Leveraging Meta-Learning for Enhanced False Data Injection Detection in Smart Grids: The ONF-ML Approach

Abstract

While digitalization promotes grid management efficiency, it also makes power systems more vulnerable to a variety of anomalies, especially false data injection (FDI) anomalies. FDI intrusions pose a serious threat to the security of smart grids. The existing approaches, like machine learning, have certain limitations, which can be addressed by proposing the optimized neuro-fuzzy meta-learning (ONF-ML) model. This model combines several machine learning classifiers serving as a two-step optimization process including hyperparameter optimization for individual classifiers and simulated annealing for tuning neuro-fuzzy parameters. Simulation results conducted on the IEEE 14-bus system using MATPOWER demonstrate the superior performance of ONF-ML in detecting FDI intrusions compared to baseline models, especially for subtle injections. In every bus, FDI intrusion has occurred and average performance metrics are considered. The results illustrate an average detection rate of 91.7% and 81.9% for intrusion samples and 99.9% and 99.8% for normal instances in cases of −3% and +3% occurrences, respectively. While baseline models illustrated critical performance degradation during robust analyses, this technique was remarkably stable, maintaining a detection rate of over 75%, outperforming the second-best technique by up to 45% in worst-case scenarios. By addressing real-world challenges such as sensitivity to noise, inflexibility and incompetence in detecting subtle intruders, the ONF-ML approach enables continuous learning from new data, ensuring adaptability to new threats. Taken together, these features make ONF-ML a practical and scalable solution to overcome the limitations of traditional FDI detection techniques and provide a path to improved smart grid security.