We spoke to REACH startup Builtrix in order to learn more about their deep learning-based energy disaggregation approach

Mojtaba Kamarlouei – CTO and co-founder at Builtrix

Today, Builtrix is a fast-growing provider of data-driven energy intelligence solutions for energy experts, consultants, and facility managers with large scale pilots and customers in Portugal, Italy, and Switzerland. Builtrix has a cloud-based solution that combines big data analytics and artificial intelligence (AI) with energy efficiency. Its’ solution enables facility and property managers, energy services providers and consultants to understand the energy consumption, analyze usage patterns as well as inefficiencies, and detect anomalies in real time.

How did you learn about REACH Incubator and what made you apply?

We know the REACH partners from years ago when we joined the EDI program in 2019. I should admit that EDI was one of our outstanding experiences with a lot of learnings around developing data-driven products and business. So, we have heard about the REACH program and its major goals from our EDI network. 

The main motivation for applying to REACH Incubator program was the challenge proposed by CERTH. This challenge is identical to one of our recent large scale pilot projects where we help our customer to reduce its cost intensive power peaks by disaggregating the central power and identifying the appliances contributing to the peak loads. Thus, putting our team into this challenge and using the business and technical resources provided by the incubator, we aim to not only develop a novel AI model but also a product and use case that can be scaled to a wider network of customers.

You are tackling CERTH’s challenge Deep Learning Methods For Building Energy Disaggregation. How do you think experimenting with and solving their data value chain challenge will set the foundations for scaling your solution?

With the expansion of smart buildings in the EU and global market, Virtual Disaggregation is becoming more appealing for the building and property owners. The main advantage of this technology is known as reducing the needs for so many IoT device installations which is a time consuming and costly process. As the name reflects, the technology promises to provide the state and consumption of main appliances in the building by disaggregating the central electricity smart meter data. Thus, the technology works like an enabler for smart meters and generates data for algorithms such as abnormality detection and demand response solutions.

With the advances of big data management systems and AI, the feasibility of Virtual Disaggregation technology is increasing. Although there is no widely adapted solution in the market, there are some innovative startups that offer their solutions to the residential and commercial sectors. The main drawback of these solutions is their adaptation to the new buildings.

Within the DeepEnD experiment, we are going to take our Virtual Disaggregation model into a higher level of accuracy and scalability by developing and embedding Transfer Learning (TL) approaches into deep learning models. This approach will increase the performance of the model in terms of both running time and accuracy.

I have to add that our data provider is not limited to CERTH. The Municipality Office of Cascais which is well-known as a green municipality in Portugal has also shared its data to be part of this challenge and test Builtrix innovation in its properties. Thus, we expect the experiment to help us in increasing the Technology Readiness Level (TRL) of our solution and at the same time help us to find more tractions with pilot results.

Please present your solution and elaborate on how it differentiates from the competition.

Builtrix’s solution for the proposed challenge is a deep learning-based energy disaggregation approach. We use state-of-the-art techniques for time series imaging (i.e., Gramian Angular, and Markov Transition Field) and then use convolution neural networks (CNN) for the deep learning of the images. Thus, the disaggregation process is performed based on the similarity of appliance’s behavior into the aggregated power images. This approach suggests better performance compared to the use of other conventional models such as Hidden Markov Models and Feedforward networks. During the experiment and as the main novelty, we include the TL approach to pre-train the CNN and transfer the knowledge from one use case to another one. This approach provides us with even higher performance and is expected to increase the scalability of the solution significantly.

Moreover, we are building the DeepEnD solution as part of Builtrix energy analytics platform where all the computation load-intensive modules are deployed on elastic computing units and based on docker containers with proper orchestration to guarantee the best practice of data ingestion, management, and analysis. This architecture enables Builtrix to scale the solution from 10 to 1000 buildings in a fraction of less than 3 days. Furthermore, the architecture is deployed using open source big data tools that can be deployed on any commercially available cloud services such as AWS, Google, or MS Azure.

Do you foresee any obstacles in successfully developing and commercializing your solution?

Of course, when we talk about machine learning (ML), the model performance is very much limited to the sample data. This has been a challenge for startups in the scaleup phase, with limited access to the data. Let’s say ML shows its best performance when the data comes from a limited world. In this regard, Virtual Disaggregation, it is very much like the self-driving car challenge. The training data is limited but the real-world data can be in any shape. This problem gets even worse when we talk about commercial buildings with various locations and embedded electrical systems.

In addition, at the same level of accuracy you reach for your ML model, the higher run-time you may need. This means, you may either need a huge computing resource to run your code for the scale of 10000 buildings or it may take months to show you the initial results.

In the case of both obstacles, the TL approach and proper time series imaging model can support ML to have better performance during the scaling phase.