摘要
利用自动监测数据,采用神经网络对水体中叶绿素a含量进行预测,是水体中叶绿素a含量预测的主要手段之一。但受梯度下降法局部搜索的限制,传统BP神经网络模型预测精度和稳定性均存在问题。鉴于此,引入全局搜索的思维进化算法优化BP神经网络权值、阈值,提高叶绿素a预测效率;并采用偏导方法对预测模型输入因子敏感性进行分析,精简模型输入因子。结果表明:在叶绿素a的BP神经网络预测模型中,引入思维进化算法可显著提高网络训练稳定性和精度,预测精度波动范围从[0.364,0.978]提高至[0.917,0.983],平均预测精度从0.950提高到0.968。利用Dimopoulos敏感性分析将模型输入因子从12因子精简为8因子后,平均预测精度从0.968降至0.962,预测精度波动范围从[0.917,0.983]变为[0.921,0.976],预测模型稳定性更好;在输入因子数目均为8条件下,基于Dimopoulos方法敏感性分析结果筛选出的输入因子组合平均预测精度明显高于基于主成分分析法筛选出的输入因子组合。研究可为基于BP神经网络叶绿素a预测模型输入因子优化提供参考,提高模型预测的稳定性。
Combining automatic monitoring data and neural network method is one of the main methods to predict the chlorophyll-a concentration in waterbody.However,the prediction accuracy and stability of the traditional BP neural network model are questionable due to the limitations of the local search with the gradient descent method.To solve this problem,the global search algorithm EMA was used to optimize BP neural network weights and thresholds to improve the chlorophyll-aprediction efficiency.The partial derivative method was used to analyze the sensitivity of the input factor in prediction model,and then to simplify the number of input factors.The results showed that EMA could significantly improve the stability and accuracy of network training in the BP neural network prediction model for chlorophyll-a concentration.The prediction accuracy ranged from[0.364,0.978]to[0.917,0.983],and the average prediction accuracy improved from 0.950 to 0.968.The predictive model was more stable using Dimopoulos sensitivity analysis to reduce the model input factor from 12 to 8.The average prediction accuracy decreased from 0.968 to 0.962 and the prediction accuracy ranged from [0.917,0.983]to[0.921,0.976].Under the condition that the number of input factors was 8,the average prediction accuracy with the input factors selected by the sensitivity analysis of Dimopoulos method was significantly higher than that with the input factors based on traditional PCA method.The study results can provide reference for input factor optimization based on BP neural network on chlorophyll-aprediction model to improve the stability of model prediction.
作者
蒋定国
全秀峰
李飞
刘伟
JIANG Dingguo;QUAN Xiufeng;LI Fei;LIU Wei(College of Hydraulic and Environmental Engineering,China Three Gorges University ,Yichang 443002,China)
出处
《南水北调与水利科技》
CAS
北大核心
2019年第2期81-88,共8页
South-to-North Water Transfers and Water Science & Technology
基金
国家自然科学基金(51709153)~~
关键词
叶绿素A
BP神经网络
思维进化算法
敏感性分析
优化
chlorophyll-a concentration
BP neural network
EMA
sensitivity analysis
optimization
