Abstract: Pseudorange bias refers to the constant bias in pseudorange measurements produced by receivers of different technical states due to the non-ideal characteristics of satellite navigation signals, and it has become one of the major error sources limiting high-precision applications of GNSS. This study investigates the pseudorange bias of the B3I signal of the BeiDou Navigation Satellite System (BDS). First, the basic principle of the collocated-receiver double-difference method is presented, and the calculation of pseudorange bias is described. To address the difficulty of estimating pseudorange bias in medium-baseline and long-baseline scenarios using conventional methods, a joint error estimation approach based on a hybrid convolutional neural network (CNN) and long short-term memory (LSTM) deep learning model is proposed. This method predicts and estimates the pseudorange double-difference residual components, including pseudorange bias. Real BDS observations from the WUH2_JFNG medium- and long-baseline with an inter-station distance of 12.9 km are used to construct the training, validation, and test datasets. The proposed model is comprehensively compared with a CNN and recurrent neural network (RNN) model in terms of prediction performance. Experimental results demonstrate that the proposed CNN-LSTM model achieves higher accuracy and better stability in the prediction task. Compared with the CNN-RNN model, the root mean square error (RMSE) value and mean absolute error (MAE) value are reduced by 10.83% and 11.10% respectively, while the R² value is improved by 0.015 9. In addition, the proportions of prediction errors within ±0.2 m and ±0.5 m are increased by 2.77 and 3.78 percentage points respectively. The proposed model provides effective technical support for subsequent joint compensation of pseudorange bias and the improvement of positioning accuracy.