A control device for controlling an HVAC system
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing HVAC systems face challenges in optimizing energy consumption and thermal comfort due to reliance on rule-based controls with default parameters, and data-driven approaches like reinforcement learning require extensive modeling and long training periods, limiting their market penetration.
Innovation Solution
A control device that combines ordinary differential equation (ODE) models with gradient-based optimization to improve prediction accuracy, integrating physics-based and data-driven components, allowing for efficient model-predictive and rule-based control strategies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If rule-based controls with default parameters are used, then the control system is simple to implement, but energy optimization potential remains untapped
Solution Approach 1:
The system automatically optimizes control parameters (such as supply air temperature, outdoor air ratio, and pre-cooling duration) by learning from historical operational data and environmental conditions, transforming static default parameters into dynamic optimized values that reduce energy consumption without requiring manual reconfiguration
Solution Approach 2:
The control system performs self-optimization by automatically analyzing its own operational data, identifying energy-saving opportunities, and adjusting control strategies autonomously, eliminating the need for continuous manual parameter tuning while capturing energy optimization potential
2Loss of energy
If model-predictive control with physics-based models is used, then energy consumption and thermal comfort can be optimized, but significant modeling effort is required for each new system
Solution Approach 1:
Instead of creating custom physics-based models for each HVAC system, the system uses copied templates of simplified thermal models that can be rapidly instantiated and automatically calibrated using historical operational data, dramatically reducing the modeling effort required for each new installation
Solution Approach 2:
The system introduces an intermediate data-driven calibration layer that bridges the gap between simplified models and complex physics-based models, using historical data to automatically adjust model parameters and achieve accurate predictions without requiring detailed system-specific modeling
3Loss of energy
If reinforcement learning is used, then purely data-driven control optimization can be achieved, but extensive training periods are required before convergence
Solution Approach 1:
The system performs preliminary action by pre-processing and feature-engineering historical operational data to create high-quality training datasets, and by implementing efficient training algorithms with appropriate regularization and hyperparameter tuning, enabling faster convergence of reinforcement learning models without requiring extensive real-time training periods
Solution Approach 2:
The system uses partial action by implementing a phased approach where a simplified model provides immediate control improvements, and the reinforcement learning model is gradually refined over time using online learning, achieving useful optimization before full convergence without requiring the complete training period
4Loss of energy
If night purge control is implemented to achieve substantial energy savings, then cooling energy demand is reduced, but adequate consideration of air handling unit and building behavior is required
Solution Approach 1:
The system performs preliminary pre-cooling of the building thermal mass during nighttime hours by controlling outdoor air intake and AHU operation, storing cooling capacity in the building structure itself, which reduces the need for active cooling during daytime while the system automatically optimizes the duration and intensity of pre-cooling based on predicted conditions
Solution Approach 2:
The system implements feedback control by continuously monitoring building temperature, outdoor conditions, and AHU performance during night purge operations, automatically adjusting control parameters to maintain thermal comfort while maximizing energy savings, and using this learned behavior to improve future night purge strategies
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to a control device (10) for controlling an HVAC system (12), wherein the control device (10) is configured to: provide a first model (18) using first model parameters giving first output variables as a function of first input variables; provide an ODE solver (24) configured to solve the first model (18) to calculate a determined trajectory of the indoor temperature and a determined trajectory of the return air temperature as a function of trajectories of the first input variables; define a training objective function determining a difference between the determined trajectories of the indoor temperature and the return air temperature and respective measured trajectories of the indoor temperature and the return air temperature; minimize the training objective function with respect to the first model parameters; and update the first model parameters in the first model (18).