Abstract
Weather temperature prediction is widely used in the fields of environmental management and disaster warning. Temperature prediction based on meteorology is more expensive and has a shorter time horizon. This paper proposes a weather temperature prediction method based on Long Short-Term Memory (LSTM) and Transformer to predict future temperature based on historical temperature information. First of all, Seasonal and Trend decomposition using Loess (STL) is used for feature decomposition to segment weather trends and noise. Then, feature extraction and prediction are carried out based on the LSTM-Transformer hybrid model. When the model is trained on real weather data, the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) reach 0.1902 and 0.2558 respectively, which is significantly better than the independent LSTM and Transformer models. These results show that the integration of LSTM and Transformer architecture can effectively improve the forecasting capability of meteorological time series analysis. Finally, the future research direction is prospected.

This publication has 14 references indexed in Scilit: