Superlearner Models for HVAC Simulation Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning techniques for simulations, particularly in complex systems like HVAC control, are resource-intensive and time-consuming, often getting stuck in local minima, which slows down the optimization process and increases the time needed to find global solutions.
Innovation Solution
The implementation of 'superlearners' that learn from the operation of digital twins and agent-based solvers to generate training examples, train lightweight machine learning models, and use these models to quickly answer questions and optimize control actions, thereby speeding up the simulation process and improving the likelihood of finding global minima.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If resource intensive algorithms are used to answer simulation questions, then accuracy is improved, but time consumption increases
Solution Approach 1:
The system performs preliminary actions by training lightweight machine learning models in advance using training examples generated from digital twin simulations. These pre-trained models can then quickly answer simulation questions without requiring resource-intensive algorithms to run in real-time, thus resolving the contradiction between accuracy and time consumption.
Solution Approach 2:
The system creates simplified copies of the complex simulation process by generating training examples from digital twin outputs and using these to train lightweight models. The lightweight models are copies that replicate the answer-generation function of resource-intensive algorithms but with significantly reduced computational requirements, enabling fast and accurate answers.
2Reliability
If resource intensive algorithms are used to optimize control actions, then optimization quality is improved, but computational resources increase
Solution Approach 1:
The system replaces expensive, resource-intensive optimization algorithms with cheap, lightweight machine learning models. These lightweight models consume significantly fewer computational resources while maintaining optimization quality, effectively substituting high-cost computational objects with low-cost alternatives that achieve the same function.
Solution Approach 2:
The system performs optimization work in advance by training lightweight models on training examples generated from digital twin simulations. This preliminary training allows the models to perform optimization tasks quickly and efficiently during actual operation, reducing real-time computational resource requirements while maintaining optimization quality.
3Device complexity
If traditional machine learning models are used, then simplicity is maintained, but they get stuck in local minima
Solution Approach 1:
The system introduces digital twins as intermediaries that generate training examples for lightweight models. The digital twins simulate complex system behaviors and provide high-quality training data, enabling simple lightweight models to learn from realistic scenarios and avoid getting stuck in local minima while maintaining their simplicity.
Solution Approach 2:
The system performs preliminary simulation and training using digital twins to prepare lightweight models before deployment. This preliminary action allows simple models to learn optimal solutions from pre-generated training examples, enabling them to achieve high optimization accuracy without the complexity of resource-intensive algorithms during actual operation.
Data Source
AI summary
Various embodiments relate to a method, apparatus, and machine-readable storage medium including one or more of the following: using a resource intensive algorithm to answer a first question of a question type; generating at least one training example from the normal operation of the resource intensive algorithm; training a lightweight machine learning model based on the at least one training example to produce answers to questions of the question type; and using the lightweight machine learning model to produce an answer to a second question of the question type.