ESB
Nominated Award:
Best Application of AI in a Large Enterprise
Website of Company:
https://esb.ie/
At ESB, we have been Ireland’s foremost energy company since our establishment in 1927. We are driven to make a difference by achieving zero carbon emissions by 2040. Our commitment to providing clean, affordable, and reliable energy directly supports the achievement of UN Sustainable Development Goals. We are a strong, diversified utility operating right across the electricity market: from generation, through transmission and distribution, to supply. We also work in other related sectors where we can make a difference, including telecommunications, electric vehicle charging, home retrofits and more.
At ESB, we believe that electricity is a social good that should be accessible and affordable for all. We also believe in acting with integrity and transparency, protecting the world around us and creating an inclusive and flexible culture that protects and empowers people. Across our diverse range of businesses, we use our capability and expertise to develop smart and sustainable energy solutions to tackle climate change, one of the biggest challenges facing society today. In ESB Analytics, we collaborate and partner with people and teams across ESB, to identify and develop analytical solutions to support the efficient running of the company in line with our Brighter Future strategy.
Integration of analytics into the wider organisation
Since its establishment in 1927, ESB has been characterised by a commitment to drive society forward and deliver a brighter future for the customers and communities we serve. This strong sense of purpose is reflected in our constant and unwavering commitment to tackling society’s biggest challenges, enhancing people’s lives, and creating new opportunities for individuals and communities to thrive. We are driven to make a difference. Digital and data driven technologies are also transforming the electricity sector, underpinning the development of smarter networks, and giving rise to new business models, including digital only suppliers offering an enhanced customer experience at a lower cost to serve.
Being Digital and Data Driven is seen as a foundational capability required to deliver the ESB Strategy for Net Zero by 2040. This will see ESB transformed into a data driven digital utility delivering excellent customer experience, an enhanced people experience, and modern business operations and processes enabled by technology.
ESB Data & Analytics sits in the strategic CIO organisation. We are arranged in a Hub and Spoke Model; Data Science, SAP BI, and Data Engineering sit within the Hub and each of the ESB Business Units has an Analytics Spoke which includes – Customer Solutions, Enterprise Services, Networks, Generation and Trading and Engineering & Major Projects, made up of a team of data analysts. With this model, we get to sit close to each business area, working collaboratively with them to understand and help solve business challenges through data. The ESB analytics team has developed many innovative solutions in the past few years as the company moves to becoming a more data-driven utility. The progress of the team was recognised at the recent Analytics Institute Awards, where ESB won Analytics Team of the Year.
Reason for Nomination:
The Commission for the Regulation of Utilities (CRU) is requesting ESB Networks to map the Low Voltage Networks in Ireland. Currently the LV network has issues with data quality. In some locations, recorded data for our overhead network is either incomplete or does not exist. In previous years this was not a big issue, but recently with the introduction of domestic low carbon technologies such as EVs and heat pumps, it has become much more important for ESB Networks to understand exactly how our customers are connected back to the grid. Previously, the only way to map the network would be to have technicians “walk the lines” and grab the location of each pole. There are 150,000km of overhead lines managed by ESB Networks connecting 2.4m customers nationwide and the CRU have imposed a strict timeline and provided incentives for the project’s completion, therefore this work required a more automated and sustainable method. It was decided to use publicly available street-level imagery with a computer vision model developed in Microsoft Azure to move along the streets of an area and identify LV poles and cables coming from the pole to the house.
Method:
Following initial conversations with business SMEs we identified test areas that we could use to train and test the model. Currently, less than 2% of the LV network has been mapped, however Dublin has a higher percentage of mapped utility poles, so we were able to procure coordinates of poles and use these as a basis to scrape images of utility poles. We developed a Python script using Selenium automation tools to display street imagery at the location of the poles and take a number of images of the poles. These images were saved and uploaded to Azure where the Pole_Identification model was trained to distinguish between utility poles, telecom poles and other street furniture such as street-lights, 5g poles, etc. We then created an LV_Detection model that would identify the LV cable going from the utility pole to a house.
Once the models were sufficiently trained, we applied the models to a location where we had been previously mapped to compare results, Loughrea, Co. Galway. Using OpenCV, we were able to identify all the roads in Loughrea from an overhead map and generate a route that the model could “walk down the street” and identify poles and the map the network. Using Selenium to automate the “walking”, we placed the model at the starting point in the road, found the closest utility pole, calculated the pole coordinates using advanced trigonometry and jump to that location in the road and, finally, run the models from three locations relative to the pole, in front, from before the pole looking up the road and from after the pole looking back the road. This provides us with the clearest view of the pole and avoids as much visual obstructions as possible. If the models confirm it is a utility pole, the coordinates are saved and the images are passed to the LV_Detection model to identify if it is part of the LV network. Once complete, the model turns to look down the road again and move to the next pole. Finally, the results are provided back to the business with the location and type of pole to be incorporated into the systems.
