Comparison of non-intrusive load monitoring algorithms based on AutoML

Non-Intrusive Load Monitoring (NILM), an extensively researched domain, relies solely on measured variables such as total load voltage, current, power, and other power-related signals to achieve load monitoring, without the need for additional metering devices. In the context of conventional machine learning methods, which often involve a complex process of model selection and parameter optimization, this study proposes the utilization of Automated Machine Learning (AutoML) techniques for load monitoring. Specifically, three different AutoML frameworks are compared in the context of NILM, utilizing load active power as the input feature. The evaluation metrics employed in this study include Mean Absolute Error (MAE), accuracy, and F
1
score. The experimental results indicate that all three approaches successfully identify five commonly encountered electrical appliances. Overall, the prediction performance of H2O and AutoGluon surpasses that of Auto-Keras.