Sistema de análise preditiva em tempo real para Smart Meters usando Machine Learning embarcado (TinyML)

Abstract

The growing global focus on energy efficiency drives the need for innovative solutions that combine sustainability, economy, and technology. In this context, smart meters play an essential role by enabling real-time monitoring of energy consumption, promoting transparency for both providers and consumers. This work presents the development of a realtime predictive analysis system for smart meters, utilizing embedded Machine Learning (TinyML) on the ESP32 microcontroller. The system is designed to operate in environments with limited connectivity, performing local data processing and reducing reliance on cloud infrastructure. The prototype uses real data collected from a refrigerator over 31 days, with continuous 24- hour measurements. Three models were tested and two implemented: XGBoost regressor for consumption forecasting, and One-Class SVM and Autoencoder for anomaly detection. The models were optimized for embedded execution: the One-Class SVM and XGBoost were converted to C++ using the micromlgen library, while the Autoencoder was adapted for TensorFlow Lite (TFLite) with pruning techniques to reduce its size and computational consumption. The results showed that the One-Class SVM achieved higher accuracy in anomaly detection, while the Autoencoder presented lower inference time and memory usage, making it an efficient solution for embedded devices. The XGBoost demonstrated a Mean Absolute Error (MAE) of 0.27 in consumption forecasting, indicating good performance in the accuracy of its predictions. These results highlight the viability of applying TinyML in energy monitoring systems, contributing to the efficient management of distribution networks, fraud detection, and the promotion of more sustainable behaviors among users.