Modeling And Prediction Of Air Temperature Time Series At Dome A In Antarctica

The near surface temperature changes in Antarctic Plateau is the baseline information for various climate studies.The remote sensing methods have great potential in obtaining some parameters related to air temperature such as ice surface temperature and other factors accurately because of its easy availability and long term monitoring characteristics.Dome A is 1,250 kilometers away from Zhongshan Station and its altitude is 4,083 meters.It is the highest point of the Antarctic Plateau which is called "the inaccessible pole".Due to its unique location and high altitude,Dome A is also an important observatory for analyzing global climate change.This study extends the time series of air temperature from the perspective of past and future,and provides reference for global climate change.Before the arrival of the Chinese Antarctic Scientific Expedition in 2005,there were no measured records about near surface temperatures in the region.In this study,the ice surface temperature,short wave radiation,long wave radiation,wind speed,cloud type and other meteorological parameters from remote sensing products and reanalysis data were used to establish a reliable machine learning model to fit the in situ data about air temperature,and then the air temperature before 2005 was deduced.In this paper,we analyze the response mechanism of different climate parameters to air temperature over a long time series from 2005 to 2020.there is a strong correlation between surface temperature and air temperature,and their relationship changes with the seasons.However,there is a more complex mechanism on air temperature between other factors and their relationship can not be obtained from simple correlation analysis.Therefore,the relevant climate parameters are put into the machine learning model such as the linear regression,random forest and deep learning to explore the potential relationship between factors so as to realize the temperature fitting.In the tests of the three models,random forest and deep learning performed better.The Mean Error(ME),Mean Absolute Error(MAE),Mean Absolute Percentage Error(MAPE)and Root Mean Square Error(RMSE)of random forest were 0.11?,2.77?,7.3% and 3.70? respectively.The ME of deep learning is 0.25?,MAE is 2.65?,MAPE is 6.99%,RMSE is 3.42?.Finally,the machine learning model with better accuracy is applied to simulate the measured temperature in 2004.On the other hand,air temperature changes at Antarctica in future are also worthy of attention,and it is crucial to find a suitable method to predict the trend of air temperature.In this study,The time series decomposition model and Autoregressive Integrated Moving Average model(ARIMA)are proposed for long term temperature prediction,A recurrent neural network model based on long short term memory(LSTM)is proposed for short term temperature prediction.In the time series decomposition model,the long term trend of air temperature at Dome A was extracted based on singular spectrum analysis.The air temperature trend satisfies the function form of 7.526 sin 0.526 1.15252.128 where time is the independent variable.Secondly,the time series decomposition model proposed in this study captures the variation law of temperature.The fitted monthly average temperature deviation is 0.0004?,and the RMSE is 2.13?.For the ARIMA model,the residuals generated by the fitted temperature series are nearly normal,and the predicted curve are sharper than those of the time series decomposition model,indicating that the model can better predict abnormal situations.In LSTM,the air temperature in2005-2019 is used as the training set,and the air temperature in 2020 is used as the test set of the model.The ME and RMSE in model were-0.33°C and 5.42°C respectively.Finally,the air temperature at 2021 predicted by the time series decomposition model,ARIMA model,and LSTM is compared with the air temperature product in MERRA-2.The temperature predicted by LSTM is very close in value to the reanalysis at the initial stage,but the deviation will increase over time.The validity of the prediction of the time series decomposition model and the ARIMA model is easily affected by extreme weather.In general,random forest and deep learning can be applied to air temperature simulation well and RMSE was controlled at about 3? in the test set.However,due to the use of multi-source data,the loss of a certain data will lead to the failure of simulation,so it is relatively limited to use this method to extend the length of time series.In terms of prediction,LSTM can predict instantaneous temperature in the short term,but for long term temperature changes in the future,LSTM has the disadvantage of error accumulation,while time series decomposition and ARIMA model lack the response to extreme events.