<p>A multi-feature fusion prediction model based on TCN-BiGRU-SE (temporal convolutional network–bidirectional gated recurrent unit–squeeze-and-excitation) is developed to address the model training challenges—due to process parameters’ variations, scarce sample data for the specific products in injection molding, and the low accuracy in multi-quality label prediction. Furthermore, deep transfer learning and multi-task learning are integrated to enhance the injection molding quality prediction. First, the TCN-BiGRU-SE model is employed to extract the shared process knowledge of the injection molding process from large-scale production data. The model-based transfer learning is then introduced through freezing partial layer parameters to reuse the knowledge, and a dimension-adaptive layer is designed to handle the differences in feature dimensionality across the datasets. Finally, a multi-task learning framework is applied to achieve the parallel prediction of the product weight and dimensions. The proposed method enables cross-product knowledge transfer, which significantly improves the prediction accuracy under few-shot data conditions. The experimental results demonstrate that, compared with existing approaches, the proposed method effectively adapts to the variations in the data distribution and achieves superior accuracy and robustness across multiple injection molding datasets.</p>