https://github.com/sanjurajveer/water-usage

Deep learning model to predict water usage in muti-family properties in USA
https://github.com/sanjurajveer/water-usage

Last synced: about 2 months ago
JSON representation

Deep learning model to predict water usage in muti-family properties in USA

• Host: GitHub
• URL: https://github.com/sanjurajveer/water-usage
• Owner: sanjurajveer
• Created: 2025-06-26T21:15:55.000Z (4 months ago)
• Default Branch: main
• Last Pushed: 2025-08-10T09:51:02.000Z (2 months ago)
• Last Synced: 2025-08-10T11:40:39.029Z (2 months ago)
• Language: Jupyter Notebook
• Size: 3.72 MB
• Stars: 0
• Watchers: 0
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# Water Usage Prediction using Deep Learning

This project is part of a capstone initiative in collaboration with Connect IOT, focused on predicting water usage in multi-family residential properties using time-series data. The project utilizes advanced data preprocessing techniques and deep learning models to achieve high accuracy in forecasting.

## 🚀 Objective

To build a robust predictive model that can estimate hourly water usage across various units in a residential complex using features like temperature, unit size, time-of-day indicators, and historical consumption patterns.

## 📂 Project Structure

- `capstone_updated_2_part_ml.ipynb`: Jupyter notebook containing the entire pipeline including:

  - Data ingestion and preprocessing

  - Feature engineering (lag features, rolling statistics, time flags)

  - Categorical encoding (label + embeddings)

  - Binary classification to detect zero vs non-zero usage

  - Deep learning regression model for non-zero usage values

  - Evaluation and visualizations

## 🛠️ Technologies Used

- Python (Pandas, NumPy, scikit-learn)

- TensorFlow / Keras

- SMOTE (for class imbalance)

- Label Encoding and Embedding for categorical variables

- Matplotlib, Seaborn for data visualization

## 🧠 Model Workflow

1. **Data Cleaning & Feature Engineering**:

   - Created lag, rolling average, and time decomposition features

   - Encoded categorical variables using embeddings and label encoders

2. **Classification Task**:

   - Trained a binary classifier to predict if water usage = 0 or not

   - SMOTE used to handle class imbalance

3. **Regression Task**:

   - For non-zero predictions, a deep learning model (ANN) was trained to predict the actual water usage amount

   - Performance evaluated using MAE, RMSE, R², MAPE

## 📊 Key Results

- Achieved over **99% accuracy** in the classification task

- Regression model for non-zero usage delivered **low error rates** with robust generalization

## 📈 Future Improvements

- Incorporate weather API integration for real-time forecasting

- Use LSTM or TCN models for improved sequence modeling

- Deploy the model via Flask or FastAPI for real-time inference

## 🤝 Acknowledgements

Special thanks to:

- Connect IOT for real-world data and domain support

- UCD Capstone Program & Faculty Advisors

- Open-source contributors whose tools and libraries made this project possible

---