改进BP神经网络模型在小康水利综合评价中的应用

Application of improved BP neural network model to comprehensive evaluation of water conservancy in a state of relative prosperity

分析 BP 神经网络应用于小康水利综合评价中存在的几个关键性问题。利用层次分析法（AHP）从100余个水利统计指标中遴选出30个具有一定代表性的指标用于构建小康水利综合评价指标体系并给出相应的分级标准；采用 LM 算法弥补标准 BP 神经网络在实际应用中存在收敛速度慢、易陷入局部极值等不足,建立了神经网络小康水利综合评价模型---LM-BP 模型；利用随机内插方法在小康水利综合评价分级标准阈值间生成训练样本和检验样本；提出网络拟合度的概念；选取网络拟合度、平均相对误差等5个统计指标用于评价模型性能。在模型达到预期的评价精度和泛化能力后,将其用于文山州小康水利综合评价,并构建传统 BP 模型、RBF 模型作为对比模型。结果表明：（a）无论是训练样本还是检验样本,LM-BP 模型的评价精度均高于传统 BP 神经网络模型、RBF 神经网络模型近一个数量级,表明 LM-BP 模型具有较高的评价精度和泛化能力,可用于文山州小康水利综合评价,模型收敛速度快、稳定性能好。（b）2010年文山州及各县级行政区小康水利综合评价为1~2级,处于起步-基本实现阶段；2020年预测评价为3级,全州基本实现小康水利。

This paper focuses on several key issues of a BP neural network when it is applied to the comprehensive evaluation of water conservancy in a state of relative prosperity. Based on the analytic hierarchy process （AHP）, 30 representative indicators were selected out of more than 100 water conservancy indicators, in order to build up a comprehensive evaluation system of water conservancy in a state of relative prosperity and grading standards as well. In practical application, the BP neural network has shortcomings, including the slow convergence and likely occurrence of local extreme values. To overcome these shortcomings, an LM-BP neural network model was established for comprehensive evaluation of water conservancy in a state of relative prosperity. In this case, training and testing samples were generated between standard thresholds using the random interpolation method. A concept of network fitness is proposed as well. The performance of the proposed model was evaluated using the network fitness, the average relative error, and three other statistical indicators. After the evaluation of the model achieved the expected accuracy and generalization ability, it was applied to the comprehensive evaluation of water conservancy in a state of relative prosperity in Wenshan Zhuang and Miao Autonomous Prefecture, and compared with traditional BP and RBF models. The results are as follows： （ a） In both the training samples and testing samples, the LM-BP model had higher evaluation accuracy than traditional BP and RBF models by nearly an order of magnitude, indicating that the LM-BP model has high accuracy and generalization capability and is applicable to the comprehensive evaluation of water conservancy in a state of relative prosperity in Wenshan Zhuang and Miao Autonomous Prefecture. In addition, the LM-BP model has the advantages of fast convergence and a high degree of＆amp;nbsp;stability. （b） In the year 2010, water conservancy in a state of relative prosperity in Wenshan Zhuang and Miao Autonomous Prefecture and county-level administrative districts was at level I to II, the initial stages. It will reach level III in the year 2020 according to the prediction, which means that the water conservancy in the whole prefecture will basically achieve a state of relative prosperity.