(The report is available at the PSRC website https://www.pes-psrc.org/kb/report/117.pdf)
By PSRC Working Group C43 Report
PSRC WG C4:
Chair: Yi Hu
Vice Chair: Adi Mulawarman
Secretary: Zheyuan Cheng
Members and Contributors:
Abder Elandaloussi, Alex Apostolov, Ali Bidram, Athula Rajapakse, Carolina Arbona, Dan Sabin, Jayaprakash Ponraj, Jean Raymond, Jörg Blumschein, Juan Piñeros, Matthew Reno, Nirmal Nair, Ratan Das, Robert Fowler, Sebastien Billaut, Sukumar Brahma, Sukumar Kamalasadan, Vahid Madani, Yu Liu, Yujie Yin
The ongoing digital transformation of the electric power industry is resulting in the availability of huge amounts of data that provides an opportunity to improve the efficiency, reliability and security of power system operations. This is becoming especially challenging as the result of the large penetration of inverter based distributed energy resources. The rapid evolution of communications, computers and artificial intelligence technology can help solve many of the challenges that we are facing. That is why the IEEE Power Systems Relaying and Control (PSRC) Committee established Working Group C43 with the task to produce a report on the Practical Applications of Artificial Intelligence / Machine Learning in Power System Protection and Control.
In practical terms, intelligence can be defined in many ways: advanced comprehension, a capacity to further the existing reasoning, demonstrative knowledge, and learned decision-making. Intelligence can be demonstrated using machines, similar to the natural intelligence shown by humans. Adaptive development is required at various stages, such as cognition, manipulation, rationalization, communication, and reaction to any common transaction. Here, the continuous learning experience facilitates the challenge of automated enhancement in overall system performance over time.
Artificial intelligence (AI) is the simulation of human intelligence in machines that think and act like humans. An effective AI application requires many skill sets such as cognition, manipulation, rationalization, communication, and reaction to be incorporated into the scheme.
AI utilizes Machine Learning (ML) and other associated techniques such as Heuristics to resolve real challenges. Various computational tools used for implementing these skill sets include search and optimization, artificial neural network, fuzzy logic, probabilistic methods for uncertain reasoning, reinforcement learning, and other supervised and unsupervised learning methods. ML is a subfield of artificial intelligence that enables the gathering and analyzing volumes of data to extract representative features based on appropriate training (learning) and develop an equation or algorithm for deriving useful information or action. As an example, text recognition within images is now a commonly used tool (OCR, optical character recognition) that is powered by ML. OCR models can search for the pertinent information using representative features to identify text predictively rather than programmatically.
ML or its earlier vintages of ‘Artificial Neural Network’ has been in practice within the power system industry, especially in operations, workforce management, and planning over a considerable period of time. Some examples are forecasting, optimization, scheduling, and unit commitment. Due to the recent digital transformation progression in the areas such as Cloud Computing (CC), Internet of Things (IoT), and the use of ML to expand the earlier techniques, an upsurge of AI or ML (sometimes used interchangeably, also in this report) is conveying additional significance in many power system applications.
The Working Group C43 report aims to introduce significant AI/ML technical concepts to a Protection and Control (P&C) knowledge base and audience. Providing the foundation for a methodical approach to incorporate this new technology within our industry while displaying representative examples of possible applications. It introduces the reader to current developments and futuristic ideas in AI/ML, with the goal to spark creativity in advancing P&C applications within the Power System Industry. For example:
- AI/ML can help detect and locate high impedance faults faster
- AI/ML can assist in predicting and isolating potential component failures by utilizing partial discharge sensor data
- AI/ML can be used to evaluate challenges to the dynamic stabilities of the power network and support decision-making by using Phasor Measurement and Control Unit data and architecture, thus mitigating the stability control challenges due to increasing low inertia resources
The report covers the practical applications of AI/ML in the protection and control of power systems. The process of applying AI/ML for solving protection or control challenges in the power system involves the design, development, validation, integration, field testing, and deployment of AI/ML models. All these applications focus on reliability, availability, dependability, security, speed, and accuracy to support the operation of a dynamic power system. The report primarily provides insights from data collection to detailed analytics and informed decision-making to support asset and network protection and anticipate potential disturbances to lessen the impacts. The protection and control applications in this report are limited to the protection of the electric network or the power assets.
