A water quality prediction model based on neural network at data-scarce sites

Abstract

Artificial intelligence technology (AI) has been widely applied in water quality prediction owing to its superior predictive capabilities. However, AI models typically require extensive datasets for parameter calibration. The predictive performance of AI models for surface water quality prediction at data-scarce sites remains to be investigated. Therefore, this study proposed a prediction framework at data-scarce water quality sites along Yellow River Basin using neural network model. By using the source domain and transfer learning hyperparameters, we have validated that transfer learning can significantly improve the prediction performance of neural network models with median improvement of 50%, effectively addressing the issues of poor surface water quality prediction at data-scarce sites. Furthermore, similarity measurement was proposed to construct model transferring the knowledge from source domain to target domain. Similarity measurement is positively correlated with the effectiveness of transfer learning. The hyperparameters of transfer learning have a significant impact on its application effectiveness. We recommend using validation samples reserved from the target domain. This approach can effectively ensure the performance of transfer learning applications.