青海湟水流域基于改进的MLP大气加权平均温度模型研究

Atmospheric weighted average temperature model based on the modified MLP in the Huangshui River basin of Qinghai

为尽可能地提高大气加权平均温度的估计精度,利用神经网络所建立的大气加权平均温度模型,却未顾及大气加权平均温度的长期变化趋势。因此,本文利用全球气候第五代再分析数据集和大气逐层数据,考虑大气加权平均温度的年际变化,基于多层感知器建立了湟水流域大气加权平均温度模型,并使用探空数据与已有的六种模型进行了比较验证分析。结果表明:本模型的年均偏差和均方根误差分别为–0.01 K、2.71 K;均方根误差相比于Bevis式、双因子、多因子、全球气压温度3(global pressure and temperature 3,GPT3)、改进的GPT3模型、谢劭峰等(2022)方法分别减小了32%、23%、15%、14%、7%、5%。验证了引入年际变化因子可进一步提高神经网络模型的精度,建立了目前湟水流域精度相对最优的大气加权平均温度模型。

In order to improve the estimation accuracy of atmospheric weighted average temperature as much as feasible,some studies have established the atmospheric weighted average temperature model using neural networks,which has a certain improvement in accuracy compared to other methods.The atmospheric weighted average temperature’s long-term trend,however,was not take into account.The European Centre for Medium-range Weather Forecasts(ECMWF)from 2010 to 2020 provided the pressure level data and land surface temperature data from 1950 to the present,which are used in this paper to establish a high-precision atmospheric weighted average temperature model for the Huangshui River basin.The atmospheric weighted average temperature model of the Huangshui River basin was established taking annual changes in Tm into accpount,and it was then verified with radiosonde data.The Tm of Huangshui River basin has significant seasonal and annual changes.Following an analysis of the correlation between Tm and Ts,Es,hs,and bs of 136562 samples,it was found that Tm is positively correlated with Ts and es.The correlation coefficient between Ts and Tm is 0.934,and the correlation coefficient between es and Tm is 0.85.hs and bs have a systematic impact on Tm.Therefore,this article establishes a new MLP neural network model for the Huangshui River basin based on Ts,Es,hs,bs,DOY,and CD.The following are the main steps:①Select Ts,es,hs,bs,DOY,and CD that are related to Tm as covariates and input them into the input layer,with Tm as the dependent variable.②Reverse optimize the model parameters using 70%of the data as the training set and 30%as the validation set,then assess the created model’s performance using the measured Tm of the xining sounding station in 2018.③Define the structure of the neural network model and use the trial and error method to obtain a hidden layer with one layer and four nodes.The activation function of the hidden layer is a hyperbolic tangent function,the activation function of the output layer is an identity function y=x,and the loss function is the sum of squares of errors.④When training the model,opt for batch training anduse the scaled conjugate gradient method to optimize the algorithm.The results show that the annual mean bias and root mean square error of the new model are-0.01 K and 2.71 K,respectively.The root mean square error is reduced by 32%,23%,15%,14%,7%,and 5%compared to the Bevis,bi-factor,multi-factor,GPT3,improved GPT3 model,Xie Shaofeng(2022).Verified that adding annual variation factors can further improve the model’s accuracy.It is currently the most accurate atmospheric weighted average temperature model in the Huangshui River basin.