The focus is AI/ML applications in P&C with updates on:
- Basics of the AI/ML technology
- Practical applications in use and Emerging application areas
- Challenges of applying the technology in power system protection and control
- Considerations in developing, validating, field testing, and implementing practical applications
- Risks, challenges and acceptance criteria for the use of AI/ML for protection
The report is not intended to be used as a technical guide but as a resource that provides information on recent advances in AI/ML applications in the P&C area of power system.
AI/ML for Protection & Control – Practical Considerations
Chapter 4 of the Report is an overview of the practical considerations for the application of AI/ML for Protection and Control.
Drivers: The existing protection and control applications have been working for many years assuming the conventional grid architecture, and they have performed exceptionally well thus far. However, modern power systems dynamics have changed, leading to protection and control facing new challenges to maintain/improve the network and asset performance metrics. Many engineers and researchers are exploring new approaches such as biology-inspired AI/ML-based solutions to the problems that cannot be properly resolved by conventional physics-based approaches. Increasing availability of large amounts of high-fidelity, rapidly sampled data, both in time domain and frequency domain is enabling the development and deployment of AI/ML-based solutions to challenging protection and control problems.
Some drivers leading the focus towards the Artificial Intelligence are:
- High penetration of Distributed Energy Resources
- Large scale inverter-based resources
- Growing need for advanced distribution protection and control applications including fault detection and location
- Wide area instabilities
- Power system reconfiguration
- Increased use of HVDC converters, FACTS devices and compensators
- Natural disasters
- Need for improving the efficiency in protection systems design and engineering
- Mitigating human errors
Limitations of existing P&C technology: Faults can occur on any electrical equipment/component due to various causes such as extreme weather, aging, impact due to repeated stress, adverse environmental conditions resulting in degradation of the insulation, internal/external damages during construction, etc. They result in outages of a specific asset or the entire network depending on the design of power system, severity, coverage of protection zone (e.g., line, transformer, reactors, buses, generators, loads).
Conventional power system protection schemes have been successfully applied to provide adequate protection functions under most power system conditions including normal operating conditions and various fault scenarios. However, there are conditions where conventional protection schemes have shown limitations such that existing schemes either could not provide adequate protections at all or the desired performance and/or dependability-security balance of the protection schemes could not be achieved. Conventional protection schemes work the best when the power system conditions that require protection actions are very different from all other conditions for which when no such actions are needed. Some protection schemes (e.g., out-of-step protection schemes, system integrity protection schemes) work better when the power system where the schemes are applied is not dynamic in terms of topology but may have difficulties to balance the dependability and security requirements if the topology of the power system is dynamic.
On the other hand, designing some protection schemes for dynamic or complex power systems is a combination of science and art. This means that there may not be any “standard” designs for such protection challenges. Solutions by different engineers could be different from each other, which can result in differences among the power systems where they are applied and/or the differences in personal experience and knowledge among the designers. The situation described is common, for example, in new transmission grid projects. New projects modify a part of the power system and they can create complex topologies because of weak short-circuit conditions with a high source impedance ratio (SIR), infeed, or special configurations like three-terminal lines. According to the regulations of some countries, in many cases, the projects do not replace the existing protection systems of existing elements, and the protection engineers would develop settings for new and existing protection systems to maintain proper protection coordination. Existing protection systems can present challenges if teleprotection is not available for distance protections or if only overcurrent relays are available. In these cases, regulatory fault clearing times are difficult to fulfill and even the performance of the backup protection coordination may be insufficient if redundancy is not available.
In protection coordination studies the major challenge when protection settings are calculated and verified is to achieve settings that are reliable for most operating scenarios of the power system (optimal settings). Power system expansion, new dynamics of equipment, and maintenance constantly add more complexity for finding the optimal settings of protection systems. Figure 1 shows the conventional focus of the reliability in protection systems.
Microprocessor-based relays are basically an industrial computer with settings as determined by the protection algorithms. Pickup, dial, zone reach, and operation time are the first to come to mind, but there are many more that require attention as well. Examples are polarization, phase selection, directionality, blocking logics, etc. Depending on the protection element or function, it may require a revision to optimize settings for proper protection coordination in all the possible operating scenarios.
In the presence of high inverter-based resources penetration, there are many challenges when using classic analytical methods to solve complex contingencies to determine optimal protection settings.
