DSI-DAB-smart-building-env-prediction

This repository contains the codebase for a collaborative research project between the Data Science Institute (DSI) and the Faculty of Design, Architecture and Building (DAB). The project focuses on optimising indoor environmental prediction in smart buildings using deep learning models.

Prerequisites

• Python 3.10
• Git

Setup

1. Clone the repository:

git clone https://github.com/your-username/DSI-DAB-smart-building-env-prediction.git
cd DSI-DAB-smart-building-env-prediction

2. Create and activate a virtual environment:

python3.10 -m venv env
source env/bin/activate  # On Windows, use: env\Scripts\activate

3. Install dependencies:

pip install -r requirements.txt

Project Structure

workspace/
├── Processed Datasets/     # Contains preprocessed CSV files
├── t/                     # Target variable specific folders
│   ├── Model Weights/     # Trained model weights
│   ├── Model Training History/  # Training history files
│   └── Performance Tracking/ # Model tracking
└── Evaluation/           # Model evaluation results
    └── Figures/          # Generated visualizations including Shapley plots

Running the Project

Training Models

The project uses Weights & Biases (wandb) for experiment tracking. To train models, run:

python experiment-run.py --device [cpu/gpu] --mixed-precision

This script performs a comprehensive grid search across multiple configurations:

• History lengths: 1 to 12 timesteps
• Prediction horizons: 1 to 12 timesteps
• Model architectures: CNN, LSTM, and CNN-LSTM hybrid
• Input types: univariate and multivariate

Options:

• --device: Choose between 'cpu' or 'gpu' (default: 'cpu')
• --mixed-precision: Enable mixed precision training for GPU (optional flag)

Evaluating Models

To evaluate trained models and generate performance metrics, simply run:

python experiment-test.py

This script will:

1. Process all models in the Model Weights directory
2. Calculate performance metrics (RMSE, MAPE, SMAPE)
3. Generate visualization plots
4. Save results to a CSV file in the Evaluation directory

Shapley Analysis

To perform feature importance analysis using Shapley values, run:

python shapley-run.py

Note: The Shapley analysis script has been specifically configured and verified for:

• History length: 12 timesteps (fixed)
• Prediction horizon: 1 timestep (fixed)
• Input type: multivariate only
• Model architectures: CNN, LSTM, and CNN-LSTM hybrid
• These parameters are currently hard-coded in the script

This script generates:

• Temporal Shapley value plots showing feature contributions across timesteps
• Feature importance bar plots based on mean absolute SHAP values
• Detailed statistics of feature impacts saved in CSV format
• Visualizations saved in the workspace/Evaluation/Figures/ directory

Output

The evaluation process generates several outputs:

• Performance metrics in CSV format
• Prediction plots comparing actual vs predicted values
• Training history plots
• Shapley value analysis plots and statistics (for multivariate models)

License

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

Contributors

• Roupen Minassian (Machine Learning Researcher, Data Science Institute)
• Adriana-Simona Mihăiţă (Associate Professor, Data Science Institute)
• Arezoo Shirazi (Senior Lecturer + Course Director, Faculty of Design, Architecture and Building)

Contact

For questions or collaboration opportunities, please contact Roupen dot Minassian @uts.edu.au