基于遗忘因子的BP神经网络水文实时预报方法  

Real-time hydrological forecasting method of artificial neural network based on forgetting factor

作　　者：袁晶[1] 张小峰[1] 

机构地区：[1]武汉大学水资源与水电工程科学国家重点实验室,湖北武汉430072

出　　处：《水科学进展》2004年第6期787-792,共6页Advances in Water Science

基　　金：国家自然科学基金资助项目(50279035);国家高技术研究发展计划(863)资助项目;世界银行资助ANFAS项目~~

摘　　要：在应用神经网络进行洪水预报时,因洪水系统随着河道上游来流、区间降雨、河床演变等因素的动态变化,其特性并不总是按照基本相同的规律变化,对这类系统的参数辨识,要求算法具有较强的实时跟踪能力,以适应模拟或预测洪水运动变化过程的要求。在BP神经网络模型的基础上,运用最小二乘递推算法,引入时变遗忘因子实时跟踪模型中时变参数的变化,建立了神经网络在非线性系统中动态系统输入、输出数据间的映射关系。计算实例表明:该法对参数的快速时变具有较快的跟踪能力和较高的辨识精度,是一种非常实用的水文实时预报方法。Flood system is usually very complex, and always changes with different inflow from upstream, local rainfall, river-bed deformation and other factors. When the back propagation (BP) neural network is applied in such system for flood forecasting, the algorithm must have ability for real-time tracing of the changes of parameters in the system. In this paper, a variable weighted forgetting factor based on recursive least-squares parameter estimation is introduced into the BP model to simulate such time-variant system. Each weight of the neural network can be real-time modified and the transitional invariable mapping relationship between input and output in the non-liner system of neural network is improved. And two examples are given to demonstrate the effectiveness of the improvement.The calculated result shows that the time-variable weights can be traced with a fast speed and agrees well with the measured data.

关 键 词：神经网络 最小二乘递推算法 时变遗忘因子 时变参数 水文实时预报 

分 类 号：P338.9[天文地球—水文科学]