Detection of False Data Injection Attacks (FDIA) Targeting Distribution Transformers in Smart Grids Using Hybrid Artificial Intelligence Methods

The attack vector is defined by the expression in Equation (1).

1 zattack = znormal + a {\rm{zattack = znormal + a}}

Here, a ∈ ℝ¹² represents the distortion vector calculated to keep the residual value below the threshold to avoid being caught by the conventional BDD (Bad Data Detection) filter. The study was simulated through four scenarios. In Scenario A, a gradual deviation ranging from 5% to 15% was applied to the voltage values. In Scenario B, large-scale deviations were created by including a sudden data injection into the current and power measurements. In Scenario C, coordinated and consistent distortions were applied where multiple data points were manipulated to support each other. Finally, under Scenario D, a persistent distortion with low amplitude (2–4%) was applied over a long period.

The selection of LSTM, Isolation Forest, and One-Class SVM as the candidate models in this study was guided by three principal criteria: the sequential and temporal nature of smart grid measurement data, the operational constraint of limited labeled attack samples in realistic deployment scenarios, and the need to benchmark a supervised deep learning approach against well-established unsupervised anomaly detection paradigms. A broader range of algorithms was evaluated at the design stage but excluded for methodologically motivated reasons. Purely supervised classifiers such as Random Forest and XGBoost require sufficiently large and balanced sets of labeled attack instances, which are inherently scarce in real-world FDIA scenarios, introducing bias and limiting generalizability. Reconstruction-based models such as LSTM-Autoencoders and Variational Autoencoders, while effective in related work, introduce additional architectural complexity and threshold calibration demands that fall outside the scope of this comparative study. Graph Neural Network-based architectures require explicit graph topology modeling and are not directly applicable to the single-node experimental framework employed here. Accordingly, LSTM was selected as the representative supervised temporal model, Isolation Forest as a tree-based unsupervised baseline, and One-Class SVM as a kernel-based unsupervised baseline, together spanning three distinct algorithmic families under a unified experimental protocol.

The results demonstrate that the LSTM model significantly outperforms the other evaluated methods in all performance criteria. This is especially supported by the fact that only the LSTM model was able to surpass the 95% recall threshold considered a critical success metric in the literature. This demonstrates the capability of the model to satisfy stringent power system safety requirements. From an operational efficiency perspective, the LSTM model’s achievement of a very low false positive rate (FPR) of 1.0% represents a strategic advantage. Compared to Isolation Forest (3.0%) and One-Class SVM (4.0%), this rate shows that LSTM is much more successful in filtering false alarms. Therefore, this proposed method reduces operator alert fatigue and intervention costs by reducing unnecessary system alerts.

Values in brackets represent 95% confidence intervals in Table 2. Wilson score method for Accuracy, Recall, and Precision; standard error propagation for F1-score; Hanley-McNeil method for AUC-ROC.

Figure 2 demonstrates the effect of stealthy FDIA manipulation on critical electrical state variables under Scenario A. The attack resulted in voltage fluctuations of up to 7.8%; however, these deviations remained below the conventional Bad Data Detection (BDD) threshold for approximately 6 hours. Despite this stealth characteristic, the LSTM-based detection framework successfully identified the anomalous temporal behavior with a detection delay of only 15–30 minutes through sequential pattern analysis.

Figure 2.

Temporal variations in voltage, current, and active power measurements at Bus 69 under normal operating conditions and FDIA attack scenarios, where the shaded red region indicates the active attack interval

LSTM achieved the highest discriminative capability with an AUC value of 0.987, outperforming Isolation Forest (0.965) and One-Class SVM (0.942), particularly in the low-FPR operational region critical for real-time smart grid protection systems. In the FPR < 0.05 region, LSTM maintained a true positive rate of 94%, whereas Isolation Forest and One-Class SVM achieved 81% and 74%, respectively. These findings indicate that LSTM provides a substantial advantage in high-sensitivity smart grid protection scenarios where minimizing false alarms is operationally critical. The ROC curve comparison is given in Figure 3.

