Abstract:A deep learning approach is proposed to detect data and system anomalies using high-resolution continuous point-on-wave (CPOW) or phasor measurements. Both the anomaly and anomaly-free measurement models are assumed to have unknown temporal dependencies and probability distributions. Historical training samples are assumed for the anomaly-free model, while no training samples are available for the anomaly measurements. By transforming the anomaly-free observations into uniform independent and identically distributed sequences via a generative adversarial network, the proposed approach deploys a uniformity test for anomaly detection at the sensor level. A distributed detection scheme that combines sensor level detections at the control center is also proposed that combines local detections to form more reliable detections. Numerical results demonstrate significant improvement over the state-of-the-art solutions for various bad-data cases using real and synthetic CPOW and PMU data sets.
What problem does this paper attempt to address?
The main problem that this paper attempts to solve is to use high - resolution power system measurement data to detect data and system anomalies. Specifically, the paper focuses on using phasor measurements up to 256 Hz and continuous wave point (CPOW) measurements up to 100 kHz to detect anomalies in the power system. At these sampling rates, power system measurements exhibit strong temporal dependencies. The paper proposes a deep - learning - based method, aiming to solve this problem through the following points:
1. **Anomaly Sequence Detection**: Traditional anomaly detection methods are usually based on the assumption of independence between samples or sample - by - sample detection, and these methods perform poorly when dealing with high - resolution data with strong temporal dependencies. The paper emphasizes the importance of anomaly sequence detection, that is, performing anomaly detection based on measurement sequences rather than individual samples.
2. **Unknown Temporal Dependencies and Probability Distributions**: The paper assumes that both normal and abnormal measurement models have unknown temporal dependencies and probability distributions. For the normal model, it is assumed that there are historical training samples; while for abnormal measurements, no training samples are available. This makes anomaly detection can only be derived from the normal model, regardless of the possible types of anomalies.
3. **Defining "Normal"**: In the power system, although there are clear nominal operating conditions, frequent transient changes occur when generators are dispatched in real - time, so there is no standard data model that can provide a clearly defined statistical test method.
4. **Three Types of Anomalies**:
- Conventional bad data, caused by sensor failures and communication errors, generating outliers.
- Data attacks, where adversaries manipulate sensor data to influence the operator's decision - making process.
- System anomalies, such as faults and operational contingencies, which will be manifested in the data.
5. **Technical Contributions**:
- **Non - Parametric Anomaly Sequence Detection**: The paper proposes a non - parametric anomaly sequence detection method at the sensor level, without making any assumptions about the abnormal data model. This method is applicable to bad data anomalies, data injection attacks and system anomalies.
- **Distributed Detection Framework**: The paper also proposes a distributed detection framework, which combines the detection results at the sensor level to form more reliable system - level anomaly detection. This framework utilizes ICA - GAN (Independent Component Analysis - Generative Adversarial Network) to convert dependent measurement sequences into uniformly independent and identically distributed (i.i.d.) samples, thereby achieving the consistency of local detectors.
6. **Experimental Verification**: The paper verifies the effectiveness of the proposed method through numerical case studies in three scenarios, including natural data anomalies, unobservable data injection attacks and system anomalies. The experiments use a real - world data set from the EPFL network and a PMU measurement data set from a larger synthetic network in North Texas.
In summary, this paper aims to fill the theoretical and practical gaps in power system anomaly detection using high - resolution measurement data, and proposes a new deep - learning - based method that can effectively detect multiple types of anomalies.