PSO-RBF耦合神经网络在水质评价中的应用

Application of hybrid PSO-RBF neural network in water quality evaluation

径向基函数神经网络（RBF）中,连接输出单元的权重是影响径向基函数神经网络学习精度的主要参数,采用粒子群算法（PSO）对模型参数（权值）进行优化,建立了PSO-RBF耦合神经网络模型,避免权重的人为设定,具有客观性。并在Matlab环境下编程计算,构建了用于地下水环境质量评价和海水富营养化评价的径向基函数网络评价模型。将模型应用于黑龙洞泉域5个地下水监测点和某海水富营养化评价。评价结果和其他方法的评价结果一致,与实际情况符合,评级结果是合理的。评价结果表明,RBF网络简便有效,收敛速度快,有较强的分辨能力,优化确定模型参数后,便可对评价样本进行评价,具有较好的实用性和通用性。

This paper is inclined to optimize the model parameters （weights） through the particle swarm optimization （PSO）, by which the weights of the connected output units can be taken as the chief parameters that may affect the learning accuracy of the radial basis function neural network （RBF）. For the research purpose, we have established a PSO - RBF coupled neural network model so as to get rid of the artificial weight sets and enhance the research objectivity. At the same time, we have also constructed the radial basis function network evaluation model for the groundwater environmental quality evaluation while the seawater eutrophication evaluation can be done by means of the programmed calculation in the Matlab environment. The aforementioned radial basis evaluation model can thus be .applied to the evaluation of the groundwater quality of the 5 groundwater monitoring points in the Helongdong spring area and the eutrophication of a certain one seawater point. For the purpose of the research, we have also calculated the weights of the output units of the RBF and substituted the optimal weight method into the RBF network to assess the water quality of the corresponding monitoring points in the said spring area. The results of the evaluation we have done prove to be the same with those gained with the other methods, and well in accord with the actual situations in accordance with the other resources. In addition, the PSO - RBF coupling model can also help to produce more accurate and reasonable evaluation results than the RBF neural network gained in the traditional least-squares methods. For the weight of RBF network model, once optimized by PSO algorithm, can get rid of the disadvantages of the artificial sets and increase the objectivity and versatility. Apart from the above mentioned advantages, the resuits can also be quantitatively represented by the real numbers between 0 and 1, with the water quality being described continuously in specific indexes. Therefore, the suggested PSO- RBF coupled neural network model can be said of popular practicability and great versatility through successful applications in a great number of evaluation cases.