To address the deficiency in loss diagnostic methods for turbines working at off-design angles of attack, a novel loss decomposition method suitable for cascade flow with large separation is proposed. The method proposed has the following advantages over existing methods: (A) It enables refined loss decomposition for cascade flows, capable of identifying the spatial range of specific regions such as shear layers and backflow regions, thereby obtaining the loss characteristics of these regions. (B) The region identification criteria in this method have clear physical meanings, rather than relying on arbitrary area division. (C) The method has good applicability and is suitable for cascade flows under various angles of attack. Validation shows that this method achieves satisfactory results. Based on this method, the loss mechanisms of a low-pressure turbine cascade at a low Reynolds number of 4.3 × 10⁴ and angles of attack of −5°, −20°, and −45° are investigated using Large Eddy Simulations (LESs). Entropy analysis quantitatively demonstrates significant differences in the composition of losses among flow regions, due to their different flow characteristics. From the perspective of flow regions, wake loss dominates total loss, while loss in backflow region is negligible. Furthermore, the variation mechanisms of loss with incidence differ among different flow regions.