TL;DR
This paper introduces a novel method for identifying specific types of abnormal bidding behavior by power suppliers in electricity markets, moving beyond simple anomaly detection. By extending the isolated forest algorithm to a "multiple isolated forest" approach, the authors can classify behaviors like extreme price or quantity declarations, offering crucial tools for market supervision and ensuring fair competition.
What Problem Does This Paper Address?
The increasing complexity and scale of electricity markets, particularly with the rise of spot markets, create opportunities for various forms of market manipulation and rule violations. Identifying these abnormal bidding behaviors is critical for maintaining fair competition, preventing market disruptions, and ensuring the stable operation of the power grid. While existing methods can detect anomalies, they often struggle to pinpoint the specific type of abnormal behavior, leaving a gap in providing actionable insights for regulators and market operators. This paper aims to fill that gap by developing a system that not only detects but also classifies abnormal bidding patterns.
Methods
The core of this paper's contribution lies in its novel approach to anomaly detection and classification.
- Multi-dimensional Abnormal Behavior Identification Index: The authors first define six key numerical indicators to characterize bidding behavior. These include weighted average declared electricity price, ratios of highest/lowest prices to the average, total declared electricity quantity, and ratios of highest/lowest declared quantities to the total. These indices are designed to capture different facets of a supplier's bidding strategy.
- Feature Coding: Based on these indices, they propose a method to encode different types of abnormal behaviors (extreme quotation, extreme declaration of electricity quantity, and complex anomalies) into a numerical format.
- Multiple Isolated Forests: The paper extends the standard Isolated Forest algorithm. Instead of a single forest for general anomaly detection, they propose using multiple, specialized isolated forests. Each forest is trained on one of the six identified dimensions of abnormal behavior. This allows for a more granular analysis and, crucially, the ability to identify the type of anomaly. The output of these multiple forests is then combined to classify the specific abnormal behavior.
Datasets
- US PJM Spot Market Data: The primary dataset used for training and evaluation is historical bidding data from the US PJM spot market. Specifically, data from January 2017 to October 2017 was utilized. The paper notes that for data prior to November 2017, PJM only provided daily-level quotation data, not hourly, which influenced the data selection.
Key Insights and Ideas
- Beyond Anomaly Detection: The most significant insight is the shift from simply detecting "anomalies" to classifying types of abnormal behavior. This moves the field from a binary "normal/abnormal" classification to a more nuanced understanding of market participant actions.
- Dimensionality is Key: By breaking down bidding behavior into six distinct, quantifiable dimensions, the authors create a rich feature set that can capture subtle deviations indicative of specific manipulation strategies.
- Ensemble of Detectors: The "multiple isolated forests" approach is a clever adaptation of an existing algorithm. By training specialized detectors for each dimension, they leverage the strengths of isolated forests for outlier detection while building a classification system.
- Actionable Insights for Regulation: The ability to identify specific behaviors like "extreme quotation" or "extreme quantity declaration" provides regulators with concrete evidence and targets for intervention, rather than just a general alert.
Key Results
The paper demonstrates the effectiveness of their proposed method through an example analysis using PJM spot market data.
- Anomaly Detection: The method successfully identified four anomalous bidding instances within the test set of 61 data points.
- Type Identification: For these four anomalies, the multiple isolated forest model correctly identified the specific types of abnormal behavior:
- January 7, 2017: Complex anomaly (quotation and quantity abnormalities)
- January 10, 2017: Extreme quotation behavior
- February 7, 2017: Extreme quantity behavior
- February 9, 2017: Complex anomaly (quotation and quantity abnormalities)
- Verification: The identified behaviors were cross-verified with historical data visualizations (Figures 3, 4, 5, 7) and detailed analysis of the index values, confirming the accuracy of the classification. For instance, on January 7, 2017, the analysis showed abnormally high index1 (weighted average price), low index2 (highest price ratio), and high index3 (lowest price ratio), consistent with extreme high-price quotation.
Limitations and Future Work
- Data Availability: The paper notes that historical data prior to November 2017 was at a daily level, which might limit the granularity of analysis for older datasets.
- Generalizability: While tested on PJM data, further validation on diverse electricity markets would strengthen the generalizability of the method.
- Future Work: The authors suggest that their method provides a foundation for improving operational efficiency and reducing deployment costs for abnormal behavior identification models. They also imply that the framework can be extended to other market members and more complex anomaly types.
Summary updated