Massive open online courses (MOOCs) have become one of the most popular ways of learning in recent years due to their flexibility and convenience. However, high dropout rate has become a prominent problem that hinders the further development of MOOCs. Therefore, the prediction of student dropouts is the key to further enhance the MOOCs platform. The traditional dropout prediction models based on machine learning are difficult to guarantee the prediction effect due to the shortcomings such as insufficient mining of feature information and not considering the influence of time series. To address this problem, in this paper, we propose the learning behavior feature fused deep learning network model (LBDL) for MOOC dropout prediction. The core of the model lies in modeling different types of information separately and incorporating them into an overall framework. In the data processing stage, the LBDL model divides the data features into video learning behavior features containing time series information and general information features. For video learning behavior features, the model uses Bi-LSTM and attention mechanisms to mine time series information, and for general information features, it uses embedding layer and fully connected layer for processing. A hidden vector containing both types of feature information can be obtained by two different modeling approaches. Then the original feature information is combined to train the gradient boosting framework LightGBM. Experiments on the MOOCCube video dataset show that the AUC and F1-Score of our model can reach 82.39% and 74.89%, respectively, which are higher than other baseline models. It indicates that the proposed LBDL model has better performance in the dropout rate prediction problem.