Non-Intrusive Load Monitoring Under Low-Frequency Smart Meter Constraints: A Synthesized 2025–2026 Multi-Model Benchmark and Edge Computed Sub-Cycle Resolution on ESP32
Abstract
In modern power distribution networks transitioning toward decarbonization, electrification, and distributed renewable generation, Non-Intrusive Load Monitoring (NILM) serves as an essential software-driven alternative to capital-intensive sub-metering infrastructure. However, contemporary Advanced Metering Infrastructure (AMI) deployed by utilities worldwide is fundamentally constrained by cellular backhaul bandwidth and Meter Data Management System (MDMS) storage economics, restricting periodic telemetry to coarse temporal profiles (1–15 minute resolutions). Such low-frequency reporting mathematically eliminates the high-frequency inrush currents, harmonic trajectories, and voltage-current phase displacements necessary to discriminate inductive and reactive consumer appliances. This paper presents a unified cyber-physical architecture bridging macroeconomic smart meter benchmarks with edge-embedded sub-cycle metrology:1. Multi-Model Benchmark (SIRED-2026): Systematic evaluation of 6 algorithms (CO, FHMM, DAE, BiLSTM, Seq2Point, SGN) across 1, 5, and 15-minute sampling intervals on a 20-building synthesized smart meter pilot.2. Edge Hardware Architecture: An ESP32 microcontroller platform operating across an exact 100 ms temporal lock window (5 complete 50 Hz cycles), eliminating a +43.0-count (+3.47 A) DC midpoint drift via integer-cycle zero-sum integration and extracting signed sub-cycle phase vectors (theta = ±0.1°).3. TinyML Embedded Roadmap: Quantized INT8 1D-CNN execution on ESP32 Core 1 requiring <45 kB SRAM, <180 kB Flash, and <28 ms inference latency. Hardware schematics, C++ firmware, and MATLAB/Simulink digital twin models are open-sourced at: https://github.com/manttii/Himshikhar-NILM-ESP32