Deep learning based multi-temporal scale precipitation modeling for spring discharge prediction, Shentou springs, China

Shentou springs, China, are one of the typical karst springs in northern China. This study focuses on Shentou springs discharge prediction based on the multi-scale precipitation. A framework called MCT that is composed of Multi-scale Convolutional network and Transformer is proposed. First, an adaptive denoising method for precipitation data is put forward. Second, a multi-scale convolutional network is designed to extract multi-temporal scale features of precipitation. Then, the Transformer is used to explore the spatial relationship between spring discharge and precipitation. Finally, the output of the coupled model is used to predict spring discharge. Results show that (1) denoising precipitation data can improve the accuracy of spring discharge prediction. (2) Precipitation and spring discharge exhibit obvious multi-temporal scale relationships, precipitation typically affects spring discharge after a lag of 6 months, and precipitation has a temporally persistent influence on spring discharge. (3) Precipitation in areas closer to the spring have an intense impact on spring discharge, while areas far from the spring but located in topographic depressions have volume-driven influence on spring discharge. • Denoising based on cosine similarity of the decomposed IMF can improve data quality. • Modeling multi-temporal scales of rainfall improves the spring discharge prediction. • Transformer self-attention finds lower-lying rains affect spring discharge largely.