Adaptively learning surrogate model for predicting building system dynamics from simulation model
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep neural network (DNN) models require a large amount of historic operational data to train, leading to delays in applying predictive models for building system dynamics, such as HVAC systems, resulting in lost energy savings and increased computational resources needed for training.
Innovation Solution
A method involving a calibrated simulation model to generate simulated data for initial training of a surrogate model, followed by re-training with updated operational data, reduces the time and computational resources required for developing accurate predictions of building system dynamics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional deep neural network models are trained using large amounts of historic operational data, then prediction accuracy is improved, but training time and computational resources increase significantly
Solution Approach 1:
The patent applies preliminary action by pre-training the neural network model using simulated data generated from a building simulation model before deploying it with real operational data. This allows the model to learn basic patterns and dynamics in advance, reducing the amount of real historical data needed and the time required for final training and deployment
2Measurement precision
If conventional deep neural network models are trained using large amounts of historic operational data, then prediction accuracy is improved, but computational resources required for training increase
Solution Approach 1:
The patent introduces an intermediary approach by using a building simulation model as a mediator to generate synthetic training data. This simulated data serves as an intermediate training source that reduces the computational burden of training on large volumes of real operational data while still achieving accurate predictions
3Loss of time
If a simulation model is calibrated and used to generate simulated data for training, then training time is reduced, but model calibration complexity increases
Solution Approach 1:
The patent applies copying by creating a simulated copy of the building system through a simulation model that replicates the actual building's dynamics. This virtual copy can be calibrated and used to generate unlimited training data without the complexity of collecting and processing large amounts of real operational data
Data Source
AI summary
Systems and methods for training a surrogate model for predicting system states for a building management system based on generated data from a simulation model are disclosed herein. The simulation model is calibrated for a building of interest. The building of interest includes building equipment operable to control a variable state of the building. The simulated data of system states are generated using the calibrated simulation model. A surrogate model is trained based on the simulated data of system states from the calibrated simulation model. System state predictions are generated using the surrogate model. The surrogate model is re-trained based on updated operational data. An updated series of system state predictions is generated using the re-trained surrogate model.