Abstract
This paper presents a comprehensive machine learning framework for anomaly detection and predictive modeling in smart grid environments, tailored to support the stability and efficiency of energy systems within smart cities. The proposed framework initially applies the unsupervised learning methods to find anomalous or unstable states of the grid with Isolation Forest and Local Outlier Factor (LOF). The performance of Isolation Forest in terms of anomaly detection with the help of Mean Squared Error (MSE) and Mean Absolute Error (MAE) is higher than LOF. Anomalies are eliminated, and subsequently predictive modeling is done based on three supervised algorithms: logistic regression (LR), support vector machine (SVM), and random forest (RF). It is shown in experimental analysis that SVM and RF perform better than LR with classification accuracies of 98 percent and 95 percent, respectively. The models are also tested on the performance measures, which include precision, recall, F1-score, confusion matrices, Receiver Operating Characteristic (ROC), and Precision-Recall (PR) curves. The framework is able to identify anomalies and forecast grid behavior, which are important to the intelligent, fault-tolerant, and data-driven energy management systems that are required in a smart city infrastructure.