Study on turbidity prediction method of reservoirs based on long short term memory neural network

Abstract

Turbidity is one of the important indicators in water quality management of reservoir. There are many factors affecting turbidity, and its time series is non-linear, making prediction difficult. Therefore, it is necessary to carry out research on reservoir turbidity prediction methods. In this study, the Long Short-Term Memory (LSTM) neural network was identified, validated and tested for the computation of turbidity in the Qingcaosha Reservoir. The model employed historical data of turbidity, water level, wind direction and wind speed over a period of 2 years at various monitoring points. Within 40 iterations of the model, the mean square error converged to less than 0.05 steadily, and the Nash efficiency coefficient of the 24 h prediction was above 0.5. It showed that the model has the characteristics of fast convergence, high stability, and accurate prediction, which meant this model can be well applied to prediction of reservoir turbidity. This study also tried to use the forecasted wind field data to improve the actual turbidity prediction of the reservoir. The results showed that the accuracy is slightly lower than the predicted result using the measured wind field data, but it was significantly higher than the prediction result using the extended wind field data at the previous time point. Therefore, using forecasted wind field data can effectively improve the accuracy of the actual reservoir turbidity forecast. The results of this study indicate that the LSTM neural network model is fast, stable, and highly accurate, indicating that it is suitable for prediction of turbidity in reservoirs and can provide support for water quality management of reservoirs.