An Adaptive Reconstruction and Branch-Interaction Network for False Data Injection Attack Detection in Smart Grids

False data injection attacks (FDIAs) pose a serious threat to smart grids because they can alter state estimates without triggering conventional bad-data detection. Although learning-based detectors reduce the need for an exact system model, many of them focus on either temporal changes or relations among measurement variables. To address this problem, this paper presents an Adaptive Reconstruction and Branch-Interaction Network (ARBI-Net). A bidirectional long short-term memory network extracts temporal features from measurement windows, while an inverted Transformer learns relations among electrical variables. Between the two branches, a variable-token reconstruction module converts the temporal representations into tokens linked to individual measurement channels. An adaptive fusion module then combines the temporal and cross-variable features for classification. Experiments are conducted on the IEEE 6-bus, 14-bus, and 118-bus systems. The ablation study shows that the complete model achieves an F1-score of 97.18%, compared with 96.98% for BiLSTM alone and 94.39% for iTransformer alone. On the IEEE 14-bus and 118-bus systems, ARBI-Net achieves F1-scores of 92.60% and 98.50%, exceeding the best learning-based baselines by 1.41 and 10.63 percentage points, respectively. It also provides a better balance between missed detections and false alarms. Further tests show that its detection performance remains more stable under changes in attack strength and measurement noise.