Precipitation Forecasting Based on the Interpretability of Neural Network Models
FAN Zhongxin
WANG Yan
WANG Ruotong
Abstract：To improve the accuracy and reliability of localized,fine-scale precipitation forecasts,the present study proposed a new neural network capable of precipitation forecasting based on KernelExplainer and clustering of interpretable Shapley additive explanations(SHAP)values.First,the output jitter in the neural network was addressed by using normalized distribution transformation.Subsequently,we estimated the deep learning neural network model comprising convolutional(CNN)layers,long short-term memory(LSTM)networks,and dense layers using the KernelExplainer.This process yielded SHAP values that represent the contributions of meteorological parameter m and time step parameter tl to forecasting results.Finally,by dynamically adjusting the model's m and tl parameters through SHAP value clustering in each rolling forecast,we managed to use the method to improve forecasting performance for non-precipitation and heavy precipitation events.Using this method,a precipitation forecasting model for the Atmospheric Observation Station of Nanjing University of Information Science&Technology was established based on observational data and numerical weather prediction model outputs from January 2018 to December 2023.Experimental results show that,compared to fixed-parameter models,multilayer ConvLSTM models,Analog-Ensemble-CNN models,and numerical weather prediction models,the proposed model reduced the mean absolute error of precipitation forecasts by 8%,7%,11%,and 19%,respectively.
Keywords：convolutionlong short-term memoryReLU activation functionKernelExplainerShapley additive explanations(SHAP)precipitation forecasting
Publication Date：2024-12-28
Online Publishing Date：2025-08-15(First online date of this platform, not the publication date of the document)
Pages：15( 1030-1044 )