LATENT: Low-Latency Anomaly Tracking in National Electricity Time-Series Using Hybrid LSTM-Regression Architectures – A Case Study of Bangladesh’s PGCB Grid

Abstract

The operational instability of national power grids in rapidly developing economies exemplified by Bangladesh’s recurrent loadshedding despite rising generation capacity demands anomaly detection systems that are not only accurate but also deployable under severe computational and data constraints. To address this unmet need, this research propose LATENT: a novel unsupervised framework that uniquely fuses lightweight Long Short-Term Memory (LSTM) forecasting with regression-based residual uncertainty quantification to enable real-time anomaly surveillance using only coarse-grained, hourly telemetry from the Power Grid Company of Bangladesh (PGCB). Unlike existing deep learning approaches that rely on high-frequency sensors or incur prohibitive latency (>500 ms), LATENT operates exclusively on publicly available generation, demand, as well as loadshedding records requiring no labeled anomalies and achieves 98.7% precision and 96.4% recall with inference latency under 120 ms on edge-compatible hardware. LATENT provides proactive early warnings up to 3 hours before major outages, validated against historical grid logs, while maintaining a model footprint below 8 MB for direct deployment on legacy Remote Terminal Units (RTUs). In addition, by reconciling high accuracy with extreme computational frugality, this work establishes the first practical blueprint for scalable, real-time grid resilience in data-scarce, resource-constrained environments offering a transformative pathway for Global South utilities striving to modernize without costly infrastructure overhauls.