Figure 3.

Receiver Operating Characteristic (ROC) curves comparing the FDIA detection performance of LSTM, Isolation Forest, and One-Class SVM models on the IEEE 118-bus dataset

The evaluations concluded that LSTM was the most successful model in the 500-sample test set. The LSTM model successfully detected 97 out of 100 FDIA attacks, missed only 3 attacks, and generated 4 false alarms. The Isolation Forest model successfully detected 93 attacks, failed to detect 7 attacks, and generated 12 false alarms during the evaluation. The One-Class SVM model, which exhibited the weakest performance, had the highest error rate with 90 correct detections compared to 16 false alarms. The findings confirm that LSTM has a statistically significant advantage over the other two models in both attack detection and normal traffic classification capabilities. The confusion matrix for FDIA detection is given in Figure 4. This is confirmed by the statistical analysis presented in Section 3.10.

Figure 4.

Confusion matrices obtained from the evaluation of LSTM, Isolation Forest, and One-Class SVM models on the 500-sample test dataset containing 100 FDIA instances and 400 normal operating samples

Each matrix presents the distribution of True Positives (TP), False Positives (FP), True Negatives (TN), and False Negatives (FN). The results demonstrate that the LSTM model achieved the highest classification reliability with the lowest number of missed attacks and false alarms among the evaluated methods.

Detailed data on the confusion matrix parameters obtained from the three different models are presented in Table 3.

The comparative results demonstrate that LSTM consistently outperformed the unsupervised baseline models across all evaluation metrics. In particular, LSTM was the only method to surpass the critical 95% recall threshold, indicating its suitability for safety-critical smart grid protection applications where missed attack detection must be minimized. A comparative performance analysis of the evaluated FDIA detection models is given in Figure 5.

Figure 5.

Comparative performance analysis of the evaluated FDIA detection models using Accuracy, Recall, Precision, F1-score, and AUC-ROC metrics on the IEEE 118-bus test system

Table 4 details the recall performance of the models in different cyberattack scenarios. The LSTM model did not fall below the 95% success threshold even in the difficult-to-detect ‘Gradual Drift’ and ‘Low Magnitude’ attack scenarios. This demonstrates the operational reliability of the LSTM. While Isolation Forest and One-Class SVM models achieved reasonable success in ‘Sudden Injection’ attacks their performance declined significantly as the attacks became more complex and diverse.

The consistent success of LSTM networks in all types of attacks reveals the indispensability of time-based modeling in anomaly detection in power systems. The systematic manipulation of data by FDIA attacks makes the effects of these interventions most visible in patterns spanning time. Therefore, the use of methods capable of analyzing changes along the time axis becomes a necessity. The LSTM-GIN model developed by Jobe et al. [22], which simultaneously evaluates both temporal and spatial correlations, confirms this result. The findings also show that it is not enough to examine only instantaneous data when ensuring the safety of power systems. In addition, it clearly shows that it is necessary to model the patterns that emerge over time.

LSTM surpassed the 95% recall barrier (a critical reference point for power system safety) with a recall of 97.0%, while Isolation Forest (93%) and One-Class SVM (90%) fell short. In terms of computational cost LSTM’s training time (42.3 seconds) requires approximately 52 times more time compared to Isolation Forest. However, this disadvantage can be largely overcome by completing the training process in an offline environment. Thus, the trained model can be directly deployed for real time inference without the need for retraining.

The vulnerability of GOOSE against attacks that compromise data integrity at the measurement layer stems from the lack of a fundamental authentication mechanism. Full integration of security solutions provided under IEC 62351 such as role-based access control and data encryption will significantly minimize the existing cyberattack surface [11, 12].

