CDS Students Leverage Meta’s Newest AI Forecasting Model to Address Energy Consumption and Climate Change

Jennifer Rodriguez-Trujillo and Max Wang’s case study compares Meta’s NeuroProphet to other existing forecasting techniques

CDS MS students Jennifer Rodriguez-Trujillo and Max Wang spent the summer working on a case study they developed using NeuroProphet, Meta’s newest deep-learning time series prediction algorithm. The data science students used the opportunity to learn how the model works and compare it to traditional existing time series models such as Exponential Smoothing and SARIMA.

The deep learning model from Meta has been cited to improve forecasting accuracy because of its ability to take in information other forecasting models leave out. The research endeavor, supervised by CDS PhD student Vishakh Padmakumar, used NeuralProphet to forecast energy consumption and energy prices in Victoria, Australia based on data collected from 2015 to 2019.

To learn more about the case study, CDS spoke with Jennifer and Max about how the project got off the ground, the challenges they faced, and how machine learning can be applied to a global crisis like climate change.

The below interview was lightly edited for clarity.

What was the inspiration for the case study?

Jennifer: We decided to work on a machine learning-related project but weren’t sure how to approach it because we have different interests. I’m interested in fashion, arts, and entertainment, whereas Max is interested in climate solutions. We decided on this case study for its ability to use time series forecasting, which is widely applicable in fashion and e-commerce as well as electricity forecasting but is also critical to climate solutions.

Max: I was browsing through Kaggle [a website with publicly available data sets and ongoing projects] and was inspired by a user who was working on a dataset containing electricity demand data in the province of Victoria, Australia. Electricity demand is a typical time series problem; how do you predict how much electricity a certain area, such as a city or a hospital, may need? This type of data has become increasingly volatile, due to many factors caused by climate change such as hurricanes, droughts, and heat waves, that put stress on the electric grid.

Was there anything surprising or challenging about the research you discovered while working with NeuroProphet?

Jennifer: One thing I found surprising was its malleability. I loved its ability to mold to the user’s needs. For instance, one trait of the model is its ability to take in change points or holidays like Christmas which would lead to a spike in electricity usage. These hyperparameters allowed us to take those particular events in mind.

Max: There are a lot of challenges you naturally face when using a new model. Whenever you encounter a bug or a problem, trying to troubleshoot is not as easy as an old model where people have probably encountered the same problem. We were still learning about the model as we were implementing it so we were coming at it with the mindset of whatever results we produced would be the result of a learning process.

Could you talk a bit more about your research’s connection to climate change and how machine learning can help us respond to global warming?

Max: Time series forecasting specifically applies to climate change in a few ways. With extreme weather becoming more common, as a grid operator using machine learning intensive forecasting could know exactly what the electricity demand is. They’d be able to both protect the grid from over-exhausting itself and ensure they have enough supply so there isn’t an electrical outage. Machine learning has also been a revolutionary tool in precise weather forecasts. As we decarbonize our grid and increasingly rely on solar and wind energy, those renewable energies depend on the weather. We’ll be able to precisely forecast how much electricity a certain wind and solar farm is going to supply.

Some other applications of machine learning in the field of climate solutions include the application of image recognition that plays a role in greenhouse gas detection. For my capstone project, I’m working with a startup on using satellite images to track methane emissions around the world.

Jennifer: Overlapping with the fashion, entertainment, and arts industry, there can be so much overproduction, especially with e-commerce. These forecasting algorithms can help optimize our production, reduce waste, and ultimately help our climate.

From your research, what recommendations would you make for the NeuralProphet model?

Jennifer: I think hyperparameter tuning. The model has the ability to be very malleable, but we came to the realization that we couldn’t build our optimal model as we anticipated with all the hyperparameters at once. When we specify all of them at once, there’s really a limitation there, and the model fails to perform as well.

Max: Talking about electricity demand forecasts, the cost of over-forecasting is not the same as the cost of under-forecasting. If you under-forecast that means you might not have enough electricity for this area which could shut down important facilities such as hospitals. There’s a significant economic, and also potential human costs to that. If NeuralProphet could incorporate not just accuracy, but also some sort of cost function that penalizes under-forecasting and less so penalizes over-forecasting, I think that would make this model more usable in the global context.

Did you have any other takeaways from the project?

Jennifer: We approached this project with the idea of testing out different machine learning algorithms and settled on a project, but at the same time we were delving into the unknown. I came into this program without a technical background so it felt intimidating. I would hope that it motivates others as well who come from non-technical backgrounds to delve into a passion project, even if you don’t know what you’re going to face.

Max: Both of us were outside our comfort zone. There are many times during the process when it seemed like we weren’t getting anywhere. I’m very grateful to have a friend as persistent and patient as Jennifer to stick with me as well as our mentor, Vishakh, who was very generous with his help and guidance. Going on an endeavor like this, even with the possibility the result might turn out differently than we hoped, the journey itself was very much worth it.

To learn more about the problem, method, and code, check out the project’s GitHub.

By Meryl Phair