基于小波非线性自回归网络的水文预测模型

Hydrological Series Prediction Model Based on Wavelet Nonlinear Autoregressive Network

不同时间尺度上的水文序列预测在水资源调配和防洪减灾决策中起着重要的作用。提出了一种基于小波分解和非线性自回归神经网络相结合的水文时间序列预测模型(WNARN)。运用Daubechies 5(db5)离散小波将水文序列数据分解为低频和高频子序列,作为非线性自回归神经网络模型(NARN)的输入变量,贝叶斯正则化优化算法用来泛化网络,训练模型对各子序列进行模拟预测,预测值经db5小波重构后得到原序列预测值。利用渭河流域三个水文站40多年的月径流量序列对所提出的WNARN模型进行验证和向前48步的预测能力测试,并与单一NARN模型的验证和预测结果进行对比。结果显示在相同的网络结构下所提出的方法能够显著提高水文序列的预测精度、预测周期及对重大水文事件的预测性,具有较高的泛化能力。

Hydrological series prediction at different time scales plays an important role in water resources allocation,flood control and disaster reduction decisions.We propose a hydrological time series prediction model based on wavelet decomposition and nonlinear autoregressive neural network(WNARN).Daubechies 5(db5)Discrete Wavelet is used to decompose hydrological series data into low-frequency and high-frequency subseries,which are taken as the input variables of nonlinear autoregressive neural network model(NARN),and Bayesian regularization(BR)optimization algorithm is used to generalize the network.The NARN is trained to model and predict each of the decomposed subseries,and the predicted values of the subseries will be reconstructed through db5 wavelet and generate the predicted value of original series.The proposed model was validated and tested by 48-step ahead forecasting using over 40 years historical monthly runoff series from three hydrological stations in the Weihe River Basin,and the validation and testing results are compared with these of the single NARN models.The results show that under the same net structure the proposed model will significantly improve the prediction accuracy,prediction period and the ability of predicting major hydrological events with a high generalization ability.