Energy Saving Bear Limited (ESB) is a company that provides energy monitoring services to buildings: they install sensors, collect energy consumption data and display information in dashboard reports to inform customers. Their monitors not only offer a way to track building energy usage, but also to understand it - ultimately leading to targeted and quantifiable savings for building owners.
ESB was seeking to enhance its energy saving monitoring service by introducing a machine learning component, with the aim of forecasting hourly sensor consumption and triggering alerts if actual usage deviated from agreed thresholds. ESB felt that manually reviewing and commenting on the monitoring reports was time-consuming and expensive. They wanted to explore if machine learning could enhance analysis speed, improve appraisals, enhance alert accuracy and discover more energy-saving opportunities.
We were contracted to help develop the predictive machine learning model and review ESB’s current dashboard monitoring reports to improve usability.
Process
The project was funded by Innovate UK through its Business Growth initiative. As part of our brief we:
- Conducted a literature review to advise ESB on the machine learning methods to deploy.
- Undertook a thorough review of the dashboard reports generated by ESB to provide feedback on how these could be improved. Specifically, the client was interested to focus on “exception reporting”, so relevant changes in consumption could be reported promptly to clients for their quick response.
- Performed an analysis of the historical electricity consumption data from a collection of buildings. This data was then used to develop a machine learning model to predict energy consumption in the future.
Solution
The proposed solution involved constructing individual XGBoost models for each building using one month of historical data and retraining monthly. This approach balances model accuracy and system complexity while capturing intra-building correlations.
Key features of the model included:
- Model level: One model per building.
- Training data: One month of historical data.
- Retraining: Monthly updates for improved accuracy.
- Feature importance: Lagged consumption data was a key predictor.
This approach provided a solid foundation for predicting energy consumption at the building level.
Impact
By leveraging a supervised learning approach, our model was trained on historical energy consumption data with known outcomes. This enabled the model to establish a robust understanding of the relationship between input factors, such as time of day and past consumption, and the target variable: energy consumption. This learning process empowers the model to predict energy usage for new, unseen data, delivering valuable insights and supporting informed decision-making.
Tell us about your data challenge