Based on the atmospheric temperature data from January 1,2015 to December 21,2020 in Yuncheng City,the Multilayer Perceptron(MLP)model and Transformer model are used to forecast the atmos-pheric temperature in Yuncheng City in this paper.Since weather data is highly time-sequential,we used two-layer and four-layer versions of the Transformer and MLP models to conduct multiple comparative experi-ments over 3-day,5-day,and 7-day periods.The results showed that the 7-day mean squared error(MSE)of the MLP-4 model was 3.2649,and the 3-day MSE of the Transformer-4 model was 5.3767,both demonstrating relatively high prediction accuracy.Moreover,the MLP model's accuracy in predicting tempera-ture was higher than that of the Transformer model.The MLP-2 model's MSEs were 3.2662,3.2996,and 3.3579,while the MLP-4 model's MSEs were 3.2674,3.2996,and 3.2649,showing minor fluctuations but re-maining stable.In contrast,the MSEs of the Transformer-2 model were 5.6225,5.9491,and 5.3892,and the Transformer-4 model's MSEs were 5.3767,6.3787,and 6.1108.Increasing the model layers and parame-ter count led to larger MSEs and overfitting.The overfitting phenomenon seen in the Transformer model was due to its excessive size(close to 4 million parameters),while the study datas only included 1,531 sets.Despite u-sing regularization methods like weight decay and dropout,the model still overfitted the 1,531 data sets,resul-ting in a decline in prediction accuracy.
temperature predictionMLP modelTransformer modelneural networks
NETL
NSTL
万方数据
