基于神经网络的城市夜间环境监控与预测

A Study on Monitoring and Prediction of Urban Night Environment Based on Neural Network

因夜间天空亮度分布具有非线性变化特点,故引入神经网络算法,建立基于时间序列的夜天空亮度预测模型,夜天空亮度预测模型可为城市光污染防治提供评价依据。文章对神经网络的原理进行了论述,建立了基于时间序列预测模型。以测试数据为训练样本集,基于MATLAB（矩阵实验室,Matrix Laboratory的简称）,采用改进的BP算法（误差反向传播算法）对网络进行学习训练,并对存在的误差进行了分析。基于时间序列BP神经网络的夜天空预测模型,当隐含层神经元数目为5,训练函数为L-M优化算法（trainlm）时,最大绝对误差可达到0.003 6 cd/m2,最大相对误差达到2.361 4%。结果表明,模型的运行结果与试验数据比较吻合,输出与目标矢量之间相关性也较好。

The purpose of a night sky brightness prediction model was for evaluation and prevention of urban light pollution.Owing to the nonlinear distribution of night sky brightness,the BP neural network algorithm based on time series theory was introduced.The principle of neural network was discussed and the prediction model based on time series theory was established.The test data as training samples were trained by improved BP algorithm with MATLAB（short for Matrix Laboratory）,and their error were analyzed.When the number of hidden neurons was five and training function was L-M optimising algorithm called trainlm,the maximal absolute error could reach 0.0036 candela per square metre while the maximal relative error being 2.361 4%.It has shown that the modelling results were consistent with the experiment data,and the correlation between the outputs and target vectors was fairly satisfactory.