基于年际增量法的卷积神经网络模型对中国夏季降水的预测研究

将前冬的500 hPa位势高度、向外长波辐射和海表温度的年际增量作为预测因子，建立基于卷积神经网络(Convolutional Neural Network, CNN)的非线性预测模型，对中国160个测站夏季降水展开预测研究，并与基于线性奇异值分解(Singular Value Decomposition, SVD)的预测模型进行效果对比。结果表明：CNN在1981—2020年的交叉检验中所回报的降水平均PS评分和距平相关系数(ACC)分别为74.33和0.12,比SVD高2.15和0.06,说明CNN比SVD在整体上对夏季降水具有更好的预测能力。其中，CNN对SVD预测较好年份的预测效果提升较为明显，对SVD预测较差的年份则改进不大。CNN对中国降水预测存在一定的系统性偏差，订正后CNN对拉尼娜年的降水预测改进较大。结果表明，基于年际增量法的CNN预测模型展示出较好的潜在应用价值。

Prediction research of summer precipitation in China by convolution neural network model based on the year-to-year increment approach

With the year-to-year increment of 500 hPa geopotential height, outgoing long radiation and sea surface temperature, a nonlinear model was established to predict the summer precipitation of 160 stations in China based on Convolutional Neural Network (CNN), and the effects with the prediction model were compared based on linear Singular Value Decomposition (SVD). Results show that CNN''s average PS score and Anomaly Correlation Coefficient (ACC) in the past 1981-2020''s cross test was 74.33 and 0.12, which was 2.15 and 0.06 higher than SVD. It reflects that CNN has better prediction ability of summer precipitation than SVD on the whole. CNN has significantly improved the prediction effect in the years with SVD''s good prediction, but not in the years with SVD''s poor prediction. CNN has some systematic deviation in China''s precipitation forecast, and after revision the forecast of rainfall in La Niña years has been improved better. Results show that the CNN prediction model based on the year-to-year increment approach reflects a good potential application value.