Load profiling and its application to demand response: A review

Yi Wang, Qixin Chen, Chongqing Kang, Mingming Zhang, Ke Wang, Yun ZhaoTsinghua 2015 · DOI

Abstract

The smart grid has been revolutionizing electrical generation and consumption through a two-way flow of power and information. As an important information source from the demand side, Advanced Metering Infrastructure (AMI) has gained increasing popularity all over the world. By making full use of the data gathered by AMI, stakeholders of the electrical industry can have a better understanding of electrical consumption behavior. This is a significant strategy to improve operation efficiency and enhance power grid reliability. To implement this strategy, researchers have explored many data mining techniques for load profiling. This paper performs a state-of-the-art, comprehensive review of these data mining techniques from the perspectives of different technical approaches including direct clustering, indirect clustering, clustering evaluation criteria, and customer segmentation. On this basis, the prospects for implementing load profiling to demand response applications, price-based and incentive-based, are further summarized. Finally, challenges and opportunities of load profiling techniques in future power industry, especially in a demand response world, are discussed.

Introduction

Traditionally, utilities categorized customers into broad buckets based on their building type. However, two neighbors in identical houses might have completely different energy habits: one might be a "night owl" using heavy appliances at 2 AM, while the other is a "9-to-5er" with a sharp evening peak.

With the global rollout of Advanced Metering Infrastructure (AMI), high-resolution consumption data is now available. The challenge is that this data is "big, noisy, and high-dimensional." This review addresses the gap between raw data and utility strategy: how to systematically identify representative consumption patterns (load profiling) and use them to predict how customers will react to price changes or incentives (demand response).

The Load Profiling Pipeline

The authors categorize the load profiling process into a five-stage pipeline, moving from raw data preparation to practical application in the field.

The five-stage pipeline for load profiling and its application.

The core technical contribution of the review is the distinction between two primary clustering approaches used to analyze these curves:

  • Direct Clustering: Applying algorithms like k-means, Fuzzy c-means, or Self-Organizing Maps (SOM) directly to raw time-series data. While standard, these can struggle with the "curse of dimensionality."
  • Indirect Clustering: A more sophisticated approach where features are extracted before clustering. This includes dimension reduction (PCA, Deep Learning) and time-series analysis (DFT, DWT, SAX, HMM).
Taxonomy of direct and indirect clustering techniques for load profiling.

Datasets and Evaluation

As a review paper, it evaluates findings across various international studies, including:

  • Non-residential customer data from medium-voltage systems (234 customers).
  • Large-scale consumer sets (e.g., 1,205 consumers for subspace clustering studies).
  • Simulated environments for testing Time-of-Use (TOU) pricing (300-customer models).
  • Regional datasets from China Southern Power Grid and utilities in Finland and the USA.

Key Insights and Results

Shape vs. Magnitude

Standard Euclidean distance often fails because it focuses on absolute values. "Weighted" distances or Dynamic Time Warping are superior for capturing the shape of a load curve, which is more critical for demand response than total energy volume.

The 30-Minute Sweet Spot

Research suggests that data sampled at 30-minute intervals is the "goldilocks" zone—sufficient for high-quality clustering without the overhead of 15-minute data, whereas 60-minute data often misses critical peaks.

Performance and Economic Impact

  • Compression: PCA can capture over 95% of the variance in a daily load curve using just three principal components, drastically reducing computational load.
  • Algorithm Robustness: While k-means is common, hierarchical clustering is more robust at handling "isolated uncommon load patterns"—the outliers that often break standard models.
  • Profitability: Simulations showed that clustering techniques with lower Clustering Dispersion Indicator (CDI) scores directly correlate to higher profit margins for retailers when designing Time-of-Use tariffs.

Limitations and Future Work

  • Scalability: Most reviewed studies capped at ~10,000 customers. Real-world "Big Data" scenarios with millions of meters require migration to cloud and distributed computing.
  • The "Energy Internet": Future profiling must integrate "multi-energy" data, including natural gas, heat, and even electric vehicle traffic patterns.
  • Prosumer Complexity: The rise of rooftop solar and home storage means a customer's "net load" is no longer just a reflection of behavior, but also of weather and battery logic. Future models need to decouple these factors to understand the human element of demand.

Summary updated

Abstract

The smart grid has been revolutionizing electrical generation and consumption through a two-way flow of power and information. As an important information source from the demand side, Advanced Metering Infrastructure (AMI) has gained increasing popularity all over the world. By making full use of the data gathered by AMI, stakeholders of the electrical industry can have a better understanding of electrical consumption behavior. This is a significant strategy to improve operation efficiency and enhance power grid reliability. To implement this strategy, researchers have explored many data mining techniques for load profiling. This paper performs a state-of-the-art, comprehensive review of these data mining techniques from the perspectives of different technical approaches including direct clustering, indirect clustering, clustering evaluation criteria, and customer segmentation. On this basis, the prospects for implementing load profiling to demand response applications, price-based and incentive-based, are further summarized. Finally, challenges and opportunities of load profiling techniques in future power industry, especially in a demand response world, are discussed.
BibTeX
@inproceedings{wang2015load,
  author = {Yi Wang and Qixin Chen and Chongqing Kang and Mingming Zhang and Ke Wang and Yun Zhao},
  title = {{Load profiling and its application to demand response: A review}},
  booktitle = {Tsinghua},
  year = {2015},
  doi = {10.1109/tst.2015.7085625},
}
Figure