基于多变量LSTM神经网络模型的PDO指数预测研究

利用1921–2020年的海平面气压、海平面高度、热含量数据以及海冰密集度作为太平洋年代际振荡（Pacific Decadal Oscillation, PDO）指数的预报要素，建立了关于PDO指数时间序列预测的多变量长短期记忆（Long Short Term Memory, LSTM）神经网络模型，对比分析了2011–2020年不同时间序列预测模型的PDO指数预测结果，最后利用多变量LSTM神经网络模型实现了2021–2030年的PDO指数预测。结果显示，多变量LSTM神经网络模型的预测值与观测值经过交叉验证后的平均相关系数和均方根误差分别为0.70和0.62；PDO未来10年将一直处于冷位相，PDO神经网络指数出现两次波动，于2025年出现最小值。相比于其他时间序列预测模型，本文采用的多变量LSTM神经网络模型预测结果误差小、拟合效果好，可以作为一种新型的预测PDO指数的手段。

Research on PDO index prediction based on multivariate LSTM neural network model

A multivariate long short term memory (LSTM) neural network model was developed for the Pacific decadal oscillation (PDO) index time series prediction using sea level pressure, sea level height, ocean heat content data and sea ice concentration from 1921 to 2020 as forecast elements of the PDO index. The PDO index prediction results of different time series from 2011 to 2020 were compared and analyzed, and finally the PDO index forecasting from 2021 to 2030 is realized by using the multivariate LSTM neural network model. The results show that the average correlation coefficient and root mean square error of the predicted value and the observed value of the multivariate LSTM model after cross-validation are 0.70 and 0.62, respectively. PDO will remain in the cold phase in the next ten years, and the PDO index will fluctuate twice, there will be a minimum in 2025. Compared with other time series forecasting models, the multivariate LSTM neural network model used in this paper has less error in forecasting results and good fitting effect, which can be used as a new method of predicting PDO index.