During the well-logging process, factors such as instrument malfunction and borehole collapse often lead to distortion or loss of density curves in certain well intervals, which in turn introduces errors in reservoir evaluation. To enhance the accuracy of reservoir evaluation, the reconstruction of density curves becomes essential. Traditional machine learning methods for curve reconstruction often fail to meet the required precision. To address this limitation, this paper proposes a novel method for density curve reconstruction that integrates Temporal Convolutional Networks (TCN), Bidirectional Gated Recurrent Units (BiGRU), and Multi-Head Attention (MHA) mechanisms. The proposed method utilizes the convolutional characteristics of TCN to capture the long-term dependencies in well-logging data, while the introduction of the MHA mechanism enhances the ability of BiGRU to selectively focus on critical features, thereby achieving precise density curve reconstruction. This method was applied to field data from the study area for reconstruction experiments. Initially, the impact of incorporating lithology indicators on the model's reconstruction capability was evaluated. Subsequently, a comparative analysis was conducted between the proposed network and Gardner's equation, multiple regression, Gated Recurrent Units (GRU), and Bidirectional Gated Recurrent Units (BiGRU). Finally, the generalization ability of the proposed network was validated through core calibration. The results indicate that the proposed density curve reconstruction method not only achieves higher accuracy but also demonstrates excellent generalization capabilities.