Multi-Modal Fusion for Substation Thermal Situational Awareness: Infrared Clustering and Time-Series Forecasting

Power transformers require reliable monitoring to prevent overheating failures. This study develops an integrated approach combining infrared thermal imaging and oil temperature analysis for improved safety assessment. Infrared images are processed using K-means clustering to automatically identify high-, medium-, and low-temperature zones based on brightness. Simultaneously, a Seasonal Autoregressive Integrated Moving Average(SARIMA) model predicts oil temperatures using historical data, accounting for seasonal patterns and short-term variations. A fusion warning system is proposed: the alarm will be triggered only when the two conditions are met: (1) the brightness ratio between high- and medium-temperature clusters exceeds dynamic thresholds adjusted for seasonal conditions; (2) real-time oil temperatures surpass predicted values by optimized margins. This dual-check method significantly reduces false warnings compared to single-factor approaches while effectively detecting true overheating events. The framework adapts to environmental changes and load variations, providing utilities with a cost-effective solution for transformer health monitoring.