In mid-March I attended a talk organized by the ClimateLink SF Meetup, titled “Rise of the Machines: AI for Energy Management.” The three presenters were Daniel Lim, Co-Founder at Crocus Energy, Pratik Parekh, Sr. Data Scientist and PM at Bidgely, and Andrew Byrnes, Investment Director with Comet Labs. In this post, I’ll summarize the main points of the talks and their connections to data science.
Crocus Energy
First up was Daniel Lim. Crocus Energy is an industrial & energy AI software firm, focusing on disrupting the commercial and industrial sector. Crocus is building a new energy management platform that uses machine learning and advanced analytics to understand how electricity is delivered to every device in large, energy-intensive facilities. Their goal is to increase a facility’s energy efficiency. Their website says they can help reduce electricity use by up to 4%.
Daniel explained that the grid was not built for demand optimization, and the energy supply has to be at least as high as the peak. Here’s how we can visualize this:
Crocus Energy takes the approach of analyzing historical demand and optimizing supply for a given facility. They call their product Power Saver™, and here is how the new time series of supply and demand looks:
In more detail, each facility has a different configuration of loads on the grid, and needs to determine the voltage set point for a given time period. Complicating the situation, there are voltage differences at different outlets, especially at the end of the line, which is often degraded. Further, machinery doesn’t work outside narrow voltage specifications. All of this provides an opportunity for ML to help, by learning the demand patterns typical of a given facility to optimize energy supply.
Daniel didn’t go into detail about the ML techniques. Briefly, Crocus Energy computes usage statistics over a time window, and then creates a weighted sum of ensembled predictions at different time periods. They then monitor the prediction error and retrain the models as needed. There is also a user in the loop, defining acceptable thresholds for operation. Some of the data science challenges include noisy data at the metering points, and reading thousands of data points while needing a fast response.
In the end, operators control the load with the help of output from the models. Customers can configure the predictions to occur at 10, 30, or 60 minute intervals. Crocus partners with existing infrastructures and electrical device control manufacturers for hardware control.
Bidgely
The next speaker was Pratik Parekh of Bidgely, an energy utility AI solutions provider. They partner with utilities and other energy providers, helping them leverage customer data to improve the customer experience and generate new revenue streams. As such, they are both a B2B and a B2B2C company.
Their main AI advantage is energy disaggregation, sometimes called nonintrusive load monitoring (NILM). This is the ability to take energy meter data and identify which appliances and devices are running during which time periods. This in turn can be used to compare usage patterns to identify which appliances might be running inefficiently or when someone is on vacation and might be able to take advantage of special rates. Utilities could leverage this information to improve their demand response programs or to offer sales on EV chargers or LED light bulbs, gaining revenue through partnerships.
Bidgley holds over a dozen patents on disaggregation, and is able to use meter data alone to identify appliance use, no sensors in the home are required. They retrieve data from three types of meters: HAN (home area network), which provides updates every 1-60 seconds, smart meters, updating every 15, 30, or 60 minutes, and older meters which are read manually at between daily and monthly intervals. Bidgley obtains ground truth data by buying it from third parties – I looked around on the web and found some disaggregation datasets.
Here’s a way to visualize the load data from a smart meter:
Here there are some easily identifiable periodic pulses such as the refrigerator or water heater:
Some other patterns are more difficult to spot, such as lighting or the “always on” baseload:
Pratik didn’t go into any detail on the techniques used for disaggregation, but a recent survey indicates that modern techniques include hidden Markov models, graph signal processing, and deep learning. In the end, the system provides information on devices identified, their kW load and the time period they were on, and a classification of “efficient” or “inefficient.” Here’s a visual example from their web page captured on 2019.04.12:
Comet Labs
The final presenter was Andrew Byrnes, an investor with Comet Labs. I rarely attend technical Meetups that include speakers from funding organizations, so I enjoyed the opportunity to hear from Andrew. Comet Labs focuses on AI startups that are “transforming foundational industries.” Andrew focuses on energy efficiency, for three motivating reasons. First, change: the current energy infrastructure is big and static. He sees entrepreneurs using automation to solve problems of change, such as EV charging or distributed energy. Problems of change provide great opportunities for data science approaches. Second, money: some energy is expensive, some is very cheap. He advised that we should build solutions that asset owners will pay for; he personally looks for ideas that over time will improve people’s lives. Third, reduce: energy has an emissions problem, we need to always be solving for lower reduced emissions. As someone passionate about reversing global warming, I couldn’t agree more.
During the question period, he mentioned that Comet Labs focuses on “series seed,” to cover the gap that often happens before Series A but after seed or angel funding. He also mentioned that people may not have heard of many of the companies they invest in because they are B2B or industry-focused companies, not B2C. He’s also very interested in EV acceleration and innovation in developing countries.
In conclusion, I was excited to learn about these three initiatives making impact on energy efficiency using data science and AI. What companies are you keeping an eye on?
Photo by Patrick Tomasso on Unsplash
