Deploying AutoML for Effective Energy Balancing
-
Purpose:
A scalable AutoML platform has been developed to optimize energy management across various building types, such as residential, industrial, and public facilities. -
Autoregression Models:
The platform primarily relies on autoregression models to forecast energy production and consumption, utilizing historical data.-
Scalability: Autoregression was chosen to enable the platform to scale efficiently across thousands of buildings.
-
Simplification: By focusing on historical consumption patterns, the need for individual models for each building or complex external data integration is eliminated.
-
-
Building Types: The platform categorizes buildings into predefined types (e.g., small residential, industrial, office buildings, public buildings), each with a portfolio of pretrained models for energy production (solar panels), energy consumtion and battery storage optimisation.
-
Model Deployment:
-
When a new building is added, the AutoML platform automatically selects the best-performing model from the portfolio based on historical data and building type characteristics.
-
The selected model is then deployed to predict future energy consumption and production.
-
-
Continuous Monitoring:
-
The platform continuously monitors model performance, comparing real-time data with predictions.
-
If a model’s performance degrades, the system automatically selects and deploys a better-performing model from the portfolio.
-
-
Self-Sustaining System:
-
The AutoML platform operates without the need for manual coding.
-
Managed by data scientists and energy managers, the platform uses Azure tools to ensure it remains effective and scalable, maintaining high performance across all deployed models.
-
This AutoML platform represents a significant advancement in the field of smart energy management, offering a scalable, efficient, and automated solution for optimizing energy production and consumption. By leveraging autoregression models and a well-structured portfolio of algorithms, the platform ensures that businesses can effectively manage their energy resources across a wide range of building types. The continuous monitoring and automatic model updates further enhance the platform's reliability, making it a vital tool for sustainable energy management in an increasingly complex and dynamic energy landscape.
The Race Towards Understanding Energy Production and Consumption
With the current movement towards green energy, understanding and optimizing energy production and consumption is becoming increasingly important, especially for companies with diverse and complex energy flows.
On an energy production level, companies are starting to move towards a hybrid approach, where they are partly responsible for their own energy production through solar panels or wind turbines. However, these systems do not produce energy at a constant rate. Rather, the energy production depends on weather conditions like sunlight availability and wind speed, posing additional challenges in estimating and optimizing the amount of energy that will be produced.
On an energy consumption level, companies struggle with mapping their energy consumption. This becomes increasingly more complex for multi-site companies, especially with different types of sites. For example, a company with several office buildings and production plants, varying in size, may find it difficult to understand energy consumption patterns.
To make it even more complex, at the intersection of energy production and consumption, energy storage systems, like batteries, can be used to temporarily store energy. The idea behind storing energy comes from the fact that companies, with their fluctuating energy production and consumption, may produce more energy than they consume at certain moments. Rather than letting this energy go to waste or selling it to grid, companies temporarily store the excess energy in batteries. At moments in time where energy consumption surpasses energy production, companies can benefit from discharging their batteries rather than using more expensive electricity from the grid. However, this charging and discharging approach is not as simple as it sounds. There’s a wide range of available strategies for charging and discharging batteries and potentially selling energy to the grid, all based on specific scenarios and preferences.
Data-driven Decision Making for Sustainability
Companies can benefit from insights into energy production and consumption in many aspects. From optimizing operations, reducing costs, and minimizing environmental impact to better planning, improved sustainability practices, and compliance with environmental regulations, the possibilities are near endless. AI-driven energy production and consumption forecasting has proven to be one of the most important drivers for generating relevant insights. Being able to accurately predict both energy production and consumption contributes significantly to the aforementioned benefits. In today’s evolving landscape, companies are looking for an AI-driven energy production and consumption forecasting solution, taking into account the intricacies of complex multi-site configurations with variable production and consumption patterns as a consequence.
Faktion’s Approach towards Smart Energy Management
The integration of AI into various business operations has become a critical factor for achieving competitive advantages. At Faktion, recognising the potential of AI to transform the energy operations of businesses, we created a framework for gathering insights into energy production and consumption, allowing businesses to forecast, balance, and optimize both energy supply and demand for complex multi-site configurations in a scalable manner.
The solution takes the form of an AutoML (Automated Machine Learning) platform. An AutoML platform is a software system that is designed to automate the end-to-end process of applying AI to real-world problems. Such platforms aim to make AI accessible to any business user, allowing them to become AI operators, taking full ownership of the proposed AI solution.
