Short-term Precipitation Correction Based on GRU Deep Learning

Abstract

To improve the accuracy of short-term precipitation forecasts, a deep learning method is proposed to correct numerical model precipitation forecast products. This method extracts spatiotemporal features from numerical model forecasts and observations using a neural network and performs corrections based on a gated recurrent unit (GRU) framework. Additionally, an atmospheric physics adaptor module is meticulously designed to address systematic errors in the intensity and displacement of numerical model forecast by leveraging physical condition mechanisms. The module plays a crucial role within the overarching model framework, which consists of three integrated components: Feature network, recurrent-revising network, and physical adaptor. The feature network extracts precipitation intensity, distribution, motion characteristics and other related atmospheric physical features from precipitation in situ and numerical model forecast data, serving as input to the recurrent-revising network. Recurrent-revising network utilizes a recurrent neural network structure to adjust grid point forecast results on a time-step basis. Deep neural networks are used to extract spatiotemporal variation features from numerical model forecast data and historical observations, learning systematic errors in the evolution processes to correct the precipitation magnitude and distribution. The physical adaptor is an atmospheric physics adaptation module, which preprocesses numerical model forecast data using frequency distribution fitting and distribution deviation correction methods. In Guangxi convective-scale model precipitation forecast data, when there are significant differences between numerous samples and the precipitation in situ, the feature correlation is low, making it a challenge to capture systematic error characteristics during neural network training. By preprocessing the samples with the physical adaptor, differences between forecasts and observations are reduced, enhancing feature correlation between training input datasets and observations, thus facilitating better neural network training and achieving superior correction skills. This method not only adheres to but also integrates fundamental atmospheric physical laws governing precipitation evolution, thereby offering a robust and innovative approach for post-processing numerical model short-term precipitation forecasts. By incorporating these physical principles into the model framework, corrected forecasts not only reflect statistical patterns but also adhere closely to the underlying physical processes driving precipitation dynamics. Experimental results in Guangxi indicate that the model demonstrates positive threat score skills across various forecast times and precipitation intensities. Specifically, for different precipitation intensities (average of all times), threat score skills for 0.1 mm·h-1, 2 mm·h-1, 7 mm·h-1, 15 mm·h-1, 25 mm·h-1, and 40 mm·h-1 are 5.67%, 3.59%, 2.18%, 1.46%, 1.01%, and 0.46%, respectively; for different lead times, threat score skills for 0-2 h, 2-4 h, and 4-6 h are 4.77%, 1.28%, and 0.91%, respectively; and the overall average threat score skill across all precipitation intensities and times is 2.21%.

Keywords

• forecast correction
• deep learning