Abstract: Accurate prediction of reservoir parameters plays a crucial role in hydrocarbon reservoir evaluation and development decision-making during oil and gas field development. Existing deep learning methods have limitations in spatiotemporal feature collaborative modeling: Conventional convolutional neural networks (CNNs) struggle to capture the multi-scale spatial correlations in well-logging data, recurrent neural networks (RNNs) demonstrate insufficient efficiency in modelling long-sequence temporal features, hindering the model’s ability to fully uncover the complex nonlinear relationships between spatiotemporal features and reservoir parameters. To address these challenges, a reservoir parameter prediction model based on MCNN-ABiGRU-AFFMHA was proposed, featuring a dual-branch parallel architecture. The first branch employed multi-scale CNNs (MCNNs) with kernel sizes ranging from 1 × 1 to 4 × 1 to extract hierarchical spatial features, effectively addressing scale sensitivity caused by reservoir heterogeneity. The second branch integrated an attention mechanism-enhanced bidirectional gated recurrent unit network (ABiGRU) to enhance the extraction of temporal dependencies in well-logging curves. Accordingly, an adaptive feature fusion multi-head attention module (AFFMHA) was designed to dynamically optimize the fusion of spatiotemporal features. Experimental results on data from three wells (A, B, and C) in the Volve Field, Norwegian North Sea, demonstrate that in intra-well prediction tasks, the model achieves outstanding results in Well A (RMSE = 0.006 08, MAE = 0.005 04, R² = 0.910) and Well B (RMSE = 0.005 49, MAE = 0.004 34, R² = 0.905). Ablation studies confirm that the MCNN, ABiGRU, and AFFMHA modules contribute 15.74%, 9.93%, and 26.96% performance improvements, respectively. Furthermore, cross-well testing results (RMSE = 0.022 6, MAE = 0.015 6, R² = 0.931) validates the model’s strong generalization capability, providing an effective spatiotemporal feature collaborative modeling approach for complex reservoir parameter prediction.