Automated Building HVAC Transfer Learning for Cold-Start Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automated buildings face a 'cold start' problem during commissioning due to the lack of historical data, leading to inefficient energy consumption and potential discomfort to residents from random HVAC adjustments, which existing machine learning techniques struggle to address effectively.
Innovation Solution
A method and system utilizing Transfer Learning (TL) to leverage data from an optimized automated building (donor building) to accelerate the optimization process of a newly constructed or retrofitted building by implementing an intelligent optimization agent with an Artificial Neural Network (ANN) and actor-critic deep reinforcement learning (DRL) to predict and adjust chiller set points.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine learning techniques relying on historical data are used, then predictive accuracy can be achieved, but the cold start problem causes significant delay in obtaining acceptable algorithm performance
Solution Approach 1:
The patent applies preliminary action by pre-training the machine learning model on aggregated historical data from multiple donor buildings before deploying it to the target building. This pre-training phase occurs in advance, allowing the model to start with pre-acquired knowledge rather than beginning from scratch, thus resolving the cold start problem and reducing the time needed to achieve acceptable predictive accuracy.
Solution Approach 2:
The patent uses copying by transferring the pre-trained model weights and parameters from donor buildings to the target building. Instead of collecting data de novo at the target building, the system copies the learned representations and knowledge from source buildings, enabling immediate predictive capability and significantly reducing the data collection timeframe.
2Productivity
If online optimization algorithms like Reinforcement Learning are deployed to address cold start, then immediate optimization can be achieved, but significant risk is incurred as the intelligent agent explores its action space causing discomfort to residents
Solution Approach 1:
The patent applies preliminary action by pre-training the reinforcement learning agent on simulated environments and aggregated data from multiple buildings before actual deployment. This pre-training phase allows the agent to learn optimal policies in advance without affecting real occupants, thereby eliminating the harmful exploration phase that would otherwise cause discomfort to residents while maintaining immediate optimization capability.
Solution Approach 2:
The patent uses an intermediary approach by introducing a pre-trained model as a mediator between the raw reinforcement learning agent and the building control system. The pre-trained model acts as a safeguard that guides the agent's actions, preventing harmful explorations that would cause occupant discomfort while still enabling the agent to learn and optimize performance.
3Quantity of substance
If data augmentation using preference maps is used to overcome cold start, then the dataset can be expanded, but the complexity of creating and maintaining aggregated datasets from multiple ACs increases
Solution Approach 1:
The patent applies universality by training a single multi-building model that can generalize across multiple buildings and HVAC systems. Instead of creating and maintaining separate aggregated datasets for each building, the universal model learns common patterns from diverse sources and can be deployed to any target building, reducing the overall complexity of data aggregation and model management while still providing sufficient training data volume.
Data Source
AI summary
A method and system for optimization knowledge transfer between automated buildings is provided. The present invention provides a system that uses transfer learning in conjunction with an intelligent optimization agent to transfer knowledge from an existing, optimized automated building, to a newly commissioning building, to reduce the adverse impact of the intelligent optimization agent's warm-up period. The method of the present invention allows transfer of knowledge gained by optimizing one automated building, to be transferred to another, newly constructed, automated building for which there is no historical energy use data. This allows the newly constructed automated building to reach optimal energy consumption much sooner, resulting in significant financial savings and a reduction in discomfort for building residents.