Tools & Obstacles:
We used a combination of Microsoft Azure Cognitive Services, native python code and Selenium automation tools to produce training and test images and also to roll the product out for all areas. Microsoft Azure is our cloud platform of choice and we have produced AI products in Azure in the past, so we have the expertise to fully exploit that product. OpenCV and Selenium were also utilised. It was a very difficult project as we needed to overcome many different obstacles to develop the model. First, we needed to become experts in identifying the different pole types and understand the network better.
The procurement of images was also a difficult task. We don’t have a store of utility pole images in ESB so we needed to scrape our own images. We attempted to make use of various APIs, however we were coming up with restrictions in terms of use and also poor resolution imagery was being returned. Deriving coordinates from 2D images was also a big challenge, however we were able to accomplish this through advanced trigonometry.
Result:
Following the Loughrea trials, we were able to demonstrate that the model was comparable to the manual inspection of the lines for visible areas. The model was able to map LV poles and LV cable from the pole to a house with 84% accuracy. The model also was able to pinpoint the poles to within 5m of its true latitude/longitude. A process that would have taken a number of days to map an area, was completed through Azure in 3 hours. This saves on workers time and provides sustainability benefits as we don’t need to send technicians out to drive around areas.
Having been delivered as an MVP, the product is now going to be rolled out to new areas to map as much of the country as possible to reach our targets. The output of the project will have many benefits. Not only will ESB meet the requirements set by the CRU but will provide ESB Networks with crucial information needed to maintain and upgrade the network. The store of images can also be built upon to identify safety aspects, such as identifying pole leaning and other damage. Damaged poles are the cause of a large proportion of outages in Ireland, particularly during storm time when they are more susceptible to failing, or worse, collapsing and being a danger to the public. Further information can be taken from the images, for example the material of the pole (wooden, steel), type of connectors, exposure to vegetation, etc which can be incorporated to a forecast model to predict when poles need to be manually inspected or replaced.
Additional Information:
I truly believe that this is an original concept that has been carried out superbly by the team. Utilising publicly available street level imagery to identify street furniture, while not a new concept academically, in industry hasn’t been trialled too often. In ESB especially, this is a new venture. Combining this imagery with advanced computer vision techniques will be revolutionary to ESB’s ability to carry out audits, predict faults and protect and upgrade the network. As we move towards Net Zero by 2040 as per the ESB strategy, the LV network will come under increased strain as the electrification of heat, transport and other daily activities comes onto the grid. By undertaking this project now, we are taking the necessary steps to protect the network and the public it serves.
The project team had to overcome a multitude of challenges to get to this point of what is an incredibly difficult task. Detecting a small black cable against real world backdrops is not an easy task for the human eye, not to mind training a machine to detect it. The team trialled many different techniques to improve the precision, some of which worked and others which did not. We worked with different image types (2d, panorama), different image processing techniques (canney-edge, hough lines transform) and image scraping techniques via APIs, partnerships, etc. The most important output of the project is that this technique can now be replicated to overcome a number of other challenges and to identify other ESB furniture such as minipillars, generators, connectors, etc. The project is a massive breakthrough for ESB and one that sets the team and company up to drive AI in the entire company and for the benefit of the country.
ESB project 2021
Nominated Award: Best Use of AI in a Consumer/Customer Service Application
Electric Ireland is the retail division of ESB (Electricity Supply Board). ESB was established in 1927 as a statutory corporation in the Republic of Ireland and most shares are held by the Irish Government. Previously known as ESB Customer Supply and ESB Independent Energy, the retail division of ESB has been rebranded to Electric Ireland in 2012.
Recognised as Ireland’s leading energy provider, Electric Ireland supplies electricity, gas and energy services to over 1.2 million households and 95,000 businesses in the Republic of Ireland and Northern Ireland. ESB is now supplying electricity and gas to homes across Great Brita in. ESB Energy launched as Britain’s newest energy provider in 2017 and our pledge is to be the most innovative, responsible and easy to deal with electricity and gas supplier in the British market. ESB has been working in Britain for over 25 years and invested nearly £2 billion in projects across the country.
Whether you’re a residential or business customer, our ambition is to always offer you the most relevant, affordable and innovative energy solutions that answer the way you use our electricity or gas services. We call this ’Smarter Living’ – a way of understanding and answering your needs so you can enjoy our energy with more control and in sustainable ways. ‘Smarter Living’ also defines our approach to supporting community initiatives and sponsorships.
