With climate change becoming a bigger and ever-looming threat to the health of the environment, it lies in the hands of companies and individuals to learn how to manage their energy usage. This is not only beneficial for emissions, but also for reducing a customer's energy bill, saving them significant amounts of money in the long run. The best way to effectively discharge stored energy can help reduce overall energy usage.
What it does
Optumenergy uses a year's worth of energy usage data from a building in San Diego to train two different models to predict the next month's energy usage. Using this data, we are also able to identify an optimal point for battery discharge - a point where the energy savings are maximized, but the threshold is minimized. We then visualize energy usage and savings in an interactive environment where data can be uploaded and processed by users in real time.
How we built it
Energy optimization has two major components: energy load forecasting and battery threshold optimization. The first problem we tackled was energy load forecasting. There was a wealth of literature we saw on the internet on how to tackle this problem, and the first method we tried was a simple AutoRegressive Moving Average model. We used the ARIMAS feature of the Pandas package in Python to make a least-squares prediction for the next month of energy usage of this given location. For another model, we used the Keras package with a Tensorflow backend in Python, creating a neural network with neural decomposition to predict the next month in the periodic pattern of the given time-series data. The second piece of the challenge was figuring out at what threshold to start discharging the battery to the grid. This was an interesting part of the challenge, and we decided to pose the problem as one of convex optimization. Here is the formulation of our convex optimization problem: https://ibb.co/dhBoQn
Using this, we were able to minimize the threshold (t), while making sure there was as much capacity in the battery as possible
For the visualization, we used stdlib to link our different pieces together in order to create a unified user experience.
What's next for Optumenergy
There are many features that go into predicting future energy usage an optimizing energy discharge. We would incorporate more data sources, including seasonal and local weather data and appliance intensity, in order to produce more accurate predictions and suggestions for discharge thresholds.