Research on PDO Index Prediction Based on Multivariate LSTM Neural Network Model

Figure 1: Schematic Diagram of LSTM Neural Network Model Structure

Figure 2: Experimental Flowchart

Figure 3: Correlation Coefficient Between Forecasting Factors and PDO Index from 1921–2020

Figure 4: Time Series of the Pacific Decadal Oscillation (PDO) Index and Forecast Factors

Figure 5: Incremental Time Window Cross-Validation Method

Figure 6: Predicted Values and Observed Values of LSTM Network Model

Figure 7: Prediction Results of Different Models from 2011−2020

Figure 8: Accuracy Evaluation Results of Different Model Predictions

Conclusion

Using sea level pressure, sea level height, heat content data from 1921–2020, and sea ice concentration as forecasting factors for the PDO index, the PDO index forecasting factors were selected through spatial correlation analysis. Furthermore, the time series data of the forecasting factors underwent 5-year low-pass filtering, and the optimal parameter combination was selected through cross-validation, establishing a multivariate LSTM neural network model for PDO index prediction. The prediction results of different models for the PDO index from 2011 to 2021 were analyzed and compared, and the model was then used to predict the PDO index for 2021–2030. The study yielded the following conclusions:

(1) The multivariate LSTM neural network model established based on multiple forecasting factor data can effectively predict the PDO index, with an average correlation coefficient and root mean square error of 0.70 and 0.62, respectively, indicating satisfactory prediction results.

(2) The prediction performance of the multivariate LSTM neural network model is superior to that of the multivariate SVR model, statistical regression model, and univariate LSTM neural network model, with smaller prediction errors and better fitting effects, making it a new method for predicting the PDO index.

(3) The low-frequency PDO index results for the next 10 years were predicted, showing that the PDO will remain in a cold phase for the next 10 years, with positive SST anomalies in the central North Pacific and negative SST anomalies in the tropical eastern Pacific and along the North American coast, with the PDO index showing two fluctuations, reaching a minimum in 2025.

As a signal of interdecadal climate change, the multivariate LSTM artificial neural network model established in this paper achieved phase prediction of the PDO index. Although previous studies were able to effectively forecast the PDO index from 1975 to 2009 (with a correlation coefficient of 0.81), their algorithms could not predict the PDO index on future interdecadal time scales. The advantage of this study lies in employing machine learning methods and utilizing forecasting factors over a larger spatial range, which to some extent overcomes the dependency of prediction results on PDO, achieving predictions for the PDO index over the next 10 years. However, the cross-validation results of this study showed a correlation coefficient between 0.6 and 0.7 with the observed values, lower than the correlation predicted by Huang and Wang. This discrepancy may be due to their prediction target being PDO winter (December, January, February) data, while this study selected data for all months; additionally, the selection of hyperparameters involves some subjectivity, and future research could appropriately increase the number of parameter combinations to reduce this error; finally, ideal prediction results over long time scales require the model to be trained on longer datasets, and the prediction accuracy over longer time scales still needs improvement.