Anticipating Building Energy Consumption and Emissions

Overview

This repository contains the code and analysis for predicting CO₂ emissions and total energy consumption of non-residential buildings in Seattle using structural and contextual features. The project aims to support the City of Seattle’s goal of becoming carbon-neutral by 2050 by providing scalable predictive models that reduce the need for costly and time-consuming annual energy audits.

Project Structure

• Notebook_exploration.ipynb
  Exploratory Data Analysis (EDA) to understand feature distributions, missing values, correlations, and potential data leakage.

• Notebook_prediction_consumption.ipynb
  Model development and evaluation for predicting total energy consumption. Includes data preprocessing, feature engineering, model training, hyperparameter tuning (via cross-validation), and performance comparison across at least four algorithm families (e.g., ElasticNet, SVM, Gradient Boosting, Random Forest).

• Notebook_prediction_emission.ipynb
  Model development and evaluation for predicting CO₂ emissions. Explores the impact of including the ENERGY STAR Score as an additional feature and compares model performances.

• Presentation_of_analysis.pptx
  Slide deck summarizing key findings, methodology, and recommendations for integrating the predictive models into the City of Seattle’s workflow.

Data

The project uses the Building Energy Benchmarking dataset provided by the City of Seattle, which contains detailed 2015–present energy use and emissions data for non-residential buildings.

Source: Seattle Building Energy Benchmarking Data (2015–Present)

Key Features:

• Building size (e.g., gross floor area)
• Primary use type (e.g., office, hotel, retail)
• Year built
• Location (address, neighborhood)
• ENERGY STAR Score (when available)
• Annual energy consumption (kBtu)
• Annual CO₂ emissions (metric tons)

Methodology

1. Feature Engineering:

   • Extract structural features (size, year built, use type).
   • Geocode addresses to derive latitude/longitude and neighborhood-level indicators.
   • Compute transformations (log-scaling, normalization) to handle skewed distributions.

2. Model Training:

   • Split data into training and hold-out test sets to prevent data leakage.
   • Evaluate at least four different model families: ElasticNet, Support Vector Machines, Gradient Boosting, Random Forest.
   • Perform grid search with k-fold cross-validation for hyperparameter tuning.

3. Evaluation:

   • Compare models using appropriate regression metrics (MAE, RMSE, R²).
   • Analyze the incremental value of including the ENERGY STAR Score for emission predictions.

4. Results & Recommendations:

   • Summarize model performance and identify the best-performing approach.
   • Discuss potential deployment strategies and data requirements for future buildings.