چكيده لاتين
In power distribution networks, forecasting load consumption and detecting feeder anomalies are critical and challenging tasks for enhancing reliability and improving system performance. With the increasing complexity of smart grids and the growing presence of renewable energy sources, traditional fault detection methods have become insufficient, highlighting the need for advanced artificial intelligence tools, particularly machine learning, in power systems. In this study, a machine learning-based system was designed and implemented for detecting distribution feeder outages in smart grids. The primary objective was to develop a model capable of processing time-series data from multiple feeders simultaneously and identifying anomalies caused by outages or abnormal load variations. Unlike conventional approaches, which are often designed for a specific feeder and require similar data and precise labeling, the proposed model leverages a combination of Transformer architecture and the Isolation Forest algorithm, providing high generalizability and reliable performance across different feeders. In this approach, the data are first divided into training and testing sets, and the neural network model is trained with automatic weight and parameter learning. Subsequently, the Isolation Forest algorithm is applied to detect anomalies and abnormal points in the test data, without requiring highly similar samples in the database. Simulation results on ten different feeders demonstrated the proposed model’s strong capability in identifying and predicting load variation trends, achieving an average R² of approximately 97%, indicating its effectiveness in learning complex network patterns and capturing sudden fluctuations in time series. Despite the presence of noisy data and long sequences of missing values, the model was designed to ignore these deficiencies and focus only on continuous data segments, preventing false anomaly detection and maintaining high accuracy. Comparison with a reference XGBoost-based method showed that the proposed model outperformed in metrics such as R² and exhibited significant generalizability even across feeders with differing test data, whereas the reference method heavily relies on similar data and feeder-specific parameters, limiting its applicability in real-world networks such as the Iranian power grid. Beyond load prediction, the proposed model accurately identified true anomalies with fewer false positives compared to the reference approach, making it suitable not only for behavioral analysis and load trend forecasting but also for proactive measures to enhance network reliability and reduce sudden feeder outages. Overall, this research demonstrates that combining advanced deep learning models with unsupervised anomaly detection algorithms offers a scalable and precise solution for intelligent monitoring and management of power distribution networks, overcoming the limitations of traditional methods based on limited and labeled datasets.