Opportunities for Artificial Intelligence in P&C: AI/ML can sometimes enhance existing power system protection and control functions that microprocessor relays currently perform, such as an AI application in addressing secondary arc (Figure 2). It enables engineers to make more data informed decisions by learning and creating models from historical and near real-time data. By leveraging the data to make protection and control decisions, the engineers can efficiently adapt their protection and control approach to account for a dynamic grid whose electric properties and characteristics are changing.
Artificial intelligence could create efficiencies by adding value and insight into various stages of the protection and control process. The main stages of the P&C process that AI could impact are as shown in Figure 3:
Each of these tasks above provide an opportunity for some aspect of AI/ML to assist either the engineers or the P&C devices react faster, be more accurate, and become more efficient. Applications and use cases that benefit from AI/ML are discussed in the subsequent sections. These discussions emphasize how each application benefits from AI/ML and how it impacts the various stages. In general, AI/ML may supplement and aid protection engineers but will never replace them.
Considerations before using AI in protection: Although, as discussed in this section, there are opportunities for use of AI to protection, before using any new method to improve existing technology and practices, a philosophical as well as a technical evaluation of the pros and cons of the new method is in order. Principles and applications of power system protection have been refined and perfected over decades of practical experience and a number of revolutions in technology. Since success of protection improves safety of equipment worth millions of dollars, and more importantly, human lives, protection engineers have traditionally preferred transparent and relatively simple physics-based models.
Interestingly, pattern creation and classification, two foundational components of ML based methods that help build AI in a system, are already an integral part of protection. Practically every relay uses a feature or a combination of features (pattern), and a rule-based classifier. A feature is typically frequency-domain transformation of time-domain measurement – current, voltage, frequency, power, impedance, harmonic content, high-frequency content, etc. A pattern is created through the combinatory logic blocks provided in numerical relays. A classifier typically has two-classes: Fault & No Fault. Separation plane for the classifier is a threshold value (for the feature) or a combination of threshold values (for the pattern). Values and combinations are determined through system analytics, physics and physics-based models, and experience. Here, experience is simply a representation of the human mind learning from past data. The difference is, the learning in ML applications happens (or experience is gained) through algorithms developed and refined by computer scientists. In applications using large sets of data, it has been shown that these algorithms “digest” the data better than a human mind can. ML based methods also tend to create and use quite complex patterns using mathematical transformations on raw data. Such transformations, when used to create a data-driven model of a physical process, often yield quantities that lose direct relationship to the physics underlying the process it is trying to model. These observations lead to two philosophical questions that need to be considered before applying AI to improve or replace an existing physics-based method:
1. If AI uses the same features/patterns used by relays, why would it be able to create a more dependable and secure classifier?
2. If AI uses more complex and abstract features/patterns that have no transparent relation to the underlying physics, and thresholds (separation planes) are created simply by learning through data, why would it work better?
Extending the thought behind these questions, if a legacy protection is not working in certain system conditions, why would an AI based method work? If it uses the same patterns the legacy relay is using, or can use, applying learning to these patterns may not necessarily yield any better result, as the patterns have failed. A question about opaque patterns to solve this problem is valid in this case as well. (Figure 4).
This discussion is against replacing physics-based time-tested methods with AI-based solutions without clear justifications, even if the existing methods fail occasionally. Unfortunately, a large number of papers published on this topic takes this approach. These papers also disregard the fact that power system protection is a system that encapsulates interdependent localized protection-schemes. Replacing one/few such schemes by a ML-based method may not reconcile with the holistic nature of power system protection. A pragmatic approach could be to complement conventional physics-based approaches with biology-inspired ML approaches to provide maximum protection coverage, and optimize dependability and security based on asset/system requirement. This report will present application examples of AI that have shown to or could potentially help improve protections where underlying physical models are not precise or absent, leading to known weaknesses.
Specific areas of AI/ML P&C application: For this report, power system protection is classified into two categories: asset protection and network protection.
The asset protection function involves detection of faults or abnormal conditions, discrimination of faulted zones or devices, identification of fault types, and making trip or block decisions, sometimes adaptively.
The network protection functions such as System Integrity Protection Scheme (SIPS), Remedial Action Schemes (RAS) and Wide Area Monitoring, Protection and Control Systems (WAMPAC) requires recognizing rare critical conditions impacting the power systems by monitoring the diverse set of variables. These systems, upon detection of critical conditions, initiate emergency control actions.
