Energy disaggregation techniques for visualization and improvement of energy efficiency in processes and buildings

Abstract

About 45% of the energy demand in developed countries is consumed in households and in public and commercial services. For this reason, the improvement of electrical energy efficiency in facilities has attracted much attention in recent decades. This recent pursuit of energy efficiency has coincided in time with the rise of data-driven models, such as machine learning or data visualization techniques, which has led many researches to incorporate data-driven solutions into energy efficiency problems, with the ultimate goal of increasing the energy awareness of the user.

An example of these data-driven approaches is the Non-Intrusive Load Monitoring (NILM) approach, which provides users with a parts-based representation of the energy demand by estimating each individual consumption, using only the energy demand measurements of the entire facility as input.

Despite the huge amount of solutions proposed to the NILM problem over the last decades, only a few have addressed the NILM problem in large non-residential buildings. Furthermore, NILM models in the literature are ranked by the accuracy of the resulting energy disaggregation, rather than by the quality of the information received by users and the improvement of their energy awareness. Bearing this in mind, the present PhD thesis has two main goals: 1) exploring the NILM problem in large non-residential buildings and how to adapt existing NILM algorithms to this scenario; and 2) improving the user feedback obtained from NILM algorithms by integrating the resulting disaggregation into interactive visualizations, following the visual analytics paradigm.

Visual analytics approaches exploit the synergies between machine learning, data visualization and interaction mechanisms to incorporate the user and its prior knowledge into an infinite loop of analysis, in which his understanding about the problem under study is significantly boosted. This paradigm is adapted to the NILM problem, defining new interaction paths with the energy disaggregation algorithms and with the facility, whereby the user can tweak the analysis according to his prior knowledge and modify his behavior with the facility in order to improve the overall energy efficiency.

In a first approach, the problem of visual analysis in NILM is addressed by arranging the unsupervised decomposition obtained from the Non-negative Matrix Factorization (NMF) algorithm in a powerful OLAP data cube structure, which not only speeds up filtering and selection interactions, but also provides output values suitable for 1D and 2D insightful data visualization techniques. Real data from the Hospital of León (Spain) are used to show that the visualizations of the NMF decomposition and the fluid interaction pathways supported by the data cube allow the user to visually establish correlations between the obtained latent patterns and important subsystems in large buildings, such as thermal comfort systems.

Seeking a more detailed disaggregation, novel deep neural networks (DNN) models are explored and adapted to our visual analytics paradigm and the nature of large buildings. A novel Fully-Convolutional denoising Auto-Encoder (FCN-dAE) model is proposed for NILM in large buildings, which outperforms the state-of-art NILM approaches on data from the hospital, and shows better computational efficiency than the previous models in terms of number trainable weights and inference time.

In order to close the loop of analysis between DNN-based NILM techniques and the user, following our visual analytics formulation, fluid interaction pathways based on conditioning mechanisms are introduced into our FCN-dAE approach. More specifically, our FCN-dAE is conditioned by the general-purpose Feature-wise Linear Modulation conditioning mechanism to add an auxiliary input, by which the user can modulate the behavior of the model according to his intentions. This leads to fluid and continuous transitions in the estimated individual consumptions that not only make the model suitable for visual analytics approaches, but may also help to understand the inner mechanisms of the DNN-based approaches.