HVAC Transfer Learning for Fast Building Thermal Model Adaptation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Building management systems (BMS) face challenges in efficiently adapting HVAC control models to new buildings with limited historical data, as existing models require extensive retraining from scratch, often resulting in inadequate performance due to insufficient data.
Innovation Solution
Implementing a transfer learning method using a deep learning model, specifically a Long Short Term Memory Network based Sequence to Sequence (LSTM S2S) model, where a pre-trained model from a source building with abundant data is adapted using a smaller dataset from a target building, enabling efficient prediction and control of thermal dynamics and energy consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If a deep learning model is trained from scratch using limited historical data from a new building, then the model can be quickly deployed, but the prediction accuracy and control performance are insufficient
Solution Approach 1:
The patent applies preliminary action by pre-training the deep learning model on abundant data from a source building before deploying it to the target building. This pre-training establishes initial weight parameters that capture general HVAC thermal dynamics patterns, allowing the model to function effectively immediately upon deployment to the new building without waiting for extensive local data accumulation.
Solution Approach 2:
The patent implements copying by transferring the pre-trained model parameters from the source building to the target building. The model structure, architecture, and learned weight parameters are copied and adapted to the new building context, enabling the target building to leverage knowledge gained from the source building's historical data without having to generate it independently.
2Measurement precision
If extensive data is collected from the new building before model training, then prediction accuracy improves, but deployment time and operational delays increase
Solution Approach 1:
The patent performs the data collection and model training action in advance at the source building, creating a pre-trained model that can be immediately deployed to the target building. This eliminates the need to wait for extensive data accumulation at the target building, thereby maintaining high deployment speed while ensuring the model has sufficient training data from the pre-training phase.
Solution Approach 2:
By copying the pre-trained model parameters to the target building, the system transfers the predictive capabilities developed from extensive source building data to the target building instantly. This allows the target building to achieve high prediction accuracy without having to undergo its own lengthy data collection and training period, thus maintaining rapid deployment.
3Measurement precision
If a model is trained using abundant data from a source building, then the model captures accurate thermal dynamics patterns, but the model lacks adaptability to the specific characteristics of the target building
Solution Approach 1:
The patent applies local quality by fine-tuning the pre-trained model using a smaller amount of target building-specific data. This adjustment modifies certain weight parameters to account for the unique thermal characteristics, building envelope properties, and operational patterns of the target building, thereby adapting the general model to local conditions while preserving the knowledge gained from the source building.
Solution Approach 2:
The patent implements parameter changes by adjusting the weight parameters of the deep learning model during the transfer learning process. The model parameters are modified from their source building values to better fit the target building's specific characteristics, enabling the model to maintain accuracy while adapting to new building conditions through controlled parameter optimization.
Data Source
AI summary
A method includes collecting a first dataset of input-output data for a first building, training a deep learning model using the first dataset, initializing parameters of a target model for a second building using parameters of the deep learning model, collecting a second dataset of input-output data for a second building, training the target model for the second building using the initialized parameters of the target model and the second dataset, and controlling building equipment using the target model. Controlling the building equipment affects a variable state or condition of the building.