Inspiration: We developed GridSync as an experimental green technology that would represent a significant improvement in the efficiency and sustainability of power grids, as well as allow us to explore several new and exciting concepts. GridSync aims to revolutionize energy load balancing and the peer to peer energy market by using blockchain technology to create a decentralized energy grid that can be overseen and optimized through the use of Machine Learning. The project addresses several key challenges in the current energy landscape, including the integration of renewable energy sources, efficient energy distribution, and complexities inherent to peer-to-peer trading.
What it does: GridSync at its core can be viewed as a service. Our backend accepts connections from nodes on an energy grid such as substations, generators, and houses, receiver production / consumption data and uploads it to a database for analysis. The machine learning division of our software maintains a long short-term memory recurrent neural network (LSTM RNN) for each node that is used to anticipate load balancing operations in real time. The networks are stored on the database with the data that trained it, saving local storage and allowing the model to be pulled and retrained as new data becomes available.
GridSync makes use of a custom block-chain (developed in house) to track and verify energy transactions on the grid. Transfers between substations, surplus listings (via smart contracts) from solar-enabled homes, and negotiations between power stations are all events that are are securely recorded on the block-chain
How we built it: We realized early on that we would need a simulation to generate data and serve as a proof of concept. GridSync uses a Flask backend and a Vue.js based frontend to serve as a simulation to generate testing data. To model the data, we used real-world figures for power costs of different devices used around the home, such as HVAC units, refrigerators, lighting, and more. We then modeled a probability that the appliances were running given a certain time of the day, time of the year, and weather outside of the home. We generated around 43,000 data points representing each hour over the course of 5 years. We pulled temperature data from NOAA readings from a Wooster, OH weather station in order to model accurate climate predictions in Ohio. These data points were then what we used to train the machine learning algorithm. The model was then harnessed to the Flask server, and an endpoint was created that would take a size-4 vector (year, month, day, temperature) and would return its predicted value for that point in time. The model blockchain was also hosted on the flask server, but in practice it could be hosted by a network of computers to be truly decentralized.
Challenges we ran into: Many of the challenges we ran into dealt with the integration of our frontend and backend. Although we got the Flask server fully running, we couldn’t find a way in time to let the two communicate with each other. The proof of concept is complete, though the user must manually enter POST requests into a utility such as Insomnia or Postman to interact with the server.
Another issue we ran into was training the dataset, as 43,000 data points proved to be too much to train in such a short amount of time. To remedy this, we collapsed the dataset by taking weekly averages for a given time. The hope was that we would still be able to capture daily and monthly trends/fluctuations in energy use, even if averaging over the week meant we lost weekly patterns.
Accomplishments that we’re proud of: One of the accomplishments that we’re proud of is implementing our own block-chain. Rather than using a pre-existing network, we found it would be fit to create our own for ease of use. It turned out to be easier than anticipated, and was a fun challenge to partake in near the beginning of the competition.
We are also proud of our idea, even if it wasn’t able to be fully fleshed out over the course of 24 hours. This is a technology that we believe could make a real impact and improve the efficiency of power grids all over the world.
What we learned: This project taught us a good deal about the inner workings of recurrent neural network training, blockchains, and smart contracts. The three of us hadn’t dealt with these technologies in an intense form in the past, and this project was a great dive into them.
What’s next for GridSync: In the future we hope to finalize the integration between the frontend and backend and be able to demonstrate the powerful predicting capabilities that our neural network has over load rates at any given time. After a full PoC is created, we can see this technology being adopted on a more widespread basis.
Log in or sign up for Devpost to join the conversation.