Great interest in Exeri at the Cired conference in Madrid
Using Smart Grid Surveillance to detect and localize failures in the Overhead Medium Voltage Grid.
The complete system approach
It is not possible to know the complete grid status from a few sensor points. Instead, we need to combine information from the complete grid to draw any real conclusions about possible faults, their type and location without alarming for normal abbreviations.
To build a complete system, not only detection devices with some in-built intelligence are needed, but also a reliable communication layer and intelligent decision support to present conclusions as a decision support.
Full-scale tests and results
Smart Grid SurveillanceTM, an implementation of this, has been tested in a project together with Skellefteå Kraft AB. Preliminary test results show that both detection and positioning of a number of faults work as expected. So far, power loss, phase to phase and phase to ground failures has been verified.
No other fault types has occurred in the tested grid so far, and thus the AI has not been able to learn to recognize these patterns. It is however known how to locally detect these failures, and as soon as the fault behaviour throughout the complete grid is also known, these failures will be added to the catalogue of problems that can be positioned in any grid where the system is applied.
To achieve this is as quickly as possible, and to achieve an even higher and faster calibration of the self learning mechanisms, work has started to inject these fault types into the live grid under test.
The process from manual to automatic fault positioning
Most repeated processes can be modelled using the OODA-loop, first described by John Boyd in 1976. This includes the process of finding and repairing faults in the power grid.
Observe – making observations of the grid and its status. Manually this is often done by customers reporting power outages or indication from a protective relay. Making this automatic means having sensors distributed around the grid to be able to know what the status is in the complete grid.
Orient – making sense of gathered information. Currently done by knowing
the grid and its workings well, or by looking at a map and reason about it. Automation means gathering collected information by secure communication
and use it all together with its context (grid topology) to analyze what has actually happened and draw conclusions from that.
Decide – make the decision on what to do to improve things. Today this is based on incomplete information and may as well be called experienced guesses. An automated decision (or, in the Smart Grid Surveillance version, decision support), makes this step much easier as conclusions drawn from the Orient phase about the exact fault position and type are available.
Act – improving things, by repairing the fault, or if it in the manual case was not yet found, send someone out to look for it. In the automated version, they are sent immediately to the correct position without waste of time.
Line vs Grid
As has been known in the telecom industry for quite some time, there is a huge difference in monitoring a single line and a complete network. The telecom sector has the huge advantage of in-built intelligence in the network itself, but the power grid will never have this possibility to exploit.
A single power line may be monitored by a single sensor, as e.g. a fault current will only have one way to go and the direction of it actually is relevant information. As soon as we add a branch, the fault current will be divided – and exactly how it will transverse this simple grid depends on many factors, including the type of fault that generates it and the currently connected load.
A normal grid contains a number of nested branches, mixed overhead/cable, breakers and distribution transformers that will also affect this. The complexity quickly rises to a level that is impossible for the human brain to understand – especially built on information from one or two points within the grid.
AI and pattern recognition
This type of multivariate problem is ideal to handle using an AI with knowledge of the grid topology and many points of information collection. When using the AI upon a complete grid and start by teaching it the beginnings (by confirming specific events like faults, failures and minor disturbances as well as their specific positions within the grid), it will quickly learn to identify not only the type of fault but also where in the topology the fault position is.
The topology may then vary between grids, but as soon as the topology information is available to the AI it can use it to identify and position faults within it.
In a practical system, transfer of real time data about the grid status to the central AI-system is a challenge. Many power grids, especially in suburban areas, do not have access to a reliable communication network of ordinary means that in itself is independent of power from the grid.
There is of course the possibility to collect information for later analysis, but the response time to any problem will only grow longer as all information needs to be collected, analyzed and reported.
There is also need of a way to filter the information from the grid, so that only what is needed by the AI is actually distributed to it. There is a fine balance of how much filtering that may be done at the sensor level, but the fact remains that the sensor itself must be intelligent enough to decide whether a piece of information is interesting enough to send to the AI. Simple sensors that only collect data will thus not fulfill the purpose in this context.
Another important question is the number of measuring points needed to cover the grid. This depends completely on what resolution of fault positioning that is wanted. If we only want to know in what branch the fault has occurred, we need to put intelligent sensors at the branching points and endpoints. If we want to have resolution on pole level (in an overhead grid) we need to put a sensor in each and every pole. Here the user needs to decide what is an acceptable balance between resolution and cost. There will probably be different needs in different parts of the same grid.
Conclusions used in Smart Grid SurveillanceTM
Our conclusion from this is that a system to fulfil efficient monitoring and fault positioning in a grid needs to have the following properties:
Minimize the data transfer by making intelligent decisions at the information collection point
Streamline the logic of the collecting unit by keeping the larger picture (grid topology) in a central unit
Make communication reliable and power efficient
Make the system (including communication) independent of the power grid itself
Use AI for advanced pattern recognition through a complex grid
For a system to be usable by grid owners, some other requirements need to be fulfilled.
System needs to be easy to install. This is achieved by making the AI adaptable (self-learning) to be able to add the system to other grids without complex configurations. It must also be possible to install the intelligent sensors without breaking power.
The system also needs to be easy to maintain, and of course to use.