This repository contains the analysis and modeling code for the project "Prediction of Relationships Between Energy Demand, Consumption, Air Temperature, and Extreme Weather Events". The project focuses on two primary goals:

1. Predicting Energy Grid Impacts during extreme weather events using Generalized Linear Models (GLMs) and non-parametric methods.
2. Establishing a Causal Relationship between increased electricity demand and average air temperature in California.

1. How can we predict impacts to the energy grid during extreme weather events?

   • Models were developed to identify "impacted" days (defined as days where energy demand exceeds the 90th percentile).

2. What is the causal effect of increased electricity demand on average air temperature?

   • An instrumental variable approach was used to identify causal relationships, accounting for confounding factors like seasonality and economic activity.

• Climate Data: Hourly weather observations aggregated to daily averages for Alameda County.
• Energy Data: Monthly energy demand and CO2 emissions from the EIA database.
• Preprocessing: Heat index calculation, feature engineering (e.g., binary heatwave indicators), and data aggregation.

1. Prediction Models

   • Logistic Regression (GLM)
   • K-Nearest Neighbors (KNN)
   • Decision Tree

2. Causal Analysis

   • Instrumental variable regression using temperature extremes as the instrument for energy demand.

• Accuracy, Precision, Recall, F1 Score, and AUC for prediction models.
• Coefficients and R² for causal inference regression.

• Prediction Models:

  • Logistic Regression achieved the highest AUC (0.954) and accuracy (92%).
  • KNN had the highest recall (56%).
  • Decision Tree had the lowest recall (37%) and showed signs of overfitting.

• Causal Analysis:

  • A causal coefficient of 0.1811 metric tons of CO2 per megawatt-hour of electricity demand was identified.
  • Temperature extremes proved to be a robust instrument for modeling energy demand.

• Limited data granularity (e.g., monthly energy and emissions data).
• Geographic and temporal aggregation reduced the specificity of results.
• Assumptions of uniform energy efficiency and proportional emissions may introduce bias.

• Incorporate higher-resolution and geospatially specific datasets.
• Explore advanced models such as RNNs for capturing temporal dependencies.
• Investigate regional variations and expand analysis beyond a single state.

• Python 3.8 or higher
• Required libraries: pandas, numpy, sklearn, statsmodels, matplotlib, seaborn

Nicholas Lemoff Keita Tanabe Areeya Tipyasothi Erica Ying