Modern power systems, serving as the lifeline of societal operations, face challenges to their stability due to equipment damage and service interruptions caused by grid faults. Existing learning-based methods of-ten rely solely on structured data as features, where random fluctuations in the data can lead to overly sensitive fault predictions. To address this issue, this study innovatively integrates multi-dimensional in-formation from both structured and unstructured data, proposing the ALLA prediction framework based on the collaborative reasoning of Attention-LSTM and large language models. By dynamically filtering key features from structured data such as current and voltage through the attention mechanism, and leveraging the temporal reasoning capabilities of large language models on unstructured texts like fault re-ports and maintenance logs, a dual-channel feature enhancement model is constructed. Experimental results demonstrate that this method achieves an average F1 score of 0.9818 across five power grid datasets, representing a 2.4% improvement over current state-of-the-art methods. This framework provides a new paradigm for resolving the sensitivity-accuracy trade-off in fault prediction under complex environments, significantly enhancing the accuracy of converter valve fault predictions.