基于PSO-BP神经网络的西洞庭湖南咀站径流预测

为建立因子少、预报周期短、预报精度高的西洞庭湖控制性水文站南咀站的月平均径流量预报模型，通过对松滋-太平水系控制性水文站安乡、澧水控制性水文站石龟山站月平均水位、流量以及沙湾站月平均水位进行相关性、因子贡献率分析，确定输入因子，借助PSO-BP神经网络对南咀站1956年1月至2005年12月各月平均径流量进行训练，获取网络结构及参数进而预测2006年1月至 2008年12月各月径流量。结果表明：① 石龟山、安乡站水位对南咀站月平均径流量影响最显著；② 汛期、非汛期的划分一定程度上可提高南咀站月平均径流量预报精度；③ 以安乡、石龟山站月平均水位、流量以及沙湾站月平均水位作为输入因子，PSO-BP神经网络预报效果最好，合格率77.8 %，预报等级为乙级；④ 基于相关性、因子贡献率分析，将安乡、石龟山站作为输入因子，预报合格率降为61.1 %，预报等级降为丙级，但仍满足预报要求。

Preliminary study on runoff forecast at Nanzui Station in West Dongting Lake based on PSO-BP neural network

To establish a monthly average runoff forecast model for Nanzui station in West Dongting Lake with less factors, short forecast periods and high forecasting accuracy, we analyzed the relationship between the monthly average water level and runoff at Anxiang Station (Songzi-Taiping water system controlling hydrological station) and Shiguishan Station (Lishui River controlling hydrological station), and the monthly average water level at Shawan Station (Muping controlling hydrological station). Furthermore, the factor contribution rate to monthly average runoff was calculated and the input factor was determined according to the calculated correlation coefficients and factor contribution rates. Based on the above analysis, we used the PSO-BP neural network to train the average monthly runoff from 1956.1 to 2005.12 at Nanzui Station to obtain the network structure and parameters for forecasting monthly runoff from 2006.1 to 2008.12. The results showed that: ① The water level of Shiguishan and Anxiang station had the most significant effect on the monthly average runoff of Nanzui station; ② The division of non-flood and flood seasons could increase the forecast accuracy of the monthly average runoff of Nanzui Station to some extent; ③Importing the variables, including the monthly average water level and runoff at Shiguishan station and Anxiang station, and the monthly average water level at Shawan station, the PSO-BP neural network had the best forecast effect with 77.8% qualified rate and B forecast grade. ④ Importing the monthly average water level of Anxiang and Shiguishan stations and by correlation and factor contribution rate analysis, the forecasting qualified rate was reduced to 61.1%, and the forecasting level was degraded to C level, but the forecasting requirements were still met.