关键词: iTransformer, LSTM, 模型融合, 多特征, 农场气温, 多步预测

Abstract: To address the nonlinear and complex characteristics of farm temperature data, based on meteorological station data from Huaxing Farm in Changji City, Xinjiang Uygur Autonomous Region,seven features including temperature, ground infrared temperature, dew point temperature, relative humidity, vapor pressure, station pressure, and sea-level pressure were selected as model input features through Spearman correlation analysis,and comparative analysis was conducted among the iTransformer-LSTM model, Transformer model, LSTM model, iTransformer model, and Transformer-LSTM model. The results showed that the iTransformer-LSTM model achieved the best performance. Compared with the optimal baseline model iTransformer, this model reduced the root mean square error (RMSE) by 13.72%, mean absolute error (MAE) by 14.12%, and mean absolute percentage error (MAPE) by 13.61%.The iTransformer-LSTM model could effectively extract time-series feature representations, capture long-term dependencies, and characterize global features and contextual information, making it suitable for multi-feature multi-step time-series temperature prediction tasks.

Key words: iTransformer, LSTM, model fusion, multi-feature, farm temperature, multi-step prediction