Spatiotemporal Anomaly Detection in Smart Grids Using Latent Representation Learning and Bidirectional LSTM

In a smart grid, anomaly detection is important for assessing the power consumption, identifying abnormalities, and maintaining accurate forecasts. However, traditional approaches often inadequately address complex dependencies, and time-series nonlinearity adversely affects stability, reliability, and accuracy. Therefore, a hybrid deep learning-assisted model that utilizes an autoencoder (AE) and a bidirectional long short-term memory (BiLSTM) model is proposed for anomaly detection. AE efficiently identifies the underlying deviation by learning latent-space-based compressed representations through encoding and decoding. BiLSTM captures the temporal dependencies of data by processing them in forward and backward directions, allowing the utilization of both spatial and temporal features for efficient anomaly detection. The proposed model is evaluated using a power consumption dataset containing anomalies. The experimental results demonstrate the effectiveness of the proposed model in terms of accuracy (99.64 %), row-wise accuracy (99.26 %), and F1-score (93.89 %), highlighting its suitability and scalability for anomaly detection in power consumption data in a smart grid environment.