Abstract: As a core infrastructure of modern society, the stability and reliability of power systems are directly related to economic development and social operation, making the efficiency of power failure emergency response increasingly critical. Accurately extracting entities and relations from massive power emergency texts and organizing them structurally is a core prerequisite for improving emergency decision-making efficiency and a key direction of current research. Existing methods have obvious limitations in handling long entities, nested entities, and long-distance dependencies in power emergency scenarios, which easily lead to error propagation, insufficient semantic comprehension, and low extraction accuracy. To address these issues, this paper proposes a multidimensional span-based attention mechanism (MSAM) for the joint extraction of entities and relations, innovatively designing a multi-dimensional relation embedding layer and an entity width embedding layer. The end-to-end joint model simultaneously performs named entity recognition and relation extraction, which can efficiently capture long and short-term dependencies in texts, enhance context understanding, and suppress error propagation. Experiments on two public datasets (CoNLL04 and SciERC) show that MSAM achieves significant performance improvements compared with state-of-the-art models, with Micro-F1 values of relation extraction task increased by 1.59% and 1.47% respectively. Tests on a self-constructed power emergency dataset (containing 3218 real emergency reports) further verify its practical applicability, with Micro-F1 values of entity recognition and relation extraction reaching 82.34% and 74.46% respectively, which can provide reliable structured information support for power emergency decision-making.