The envisioned AutoML platform allows companies to create a portfolio of energy production and consumption forecasting algorithms, allowing to efficiently select and apply the appropriate forecasting model to specific plants, installations, or devices. Our approach extends beyond merely creating a repository for forecasting models models. The framework supports both the development and operational stages of AI applications. This holistic approach underscores our belief that the ability to monitor, operate, manage, and enhance energy production and consumption forecasting algorithms is as crucial as the initial development of these algorithms. This vision materializes through a dual-layered approach, consisting of a modelling platform and an operations platform, as outlined by the figure below.
Building Block 1: Modelling Platform
The envisioned modelling platform is designed for reproducible experimentation and efficient model building. The platform serves as the foundation for developing a portfolio of advanced forecast algorithms. It is equipped with all the necessary tools for seamless model operation and maintenance. Moreover, the flexibility of such a platform allows for the effortless integration of new or updated models.
The modelling platform takes data from three different sources. First of all there is asset-specific data concerning the configuration of energy producing and/or consuming assets (e.g. an office building with solar panels or an industrial site with a wind turbine). Besides that, a historical dump of energy production and consumption data is added for model training. External data sources such as energy market data or weather data can be added to extent the model. Essentially, a wide range of forecasting algorithms can be added and trained ranging from simple statistical algorithms, to machine learning models and neural network approaches.
A model selection component is added for smooth new asset onboarding. In essence, companies can add new configurations (or update old ones) to the system and by comparing the new asset’s characteristics with the ones in the asset database, similar asset configurations can be found and the best forecast algorithms are recommended based on similarity. After model training and evaluation, the best performing model for a specific asset is deployed and can now be used in real-time.
Building Block 2: Operations Platform
The operations platform is complementing the modelling platform and represents a solution for deploying and monitoring the forecasting algorithms from the modelling platform. This platform is the engine that drives energy production and consumption forecast, leveraging the forecasting models built on the AutoML platform through an API. It allows to provide real-time insights and control over the energy producing or consuming assets within a company, ensuring optimal performance and efficiency.
Moreover, the model monitoring component allows to track performance of the forecast. This feedback can be used in two different ways. On the one hand, the feedback can be looped back into the modelling platform for algorithm improvement or reselection. On the other hand, an analytics module allows to process the performance and extract relevant insights through dashboards.
Besides the forecasting algorithm portfolio, the AutoML platform allows to easily add additional features to the modelling platform and deploy them in the operations platform. Obviously, a lot of different features can be thought of to extend the AutoML platform and gain additional insights into energy production and consumption. Possible examples are:
-
Battery optimization models to take into account the presence of batteries for charging and discharging excess energy and optimize based on specific strategies.
-
Anomaly detection models to scan for patterns that are not expected, which could be traced back to suboptimal performance somewhere in the system (e.g. solar panels not producing enough energy or peaks in energy use as a result as a result of malfunctioning devices).
-
Pattern recognition models for identifying common production or consumption patterns to make informed decisions towards energy reduction strategies.
-
Drift detection algorithms to check for changes in energy production or consumption patterns and distributions, indicating the need for model updates.
Faktion’s Insights for Deploying AutoML Platforms
Having built AutoML platforms for a wide range of customers, at Faktion, we believe that by building an AutoML environment, scalability and manageability of algorithms is taken to the next level, enabling:
-
Continuous monitoring: with an AutoML platform, AI applications are not static. They are constantly monitored, and insights gained from their performance are used to make iterative improvements. This ensures that your AI solutions remain effective and relevant over time.
-
Operational efficiency: allowing to manage and operate AI applications seamlessly across various energy assets translates to significant operational efficiencies. It reduces the time and resources required to maintain these applications, allowing your team to focus on strategic initiatives.
-
Scalability: perhaps most importantly, an AutoML environment facilitates scalability. It provides the framework needed to expand the AI applications to new areas of your business / other asset (types) ensuring that the benefits of AI are realized across the board.
-
Internalized Capability: This platform empowers companies to internalize the capability of managing, monitoring, and scaling AI applications. Data specialists and energy managers can handle these tasks independently, without the need for ongoing involvement from ML or MLOps engineers, thus building long-term self-sufficiency in AI-driven energy management
The potential of building and deploying forecasting algorithms in an AutoML environment can reap many benefits for companies with regards to proper energy management of assets. In essence, it allows business users to take full ownership of the AI solution throughout its whole lifecycle, eventually reducing the need to rely on technical profiles or external partners for model management, leading to more efficient processes, cost savings and bottom line improvement.