Non-Intrusive Load Monitoring for Feeder-Level EV Charging Detection: Sliding Window-based Approaches to Offline and Online Detection

Cameron Martin, Fucai Ke, Hao WangOpenReview 2023 · DOI · OpenReview

TL;DR

This paper tackles the challenge of detecting electric vehicle (EV) charging at the feeder level, a crucial task for grid management that has been under-researched. They introduce a novel sliding window-based feature extraction method combined with classical machine learning models (XGBoost and Random Forest) to achieve high accuracy in both offline and real-time online EV charging detection.

What Problem Does This Paper Address?

As electric vehicles (EVs) become more prevalent, their charging can significantly impact electricity distribution networks, potentially leading to transformer overloading and reduced lifespan. Effective management of EV charging requires understanding when and where it's happening. While Non-Intrusive Load Monitoring (NILM) techniques have been used to detect EV charging in individual households, there's a significant gap in detecting it at the feeder level. A feeder aggregates the electricity consumption of multiple households, making it much harder to isolate individual EV charging events due to the combined "noise" of various appliances. Distribution network operators, who manage these feeders, need this information for grid planning and stability, but existing methods are often household-specific or too complex for real-time application.

Methods

The core of this paper's contribution lies in its sliding window-based feature extraction approach. Instead of relying on complex deep learning models, they opt for a more interpretable and lightweight solution using classical machine learning.

Here's a breakdown:

  • Sliding Windows: They employ three types of sliding windows:
  • Forward-looking: Analyzes data from the current time step into the future. Useful for offline analysis when all data is available.
  • Centered: Assesses both past and future data relative to a target time step.
  • Backward-looking: Analyzes only past data up to the current time step. This is crucial for online (real-time) detection, as it doesn't require future information.
  • Feature Extraction: Within these windows, they extract a set of features including:
  • Basic time-series statistics: mean, variance, standard deviation, minimum, and maximum.
  • Peak Counting: A novel feature that counts the number of power spikes within a window, which can indicate charging events.
  • Machine Learning Models: They use two classical models for classification:
  • XGBoost: The primary model, chosen for its speed and high performance in structured data problems.
  • Random Forest: Used as a baseline for comparison.
  • Offline vs. Online Detection:
    • Offline: Utilizes all three window types to extract features.
  • Online: Primarily uses backward-looking windows of two different lengths (360 and 90 samples) and peak counting to capture charging events in real-time. The use of multiple window lengths is to accommodate the varying durations of EV charging sessions.

Datasets

The authors utilize data from the Pecan Street Dataport [29]. Since labeled feeder-level data was unavailable, they generated synthetic feeder-level load data by aggregating the load data from multiple households (specifically, 85 households were used to generate 22 synthetic feeders). The EV charging labels for the synthetic feeders were derived from the original household meter data, where a feeder was labeled as having an EV charging event if one or more households within it had an active charging session. The data was partitioned into 75% for training, 15% for validation, and 10% for testing.

Key Insights and Ideas

  • Feeder-level detection is a distinct and important challenge: The paper clearly articulates why aggregating data from multiple households makes EV charging detection significantly harder than at the individual meter level.
  • Sliding windows are powerful feature extractors: The choice of window type (forward, centered, backward) and their combination is critical for adapting to different detection scenarios (offline vs. online).
  • Peak counting adds valuable information: This feature, in addition to standard statistics, helps differentiate EV charging from other household loads.
  • Classical ML can be highly effective: The paper demonstrates that sophisticated deep learning isn't always necessary. XGBoost and Random Forest, when fed with well-engineered features, can achieve state-of-the-art results for this specific problem.
  • Online detection is feasible and accurate: The backward-looking window approach enables real-time detection with impressive accuracy, which is vital for operational grid management.

Key Results

The paper reports strong performance for their proposed method:

  • Offline Detection: Achieved an F-Score of 98.88% using XGBoost.
  • Online Detection: Achieved an F-Score of 93.01% using XGBoost.
  • Window Length: Performance generally improves with longer window lengths, with diminishing returns beyond 360 samples (approximately 6 hours). A window length of 360 was chosen as a balance between performance and computational cost.
  • Model Comparison: XGBoost consistently outperformed Random Forest in both offline and online scenarios, particularly in recall and F1-score for online detection.

Limitations and Future Work

  • Synthetic Data: The feeder-level data used was synthetically generated due to the lack of real-world labeled feeder data. While they simulated the Australian scenario with 3 households, they plan to explore different scenarios with a broader range of households in future work.
  • Future Work: The authors intend to delve deeper into NILM for feeder-level detection with a wider variety of household configurations and further examine the effectiveness of their approach.

Summary updated

Abstract

Understanding electric vehicle (EV) charging on the distribution network is key to effective EV charging management and aiding decarbonization across the energy and transport sectors. Advanced metering infrastructure has allowed distribution system operators and utility companies to collect high-resolution load data from their networks. These advancements enable the non-intrusive load monitoring (NILM) technique to detect EV charging using load measurement data. While existing studies primarily focused on NILM for EV charging detection in individual households, there is a research gap on EV charging detection at the feeder level, presenting unique challenges due to the combined load measurement from multiple households. In this paper, we develop a novel and effective approach for EV detection at the feeder level, involving sliding-window feature extraction and classical machine learning techniques, specifically models like XGBoost and Random Forest. Our developed method offers a lightweight and efficient solution, capable of quick training. Moreover, our developed method is versatile, supporting both offline and online EV charging detection. Our experimental results demonstrate high-accuracy EV charging detection at the feeder level, achieving an F-Score of 98.88% in offline detection and 93.01% in online detection.
BibTeX
@misc{martin2023non,
  author = {Cameron Martin and Fucai Ke and Hao Wang},
  title = {{Non-Intrusive Load Monitoring for Feeder-Level EV Charging Detection: Sliding Window-based Approaches to Offline and Online Detection}},
  year = {2023},
  eprint = {2312.01887},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2312.01887},
  howpublished = {OpenReview},
}
Figure