Water Quality Prediction Method Based on Transfer Learning and Echo State Network

Due to the adjacency effect of space, the water quality information of adjacent monitoring points in the same water area is relevant. However, traditional prediction methods use the historical water quality information of only one monitoring point rather than other nearby monitoring points. It is important to note that the sample quantity has a great influence on the prediction accuracy of the prediction methods which require lots of training samples. In addition, the water quality information is time series. Whether the time sequence of water quality information can be effectively used has direct impact on the prediction accuracy. Therefore, this paper proposes a water quality prediction method based on transfer learning and echo state network (ESN). First, we transfer the features of water quality samples of the nearby monitoring point to the prediction monitoring point based on transfer component analysis, and further select the high-quality samples based on similarity and time sequence. In this way the sample quantity of the prediction monitoring point is increased on the premise of ensuring the sample quality. Second, we construct a water quality prediction model based on ESN which takes the advantage of the time sequence of water quality information, improving the prediction accuracy on the premise of ensuring the training time. Finally, the proposed method is applied in the actual water quality dataset of Hong Kong. The experiment results demonstrate that the proposed method can expand the quantity of training samples and further improve the accuracy of water quality prediction by making full use of the space-time characteristics of water quality information.