Accurately predicting rapid weather changes is essential for meteorological services and disaster prevention, an area typically addressed by radar nowcasting. However, achieving accurate and stable predictions over extended forecasting horizons remains a challenging task due to the increasing uncertainty and error accumulation inherent in spatiotemporal sequence modeling. To address this challenge, this paper proposes MRDF-Net, a novel spatiotemporal sequence prediction model that integrates a multidimensional reconstruction convolution module with a dynamic force unit module to enhance forecasting accuracy and stability. The reconstruction convolution module adopts a dual reconstruction strategy across spatial and channel dimensions, which effectively reduces feature redundancy while preserving sensitivity to complex meteorological patterns. The dynamic force unit module, on the other hand, simplifies nonlinear operations in the self-attention mechanism to improve computational efficiency and feature representation. Experimental results demonstrate that MRDF-Net achieves state-of-the-art performance on standard short-term forecasting tasks, as measured by the Critical Success Index (CSI) and Heidke Skill Score (HSS). More notably, the model maintains its superior predictive capability in extended two-hour forecasting scenarios. MRDF-Net effectively alleviates the echo weakening commonly observed in other models by better preserving strong echo regions, resulting in more accurate predictions of detailed spatial structures. These results highlight the strong potential of MRDF-Net for operational meteorological forecasting.