Tsetlin Machine Enables Real-Time NILM on Microcontrollers

Non-Intrusive Load Monitoring (NILM) systems are valuable for disaggregating total household energy consumption into individual appliance usage, requiring only a single aggregate meter. However, traditional NILM approaches often rely on computationally intensive optimization algorithms that process data offline, making them unsuitable for on-device deployment where privacy concerns mandate local data processing on resource-constrained hardware. This research introduces a novel NILM framework built upon the Tsetlin Machine (TM), specifically tailored for real-time applications on microcontrollers (MCUs). By reformulating the NILM problem as a classification task, the Tsetlin Machine can efficiently determine the active status of each appliance. The proposed system demonstrates impressive performance, achieving an average precision of 90% and recall of 96% for two-appliance classification, and 77% precision and 80% recall for four appliances on the REDD dataset. Crucially, the trained model requires only 18 KB of flash memory and exhibits an inference latency of 0.43 ms on an ESP32, proving its viability for privacy-preserving, embedded NILM applications on edge devices.