Abstract: In energy conservation research, energy disaggregation becomes an increasingly critical task, which takes a whole home electricity signal and decomposes it into its component appliances. While householder's daily energy usage behavior acts as one powerful cue for breaking down the entire household's energy consumption, existing works rarely modeled it straight forwardly. Instead, they either ignored the influence between users' energy usage behaviors, or modeled the influence between the energy usages of appliances. With ambiguous appliance usage membership of householders, we find it difficult to appropriately model the influence between appliances, since such influence is determined by human behaviors in energy usage. To address this problem, we propose to model the influence between householders' energy usage behaviors directly through a novel probabilistic model, which combines topic models with the Hawkes processes. The proposed model simultaneously disaggregates the whole home electricity signal into each component appliance and infers the appliance usage membership of household members, and enables those two tasks mutually benefit each other. Experimental results on both synthetic data and four real world data sets demonstrate the effectiveness of our model, which outperforms state-of-the-art approaches in not only decomposing the entire consumed energy to each appliance in houses, but also the inference of household structures. We further analyze the inferred appliance-householder assignment and the corresponding influence within the appliance usage of each householder and across different householders, which provides insight into appealing human behavior patterns in appliance usage.
Introduction
Energy disaggregation, also known as Non-Intrusive Load Monitoring (NILM), is the process of decomposing a single aggregate electricity signal from a household into the consumption of its individual component appliances. While this task is vital for providing consumers with actionable energy-saving insights, traditional methods often fall short by focusing exclusively on electrical signatures or isolated appliance state transitions. These approaches typically ignore the primary driver of energy consumption: human behavior. Human daily routines are inherently structured and social; the use of one appliance often triggers or follows the use of another. This paper addresses this limitation by directly modeling the influence between householders' energy usage behaviors.
Methodology: The ParaHawkes-LDA Framework
The researchers propose ParaHawkes-LDA, a novel probabilistic model that treats energy disaggregation and household structure analysis as interdependent tasks. The framework integrates two sophisticated modeling techniques:
- Latent Dirichlet Allocation (LDA): Borrowed from natural language processing, LDA is used here to identify "behavioral topics." In this context, a topic represents a specific household member's usage profile, such as a "housework-heavy" pattern versus an "IT-centric" pattern.
- Parametric Hawkes Processes: To capture the temporal dynamics of appliance usage, the model employs Hawkes processes—a class of self-exciting point processes. This allows the model to quantify how the occurrence of one event (e.g., turning on a stove) increases the "intensity" or likelihood of subsequent related events (e.g., using a dishwasher or a specific light).
By employing a variational Expectation-Maximization (EM) algorithm, the model jointly infers appliance usage and household structure, allowing these two tasks to mutually inform and improve each other.
Experimental Evaluation
The effectiveness of ParaHawkes-LDA was evaluated using both synthetic data and four prominent real-world datasets:
- Smart*: High-resolution data from three homes with over 50 appliances.
- REDD (Reference Energy Disaggregation Dataset): A standard benchmark containing aggregate and circuit-level data for six houses.
- Pecan Street: Data from over 450 homes, including ground truth for the number of residents.
- CER (Irish Smart-Grid): A large-scale dataset of over 4,000 homes used to validate household structure inference.
Key Results and Behavioral Insights
ParaHawkes-LDA consistently outperformed state-of-the-art baselines, including AFAMAP and NIALM, across all real-world datasets. As illustrated in Figure 1, the model achieved significantly lower Mean Absolute Error (MAE) in decomposing total energy consumption and attributing it to specific household areas and appliances.
A key strength of the model is its ability to perform "behavioral fingerprinting." By analyzing the inferred appliance-householder assignments, the model identified distinct user types within a single household. For example, Figure 6 shows two discovered patterns: Pattern 1 is characterized by high probabilities of kitchen appliance usage (refrigerator, microwave, toaster), while Pattern 2 is dominated by electronics (monitor, Wii console).
Furthermore, the model proved to be at least 7% more accurate than two-step baseline methods in estimating the number of residents and their specific usage memberships, demonstrating that understanding household structure is a powerful cue for accurate energy disaggregation.
Conclusion
The ParaHawkes-LDA model demonstrates that incorporating human behavioral dynamics into energy disaggregation significantly reduces signal ambiguity and improves accuracy. By modeling the "who" behind the "what," the framework provides a more holistic understanding of household energy consumption. Future research could focus on reducing the computational complexity of modeling dense appliance interactions and incorporating features that capture seasonal or long-term behavioral shifts.
Summary updated