Physics-Aware Neural Network Control for HVAC Energy Efficiency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional HVAC systems rely on PID control, which is inefficient due to unmeasured thermal parameters like wall temperature, leading to increased energy consumption and discomfort, and Physics Informed Neural Networks (PINNs) are challenging to implement for HVAC control due to high-dimensional input spaces and unknown exogenous inputs.
Innovation Solution
A method and system using a neural network trained with time series data, split into time slots with constant exogenous variables, to predict future system states and generate control signals, incorporating physics-informed loss functions to address unmeasured parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If PID control is used for HVAC systems, then ease of implementation is improved, but energy efficiency deteriorates
Solution Approach 1:
The patent replaces the traditional PID control mechanism with a neural network-based control system. The neural network is trained offline using historical data and physics-based thermal models, then deployed for real-time control decisions. This substitution maintains ease of implementation through simple firmware integration while dramatically improving energy efficiency by leveraging learned thermal dynamics and predicting optimal control actions.
2Measurement precision
If additional sensing parameters are deployed to improve control intelligence, then control accuracy is improved, but device complexity increases
Solution Approach 1:
The patent creates a virtual copy of the physical HVAC system through a neural network model trained on historical data and thermal physics. This digital twin captures the thermal dynamics and relationships between variables without requiring additional physical sensors. The model replicates the behavior of unmeasured parameters (like wall temperatures) based on available sensor data, achieving high control accuracy without increasing hardware complexity.
3Measurement precision
If Physics Informed Neural Networks are used to obtain unmeasured parameters, then measurement precision is improved, but difficulty of detecting and measuring increases
Solution Approach 1:
The patent performs preliminary training of the neural network offline using historical time-series data and physics-based thermal models. During this offline phase, the network learns to predict unmeasured parameters (wall temperatures, thermal loads) from available sensor data. Once trained, the model can be deployed in real-time with minimal computational overhead, making the complex physics-informed approach practical for implementation without requiring complex real-time solving during operation.
Data Source
AI summary
Use of Physics Informed Neural networks (PINNs) to control building systems is non-trivial, as basic formalism of PINNs is not readily amenable to control problems. Specifically, exogenous inputs (e.g., ambient temperature) and control decisions (e.g., mass flow rates) need to be specified as functional inputs to the neural network, which may not be known a priori. The input feature space could be very high dimensional depending upon the duration (monthly, yearly, etc.) and the (min-max) range of the inputs. The disclosure herein generally relates to Heating, Ventilation, and Air-Conditioning (HVAC) equipment, and, more particularly, to method and system for physics aware control of HVAC equipment. The system generates a neural network model based on a plurality of exogeneous variables from the HVAC. The generated neural network model is then used to generate the one or more control signal recommendations, which are further used to control operation of the HVAC.