Skip to content

readme.md

Latest commit

 

History

History
116 lines (78 loc) · 4.24 KB

readme.md

File metadata and controls

116 lines (78 loc) · 4.24 KB
Raw

US Datacenter Power Demand Estimation Algorithm

Python 3.8+
This repository contains a simplified, self-contained implementation of a machine-learning-based algorithm to estimate the power demand of datacenters in the US electrical grid, considering both datacenter infrastructure and surrounding business contexts.

The model integrates server hardware characteristics, facility design parameters, business presence (manufacturing, technology hubs, retail, etc.), local grid infrastructure, and environmental conditions to accurately predict power demand with high confidence.

Features

  • Synthetic datacenter dataset generator calibrated to US state-level statistics
  • Single-facility power demand prediction with uncertainty estimates
  • Multi-model machine learning training pipeline (Random Forest, Gradient Boosting, Neural Network)
  • Neural Network achieves top performance:
    • ( R^2 \approx 0.96 )
    • MAPE ( \approx 12.6% )
  • Command-Line Interface (CLI) for data generation, model training, and making predictions
  • Basic publication-quality visualizations to analyze data and model results
  • Comprehensive test suite to ensure correctness
  • Easy configuration through config.json
  • Lightweight dependencies for quick installation

Quick Start

Install Dependencies

pip install -r requirements.txt

Generate Synthetic Dataset

python dcpower.py generate

Train Machine Learning Models

python dcpower.py train

Predict Power Demand for a Single Datacenter

python dcpower.py predict --servers 5000 --rack-density 15 --gpu 0.4 --pue 1.3 \
--sqft 200000 --cooling Liquid --state Virginia --grid-capacity 850 --temp 65

Run Unit Tests

pytest

Repository Structure

  • datacenter_power_predictor.py — Core algorithm and ML training code
  • data_utils.py — Utilities for synthetic dataset generation
  • train_models.py — Training and evaluation pipeline
  • dcpower.py — Command-line interface for all operations
  • visualization_utils.py — Plotting functions for visualization
  • test_datacenter_predictor.py — Pytest suite to validate functionality
  • config.json — User-configurable parameters for training and data
  • requirements.txt — All Python package dependencies
  • README.md — Project overview and instructions
  • setup.py — Installation script for pip-install

Research Highlights and Results

  • Machine Learning Accuracy: Neural Networks achieved an ( R^2 ) of 0.96 and MAPE around 12.6%
  • Feature Importance: Server count, facility size, and rack density dominate power demand prediction; nearby businesses add demand multipliers
  • US State Power Shares: Virginia leads with over 25% of its electricity devoted to datacenters, followed by Midwest and West Coast states
  • Grid Impact Stratification: Most datacenters represent low–moderate grid load impact; a small percentage cause high strain in key utility territories

Getting Help

For questions or issues:

License

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

Acknowledgments

Thanks to the U.S. Department of Energy, Lawrence Berkeley National Laboratory, and Energy Systems Integration Group for data and inspiration.

References

  1. INET-TUB Datacenter Study
  2. Thesis on Datacenter Energy Use
  3. NREL Report on Datacenters
  4. ScienceDirect Research Article
  5. IEA Report 2025
  6. DOE/LBNL Guidelines
  7. A2EI Data Release