To ensure the secure and stable operation of smart grids

fast and accurate detection of false data injection attacks (FDIA) is critical. Existing data-driven FDIA detection models primarily rely on fixed discrimination thresholds for anomaly identification. However

this approach has notable limitations: attackers can iteratively probe and analyze model responses

gradually adjusting the magnitude of injected attacks to bypass detection

thereby reducing detection accuracy. To address this issue

this paper proposes a FDIA detection model based on correlation discrepancy. First

a detection framework centered on data correlation discrepancies is designed. Second

a position-aware correction factor is embedded to constrain attention scopes

enabling prior correlation extraction with enhanced positional awareness. Then

leveraging the fine-grained and multi-scale characteristics of measurement data sequences

a dual-stream granularity alignment method is developed to capture sequential correlations. Finally

topological correlations are incorporated to define correlation discrepancies

and an adversarial discrimination criterion is formulated through adversarial training to amplify the distinguishability between normal and attacked measurements

resulting in an effective discrimination criterion. Experimental results demonstrate that the proposed model achieves superior detection accuracy and robustness compared with existing methods and performs well under injection attacks of varying magnitudes.