基于幅频分离的气候时间序列预测试验

从气候波动的瞬时频率与瞬时振幅出发，结合最小二乘支持向量机技术，提出了基于幅频分离技术的气候时间序列预测方法，并对南京地区降水距平进行了30候的预测试验。结果表明，幅频分离预测法能够对所有模态的振幅和高频模态的瞬时频率进行较好的预测，而预测的瞬时频率累积误差会对模态分量的预测距平相关性产生敏感影响，该新方法能够显著提高气候序列高频模态的预测效果。对于气候序列的低频模态分量，集合经验模态分解的边界效应会对瞬时频率的求解产生较大误差，使得序列边界区的幅角计算不准确，导致对低频模态的最终预测效果不理想。对气候序列的高频分量采用幅频分离并进行最小二乘支持向量机预测，而对其低频分量仅采用最小二乘支持向量机进行直接预测，可同时提高高、低频分量的预测效果，并最终提高整个气候序列的预测准确性。该分频预测方法可以使南京降水预测的30候距平相关保持在0.4以上。

A Predicting Test on Climatic Time Series Based on Amplitude-Frequency Separation

From the perspective of the instantaneous frequency and amplitude of climatic wave series and by virtue of the technique of least square support vector machine (LS-SVM), a new prediction method of climatic series is proposed based on the separation of amplitude and frequency. A 30-pentad prediction test on Nanjing precipitation is conducted using this method. The results show that, the new prediction method based on the amplitude-frequency separation presents good prediction accuracies on both the amplitudes of all modes and the frequencies of higher frequency modes. The accumulated errors of predicted instantaneous frequencies have highly sensitive impacts on the anomaly correlations of modes. This method can distinctly improve the prediction of higher frequency modes. For the lower frequency modes, the boundary effect of ensemble empirical mode decomposition (EEMD) causes remarkable errors on the calculation of instantaneous frequency, which subsequently leads to inaccurate argument and eventually results in unsatisfactory prediction on modes of lower frequencies. Thereby, implementing both amplitude-frequency separation and LS-SVM for the prediction of higher frequency modes of climatic series while merely using LS-SVM for the prediction of lower frequency modes can give perfect predictions on components of both higher and lower frequencies, and ultimately improve the prediction of the whole climatic series. The test implementing this frequency-based prediction method on prediction of precipitation in Nanjing shows that the anomaly correlation remains greater than 0.4 in its 30-pentad prediction.