Systems and methods for predicting occupancy for one building using a model trained at another building
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing building control systems face challenges in accurately determining occupancy without relying on costly and complex occupancy sensors, and methods that use access cards can be unreliable.
Innovation Solution
A method for training a model using environmental data from sensors in a training building to predict occupancy in a different use building, applying normalization factors to account for differences between buildings, allowing control systems to operate based on estimated occupancy values without requiring occupancy sensors in the use building.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If occupancy sensors are installed to determine occupancy, then occupancy data accuracy is improved, but system cost and complexity increase
Solution Approach 1:
The patent creates a virtual copy of occupancy sensing capability through machine learning models that replicate the function of physical occupancy sensors. The model is trained on environmental data from a training building with occupancy sensors, then deployed to a use building to predict occupancy without requiring physical sensors there, thus copying the sensing function through software rather than hardware.
Solution Approach 2:
The patent replaces the mechanical/physical occupancy sensor system with a computational machine learning model. Instead of using physical sensors to detect occupancy directly, the system uses environmental sensors (temperature, humidity, pressure) combined with a trained machine learning model to infer occupancy, substituting a mechanical detection system with an information-processing system.
2Measurement precision
If occupancy sensors are installed to determine occupancy, then occupancy data accuracy is improved, but system cost increases
Solution Approach 1:
The patent creates a virtual copy of occupancy sensing capability through machine learning models that replicate the function of physical occupancy sensors. The model is trained on environmental data from a training building with occupancy sensors, then deployed to a use building to predict occupancy without requiring physical sensors there, thus copying the sensing function through software rather than hardware.
Solution Approach 2:
The patent replaces the mechanical/physical occupancy sensor system with a computational machine learning model. Instead of using physical sensors to detect occupancy directly, the system uses environmental sensors (temperature, humidity, pressure) combined with a trained machine learning model to infer occupancy, substituting a mechanical detection system with an information-processing system.
3Loss of information
If access cards are used for occupancy data, then occupancy information is obtained, but reliability decreases in some situations
Solution Approach 1:
The patent replaces the mechanical/physical occupancy sensor system with a computational machine learning model. Instead of using physical sensors to detect occupancy directly, the system uses environmental sensors (temperature, humidity, pressure) combined with a trained machine learning model to infer occupancy, substituting a mechanical detection system with an information-processing system.
Solution Approach 2:
The patent introduces environmental parameters (temperature, humidity, pressure) as intermediary measurements that indirectly indicate occupancy. Rather than directly counting people or tracking access cards, the system measures environmental changes caused by occupancy and uses a machine learning model to translate these intermediary measurements into occupancy estimates, providing a more reliable indirect measurement method.
4Device complexity
If a model is trained using data from a training building and applied to a different use building, then the need for occupancy sensors in the use building is reduced, but accuracy may be affected by building differences
Solution Approach 1:
The patent adjusts the machine learning model by incorporating building-specific parameters such as building geometry, HVAC system characteristics, and environmental sensor locations when deploying from a training building to a use building. This parameter adaptation allows the model to account for differences between buildings while maintaining occupancy prediction accuracy without requiring retraining with occupancy sensor data from the new building.
Solution Approach 2:
The patent tailors the general occupancy prediction model to local building characteristics by incorporating site-specific parameters and adjusting model weights based on the unique features of each building. This local adaptation ensures that the model accounts for building-specific environmental responses to occupancy while maintaining the benefit of not requiring occupancy sensors in the new building.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A Building Management System (BMS) may be controlled in accordance with predicted occupancy using a trained model. A model is trained by providing the model with time stamped environmental data and corresponding time stamped occupancy data pertaining to a training building, wherein the time stamped environmental data is derived from one or more environmental sensors of the training building and the corresponding time stamped occupancy data is derived from one or more occupancy sensors of the training building. Once trained, the trained model is provided with time stamped environmental data for a use building that is derived from one or more environmental sensors of the use building. Occupancy data for the use building is not required. The trained model outputs a predicted occupancy value that represents a predicted occupancy count in the use building, and the BMS of the use building is controlled based at least in part on the predicted occupancy value.