Prediction of dissolved oxygen concentration in aquatic systems based on transfer learning

Prediction of dissolved oxygen (DO) concentration pattern is important for aquatic system management and environmental monitoring. The large amounts of experimental data needed often limit the ability to develop a reliable DO prediction model for a given aquatic system. In this research, deep learning and transfer learning techniques were applied to take advantage of a large available dataset for one aquatic system in predicting DO concentration trend in another (target) system for the first time. A pre-training DO prediction model incorporating deep learning algorithms of ResNets, BiLSTM, and Attention was established based on the large dataset. The knowledge obtained and retained by the pre-training model was then transferred to develop a DO prediction model for the target system with a much smaller amount of available data. To show the benefits of transfer learning, a DO prediction model of the same structure was developed for the target system with its own data without transfer learning from the first system. The root mean square error (RMSE), mean absolute percentage error (MAPE), coefficient of determination (R-2), index of agreement (d), and Nash-Sutcliffe efficiency coefficient (NSE) were used to measure the performance of the models. The results showed that the model structure used was useful in learning and retaining knowledge from the first system. In terms of all performance measures, transfer learning improved DO time series prediction for the target aquatic system and allowed development of a prediction model for the target system without a large set of measured data.

Automated summary

The research successfully applied deep learning and transfer learning techniques to predict dissolved oxygen concentration trends in a target aquatic system using a large available dataset from another system. Transfer learning improved the predictions for the target system and enabled the development of a prediction model with limited measured data.