Deep learning tool: reconstruction of long missing climate data based on spatio-temporal multilayer perceptron

Abstract

AbstractLong-term monitoring of climate data is significant for grasping the law and development trend of climate change and guaranteeing food security. However, some weather stations lack monitoring data for even decades. In this study, 62 years of historical monitoring data from 105 weather stations in Xinjiang were used for missing sequence prediction, validating proposed data reconstruction tool. First of all, study area was divided into three parts according to the climatic characteristics and geographical locations. A spatio-temporal multilayer perceptron (MLP) was established to reconstruct meteorological data with three time scales (Short term, cycle and long term) and one spatio dimension as inputing (rolling predictions, one step predicts one day), filling in long sequence blank data. By designing an end-to-end model to autonomously detect the locations of missing data and make rolling predictions,we obtained complete meteorological monitoring data of Xinjiang from 1961 to 2022. Seven kinds of parameter reconstructed include maximum temperature (Max_T), minimum temperature (Min_T), mean temperature (Ave _ T), average water vapor pressure (Ave _ WVP), relative humidity (Ave _ RH), average wind speed (10 m Ave _ WS), and sunshine duration (Sun_H). Contrasted the prediction accuracy of the model with general MLP and LSTM, results shows that, in the seven types of parameters, designed spatio-temporal MLP decreases MAE and MSE by 7.61% and 4.80% respectively. The quality of reconstructed data was evaluated by calculating correlation coefficient with the monitored sequences of nearest station,determining the applicable meteorological parameters of the model according to the results. Results show that,proposed model reached satisfied average correlation coefficient for Max_T, Min_T, Ave _ T and Ave _ WVP parameters are 0.969, 0.961, 0.971 and 0.942 respectively. The average correlation coefficient of Sun_H and Ave _ RH are 0.720 and 0.789. Although it is difficult to predict extreme values, it can still capture the period and trend; the reconstruction effect of 10 m Ave _ WS is poor, with the average similarity of 0.488. Proposed method is applicable to reconstruct Max_T, Min_T, Ave _ T and Ave _ WVP, but not recommended to reconstruct Sun_H, Ave _ RH and Ave _ WS.