"Powering" the Clean Energy Transition with AI at Greener
Greener Power Solutions, the new promising player in the clean energy transition, built a fun group of people to work with. We were approached by their team in the beginning of 2022 to help with our Machine Learning expertise, and contribute to their mission in reducing the carbon footprint in the Energy Market.
Driven by a bold ambition to work towards, the company was founded 4 years ago with an idea that never stops growing. Greener’s mission is to eliminate carbon footprint from temporary and emerging energy markets, which they currently do by introducing large batteries to the energy sector. The scope and the impact of the new company kept on growing throughout the years, as did their international reach and their produced data.
One of the main core business propositions of Greener, is to provide Smart Microgrid Controllers to the energy market to serve the power demands in events, off-grid configurations, and various other settings where an energy regulation or provision is needed. These smart batteries are interconnected, and can work in groups to serve the power demands. They also keep logs on the energy demanded or produced in each of these projects, along with their profiles and regimes, eventually becoming big time series data. This is where we came into the picture.
Within the Greener Dev team, Sam Samoud was our BigData Republic Ambassador to the group. Together, we worked on using that time-series data to build an energy demand forecasting module, that will allow Greener to strengthen its clean energy contribution.
By predicting what the energy demand will look like in a new event or festival, Greener will be able to deliver a more tailored power provision to their clients. Forecasting the output power of their batteries in the upcoming hours, increases supply reliability and avoids any unwanted power outages or bottlenecks. That’s why we made it our mission to create an R&D focus to bring live predictions of the power demands, using historic time-series data that was collected from previous projects.
The final results of our collaboration got great feedback from Greener. Together we were able to achieve important milestones. To get there however, we had some nice challenges to overcome.
The users of Greener’s batteries operate in various sectors. These final clients can be festival or concert venues, construction sites, sport venues, catering providers, industrial clients, the list goes on. These clients have different energy needs. They also have different power demand timespans. This means that time-series data collected from Greener’s projects have different individual characteristics and can span from a couple of days to a couple of months, depending on the case. This is different from the conventional forecasting applications, where you have one continuous time-series to train from. An example of combining data from these separate “episodes” into one big time series can look like the following:
One of the key challenges in time series forecasting is dealing with stationary data, which is the case here. This is when data exhibits no systematic changes over time. It can make it difficult to accurately forecast future values in the series, as there are no clear patterns or trends to follow. One particular difficulty with time series forecasting on stationary data is dealing with variable cycle lengths. A cycle is a repeating pattern in the data, such as the seasonal or daily fluctuations of power usage in one event. The length of a cycle can vary, however, making it difficult to accurately forecast future values in the series.
One approach to dealing with variable cycle lengths is to use Fourier analysis, which decomposes the time series into a series of sinusoidal functions with different frequencies.
This allows the individual cycles to be identified and forecasted separately. However, this approach can be computationally expensive and may not always provide accurate forecasts, particularly for short or noisy time series.
Another approach is to use machine learning algorithms, such as recurrent neural networks, which can learn to identify and forecast cycles in the data. This can be more efficient and accurate than Fourier analysis, but it requires a large amount of training data and can be sensitive to hyperparameter tuning.
The backlog of our team got structured in this way. Careful evaluation of the data and the prediction methods were carried out. The combination of the different approaches led to a successful now-casting model for the desired milestone. It brought more clarity and insights about the improvement points that can be implemented down the line.
Within Greener, Data represents a core element of the business, and we had the pleasure to bring our technical expertise to help shape the company’s Data Strategy. We can't wait to start the new year together and achieve even more wins.