Household Energy Forecasting: Optimized LSTM on ESP32-S3

About the Project

As the demand of global electricity increases, the necessity for accurate and real-time energy monitoring has evolved from a luxury to an essential sustainable requirement. However, current residential monitoring infrastructures predominantly rely on cloud-centric architecture or high-performance computing, which introduce significant latency, data privacy vulnerabilities, and prohibitive infrastructure costs. This paper proposes a novel conceptual framework for an autonomous, edge-native energy forecasting system. Diverging from server-dependent models, this framework leverages an optimized Long Short-Term Memory (LSTM) network deployed directly onto a resource-constrained 32-bit microcontroller (ESP32-S3). The methodology enables offline, on-device short-term load forecasting (STLF) for a 24-hour day-ahead horizon using a recursive sliding-window approach. By employing a symmetric affine quantization scheme, the deep learning model is compressed to occupy 64 KB of flash memory and a 45 KB SRAM tensor arena, with an inference latency below 140 ms for a full 24-step recursive forecast. Trained on three months of real household consumption data and evaluated across the full test set under an identical protocol, the model achieves a Mean Absolute Percentage Error (MAPE) of approximately 32.8% (mean over five random seeds, standard deviation 2.5 points) for autonomous day-ahead forecasting and 25.6% for one-step-ahead prediction when live measurements are incorporated. Under an identical protocol the proposed model is statistically comparable to a classical ARIMA baseline (differences within one standard deviation on every metric) and outperforms GRU, seasonal-naive and persistence baselines; all models exhibit a coefficient of determination at or near zero, reflecting the intrinsic difficulty of single-household load. The principal contribution is deployability: INT8 quantization reduces weight storage roughly fourfold with negligible accuracy cost, enabling autonomous on-device forecasting that classical models requiring on-device re-estimation cannot provide. This research bridges the dichotomy between complex deep learning and embedded systems, delivering a scalable, privacy-preserving solution for sustainable energy management.