Xin-zhi Zhou, one of the authors of a new Journal of Hydroinformatics paper, has written a blog post summarising the article. 

Find out more and access the paper below: 

Daily runoff forecasting based on data-augmented neural network model

Xiao-ying Bi, Bo Li, Wen-long Lu and Xin-zhi Zhou

https://doi.org/10.2166/hydro.2020.017

Author summary

We all know that water is the source of life. Water is very important for human beings, but too much water can be dangerous for humans. The flood disaster has been happening all the time, the accurate flood prediction can play a warning role to people. For decades, a large number of researchers have done countless research on flood prediction. This paper adopts a neural network method for flood forecasting.  The proposed neural network model and data processing method can effectively improve the accuracy of flood prediction. 

The model, CAGANet, is a combination of convolutional layer, GRU, autoregressive model, and attention mechanism, with each module contributing to the final model's predictive performance. In the model, CNN Layer has the function of extracting data features, GRU Layer has the function of memorizing data information, Attention Layer is used to focus on important features of the data, and Autoregressive is used to obtain the linear characteristics of the data. In addition, based on the analysis of the characteristics of the data, the data interpolation is proposed. This simple method is portable to improve the prediction accuracy of the daily flow and is applicable to other neural network models.

Based on the social problem of flood disaster, the model and method proposed in this paper have strong robustness and portability, which is helpful to accurately predict flood disaster. We cannot prevent the occurrence of disasters, but we can avoid more loss of life and property by predicting and warning the occurrence of disasters.