Accurate Household Occupant Behavior Modeling Based on Data Mining Techniques

Márcia Baptista, Anjie Fang, Helmut Prendinger, Rui Prada, Yohei YamaguchiAAAI 2014 · DOI

Introduction

To develop effective smart grid technologies, engineers require precise simulations of residential energy demand. A critical component of this demand is human behavior—the timing of activities like cooking, bathing, or using electronics. Traditionally, researchers have relied on Markov Chain (MC) models to simulate these behaviors. However, MC models suffer from two primary drawbacks: they often treat household members as independent agents, ignoring social coordination, and they tend to "overfit" the training data, merely replicating recorded sequences rather than generating a realistic variety of daily routines.

This paper presents a novel approach using a Nearest Neighbor (NN) search mechanism combined with association rule mining to create a more flexible and realistic multi-agent simulation of household activities.

Methodology

The proposed system treats each household occupant as an autonomous agent. Instead of relying on transition probabilities, agents determine their next action by searching a database of real Time-Use (TU) data from Osaka, Japan.

Nearest Neighbor Search and Generalization

The model employs a specialized distance function to compare the current state of the household with historical records. This function accounts for both numerical data (e.g., duration of an activity) and categorical data (e.g., the specific room or activity type). To avoid overfitting, the researchers utilized Fixed-Radius Near Neighbors (FRNN). By selecting states within a certain "distance" rather than seeking an exact match, the model can generate plausible new behavior sequences that were not present in the original logs.

Social Coordination via Association Rules

To address the lack of coordination in traditional models, the authors integrated the Apriori algorithm. This data mining technique identifies "leader" occupants whose actions strongly correlate with others (e.g., a parent starting to cook often precedes a child setting the table). When a leader-follower relationship is detected, the model synchronizes the agents' behaviors, leading to more accurate simulations of joint activities.

Key Results

Enhanced Pattern Diversity

A major success of the NN approach is its ability to generate a diverse range of daily routines. While Markov models often struggle to move beyond the specific patterns found in training data, the NN model produced an average of 228 new, unique behavior transition patterns per occupant. This diversity is crucial for testing the robustness of energy systems against varied human behavior.

Figure 4: Comparison of unique behavior patterns generated across different households (A, B, C, D) and occupants. The grey bars represent the significantly higher number of patterns generated by the proposed model.

Temporal Flexibility and Stability

Despite its increased flexibility, the NN model maintains the statistical accuracy of traditional methods. It successfully replicates the probability of specific activities occurring at certain times of the day while allowing for minor, realistic temporal shifts (e.g., starting an activity a few minutes earlier or later than recorded in the source data).

Figure 6: Temporal distribution of patterns over a 24-hour period, comparing the Markov Chain (MC) model (top) with the Nearest Neighbor (NN) model (bottom). Both models show similar peak activity times, but the NN model offers greater density and variety.

Conclusion

The Nearest Neighbor approach, enhanced by association rule mining, offers a significant advancement over traditional Markov Chain models for household behavior simulation. By prioritizing generalization over raw replication and explicitly modeling social coordination, the system provides a more realistic foundation for energy demand forecasting. Future work will focus on refining leadership dynamics and directly translating these behavioral patterns into specific electrical load profiles.

Summary updated

BibTeX
@inproceedings{baptista2014accurate,
  author = {Márcia Baptista and Anjie Fang and Helmut Prendinger and Rui Prada and Yohei Yamaguchi},
  title = {{Accurate Household Occupant Behavior Modeling Based on Data Mining Techniques}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2014},
  doi = {10.1609/aaai.v28i1.8908},
}
Figure