HydroSense: A Machine Learning-Based Water Level Prediction and Water Resource Management System for Chennai Reservoirs

HydroSense is an end-to-end machine learning project aimed at predicting the 7-day future water levels of Chennai's four major reservoirs (Poondi, Cholavaram, Red Hills, Chembarambakkam) in India, and providing decision support for urban water resource planning through an interactive web application. Addressing Chennai's water crisis, the project combines 16 years of historical hydrological data to help water authorities identify water supply risks in advance and optimize scheduling decisions.

As India's fourth-largest city, Chennai relies heavily on its four major reservoirs for water supply, but has faced frequent water crises in recent years (such as the 2019 "Day Zero" crisis). The causes of the predicament include: unstable monsoons, rising water demand due to rapid urbanization, extreme weather caused by climate change, and lack of data-driven support in traditional management. The HydroSense project was created to address these issues, enabling accurate water level prediction and risk early warning.

The project uses daily data (including water levels and rainfall) from 2004 to 2020 provided by the Chennai Metropolitan Water Supply and Sewerage Board (CMWSSB). Feature engineering strategies include:

• Aggregated features: Total water level (TOTAL_LEVEL), total rainfall (TOTAL_RAINFALL)
• Lag features: Total water level 7/30 days ago
• Rolling statistics: 7/30-day rolling average rainfall, 7-day rolling standard deviation of water level
• Time encoding: Sine-cosine encoding of month/day, calendar features like year/quarter/day of the year

The project compared 9 models (Linear Regression, Ridge/Lasso, Decision Tree, Random Forest, Gradient Boosting, Extra Trees, XGBoost, SVR, etc.). Training used a time-series-friendly split (first 80% for training, last 20% for testing, no shuffling), with evaluation metrics including MAE, RMSE, and R². The best-performing model was selected and saved, and it was found that lagged water level features and rolling average rainfall are key predictors.

The application is developed based on the Streamlit framework, with core features including:

1. Parameter input: Enter current rainfall, historical water levels, etc. via sliders
2. Real-time prediction: Calculate water levels 7 days later
3. Visualization: Display predicted values, capacity ratio, and risk levels (Critical/Low/Medium/Good/Full)
4. Demo mode: Interface can still be displayed without model files

Usage: After installing dependencies, run streamlit run app.py to access the local address.

The project's prediction target is the combined water level of the four major reservoirs over the next 7 days (unit: MCft), with a total capacity benchmark of approximately 320 MCft. Key achievements include: selecting the best model via time-series test sets, and identifying lagged features and rolling average rainfall as the most important predictors. This project is a course project for III AIML-A (Batch 2028), with a tech stack including Python, Scikit-learn, XGBoost, and Streamlit.

Future expansion directions for the project:

• Data layer: Integrate real-time APIs, fuse satellite remote sensing/groundwater/meteorological data
• Model layer: Try deep learning models like LSTM/Transformer, multi-output prediction, uncertainty quantification
• Application layer: Develop mobile PWA, cloud deployment, integrate SMS/email alerts

HydroSense demonstrates the application value of machine learning in infrastructure issues, transforming into a practical tool through a complete process (data exploration → feature engineering → model training → deployment). It serves as an excellent end-to-end example for learners and provides a reproducible framework for practitioners. More importantly, it reminds us that technology should serve people and contribute to sustainable water resource management.