Subseasonal Forecasting and Analysis of China Temperature Using an Explainable Deep Learning Model

ABSTRACT Reliable subseasonal temperature forecasting plays an important role in preventing extreme temperature events and reducing damages from such events. In this study, we develop an explainable deep learning approach by integrating three consecutive modules, including the long‐short‐term‐memory‐based encoder module, fully‐neural‐network‐based decoder module, and expected‐gradient‐based explanation module, to effectively forecast subseasonal temperature in China and discover its primary influencing factors. The first two modules constitute an encoder‐decoder, which can forecast the 2‐week average temperature of weeks 3–4 and weeks 5–6 in advance, while the explanation module computes attribution values for the input features to quantify their importance. The proposed model is examined with 957 grid points over China and compared with the multitask Lasso, multiple linear regression, and Climate Forecast System version 2. The results of forecasting demonstrate that our model performs best overall among four models in the test period. The results of feature attribution show that ocean‐based and land‐based variables are much more significant than atmosphere‐related variables, and soil moisture is the most significant variable for subseasonal temperature forecasting in China, which is consistent with existing knowledge of physics. The study's significance lies in enhancing the accuracy of subseasonal temperature forecasting in China and providing explainability through the identification of influencing factors.