Effective and scalable smart residential energy monitoring systems

Abstract

The growing interest for active end-user contribution in reducing the carbon footprint brings several new challenges that traditional solutions are insufficient to handle. In this collective effort, novel solutions, such as the accessibility of detailed information on users’ electrical energy consumption or the implementation of smart energy management systems, must be applied to achieve further energy savings. However, these solutions require effective and scalable methods to monitor the power consumption of appliances and predict household power demand while maintaining systems with limited intrusiveness and low cost. In this thesis, we propose methods for appliance load monitoring (ALM) and household shortterm load forecasting (STLF). These methods are designed to address challenging problem characteristics that arise in real use cases. As an effective solution for demand response programs, semi-intrusive load monitoring (SILM) is used to get granular power measurements at the level of individual appliances in buildings. Hall effect sensors (HES) on each wire attached to a circuit breaker in distribution panels are one means of providing SILM. However, HES are greatly affected by crosstalk noise generated by neighboring wires, up to 35% of interfering signals. As a first contribution, we propose a blind source separation (BSS) approach to remove crosstalk noise and make SILM measurements accurate for home energy management systems. The proposed BSS approach is designed to deal with several challenging problems such as sparse mixing matrices, low mixing coefficients, and imbalanced signal levels. Experiments show that the proposed BSS approach achieves state-of-the-art performance, making the SILM a reliable low-cost alternative to intrusive ALM. On the other hand, non-intrusive load monitoring (NILM) is a technique that uses a single sensor to measure the total power consumption of a building. Using an energy disaggregation method, the consumption of individual appliances can be estimated from the aggregate measurement. Recent energy disaggregation algorithms have significantly improved the performance of NILM systems. However, the generalization capability of these methods to different houses as well as the disaggregation of multi-state appliances are still major challenges. As a second contribution, we propose an energy disaggregation approach based on the variational autoencoders (VAE) framework. The probabilistic encoder makes this approach an efficient model for encoding information relevant to the reconstruction of the target appliance consumption. Experiments show that the proposed model accurately generates more complex load profiles, thus improving the power signal reconstruction of multi-state appliances. Moreover, the regularized latent space of the VAE framework strengthens the generalization capabilities of the model across different houses. Finally, as a third contribution, we propose a two-stage approach for household STLF. However, the electrical load at the household level is highly impacted by user behavior. Hence, load forecasting at such a level of granularity is challenging due to the high uncertainty caused by the difficulty of expecting the user behavior. Therefore, the proposed approach leverages past and future appliance load information estimated from a NILM method. This granular information is then augmented through a household load predictive model. In the light of the experimental results on real-world data, identifying the running appliances as well as predicting their consumption is relevant information, and consequently has a significant impact on the performance of the household STLF. This thesis shows challenges and solutions for monitoring and forecasting electricity consumption at a low granularity level in real use cases in buildings. We propose a SILM solution and a NILM solution to address some of these challenges for the ALM task. Moreover, we propose a STLF approach that leverages appliance load information to predict more accurately the household load demand. We validate all proposed methods on real-world data sets and identify future research directions and remaining problems.

Date	20 Dec 2022
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Ghyslain Gagnon (Supervisor) & Mohamed Cheriet (Co-supervisor)