基于贝叶斯正则化神经网络的径流长期预报

针对神经网络用于径流长期预报时,网络结构过于复杂而易出现过拟合的问题,采用主成分分析和贝叶斯正则化神经网络对预报模型进行改进.首先利用主成分分析对输入因子进行降维和优化,然后通过贝叶斯正则化对网络权值的限制来简化网络结构,从而有效地抑制过拟合.对嫩江流域江桥站年平均径流的仿真结果表明,贝叶斯正则化神经网络结合主成分分析的预报方法,可以显著地提高泛化能力和预报精度,而且网络收敛也比较稳定.

Long-term prediction of runoff based on Bayesian regulation neural network

Aiming at the too complex structure of artificial neural network when applied for long-term prediction of runoff, which may cause the overfitting problem, the model is improved by using the techniques of principal component analysis and Bayesian regulation. First, principal component analysis is used for dimensionality reduction and optimization; then, Bayesian regulation is applied to simplify the network by limiting the weights, so the generalization capability of the neural network is enhanced. The simulation results of runoff from Jiangqiao Station at Nenjiang show that the proposed model has a remarkable improvement in the generalization capability and prediction accuracy, and the neural network can converge stably.