Neural Networks: From Theory to Thermal Analysis 🏠

Model Description

This repository contains a series of trained feedforward neural network models for predicting heat influx in building designs. The models were developed using different architectures and training approaches to optimize performance for architectural applications.

Key Features

Multiple model architectures (1, 3, and 5 hidden neurons)
Different optimization techniques (SGD vs Adagrad)
Thoroughly validated performance metrics
Practical architectural applications
Comprehensive analysis tools

Model Details

Architecture: Multi-layer perceptron with sigmoid activation
Input Features: Insulation, East/South/North orientation
Output: Heat flux prediction
Best Model Performance:
- MSE: 0.002905
- R²: 0.958846
- Architecture: 5 hidden neurons with SGD optimizer

Training Data

The models were trained on a comprehensive dataset containing:

Insulation values
Building orientation parameters (East, South, North)
Corresponding heat flux measurements
Data was normalized using MinMax scaling

Usage

import tensorflow as tf

# Load the model
model = tf.keras.models.load_model('best_heat_flux_model.keras')

# Prepare input data (normalized values)
input_data = [[insulation, east, south, north]]

# Get prediction 
prediction = model.predict(input_data)

Input Format

All inputs should be normalized to [0,1] range using MinMax scaling:

Insulation: Building insulation value
East: Eastern orientation factor
South: Southern orientation factor
North: Northern orientation factor

Model Variants

Best Performing Model (best_heat_flux_model.keras)
- 5 hidden neurons
- SGD optimizer
- Learning rate: 0.1
- Momentum: 0.9
Alternative Models
- FFNN baseline (best_ffnn_model.keras)
- Adagrad variant (best_heat_flux_model_adagrad.keras)

Performance Analysis

Performance comparison of different architectures:

Hidden Neurons	Best Trial	Learning Rate	Momentum	Test MSE	Test R²
1	A	0.1	0.1	0.004940	0.897111
3	B	0.1	0.9	0.003501	0.950399
5	B	0.1	0.9	0.002905	0.958846

Applications

The models are particularly useful for:

Architectural design optimization
Heat efficiency analysis
Building orientation planning
Insulation requirement calculations

Limitations and Recommendations

Models are trained on normalized data - inputs must be scaled appropriately
Best suited for standard building configurations
Performance may vary for extreme or unusual architectural designs
Recommended to validate predictions against local building codes

Citation

If you use these models in your work, please cite:

@software{heat_flux_nn,
  title = {Neural Networks for Architectural Heat Flux Prediction},
  author = {Chan Meng},
  year = {2024},
  publisher = {Hugging Face},
  url = {https://huggingface.co./ChanMeng666/heat-flux-perceptrons-neural-networks}
}

License

This project is licensed under the Apache 2.0 License - see the LICENSE file for details.

Contact

For questions or feedback, please open an issue in the repository or reach out through Hugging Face.