Context
Today, more and more people have started installing infrastructures such as solar panels or small wind turbines in their homes. They can then start to produce electricity for their own consumption, reducing their dependencies on the traditional power grid. When conditions are optimal, it is even possible for them to produce more energy than they might need. If a neighborhood was able to connect their local production of electricity, they would be able to create their own, small, power grid. They would be able to share their energy to further reduce their dependency on the traditional grid and reduce the amount of power they would lose by not using it. The home owners would even be able to sell some power to the main grid if they produced more than they could handle at any point in time. Yet this unique setup would expose a problem only present to large groups today: the real time management of a shared hardware infrastructure. It is a system that needs to handle both an overflow and a shortage of power and buy or sell electricity from the main power grid.
What it does
Our solution offers a visualization of the energy production and consumption of an household. Through our web interface you can see your past usage per period, as well as a forecast of how much energy your are expected to use. We use several forecasting techniques including ARMA and a recurrent neural network. This allows us to have a thinner margin of error, comparing the two results and calculating a result closer to reality. The forecast can alarm the house owners that they are going to rely heavily on the main power grid in the next days if they do not reduce their consumption, highly increasing the amount they will be billed at the end of the month. It could also alert them that the weather conditions in the following days will highly increase their power output, potentially damaging their equipment unless they find a way to reduce the production or increase the power consumption.
How we built it
Our interface relies on a nodeJS backend and predix UI front. Our forecasting models rely on a neural network created from the TensorFlow library as well as a statistics model based on time series, ARMA. We have frequent datapoints, around every 10 minutes, for two and a half years. This implies that both our models can have a rather precise and small timestep. We could predict the power production and consumption down to every couple of minutes, up to every week for the next six to seven months, if we pooled the data enough to have a relevant time scale. We use new data from captors placed in the houses to correct our predictions and continuously adapt to the users' evolving needs, according more importance to the newer informations than to the oldest, through attention biases. We are running an UAA service to login to the interface, as well as a timeseries service to manage our data sets. Our models are based on python micro services.
Challenges we ran into
We ran into a couple of challenges. Firstly, our login service has difficulties redirecting to the desired pages. We have functioning pages, but they are not linked together. We use plenty of python applications, which are complicated to incorporate into the predix devbox, as none of their dependencies have been previously installed. Finally we have had a few issues with Predix, which have held us back for a great amount of time (more than half the available time at deadstops).
Accomplishments that we're proud of
We are proud to have risen up to the challenge GE offered us, and to be able to present a solution at the end of the two days. We worked as a team through the whole challenge, helping each other through the hard parts, while still keeping up a good mood when we were stuck with no clear solution ahead.
What we learned
We have learned about team working with people from different backgrounds than us: we did not know at least half of the other teammates when we started out and bounded over the two days of challenge. We learned about machine learning, statistics (for some of us), how to do a backend and frontend, as none of us had these skills prior to the hackathon.
Log in or sign up for Devpost to join the conversation.