On-Device Edge AI for Monthly Energy Forecasting on ESP32-S3 Using 3-Day Windows: Hardware-Constrained Feasibility With Harmonics and Offline Interharmonic Benchmarking

This paper investigates the feasibility of monthly energy consumption forecasting on microcontroller-class edge devices under strict memory and energy constraints. Real-world data were collected over 30 consecutive days from two sites—a residential single-phase installation and an industrial three-phase feeder—using an ESP32-S3 microcontroller interfaced with an ADE7880 energy-measurement IC. Models are trained from contiguous 3-day windows to predict same-month targets, and a non-overlapping triplet split is used to prevent data leakage. Three input regimes are evaluated:No Harmonics, Harmonics, andInterharmonics. Interharmonics are obtained from an external power-quality analyzer and are evaluated exclusively offline as a non-deployable analytical upper bound; on-device deployment is restricted to ADE7880-supported regimes. Dimensionality reduction and resampling are applied offline, and only the quantized forward pass is executed on the ESP32-S3. We benchmark inference-only latency, memory footprint, and device-side energy per inference using an INA226 sensor. Results show that accurate monthly forecasting from short 3-day windows under extreme data scarcity and hardware-constrained conditions is feasible on MCU-class hardware, achieving up toR2≈ 0.93 in the single-phase case with sub-millisecond latency and sub-millijoule energy per inference. Cross-site performance degradation is observed and discussed as domain shift. This work is explicitly scoped as a site-specific feasibility study; multi-season and multi-site validation is left for future work to assess generalization.