Neural network methods for defining system topology
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural networks struggle to model complex building systems due to the difficulty in troubleshooting hidden layers and the computational complexity of managing multiple materials and interactions, making it challenging to create accurate thermodynamic models for buildings.
Innovation Solution
A neuron model creation system that discretizes structure elements into equal-length neuron strings, allowing for parallel processing and efficient representation of building components, including walls, ceilings, and air states, which are linked to form concatenated neuron strings.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional neural networks are used to model building systems, then the model can capture complex material interactions, but the hidden layers make troubleshooting difficult and the computational complexity increases significantly
Solution Approach 1:
The building model is segmented into multiple independent neuron strings, each representing a specific building component (wall, ceiling, floor, etc.). Each neuron string processes thermal interactions independently, allowing the complex building model to be divided into manageable segments that can be computed separately and then combined, reducing overall computational complexity while maintaining model accuracy.
Solution Approach 2:
The patent transforms the traditional 2D neural network architecture into a 1D neuron string structure. By organizing neurons into linear sequences that represent physical building layers, the model simplifies the computational dimensionality while preserving the ability to capture complex thermal interactions through the sequential arrangement of neurons representing different material layers.
2Reliability
If PID controllers are used to model each building component, then the thermodynamic behavior can be captured, but the number of controllers required becomes unmanageable
Solution Approach 1:
Multiple PID controller functions are merged into a single neural network-based neuron string. Instead of requiring separate PID controllers for each wall, ceiling, and floor component, the neural network integrates these control functions into unified neuron strings that process thermal interactions for entire building envelopes, dramatically reducing the number of controllers from dozens to a manageable few.
Solution Approach 2:
The neuron string structure serves multiple functions simultaneously: it models thermal conduction through materials, captures convective heat transfer, represents radiative effects, and performs control calculations. This multi-functional approach replaces the need for separate specialized controllers for each thermal phenomenon, reducing overall system complexity while maintaining thermodynamic modeling accuracy.
3Measurement precision
If detailed material layers are modeled individually, then the thermodynamic interactions are accurate, but the computational time required becomes excessive
Solution Approach 1:
The neuron strings are pre-configured with material properties and thermal interaction parameters during the model setup phase. By preparing the neural network structure in advance with all necessary material characteristics embedded in the neuron connections and weights, the actual thermal simulation requires only forward propagation calculations, significantly reducing computational time during runtime while maintaining accurate representation of detailed material layers.
Data Source
AI summary
A neural network in one embodiment is built by decomposing a structure into different building materials creating neurons that represent building materials and open spaces in a structure. Subsystems in the building have their neurons concatenated together to create same length neuron strings. In some embodiments, neurons in a short neuron string are split to make longer neuron strings. In some embodiments, neurons are added to some neuron strings to represent inside features, air features, and outside features.