AI-Powered Server Efficiency Optimization

This project presents a Python-based AI model designed to enhance server efficiency for Microsoft data centers. The model is trained using cpu-only and considers various real-time environmental, operational, and human activity factors to make smart, energy-efficient decisions.

---

🧠 AI Case Study Overview

In large data centers, energy consumption and cooling costs are critical factors. We built an AI model that:

• Monitors real-time metrics
• Makes dynamic decisions
• Learns over time
• Controls server operations and environment adaptively

The model was trained on simulated and real-world-style datasets, taking into account various influencing factors.

---

📥 Input Features (Used for AI Training)

Category	Input Feature	Description	Example Value
Server Metrics	CPU Usage	Current server workload	75%
Cooling Demand	Internal Temperature	Inside data center temperature	30°C
Weather	External Temperature	Outside air temperature	25°C
Humidity	External Humidity	Affects cooling efficiency	60%
Power Usage	Current Power Draw	Total energy usage now	120 kW
Renewable Energy	Solar/Wind Availability	Renewable % available now	40% solar
Electricity Cost	Real-Time Grid Price	Cost of electricity per kWh	$0.15
Human Activity	Occupancy Count	Number of people inside	10
Time Context	Day & Time	Time of day and weekday/weekend	Monday, 3 PM
Server Requests	Active Users/Requests	Number of active users on systems	2500
AI Usage	AI Task Load	Is there a big AI training job coming?	Yes

---

🎯 AI Model Outputs (Decisions & Predictions)

Output	Description	Example Decision
Workload Scheduling	Move tasks to energy-efficient time slots	Move backups to 2 AM
Cooling Adjustment	Control AC based on server & room temperatures	Reduce cooling at night
Energy Source Choice	Pick between grid and solar/wind energy	Use solar during daytime
Power Distribution	Smartly shift loads across zones/servers	Shift AI load to cool zone
Forecast Demand	Predict usage spikes or drops	Spike at 6 PM → prep cooling
Optimize Cost	Reduce energy cost without hurting performance	Turn off lights in empty zones

---

🧪 Model Training

• Training Environment: Google Colab / Local
• Language: Python
• Libraries Used: Scikit-learn,Tensorflow, Pandas, Numpy, Matplotlib, Seaborn
• Model Type: Multi-output Decision/Regression Model (e.g., Random Forest or Neural Network)
• Learning Strategy: Continual learning support with incremental updates

---

📊 Dataset Creation

To train the model effectively, we generated a dataset combining real-world-inspired synthetic values, including:

• Server logs (CPU usage, power draw, etc.)
• Environmental sensors (temperature, humidity)
• Occupancy simulation
• Time-based patterns (weekdays vs weekends, peak hours)
• AI usage simulation (e.g., workload surges during training jobs)

---

Model Overview

The DataCenterModel is a multi-output neural network designed for optimizing data center operations using deep learning. It utilizes TensorFlow/Keras to make multiple simultaneous predictions for various aspects of data center management.

Model Architecture

• Type: Multi-output Neural Network
• Framework: TensorFlow/Keras
• Architecture Type: Feed-forward Neural Network with branching outputs

Layer Structure

1. Input Layer

   • Dimensions: (None, 12) - 12 input features

2. Shared Layers

   • Dense Layer 1: 256 neurons, ReLU activation
   • Batch Normalization
   • Dropout (0.2)
   • Dense Layer 2: 128 neurons, ReLU activation
   • Batch Normalization
   • Dropout (0.2)

3. Output Branches (6 parallel branches) Each branch contains:

   • Dense Layer: 64 neurons, ReLU activation
   • Dense Layer: 32 neurons, ReLU activation
   • Output Layer: 1 neuron (linear activation)

Training Parameters

• Optimizer: Adam
• Learning Rate: 0.001
• Loss Function: Mean Squared Error (MSE)
• Metrics: Mean Absolute Error (MAE)
• Batch Size: 32
• Epochs: 100
• Validation Split: 20%
• Early Stopping: Yes (patience=10)

Data Preprocessing

• Feature scaling using StandardScaler
• Label encoding for categorical 'Day' feature
• Target variable scaling using StandardScaler

Model Features

1. Continued Training Support

   • Option to continue training from previous state
   • Model checkpointing capability

2. Serialization

   • Model saving/loading functionality
   • Scaler persistence for consistent predictions

3. Custom Loss Functions

   • custom_mse: Custom Mean Squared Error
   • custom_mae: Custom Mean Absolute Error

Usage Guidelines

1. Initialization

   model = DataCenterModel(load_previous=False)

2. Training

   history = model.train(data_path='data_center_dataset.csv', 
                        epochs=100, 
                        batch_size=32)

3. Prediction

   predictions = model.predict(input_data)
   model.interpret_predictions(predictions)

Decision Thresholds

All outputs use a 0.5 threshold for binary decisions:

• Values > 0.5: Positive action recommended
• Values ≤ 0.5: No action needed

Performance Considerations

• Uses batch normalization for training stability
• Implements dropout (0.2) for regularization
• Early stopping to prevent overfitting
• Scalable architecture for varying data sizes

---

📂 Folder Structure

/
├── model/
│   ├── data_center_dataset_1200.csv
|   ├── trained_datacenter_model_scalers.pkl
│   └── trained_datacenter_model.h5
|
├── enhanced_model.py
├── model_evalution.py
├── dataset_maker.py
|
├── datasets/
│   ├── data_center_dataset_10000.csv
|   ├── data_center_dataset_200000.csv
|
├── images/
|   
└── README.md

---

🔮 Future Enhancements

• Predictive maintenance module
• Integration with computer vision for occupancy estimation
• Energy credits and billing optimization
• Visual dashboards for admin decision-making

---

📧 Contact

For contributions, questions or collaboration:
Antarip Kar
📧 akantarip30@gmail.com

---

✨ This project is a case study given by TKS