基于趋势辨识理论的神经网络及其在水文时间序列预报中的应用

训练样本在数量级上的差别和分配的不均匀会导致网络收敛缓慢,且训练结果偏向样本比重较大的那一方。由AR模型在水文时间序列的较好应用可知,水文时间序列中趋势项占有绝对优势。因此以趋势辨识理论对样本进行规范化,使样本规范化到同一数量级,同时时间序列的趋势保持不变。此外输出层不经过非线性处理,以保证网络有更大的预报空间。经黑河流域实测流量资料验证,基于趋势辨识理论的神经网络在水文时间序列预报中训练速度较快,预报效果较好。

ANN Based on Trend Identify and the Application in Hydrological Time Series Forecast

The difference in order of magnitude and the distribution uniformity of the samples trained resuled in slow of ANN convergence,and the training results leaned to the bigger proportion part.Based on the effective application of AR Model in hydrological time series forecast,that trend item could be the biggest proportion in hydrological time series.So trend identify was used to rule the samples.The samples were the same order of magnitude,but their trends are invariable.Moreover the output did not dealt with non-linear function to ensure the greater forecast space of ANN.The method was validated by the data of Heihe basin.For ANN based on trend identify,training velocity is rapid and forecast effect is better in hydrological time series forecast.