Abstract: Porosity is a critical parameter for characterizing reservoir physical properties. The accuracy of porosity prediction holds significant implications for hydrocarbon exploration and production, as well as the geological modeling and fine-scale evaluation of hydrocarbon reservoirs. Considering the characteristics of tight sandstone reservoirs, such as low porosity, low permeability, and strong heterogeneity, this study proposed a reservoir porosity prediction model that integrates the convolutional neural network (CNN) and bidirectional long short-term memory(BiLSTM). The CNN-BiLSTM model is used to extract features from log data and establish complex nonlinear correlations between these features and porosity parameters. First, using petrophysical models, sensitive elastic parameters, which are significantly correlated with porosity, are selected as input parameters. Second, the spatially dependent features are extracted from the inversion data using the CNN. Third, the temporal dependencies throughout the data sequence are captured by BiLSTM. These processes contribute to effective porosity prediction. The test results of three wells in oilfield A, Kaiping sag, demonstrate that compared to conventional deep learning methods, the CNN-BiLSTM model achieved higher accuracy and lateral resolution in porosity prediction, verifying its effectiveness. Furthermore, the CNN-BiLSTM model was applied for the 3D reservoir porosity prediction by calculating the porosity data volume from elastic parameter volumes obtained through reservoir inversion. Overall, this study provides a novel approach for predicting reservoir porosity in the sparse well area of oilfield A, offshore Kaiping sag.