Page 146 - 《高原气象》2025年第3期

高 原 气 象 44 卷
704

Using Deep Learning to Improve Short-term Climate Prediction of

Summer Precipitation in Southwestern China

ZHANG Haoyuan， QIAO Panjie， LIU Wenqi， ZHANG Yongwen
（Data Science Research Center， Faculty of Science， Kunming University of Science
and Technology， Kunming 650500， Yunnan， China）

Abstract： In recent years， Southwestern China， including Yunnan， Guizhou， Sichuan， and Chongqing， has
been frequently hit by flood disasters caused by climate change， resulting in severe casualties and enormous prop‐
erty losses. The occurrence of these disasters is closely related to abnormal precipitation. Although traditional sta‐
tistical methods and atmospheric models have achieved certain effectiveness in precipitation forecasting， effec‐
tive approaches for dealing with the complex spatiotemporal characteristics of precipitation data are still lacking.
With the development of machine learning technology， the convolutional long short-term memory network （Con‐
vLSTM）， which integrates convolutional neural networks （CNN） and long short-term memory networks
（LSTM）， has shown outstanding performance in addressing spatiotemporal sequence problems， particularly in
the field of precipitation forecasting. In order to more accurately predict the summer precipitation in the south‐
western region of China for the next year （short-term climate prediction of precipitation）， this study constructed
a dataset by integrating global sea surface temperature and precipitation data in Southwestern China. The ConvL‐
STM was used for training and named SST-ConvLSTM. This model not only captures the spatiotemporal charac‐
teristics in real precipitation data but also learns some information from global sea surface temperature data，
thereby enhancing the accuracy of short-term climate prediction of precipitation. The results show that compared
to ConvLSTM that does not consider sea surface temperature and a traditional atmospheric model， SST-ConvL‐
STM model has significant advantages in short-term climate prediction of summer precipitation in Southwestern
China.（1） Numerically， the predictions of the SST-ConvLSTM model are closest to the real precipitation data，
with similar trend changes. In contrast， both ConvLSTM and the traditional atmospheric model show certain de‐
viations in their predictions.（2） Spatially， the SST-ConvLSTM model also performs well. Its predictions are
consistent with the spatial distribution of real precipitation data and accurately reflect the spatial distribution of
precipitation.（3） In model evaluation， three evaluation metrics were used to assess the performance of the SST-
ConvLSTM model. The results show that the SST-ConvLSTM model performs well in all evaluation metrics and
achieves the best scores. These findings provide important references and insights for future research on precipita‐
tion prediction in Southwestern China.
Key words： precipitation forecasting； deep learning； ConvLSTM； sea surface temperatures； Southwestern
China