基于正则极限学习机的非平稳径流组合预测

Hybrid forecasting model for non-stationary runoff based on regularized extreme learning machine

针对径流时间序列固有的强非线性和非平稳性特征,提出了一种将集合经验模式分解(EEMD)、样本熵(SE)和正则化极限学习机(RELM)相结合的非平稳日径流预测方法(ES-RELM)。为充分提取径流序列的局部信息以提高预测精度,利用EEMD-SE将径流序列分解为一系列差异度明显的子序列,然后根据各子序列的迥异特征构建了不同的RELM模型对各子序列进行预测,最后将各个子序列的预测结果叠加从而得到最终预测结果。将该模型应用于金沙江下游控制站屏山站的日径流预报中,与九种模型对比结果表明,该方法能有效提高日径流预报精度,是一种高效稳定的径流预报模型,为实现高精度实时径流预报提供了可能。

In this study, a novel hybrid model, named as ES-RELM, based on ensemble empirical mode decomposition（EEMD）, sample entropy（SE）, and regularized extreme learning machine（RELM）, is developed for daily runoff forecasting featured with nonlinearity and non-stationarity. To extract more reliable information from runoff time series, EEMD-SE is used to decompose the runoff series to a set of sub-series with different complexity, then each sub-series is forecasted independently by a different RELM model, and finally all the sub-series forecasts are combined into an overall forecast of the runoff time series. It is applied in a case study to forecast the daily runoff at Pingshan station, a control station of the lower Jinsha River, and compared in detail with nine other models. Results indicate that our ESRELM effectively improves the accuracy of daily runoff forecasting and is an efficient, stable forecasting model, thus laying a basis for high-precision real-time runoff forecasting.