Regardless of the focus, design of the protection function/algorithm, system analysis, selection of settings, and coordination often require trade-offs between the security (no false tripping), speed of operation, and the dependability (no missing operations).
The more secure the relay is (both the algorithm and its settings), the more it tends to be less sensitive or operate slowly. And vice versa, the faster the relay is, the more it tends to operate falsely. The design and coordination of asset and network protection functions to achieve desired reliability and cost objectives may become a complex optimization problem in modern power systems. As the grid becomes more dynamic, the optimization objective becomes a moving target, and static settings and assumptions may not be sufficient.
Asset Protection: The limitations of classical asset protection techniques may arise from the extent of information available through the approach of using the same measurements used by existing relays (e.g., sampling frequency) and/or limitations of the relaying algorithms used for detection and discrimination. These limitations can manifest in the form of less-than-ideal speed and sensitivity to high impedance faults, or adaptability to various conditions such as high source impedance, series compensated lines, inverter-based sources, etc. To mitigate these limitations of the classical relaying principles, it is possible to improve and extend the range of measurements available for decision-making and improve the recognition process itself. For example, the high frequency transient signals contained in the input signals can be utilized to detect and discriminate high impedance, arcing and evolving faults, which are challenging to traditional protection function based on power frequency phasors. However, circuit laws applied to simplified models cannot characterize the relationship between the high frequency content and the fault location/severity. In such situations, AI/ML techniques can be used to train classifiers to make decisions based on various signatures contained in the high frequency components of the input measurements.
On a case-by-case basis, engineers can consider leveraging AI/ML models to augment the selectivity, security, and speed of classical relaying principles by:
- Optimizing settings and utilizing hybrid approaches that use multiple relaying principles
- Automatically correcting CT and VT signals
- Employing non-conventional algorithms
Network Protection: With respect to network protection, significant advances have been made in the field of on-line security assessment technologies that can determine critical contingencies, transfer limits, and remedial measures. However, conventional numerical techniques based on offline studies are usually time consuming and therefore are not always suitable for real-time applications. Also, these methods suffer from the problem of misclassifications or/and false alarms. Many studies show that AI/ML techniques can be applied to develop effective RAS and WAMPAC systems to detect critical or abnormal network conditions in real-time and propose or automatically activate remedial actions. (Figure 5).
Post Event Analysis: Another context where AI/ML can be very helpful is post event analysis by using data analytics. Modern data mining techniques can be essential when it is required to analyze large data sets such as digital fault recorder data, synchrophasor data, utility meter readings, weather data, etc. and their combinations. Analysis of large structured or non-structured data sets is needed in many situations such as sequence of event analysis.
Summary of Potential P&C Areas for Application of AI/ML: Table 1 provides some examples of protection and related functions where application of AI/ML techniques could be helpful. The current intelligent electronics devices (IEDs) employ classical protection functions to achieve many of these functions with high degree of success, but there is some room for improvement in terms of speed, sensitivity, reliability, and adaptability as mentioned earlier. On the other hand, classical algorithms are inadequate or too complex to apply in real-time for some functions such as system stability prediction and control, islanding detection, controlled islanding, arc detection, etc. In these applications AI/ML can play a crucial role.
Key Conclusions/Recommendations of The Report
The two existing examples included in the report are clear evidence that the technology could be applied successfully to solve practical power system protection and control problems. The seven emerging examples highlighted the great potential of applying AI/ML Technology to help solve some of the power system protection and control challenges.
It is also very clear that practical and widespread application and deployment of AI/ML technology is still at a very early stage. Despite a large number of research and development efforts that have been reported to date, many of such efforts are still a long way from being applied mainstream.
Given the large number of on-going R&D efforts, the body of the knowledge in this area is expected to grow and expand rapidly in the years to come. Being an emergent area, this report does not attempt to addressing details of all aspects within a limited timeframe.
Therefore, this report may become a living document that needs updating on a regular basis (e.g., every 3-4 years) to help incorporate new knowledges/insights and the latest progress in the practical application of the AI/ML technology in power system protection and control. In fact, the PSRC working group C43 has been approved to continue its effort to produce an updated report to incorporate latest development and progresses in next 3-4 years.
