基于奇异谱和混沌支持向量机模型预测三峡水库月径流

Monthly Runoff Prediction Based on Singular Spectrum Analysis and Chaotic Support Vector Machines

摘要: 径流时间序列在一定程度上可看作是一种被噪声污染的一些准周期信号的组合。为了提高径流预测精度，利用奇异谱分析方法对三峡水库1882~2010的入库径流资料进行预处理，得到重建序列，并分别运用季节性一阶自回归、小波神经网络和混沌支持向量机模型对原始和重建序列进行模拟预测，分析比较三个模型的预测精度。结果表明奇异谱分析法不仅可以浓缩主要信息和减小误差，也能够明显地提高月径流预测精度；基于奇异谱分析方法的混沌支持向量机模型的模拟预测精度最高，检验期模型的确定性系数高达86.9%，年均最大、最小月径流相对误差分别为9%和–7%。

Abstract: The runoff time series is often assumed as a combination of quasi-periodic signals contaminated by noises to some extent. The singular spectrum analysis method is used to preprocess inflow data of the Three Gorges Reservoir from 1882 to 2010 and a new reconstructed sequence is obtained. The seasonal autoregressive model, wavelet neural network model and chaotic support vector machines are applied to simulate and predict the original and reconstructed inflow data series. The results show that the singular spectrum analysis method for data preprocessing not only can concentrate the main information and reduce errors, but also significantly improve the accuracy of the monthly runoff prediction. The chaotic support vector machines coupled with singular spectrum analysis performs best among these models, which determination coefficient (DC) attains to 86.9%, relative errors of annual average maximum monthly inflow (RE_max) and minimum monthly inflow (RE_min) are equal to 9% and –7% during the testing period, respectively.