The main limitations of this study can be addressed under three fundamental headings. The first is the domain shift problem that arises when transitioning from a simulation environment to real world conditions. The second limitation is that only bus 69 was examined within the scope of the analysis. This leaves it unclear to what extent the findings can be applied to more complex and multi-node networks. The third and final limitation is that the 20% attack intensity in the 2000-sample dataset does not match the low-frequency event distribution in real-world operations. The extent to which findings obtained from a single node can be transferred to heterogeneous network topologies remains an unresolved issue in the literature. To overcome the disadvantages arising from these limitations, future studies aim to utilize privacy-preserving collaborative learning models, incorporate Explainable Artificial Intelligence (XAI) techniques for transparency and cross-validate findings with real network data. It is anticipated that these improvements will significantly strengthen both the technical reliability and generalizability of the current methodology to different scenarios [25].

A particularly significant limitation concerns the exclusive use of synthetically generated data. The absence of publicly available labeled FDIA datasets specific to IEC 61850-based distribution transformer environments reflects a field-wide constraint: real-world attack events in operational power grids are rarely labeled or released publicly due to critical infrastructure confidentiality requirements. Nevertheless, external validity is well-founded. Future work should prioritize cross-validation against publicly accessible benchmark datasets (including the MSU Power System Attack Dataset, the SWaT dataset, and HIL-based ICS datasets such as that used in [15]) even where direct feature-space alignment requires adaptation. Hardware-in-the-Loop (HIL) testbed experiments and, where feasible, collaboration with utility operators to access anonymized operational data, represent the most direct pathway toward establishing the real-world generalizability of the detection framework proposed in this study. Until such validation is conducted, the performance metrics reported herein should be interpreted as indicative of the method’s capability under controlled simulation conditions consistent with the IEEE 118-bus benchmark standard, rather than as guarantees of operational performance in heterogeneous real-world grid environments.

To formally assess whether the observed performance differences between models are statistically significant rather than attributable to chance, McNemar’s test was applied to all pairwise classifier comparisons on the shared 500-sample test set. McNemar’s test is appropriate for this setting as it operates on paired binary predictions from the same samples without assuming normality of error distributions. The test statistic is computed from the off-diagonal elements of the disagreement contingency table: the number of samples correctly classified by model A but not model B (b), and vice versa (c), as χ² = (|b - c| – 1)² / (b + c).

The contingency tables were constructed by comparing the per-sample binary classification outcomes of each model pair across all 500 shared test instances. For each sample, a correct prediction was coded as 1 and an incorrect prediction as 0; the off-diagonal entries b and c in each McNemar table reflect the number of instances where the two models disagreed in their correctness, enabling a paired comparison that is not derivable from individual confusion matrices.

The results indicate that LSTM significantly outperforms both Isolation Forest (χ² = 10.083, p = 0.0015) and One-Class SVM (χ² = 17.053, p < 0.001) at the α = 0.05 significance level. The difference between Isolation Forest and One-Class SVM was also statistically significant (χ² = 5.143, p = 0.023), indicating that Isolation Forest demonstrates a reliably superior detection capability over One-Class SVM on this test set.

To assess result stability with respect to stochastic training variability, the LSTM model was retrained across five independent runs with different random seeds. McNemar test results are given in Table 5. Mean performance metrics and standard deviations are reported in Table 6. The low variance across runs (Recall: 96.8% ± 0.4%, F1-score: 0.963 ± 0.003, AUC-ROC: 0.985 ± 0.002) confirms that the reported results are robust to initialization and are not attributable to a favourable random seed.

To further address the concern regarding generalizability across different data partitions, a 5-fold Walk-Forward Validation was applied as the time-series-appropriate equivalent of standard k-fold cross-validation. In this scheme, each fold uses all temporally preceding samples as training data and the next sequential block as the test set, thereby preserving chronological order and preventing temporal leakage. The 2000-sample dataset was divided into 5 chronological blocks of 400 samples each; for fold k, samples from blocks 1 through k serve as training data and block k+1 as the test set. The cross-validated performance metrics for all three models are presented in Table 7. The low standard deviation across folds for LSTM (Recall: 96.5% ± 0.6%) confirms that the reported results generalize reliably across different temporal segments of the data and are not an artefact of a single favourable train/test partition.