The Arundo team travels to the GE Foundry in Paris after winning the GE Predix Hackathon in October 2016.
A week with the GE Digital EU Foundry in Paris to uncover power optimization and scheduling with the Predix platform.
By Piotr Janiszewski, Data Scientist (with contributions from Tucker Cowie, Software Engineer)
OSLO, NORWAY – Following the fantastic performance of the Arundo team that won GE’s Minds + Machines Hackathon back in November 2016, Arundo was invited to GE Digital Foundry in Paris to further develop the winning Power Scheduler application. During a one week workshop in early March, the team confirmed requirements with industrial users, defined several use cases and addressed them with a Predix application built from scratch.
Early week brainstorming developed scenarios around the previously explored case of power scheduling. For example, an oil rig operator who only has access to power that is produced onsite must optimize the consumption schedule given power availability constraints. In other cases where a large industrial power consumer is charged by an electricity provider based on daily peak load, operators may be interested in spreading their load over the day to reduce peak load. After exploring some of these use cases, we honed in on a simple subset that was reasonable to deliver on within the week.
The team lineup remained mostly unchanged from the original Hackathon team. Ting Chen, Tucker Cowie and Martin Lee were responsible for full stack development, Jeff Jensen for architecture and Stuart Morstead for business insight. Jeff’s electrical engineering background came in very handy. I joined the team to strengthen it on the data science side. GE invited several internal stakeholders to provide context for the workshop. Among them were representatives from GE Grid Solutions, DotVision, a producer of devices for collecting electrical power data, and facility managers of a large office building in Paris. The mix of Arundo’s expert team and GE’s industrial stakeholders led to insight-generating discussions that further focused out objectives.
In the workshop, we worked with power consumption data from a large office building in Paris. The dataset contained measurements of power, voltage, current, power factor from about 20 three-phase circuits sampled at 0.2 to 1 Hz frequency, and more. Approximately 80M data points were collected per day and continuously streamed into a Predix time series database.
Working with industrial data rarely comes without challenges and we quickly discovered that some of the nearly 500 signals had frequent time periods without any data ﹣ something traced to a dropping network connection. This gave rise to our first use case which was to provide a facility manager with a quick dashboard-like overview of data quality. After receiving this feedback during the week, the team that installed the sensors was able to react quickly and fix the problem to enhance the data quality.
The second use case was to boil down the multitude of available signals to a single measure of power usage efficiency for each circuit. The measure was based on a combination of power usage and power factors across the phases. The power factor indicates, to an extent, how much power is being used versus how much power is actually being supplied. By looking at the combined power factors across multiple phases, we can determine a general efficiency across a larger sets of electronics. This provides a very quick high level overview of where inefficiency may exist in a system. This is important as it can enable identification of 1) poorly performing equipment and 2) potential danger to equipment from reflective power being pushed into circuits from that may in fact damage all equipment.
Finally, we worked on a scheduler application that would reduce power consumption to a predefined level. As mentioned, office buildings provide few levers to control power consumption. We assumed, however, that each circuit has some flexibility in power consumption, e.g. thermostats can be set 1° higher, lights can be dimmed a notch, or the EV charger can operate at a reduced power. Assuming the facility manager can set limits on these flexibilities, we developed an algorithm that first predicts power usage for the next 24 hours and then suggests time variable power reductions on individual circuits to meet the target power level
Our MVP solution contained three main segments. For the front end, we implemented a dynamic website with several data visualizations. Secondly, we also deployed a set of idiomatic REST APIs to GE Predix for Arundo Enterprise to consume; this ultimately means any HTTP app could employ the services we wrote. Finally, for the core of the system architecture, we deployed Python microservices containing the algorithms to compute data quality, power efficiency, and optimum power scheduling. By separating the application logic from the algorithms, the team could work efficiently in parallel and deliver on time without disturbing each other.
All in all, it was a busy week, but there is great satisfaction in delivering a working product in a compressed timeframe AND we were able to celebrate building the first interoperable Arundo - GE Predix application in that our models are modular and transportable across platforms!