基于支持向量回归和高斯过程回归的水文时间序列特征提取方法

水文时间序列受多种环境因素影响,表现出明显的综合性,传统的利用单一神经网络进行特征提取解释性不足。提出一种基于支持向量回归和高斯过程回归的水文时间序列特征提取方法。首先,罗列水文时间序列候选特征,将特征组合等价于0-1规划,并将各特征组合分别进行支持向量回归与高斯过程回归建模;其次,利用遗传算法演化求解一组最优特征组合,使得支持向量回归和高斯过程回归输出误差同时最小;最后,为了证明所提方法的高效性与准确性,以屯溪流域水文时间序列数据为对象进行验证。实验结果表明,基于支持向量回归和高斯过程回归特征提取方法的水文时间序列预测结果优于传统神经网络特征提取方法。

Feature Extraction Method of Hydrological Time Series Based on Support Vector Regression and Gaussian Process Regression

Hydrologic time series is affected by many environmental factors and shows obvious comprehensiveness. The traditional use of single neural network for feature extraction is insufficient. A feature extraction method of hydrological time series based on support vector regression and gauss process regression is proposed. First, the candidate characteristics of hydrological time series were listed, the feature combinations werre equivalent to 0-1 programming, and each feature combination was modeled by support vector regression and gauss process regression respectively, and a set of optimal feature combinations were solved by genetic algorithm evolution to minimize the output errors of support vector regression and gauss process regression simultaneously. In order to prove the efficiency and accuracy of the proposed method, the hydrologic time series data of Tunxi Basin were taken as the object for verification. The experimental results show that hydrological time series prediction based on support vector regression and gauss process regression feature extraction method is superior to the traditional neural network feature extraction method.