论文部分内容阅读
为改善常规BP神经网络的性能,根据Nguyen-Widrow初始化规则对网络层的权值和阈值进行了初始化,利用黄金分割法对隐层节点数所在区间进行了寻优,并采用Levenberg-Marquardt优化算法改进了BP神经网络模型,然后利用经隐层单元优化的LM-BP网络模型对某流域的年径流量进行了预测检验。结果表明:经隐层单元优化的LM-BP网络收敛速度快;2001—2010年年径流量预测结果的相对误差均小于20%,合格率为100%。
In order to improve the performance of the conventional BP neural network, the weights and thresholds of the network layer are initialized according to the Nguyen-Widrow initialization rules. The golden section method is used to search the interval where the hidden nodes are located, and the Levenberg-Marquardt optimization algorithm BP neural network model is improved, and then the annual runoff of a river basin is tested by using LM-BP network model optimized by hidden layer unit. The results show that the LM-BP network optimized by hidden layer unit has a fast convergence rate. The relative errors of annual runoff forecast results from 2001 to 2010 are both less than 20% and the pass rate is 100%.