AI Case Study
Enedis reduces high-tension electrical grid outage with predictive maintenance using supervised learning
French power supplier Enedis prevents outages throughout its network of more than 400,000 generators using DCbrain's Deep Flow Engine. The technology integrates historical data of the topology and pre-mature ageing of assets with machine learning. As a result the company can use the technology to predict where maintenance is required and avoid a failure.
According to DCBrain, "the idea was to integrate into classical machine learning model a new piece of information: the history of the topology and the pre-mature aging of assets due to electrical impacts. In fact, operational teams had the feeling that an incident on the electrical network would impact and age any asset that would be electrically adjacent to the incident.
Thus, the objectives of the project were as follow:
* Gather within a unique model all the patrimonial references, steering, and flows dataset of the grid in order to reconstruct the history of the topology
* Propagate incidents according to the steering policy in place at the moment the incident took place
* Use that model to identify the root cause leading to these incidents on the cables.
DCbrain created a unique technology, the Deep Flow Engine, that has the capacity to modelize under the form of a graph of flows all of the flows circulating on the grid.
* The links matching the elements represent the cables and they carry with them the measured flows at the given time-stamp. They also carry all their physical and electrical characteristics
* The nodes, elements of the network, carry both their steering and state history.
Such a model therefore allows the users to browse, thanks to computing capabilities, an evolutive representation of a network. It becomes quite simple to calculate, for example, intensity propagation or evaluate the impact of an outage by identifying all the «children nodes» or downstream nodes from that outage.
Our Graphe Approach, coupled with our capacity to deal with large masses of data allows to regroup different equipment data bases within a single model. In this case, the physical data-base (equip- ment/assets), the electrical data-base, the steering and incident history data-base."
"These recent approaches to de ne maintenance plans are now being used to challenge the statistical (or even empirical) methodologies in use."
"DCbrain's Deep Flow Engine has the capacity to modelize under the form of a graph of ows all of the flows circulating on the grid. This data treatment process follows 3 main stages: Data crunching: cleaning, extraction of the meaning imbedded into data lines and formatting, Graph mining: data treatment with the graph technology in order to represent the different layers of the Grid, Machine learning: preparation and set up of self-learning algorithms in order to determine reliability scores for cables.
Once all the data is collected, cleaned, processed, assembled in the transverse and enriched reposi- tory, we can use them to train a supervised predic- tive model."
"Enedis (Formerly ERDF) is a subsidiary of EDF (Electricité de France) in charge of addressing the country’s needs in electricity, which is done today mainly through nuclear power plants. Enedis is the French main electrical DSO (Distribituion System Operatpor). With 35 million customers, 95% of the French distribution network, ENEDIS is the largest DSO on the European Continent."
Aggregation of different datasets, such as electrical measures, asset information, steering journals, incidents logs among others.