The story behind the solar survey: an insight into how this ground-breaking technology was created

Ever wondered how Absolar’s solar survey technology was developed?



solar survey

With our solar survey now live and in action, we’re taking you back a step, and running you through the 2021 development journey of our state-of-the-art solar survey technology.


Prior to 2021, Absolar was able to carry out remote solar surveys using our LiDAR and Geospatial Information System analysis techniques, developed over a number of years with assistance from Universities and the Royal Academy of Engineering.


However, we identified a gap in the technology to accurately calculate where and how many solar panels could be placed on a rooftop, free from any obstruction and in a layout commonly adopted by solar installers.


For this, we utilised AI-powered computer vision to identify rooftops and their features from satellite images as explained throughout this article.


From the initial concept, through to product development and testing, find out how the Absolar solar survey was created to help thousands of businesses, homeowners and local authorities to find their solar potential.



Why solar?


As modern technology continues to advance, the amount of energy being consumed in our homes and workspaces is ever-increasing, putting more and more pressure on the National Grid.


As a result, UK energy providers are burning fossil fuels to meet the demand, which is releasing large quantities of carbon dioxide into the atmosphere, alongside other greenhouse gasses.


After this electricity has been generated, some of it is lost in transmission and distribution, culminating in a total of an ~8% loss.


Generating electricity locally through rooftop solar helps overcome these problems.

It generates clean energy, and there are minimal transmission losses due to the proximity of the power generation to its use.


Particularly for commercial and office buildings, most power is used during the day while people are at work.


Solar panels work to generate electricity at this time, meaning that there is minimal need for battery energy storage to capture the excess power.



The need to find solar potential


At Absolar, we’re passionate about making information about solar panels, and their potential, more accessible to the public.


Whilst a solar radiation model had previously been developed, more work was required to assess these buildings at scale, more accurately than ever before.


We set out to create a solution that offered a method to automatically analyze rooftops to assess their capacity for solar panels, and where to place them for optimal power generation.


With such a method, we could perform surveys at scale, allowing local authorities, commercial property owners and homeowners to access information about whether solar power is a worthwhile investment for them.



Creating automation


After performing solar surveys primarily by hand, it became obvious that there were many aspects of the process that could be automated in the future.


The solar radiation model we use to analyze the potential was brilliant for predicting potential power output, but the manual part of the process involved outlining the roofs to show where to place panels.


This manual process was too labour-intensive for surveying larger portfolios or cities. An automated solution was needed.


Our proposed approach was to use machine learning to break roofs down into individual segments.


Machine learning involves creating advanced statistical models that can learn from the data that trains it, to create insights from future data that it processes.



labelling process

A SPRiNT-aided project


To aid in the development, a SPRiNT project was carried out with the help of three current University of Southampton university academics seconded to Absolar in an internship.


The SPRiNT project aimed to develop the computer vision model that we needed in order to automate our solar survey process.


Data Science interns, Neishka Srivastava and Ivan Durán Pérez, were responsible for training a model suitable for the task, whilst Cameron Elliott, an image processing intern, was responsible for generating the data to train the model and process the results into useful information.


From here, we were able to develop the computer vision models.


To achieve this, the first step involved labelling 2500 images of roofs. This involved manually outlining the roof segments, solar panels and other rooftop objects, such as vents, skylights, chimneys.


This took approximately one and a half weeks split between all three interns. These labelled images could then be processed to train our computer vision models.



Developing three image segmentation models


In order to achieve our objectives, it became obvious that three image segmentation models were going to be required.


The first model aimed to break down the roof into segments, by highlighting the edges and folds of the roof, as well as showing where the roof surfaces were located.


The second model aimed to detect solar panels within the image, and the third aimed, to detect obstacles on the roof like skylights, vents and chimneys.


Transfer learning


When developing each model, several different pre-trained models were re-trained to detect the features of interest.


The process of using a pre-trained model and retraining it is called transfer learning, and it cuts the design and training time drastically as well as increasing accuracy.


The pre-trained models we tried for each purpose were VGG-19, ResNet50, DenseNet121 and EfficientNetB0.


The EfficientNetB0 and ResNet50 models were not selected for any of the three purposes, as they were the worst-performing.



masks picture

For roof obstacle detection and roof segmentation, DenseNet121 was used, and for solar detection, the VGG-19 model was the best performing.

Further development, after model selection, involved parameter tuning and varying other parts of the training to ensure that we were obtaining the best possible results.


Post-processing


Once the models were optimally trained, the post-processing was developed.

This involved simplifying the results into geometric shapes.


This made processing and geo-referencing the data to visualize it on a map, far easier further on in the process.



different model outputs

The roof segments showed us where it was possible to place solar panels.


Following this, the next step was to develop a program to automatically tile these shapes to calculate how many panels could fit into each segment.


From this, we could then calculate the predicted output of the panel depending on the location of its placement.


This is where a major obstacle was encountered.


roof segments

Overcoming obstacles


The satellite imagery was not geographically aligned with the national LiDAR coverage, which meant that we were not generating accurate radiation information for some of the panels.


The national LiDAR programme covers the entire United Kingdom, and provides 3D information about the topography of cities and terrain.


This provides the opportunity to analyze 3D building information, allowing us to predict the solar radiation falling on different surfaces.


However, the LiDAR is not perfectly aligned with satellite photography data, meaning that the solar radiation predictions were not well aligned with the roof segment breakdown from the image segmentation model.


This resulted in incorrect predictions for some solar panels when running the model, heavily affecting our data quality.


LiDAR misalignment

The solution?


To overcome this problem, we had to find out exactly how misaligned the LiDAR was in the east and north direction in meters, so that we could adjust our radiation calculations accordingly.


A process was developed to automatically calculate the misalignment using existing assets. The beauty of this is that it added minimal time to the overall development process.


This was a key milestone in the process as, otherwise, the data we would be generating at the end of the project would not be satisfactory. '


We were back on track and, once the research and development for this process had finished, we could perform surveys on entire cities at once.


Product testing


To test the process, Gosport was surveyed.


Why Gosport?


It was an ideal location to test our new-found product, simply because it’s an area with a variety of property types within a relatively small population.


Testing, testing


We stumbled across a few issues regarding data querying times and cleaning data inputs, which needed to be ironed out on the way, but soon we had the rooftop solar potential for all of Gosport.


This was ground-breaking.


The quality of the generated data was reviewed by visualizing the panels on maps and by inspecting the robustness of the roof segmentation.


panels on roof

The final stage: process automation


The next stage in the process was automating the entire process, with the aim of being able to name a city, press a button and the results would all be generated without any manual input.


This would save a great deal of time, whilst presenting highly accurate solar survey data.

The automation of this process is known as data pipelining and is a proven way to efficiently process vast quantities of data.


The final version of the pipeline was ready in early January, with Southampton and Lewes being two of the first cities to be surveyed using this pipeline.


What previously took an entire day was now taking a few minutes, saving vital time and resources.


Looking to find your solar potential?


Get in contact with us today to request a solar survey of your building or local authority area.




57 views0 comments
MicrosoftTeams-image (7)_edited.jpg

Click here to find out how Absolar can find your solar potential, whether for a building, a portfolio or a city.