For the problem of accuracy reduction in hot tandem rolling caused by factors such as oil film thickness and roll eccentricity, a thickness error prediction model for hot tandem rolling strips based on dual-channel feature fusion was proposed, which was composed of two parts, namely, spatiotemporal feature extraction and multi-scale feature extraction. For the spatiotemporal feature extraction part, a Frontal Spatio-Temporal Matrix (FSTM) was constructed based on the spatial correlation of adjacent rack plate thicknesses, and the spatiotemporal correlation features of the FSTM were extracted by the Convolutional Long Short-Term Memory Network(Conv-LSTM). For the multi-scale feature extraction part, the current rack rolling thickness data was decomposed by Discrete Wavelet Transform (DWT) to obtain trend item data and detail item data, and Autoregressive Integrated Moving Average Long Short-Term Memory Network (ARIMA-LSTM) and Bi-directional Long Short-Term Memory Network (BiLSTM) were used to extract features from the trend item data and detail item data, respectively. The above features were fused and input into the fully connected layer for regression prediction, the thickness error prediction value of the hot tandem rolling strip was obtained. Experimental results show that the dual-channel feature fusion model effectively improves the accuracy of thickness error prediction, validating the effectiveness of the model.