In industrial production processes, online monitoring of critical variables can be achieved by data-driven soft sensor modeling methods. However, due to high dimensionality and high degree of temporal correlation in modern complex industrial data, traditional soft sensor models face challenges in accurately predicting these variables. To address above issues, this paper proposes a novel soft sensor model, AE-LSTM, which combines Autoencoder (AE) with long short-term memory neural networks (LSTM). On the one hand, AE is used to extract features from the input data to reduce the data dimensionality. On the other hand, LSTM is used to build a soft sensor model for the extracted feature variables to capture the dynamic nature of the data.The proposed method is applied to predict the acetic acid content at the top of a normal boiling tower in the industrial production process of purified terephthalic acid (PTA). The results demonstrate that the prediction accuracy of AE-LSTM surpasses BP, LSTM, and PCA-LSTM by 30%, 18%, and 12%, indicating higher prediction accuracy compared to traditional methods.