PSO-BP神经网络大气加权平均温度模型构建

Atmospheric weighted mean temperature model based on PSO-BP neural network

针对现有T_(m)模型建模方法多为基于最小二乘线性回归方法以致于模型精度有待提高的问题,该文以中国西北地区2015—2017年的24个探空站的探空数据作为实验数据,在中国西北地区使用粒子群优化BP神经网络(PSO-BP)回归方法建立大气加权平均温度(T_(m))模型:将地表温度、水气压、纬度、高程和时间变化等影响因素作为模型输入因子,将数值积分法所计算得到的T_(m)作为学习目标,利用神经网络模型进行迭代训练得到中国西北地区的T_(m)。以2018年探空站T_(m)数据为参考值,对PSO-BP模型精度进行验证,并与Bevis模型、GPT3模型和中国西部地区T_(m)模型进行比较。结果表明,PSO-BP模型的年均RMSE和年均bias分别为2.71 K和0.35 K,相比Bevis模型、GPT3模型和中国西部地区T_(m)模型年均RMSE分别降低了1.36 K(33.4%)、1.81 K(39.5%)和1.78 K(39.1%),年均bias分别下降了0.70 K(87.7%)、1.04 K(83.2%)和0.44 K(55.7%)。因此,PSO-BP模型在计算中国西北地区T_(m)方面有明显优势。

In view of the fact that most of the existing T_(m)modeling methods are based on linear regression and the model accuracy needs to be improved,this paper takes the radiosonde data of 24 sounding stations from 2015 to 2017 in northwest China as experimental data,and uses the particle swarm optimization BP neural network(PSO-BP)regression method to establish the atmospheric weighted average temperature(T_(m))model in northwest China:Surface temperature,hydropressure,latitude,elevation and annual date were taken as input factors of the model,T_(m)calculated by numerical integration method was taken as learning target,and T_(m)in northwest China was obtained by iterative training using neural network model.The accuracy of PSO-BP model was verified with T_(m)data of sounding station in 2018 as reference value,and compared with Bevis model,GPT3 model and T_(m)model in western China.The results showed that the average annual RMSE and bias of PSO-BP model were 2.71 K and 0.35 K,respectively,which were 1.36 K(33.4%),1.81 K(39.5%)and 1.78 K(39.1%)lower than those of Bevis model,GPT3 model and T_(m)model in western China,respectively.The average annual bias decreased by 0.70 K(87.7%),1.04 K(83.2%)and 0.44 K(55.7%),respectively.Therefore,PSO-BP model has obvious advantages in calculating T_(m)in northwest China.