Reason for Nomination:
Project Strategy & Objectives
As part of the “Smarter Living” approach a project team was put together in order to understand the driving factors behind the Repeat contacts across the Electric Ireland Residential customer base. A review of all methods of contact by customers was required to determine contact method preferred, analyse what the most frequent reason for contact is and look at suggestions to improve. This was always with a view to how to interact with our customers in an agile and timely manner in order to provide a better-quality service.
Approach:
A single source table was designed to be a view for ever y customer and included all their incoming & outgoing triggers and contact activities. An Impala table was used for the single source table by creating union of queries that draw data from tables in our Hadoop infrastructure across multiple data sources, which included Telephony data, CRM and Billi ng transactional data and customer segmentation model. The queries have been constructed to allow changing date ranges to be amended to create versions of the table for different time periods, and in a manner conducive with the addition of new triggers and contact methods. The resulting table is structured with a row for every occurrence of each trigger and contact method and can be sorted by customer and activity date to get a chronological view of each customer’s activity.
Feature Engineering:
By utilising Cloudera Data Science Workbench and Hadoop, the process of feature engineering was made easier and enabled us to understand more about how each trigger relates to contacts. This assists in analysing and discretising continuous features into buckets. For continuous variables, feature engineering enables the selection of discretisation buckets that capture the variation in propensity across the range of continuous values. This means that when running the conditional probability and profiling code, the various facto r levels provide visibility of potential variation across propensities for different regions of the continuous variable. We can do this by plotting the propensities across the continuous variable and then using the visual output to decide on where to discretise from. In plotting the data for positive and negative value it can appear to be symmetrical, the buckets contain the absolute value of a feature selected during the study. Were the positive and negative values to show quite different trends, separate buckets for positive and negative values could be created to incorporate the different t rends.
Feature Selection:
The feature selection process involved the analysis of the model running when a feature isn’t present. If a feature’s removal led to an increase in the predictability of the model, the feature was removed from the model. The code recommended the feature that, when removed, caused the greatest increase in predictability. The feature was then removed from the dataset, and the feature selection code run once again. This process was continued iteratively until all features that caused a decrease in predictability have been removed from the model. Hidden nodes can be generated as an additional way of increasing predictability.
These are new nodes that can be created as an amalgam of existing nodes that have a link between them in the Bayesian Network. If a combination of nodes increased the predictability of the model, the two nodes can be combined into a hidden node, and their originals removed. The code tested the change in predictability of the model with each combination of nodes that have a link in the Bayesian Network to see whether the joining of any pair s increased predictability. The combination of nodes that increased predict ability the most were then selected and combined.
Model Visualisation:
The model code outputs the Bayesian Network links into a connection matrix file. This file was then inserted into the visualisation code alongside the parameters to create the visualisation. The visualisation can be modified to amend the visual elements of the nodes and links to ensure that the output is visually coherent, and the sizing of the various elements is conducive with the business area where the visualisation will be utilised. Th e main element to define within the visualisation is the node size. This should ideally be defined based on the relative opportunity sizes for each node. The node colouring is also an element that was considered as this can also add meaning to the visuals.
Additional Information:
Conditional Probabilities:
The conditional probability code calculated the propensity for each individual feature to drive repeats. The generated probabilities were then outputted for each feature. These were then used to calculate the additional yearly repeats for each feature.
Profiling:
The impala query for the profiling table created a row for each customer with each column denoting a customer profiling element. The profiling elements were overlaid on to the customer data to provide visibility of the repeat propensity for various customer segments for each trigger. The profiling prep code conducts the preparatory work for this process by pulling in the profiling elements and joining them onto the observations. The profiling code then calculated the propensities for each profile overlaid onto each trigger, and outputs this for each trigger. These outputs were then used to calculate the additional yearly repeats (opportunity sizes) for each customer segment for each trigger.
Impact for the Business:
The model can be run and compared for different date ranges allowing the Electric Ireland business team to identify different triggers and events based on time dependency. This can assist in analysing the change in propensity for trigger to drive repeats when a change is made to the customer experience/engagement process. This can be done by re-running the single source creation code for different date ranges, or by creating a multi-variable matrix with large date range and selecting specific dates to draw a single sample. The modelling, profiling and conditional probability code can then be run on the samples from each date range to yield which triggers were most impactful in each date range.
Evidence of Success:
Improved understanding of the drivers to repeat contact means the ability to determine who, when and how customers will contact Electric Ireland. This means an increase in proactive outbound contacts through cheaper and more efficient channels, reducing cost to serve and increasing the customer satisfaction and improving the journey for the customers. Understanding of the customer journey for those who repeat contact improves the services provided for all customers as Electric Ireland can place the critical and necessary resources where needed and at the right time.