基于MODIS遥感产品和神经网络模拟太阳辐射

Simulation of Solar Radiation Based on Neural Network and MODIS Remote Sensing Products

现有的神经网络模拟太阳辐射的模型很少考虑云、气溶胶、水汽对太阳辐射的影响,采用MODIS提供的气溶胶、云、水汽高空大气遥感产品和常规气象数据,输入LM(Levenberg-Marquardt)算法优化后的BP(Back-Propagation)神经网络模型(简称LM-BP)模拟了和田、西宁、固原、延安4个辐射站点的太阳辐射月均值。验证结果表明:神经网络模型中加入气溶胶、云、水汽之后,4个辐射站点的R2均大于0.90,且各项误差指标均小于仅用常规气象站点数据模拟的太阳辐射结果。

Climate change is a major global issue of common concern of the international community, over the past century, the earth experienced a temperature rise, while solar radiation is an indicator of climate change. At the same time, solar radiation data is an important parameter about crop models, hydrological models and climate change models, many Artificial Neural Network ensemble models are developed to estimate solar radiation using routinely measured meteorolological variables, but it do not consider cloud, aerosol, and water vapor influence on solar radiation. In this article, we use cloud, aerosols, atmospheric precipitable water vapor from MODIS atmosphere remote sensing products and conventional meteorological data including air pressure, Air temperature, vapor pressure, relative humidity, and sunshine duration, and we analyze the relationship between solar radiation and meteorological data. In terms of conventional meteorological data, we make the selection of variables, the redundant variables are proposed. Then, BP artificial neural network model optimized by LM （Levenberg-Marquardt） algorithm （referred to as LM-BP） is used to stimulate solar radiation. This LM algorithm has fast local convergence feature about Gauss-Newton method, but also has global search feature about gradient descent method, which allows error along the direction of deterioration to search, and greatly improving the convergence rate and generalization ability of the network. Therefore, this article use LM-BP model to predict monthly mean daily global solar radiation from 2010 to 2013 about Hetian, Xining, Guyuan, Yan＇ an radiating station using only conventional meteorological data（referred to as A） and using MODIS atmo- sphere remote sensing products binding conventional meteorological data（referred to as A ＋ ） respectively. Then, we validate performance of the model with measured data about radiation station. The results show that cloud amount, cloud optical thickness, aerosol optical depth, and atmospheric precipitable water vapor these factors are added to the established model, the degree of matching simulated solar radiation and actual observations is more higher. And correlation determination （R2） for 4 radiation station are 0.90 or higher, while error indicators are small. This article showed that the use of LM-BP neural network model, combining with remote sensing data and conventional meteorological data to simulate solar radiation is a reasonable and effective way to simulate solar radiation.