Predictive Cooling Optimizer for Data Centers: Temperature-Aware Chiller Scheduling to Cut Energy Use

The project addresses energy inefficiency in data center cooling systems by developing a temperature-aware predictive model that optimizes chiller scheduling to reduce energy consumption while maintaining thermal safety. Traditional reactive cooling systems respond to temperature changes after they occur, leading to energy waste and suboptimal chiller operation.

Project Methodology

The methodology began with comprehensive data preprocessing of 13,615 HVAC samples, including outlier detection using IQR, normalization via MinMaxScaler, and chronological 80-20 train-test splitting to preserve temporal integrity. Feature engineering created 46 enhanced features encompassing lag features (16), rolling averages (12), cyclical temporal encodings (6), and interaction features (4), capturing complex system dynamics.

Two XGBoost regression models formed the core prediction engine: an Energy Prediction Model achieving R² = 0.9891 with MAE of 1.222 kWh, and a Temperature Forecasting Model achieving R² = 0.6853 with 89.24% predictions within ±1°C tolerance. Both models demonstrated efficient training times (2.12 seconds for energy, 1.87 seconds for temperature) suitable for real-time deployment.

A PredictiveCoolingOptimizer class integrated both models, enabling system-wide optimization through constraint-based temperature management and energy minimization strategies.

Testing

There are a total of 11 tests conducted under five categories. Here's the details:

Tests	Target	Status
Unit Tests	Energy and Temperature models, Optimization Engine	✅ Passed
Integration Tests	End-to-End Pipeline, System integration	✅ Passed
Functional Tests	Accuracy, Response Time and Logic	✅ Passed
White Box Test	Hyperparameters, Feature Engineering	✅ Passed
Black Box Test	Boundary Values, Output Consistency	✅ Passed
	Tests Passed	11/11
	Tests Failed	0/11
	Success Rate	100.0%

Procedure for Execution

The execution of the program is as simple as it can get. Here's a step-by-step procedure on how to do so.

1. Install necessary libraries:

pip install xgboost streamlit

2. Navigate to the project folder in Command Prompt/Powershell:

cd <your-file-path>

3. Run the application using this command:

streamlit run 1_Home.py

Tools used

1. Anaconda Jupyter (for model training and testing)
2. Streamlit library (for frontend implementation)
3. Joblib (for handling of .pkl model files)
4. Numpy
5. Pandas
6. Sklearn
7. XGBoost

Algorithms Used

1. Prediction Algorithms

• XGBoost (Extreme Gradient Boosting)
• Random Forest Regressor

2. Supporting Algorithms

• Min-Max Normalization
• Rolling Average (for feature engineering)

The tool successfully demonstrates feasibility for software-based predictive cooling optimization, with models ready for deployment in interactive web applications using Streamlit for user-freindly interface and stakeholder presentation.

License

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