<p>Blast furnace ironmaking is a core process in steel metallurgy, where accurate prediction of the molten iron temperature is crucial for ensuring product quality, enhancing productivity, and reducing energy consumption. Aiming at the strong nonlinearity and time series dependency of blast furnace molten iron temperature, this paper proposes a soft measurement method by integrating the autoregressive integrated moving average (ARIMA) model with a Transformer-optimized long short-term memory (LSTM) model in deep learning. First, the data preprocessing stage includes identifying missing values and outliers, followed by stationarity testing and differencing. Then, ARIMA parameters are optimized by calculating Bayesian information criterion (BIC) values for parameter combinations and performing Ljung–Box tests on the residuals. Finally, linear and nonlinear components are predicted separately, a transformer-optimized LSTM model is constructed, and an ARIMA–Transformer-LSTM hybrid model is developed for final prediction. The results show that the model achieves excellent evaluation metrics: MAE = 0.0030, RMSE = 0.0036, and <i>R</i>² = 0.9793, outperforming both single models and their optimized models, with optimal performance at the 10-h prediction horizon.</p>