Multi-Feature Extraction Method of Power Customer’s Portrait based on Knowledge Map and Label Extraction

Wentao Liu, Liang JiInternational Journal of Advanced Computer Science and Applications 2022 · DOI

TL;DR

This paper introduces a robust framework for creating "portraits" (detailed behavioral and demographic profiles) of power grid customers by combining structured knowledge maps with natural language processing (NLP). By transforming massive, fragmented utility data into searchable, labeled visualizations, the researchers provide a path toward highly personalized energy services and more efficient grid management.

What Problem Does This Paper Address?

China’s power grid manages data for nearly 500 million customers, ranging from smart meter readings to service call logs. However, this data often exists in "islands"—disconnected databases that make it difficult to see the "human" behind the meter.

Existing profiling methods often struggle to process this volume of data quickly or fail to extract meaningful emotional and behavioral nuances. Without a clear "customer portrait," utility companies cannot effectively recommend services, evaluate customer satisfaction, or identify sensitive users who might need specialized support. This paper aims to bridge that gap by creating a multi-dimensional model that turns raw data into human-readable labels.

Methods

The authors propose a three-stage construction model:

  1. Knowledge Map Construction: The system collects data from official websites, databases, and literature. To handle the massive scale, they use a timestamp incremental data extraction model with a variable time window. This ensures data consistency even if the extraction process is interrupted by system exceptions.
  2. Multi-Feature Analysis (The NLP Engine):
    • Chinese Word Segmentation: Uses an N-ary grammar model to statistically determine the most likely word boundaries in text.
  3. Vocabulary Weighting: Employs TF-IDF (Term Frequency-Inverse Document Frequency) to identify which words are unique and characteristic of a specific customer versus common filler words.
  4. Emotion Calculation: Uses a Latent Dirichlet Allocation (LDA) model to determine the sentiment behind customer interactions, categorizing features like density, range, and polarity.
  5. Label Extraction and Portrait Generation: Features are distilled into four categories:
  6. Basic Statistics: Age, gender, education, contact info.
  7. Electricity Category: Consumption patterns and payment types.
  8. Contact Preferences: Preferred channels (App, phone, business hall).
  9. Grid Sensitivity: How a customer reacts to power issues (ranked from Class A to D).

The final portrait is represented as a vector in a multi-dimensional space, which can then be visualized using tools like "Tagul" to create label clouds.

Datasets

The researchers utilized an open dataset from the 2020 "Customer Portrait" competition. This dataset includes real-world power usage data from millions of grid customers in a Chinese province during the year 2019. The data was categorized into three main types:

  • Power system work orders (6 fields).
  • Registration forms (3 fields).
  • Smart meter acquisition data (7 fields).

Key Insights and Ideas

  • Breaking the "Data Island": The core innovation isn't just analyzing data, but using a Knowledge Map to describe the relationships between entities (customers, meters, locations) in a structured form before applying NLP.
  • Vectorized Portraits: By defining the portrait as a vector ($d = \langle D_e, S_p, I_n, l_a \rangle$), the system can mathematically calculate the "similarity" between different customers. This allows the utility to say, "Customer A is very similar to Customer B; therefore, they might appreciate the same service plan."
  • Handling Exceptions: The use of a variable time window for data extraction is a practical engineering solution for big data environments where network or system hiccups are inevitable.

Key Results

The method was evaluated on both accuracy and efficiency:

  • Performance Metrics: The system achieved high scores across Accuracy, Precision, Recall, and F1-value (all appearing to be above 80-90% based on the provided radar chart).
  • Efficiency:
    • Emotion calculation is extremely fast, taking less than 1.31 seconds for 10,000 text samples.
  • Word segmentation is the bottleneck, taking about 97.85 seconds for 10,000 samples, though the authors note this only needs to happen once during the initial stage.
  • Visualization: The paper demonstrates a successful portrait for a sample customer ("Li Boming"), showing a clear cloud of labels including "Monthly income of 250,000 yuan," "College degree," and "Sensitive to power consumption period."

Limitations and Future Work

While the results are promising, the authors acknowledge a few areas for growth:

  • Sample Scale: The current study focuses on a single province; future work aims to expand this to a national scale.
  • Unstructured Data: The researchers plan to integrate more advanced topic mining methods like PLSI (Probabilistic Latent Semantic Indexing) and BTM (Biterm Topic Model) to better handle messy, unstructured text data from various sources.
  • Dynamic Portraits: The paper primarily focuses on a static snapshot; future iterations could look at how these portraits evolve over time as customer behavior changes.

Summary updated

Abstract

In order to realize the visualization of power customer characteristics and better provide power services for power customers, a multi-feature extraction method of power customer’s portrait based on knowledge map and label extraction is studied. The power customer’s portrait construction model is designed, which uses the knowledge map construction link to collect the power customer related data from the power system official website and database, and clean and convert the data; In the multi-feature analysis section, natural language processing technology is used to analyze the characteristics of power customers through Chinese word segmentation, vocabulary weight determination and emotion calculation; Based on the feature analysis results, the portrait label is extracted to generate the power customer’s portrait. The power customer’s portrait is used to realize the application of power customer’s feature visualization, power customer recommendation, power customer evaluation and so on. The experimental results show that this method can effectively construct the knowledge map of power customers, accurately extract the characteristics of power customers, generate labels, and realize the visualization of power customer’s portraits.
BibTeX
@article{liu2022multi,
  author = {Wentao Liu and Liang Ji},
  title = {{Multi-Feature Extraction Method of Power Customer’s Portrait based on Knowledge Map and Label Extraction}},
  journal = {International Journal of Advanced Computer Science and Applications},
  year = {2022},
  doi = {10.14569/ijacsa.2022.0131114},
}
Figure