Deep learning-based subseasonal to seasonal precipitation prediction in southwest China: Algorithm comparison and sensitivity to input features
The prediction of precipitation at subseasonal to seasonal (S2S) timescales remains an enormous challenge because of the gap between weather and climate predictions. This study compares three deep learning algorithms, namely, the long short-term memory recurrent (LSTM), gated recurrent unit (GRU), a...
| Published in: | Earth and Planetary Physics |
|---|---|
| Main Authors: | , , , , , |
| Format: | Article in Journal/Newspaper |
| Language: | English |
| Published: |
Science Press
2023
|
| Subjects: | |
| Online Access: | https://doi.org/10.26464/epp2023049 https://doaj.org/article/e14503f4ed464ee6b282dc0352328099 |
| Summary: | The prediction of precipitation at subseasonal to seasonal (S2S) timescales remains an enormous challenge because of the gap between weather and climate predictions. This study compares three deep learning algorithms, namely, the long short-term memory recurrent (LSTM), gated recurrent unit (GRU), and recurrent neural network (RNN), and selects the optimal algorithm to establish an S2S precipitation prediction model. The models were evaluated in four subregions of the Sichuan Province: the Plateau, Valley, eastern Basin, and western Basin. The results showed that the RNN model had better performance than the LSTM and GRU models. This could be because the RNN model had an advantage over the LSTM model in the transformation of climate indices with positive and negative variations. In the validation of test datasets, the RNN model successfully predicted the precipitation trend in most years during the wet season (May–October). The RNN model had a lower prediction bias (within ±10%), higher sign accuracy of the precipitation trend (~88.95%), and greater accuracy of the maximum precipitation month (>0.85). For the prediction of different lead times, the RNN model was able to provide a stable trend prediction for summer precipitation, and the time correlation coefficient score was higher than that of the National Climate Center of China. Furthermore, this study proposed a method to measure the sensitivity of the RNN model to different input features, which may provide unprecedented insights into the nonlinear relationship and complicated feedback process among climate systems. The results of the sensitivity distribution are as follows. First, the Niño 4 and Niño 3.4 indices were equally important for the prediction of wet season precipitation. Second, the sensitivity of the snow cover on the Tibetan Plateau was higher than that in the Northern Hemisphere. Third, an opposite sensitivity appeared in two different patterns of the Indian Ocean and sea ice concentrations in the Arctic and the Barents Sea. |
|---|