It is believed that the Tibetan Plateau plays an important role in shaping the global atmospheric circulation and climate change. Due to the lack of observations in its surrounding area, previous studies mostly relied on numerical and empirical models to characterize the atmosphere over the Tibetan Plateau. For example, NRLMSISE-00 is a neutral atmospheric empirical model that has been widely used. In this study, NRLMSISE-00 model temperature data were compared with radiosonde temperature profiles obtained over Golmud, which is located in the northeast part of the Tibetan Plateau. Certain degree of deviations has been observed, especially in the warm season (here we defined from May 16 to October 15), with the temperature difference up to 19 K. A three-layer feed-forward neural network (NN) has been built to improve the NRLMSISE-00 model. Radiosonde temperature profiles from year 2013 to 2014 were used as training database for this NN. The radiosonde temperatures in year 2015 were then used to evaluate this method. Our results showed that the deviations between the NRLMSISE-00 model and the in situ radiosonde data have been significantly reduced. This study demonstrates the feasibility of NN for improving the atmospheric model accuracy using radiosonde data as training samples.

中文翻译：

---

青藏高原格尔木无线电探空仪改进NRLMSISE-00模型的方法

相信青藏高原在塑造全球大气环流和气候变化方面具有重要作用。由于缺乏对周边地区的观测，以往的研究主要依靠数值模型和经验模型来表征青藏高原上空的大气。例如，NRLMSISE-00 是一种被广泛使用的中性大气经验模型。在这项研究中，NRLMSISE-00 模型温度数据与在位于青藏高原东北部的格尔木上获得的无线电探空仪温度剖面进行了比较。已经观察到一定程度的偏差，特别是在暖季（这里我们定义为 5 月 16 日至 10 月 15 日），温差高达 19 K。 建立了三层前馈神经网络改进 NRLMSISE-00 模型。2013 年至 2014 年的无线电探空仪温度剖面用作该神经网络的训练数据库。然后使用 2015 年的无线电探空仪温度来评估该方法。我们的结果表明 NRLMSISE-00 模型与原位无线电探空仪数据之间的偏差已显着减少。本研究证明了 NN 使用无线电探空仪数据作为训练样本提高大气模型精度的可行性。