Solarian

Inspiration

As Singapore moves toward being a SmartNation, the demand for sustainable energy increases. With SG facing limitations such as lack of space and natural resources, our most viable renewable energy option will be Solar Energy. Being in the tropical sun belt, Singapore enjoys an average annual solar irradiance of 1,580 kWh/m2/year.

As such, technologies that boost the use of solar power will ultimately help SG transition to a fully sustainable Smart Nation

What it does

Solarian allows property owners to identify all solar panel placement spots on their property in just a few simple steps. This is done by running a satellite image of the target property through our pre-trained computer-vision-based UNett model.

How we built it

Users input a location via an address or using longitude/longitude values and choose the image 'patch' that best represents the building users are trying to target. The app will return information such as surface area of a flat roof (for installation of solar panels) and money saved per time period after installing solar panels at a specified 'patch'.

Challenges we ran into

Getting accuracy for the actual commercial and industrial use due to lack of data and optimisation of model, unable to allow users to specify exactly what buildings to run through the model, integration hell trying to integrate everything together, certain quality of life functionalities are not implemented on time , weather data is categorised into regions ('North','South','East','West','Central') of Singapore currently only

Accomplishments that we're proud of

Excellent accuracy of ~0.87 on training model using Standard UNett for image segmentation over 100 epochs, implemented a decent frontend UI on StreamLit given the lack of time and experience , managed to establish a direct relationship between 'regions of Singapore' and 'money saved through the installation of solar panel'

What we learned

Frontend: StreamLit, Backend for our Machine-Learning Model: OpenCV, Keras, TensorFlow, segmentation_models, Other libraries used: pandas, patchify, sckitlearn, geopy, pillow, matplotlib

What's next for Solarian

Fine-tune the model and feed more data for training the model to get extremely high accuracy for actual use, Fully implement all front-end quality of life features such as an automatic screenshot of the map, Expand the usage of this app outside of Singapore, Give more accurate metrics of and conversion of pixels to the surface area, Addition of more features in Computer Vision aspect such as calculating angle of the roof