Well logging is a significant method of formation description and resource assessment in exploration and development of oil, natural gas, minerals, groundwater, and sub-surface thermal energy, as well as geotechnical engineering and environmental research. However, the shortage problem of well logging data always exists because well logs can only be measured through a drilling process involving costly and time-consuming field trials. To address this issue, bidirectional long short-term memory (BiLSTM), attention mechanism, and convolutional neural network (CNN) were coupled to build hybrid neural networks for predicting missing well logs. The proposed architecture is a structure of two branches. One branch uses CNN to capture the spatial properties of well logs, and the other one conducts the feature selections by utilizing two-layer BiLSTM with attention mechanism. The spatio-temporal correlations from two branches are merged to forecast the target well logs. The performance of the proposed method is evaluated within a highly heterogeneous reservoir at the Gangdong oilfield in China. In our experiments, six models were trained and used for generating synthetic well logs including compensated neutron logs (CNL), acoustic (AC), spontaneous potential (SP), gamma-ray (GR), density (DEN), and formation resistivity (RT). Moreover, traditional machine learning models, CNN, BiLSTM, and other deep learning benchmark models were developed to compare with the presented models. Results show that the proposed method achieves higher prediction accuracy because it takes into account the spatio-temporal information of well logs.