Cyber-resilient machine learning framework for accurate individual load forecasting and anomaly detection in smart grids

With the evolution of smart grids, accurate and secure predictions of the electricity load become crucial for efficient energy management and reliability. In this paper, a scalable and cyber-resilient methodology for electricity consumption forecasting on individual smart meter level based on machine learning and anomaly detection schemes is proposed. The proposed technique utilizes K-MEANS Clustering and Neural Networks (KMEANS-NN) to enhance Individual Load Forecasting (ILF) with reduced computational complexity and high prediction accuracy. A Principal Component Analysis based One-Class Support Vector Machine (PCA-OCSVM) model is employed as an Anomaly Detection Scheme (ADS) to identify the false data injection attacks in smart meter telemetry. The system uses five months of real-world data from [Formula: see text] smart meters gathered under the supervision of Electrical Distribution Sector (EDS) of Suez Canal Authority (SCA) in Egypt. KMEANS-NN strategy reduces significantly MAAPE by up to [Formula: see text] and cuts computational time from days to minutes. It improves forecasting accuracy across four proposed models: ARIMA, CTREE, MLP and NNETAR. To assess the cyber-security profile, [Formula: see text] of the dataset is orchestrated with scaling, ramping and random cyber-attack simulation. Proposed ADS achieves [Formula: see text] overall accuracy, [Formula: see text] sensitivity, [Formula: see text] precision, [Formula: see text] specificity and F1-score of[Formula: see text], whereas it's [Formula: see text] accurate on clean data. This integrated model offers accurate, efficient, and secure load forecasting presenting good potential for its deployment in large-scale smart grid environments.