Temperature Prediction Method Based on XGBoost-LSTM-Transformer Model

Time prediction is challenging due to the influence of multiple complex factors such as geographical location, altitude, seasonal changes, climatic conditions, and human activities. To enhance temperature prediction accuracy, a new deep learning ensemble model (Xgboost-LSTM-Transformer, XLT model) is proposed. This model predicts the average daily temperature for the upcoming week. The XGBoost model selects the meteorological features most relevant to temperature changes. Then, the processed meteorological data are input into the LSTM and Transformer network models for prediction. Finally, the prediction results of the two models are merged through a gating mechanism to yield the final prediction result. The proposed method is validated using a dataset of 11 years of average daily temperatures from 34 provincial capital cities in China. Compared to LSTM, BiLSTM, Transformer, Informer, and LSTM-Informer models, it exhibits a lower mean squared error (MSE) and mean absolute error (MAE), as well as a higher R-squared value. The results demonstrate that the ensemble model offers greater temperature prediction accuracy than the other single models.