Northern Gas Networks and CKDelta at Neo4j GraphSummit London 14Nov23.pptx
Download as PPTX, PDF1 like294 views
This document discusses using machine learning and graph databases to help manage water ingress in gas networks. It presents a recommendation system that uses a graph model of the gas network to identify locations at risk of water ingress. Features like centrality scores, community detection, and pipe embeddings are extracted from the graph and used in a machine learning model. The model achieves higher success than traditional methods at predicting whether syphons need to be pumped and the amount of water extracted. Other potential applications are also identified, such as estimating syphon pumping frequency and early detection of water ingress issues.
Northern Gas Networks and CKDelta at Neo4j GraphSummit London 14Nov23.pptx
- 1. Enhancing operational efficiency in gas networks
- 2. 2 There are multiple ways water can ingress a gas pipe and when or where these incidents occur, is difficult to predict through traditional means. This could lead to a loss of supply for customers. Business challenge Water ingress into gas pipe is a highly disruptive occurrence in gas networks. Water ingress
- 3. 3 Water ingress Syphon pumping frequency estimation Early detection of water ingress issues Detecting the location of water ingress Project goal Forecast whether functional locations must be pumped in a short period of time Water ingress management
- 4. 4 CKDelta - Confidential 1. Executive summary 2. Recommendation System Overview 3. Graph features 4. Model Results Agenda
- 5. 5 Achievements • Used up-to-date operational data that can be scaled to run regularly. • Built a machine learning model with a higher success rate than BAU • Usage of Graph Databases to accurately model the gas network Executive summary
- 6. 6 CKDelta - Confidential Recommendation system overview
- 7. Use Neo4j to translate pipe connections as features of individual pipes ✔ The model, now, also relies on pipe connection properties such as height gradient and material change, instead of only the pipe height, material, and proximity. Pipe features are limited to explaining the properties of a single pipe Pipe features explain all incoming/outgoing connections and the properties of the connections Without Neo4j With Neo4j The game changer!
- 8. 8 Nodes: • Work Order • Syphon • Functional Location • Pipe • Transmission Pipe • Meter • Work Centre • Water In Service • Planned Interruption • Unplanned Interruption Relationships: • Pumped • Attached To • Connected To • In Area • Supplies To • Nearest Feeder • Reports Graph model for syphon pumping within gas network
- 9. • Find all Syphons Upstream from a Customer Meter with a reported Water Ingress Issue • Find all pipes at Elevation peaks from a Syphon Location • Find all Network Dips around a Customer Meter with a reported Water Ingress Issue Cypher Query Neo4j Browser Output Syphon Upstream from Water-in-Service-Repair Event Questions we ask Neo4j
- 10. Shortest Paths between Street with Syphon Upstream and Water-in-Service-Repair Event *Paths extracted from 0.13 Million Nodes and relationships Get complex connections paths within seconds!
- 11. 11 CKDelta - Confidential Water-in-service repair: Syphon may be full if a customer meter downstream reports a disruption in supply. Unplanned interruptions: Syphon may be less prone to be full if a customer meter upstream has been replaced/fixed. Graph features: network event
- 12. 12 CKDelta - Confidential Centrality of a node Importance of each node within the graph, based on the number incoming relationships and the importance of the corresponding source nodes. E.g. Google search page rank algorithm - a page is only as important as the pages that link to it. In a gas network, the centrality score per location based on water flows is used to quantify the risk of finding water at that location. Graph features: topology
- 13. 13 Community detection in a network Find densely connected group of nodes within a community, compared to how connected they would be in a random network. In a gas network, the communities found are used by the model but can also help to define constraints when recommending the siphoning operation. For example: avoid too much siphoning in the same community. Graph features: network topology
- 14. 14 In a gas network, the pipe embeddings are used for efficient feature engineering, by transforming all connections to a Pipe into vectors of numbers, that may be used as inputs to the model. Graph features: graph embeddings Pipe Embeddings are used to transform the neighbourhood information of every node in the network into a low-dimensional vector of numbers associated with each node.
- 15. Model results Successful syphon if either: ❖ The water extracted value is Yes ❖ Amount of water extracted is greater than zero
- 16. 16 Other use cases identified Water ingress Syphon pumping frequency estimation Early detection of water ingress issues Detecting the location of water ingress
- 17. Thank You