition using LOESS (STL) technique. Subsequently, the obtained seasonal component, trend component, and residual component are fed into the Long Short-Term Memory network model (LSTM) for prediction. In order to address the impact of LSTM network initialization parameters on prediction performance, we introduce an optimization approach based on the Fruit fly algorithm utilizing Gaussian function. Finally, the predicted values of each component are reconstructed to the DO prediction values. To validate the effectiveness of the proposed model, we conduct simulation tests utilizing water quality data from three sections of Haihe River. Results demonstrate that the combined model exhibits a favorable prediction effect on the DO concentration of the three stations, characterized by minimal errors and strong generalizability.

Key words: water quality prediction, STL, fruit-fly algorithm, LSTM, dissolved oxygen