Energy Price Forecasting

Verizon AI Studio Challenge – Fall 2024

---

Table of Contents

1. Business Focus
2. Exploratory Data Analysis (EDA)
3. Data Preparation and Validation
4. Modeling Approach
5. Key Findings and Insights
6. Challenges
7. Acknowledgments

---

Business Focus

This project tackles the challenge of predicting electricity prices across U.S. states and sectors to optimize Verizon’s energy management strategies. With over $1 billion spent annually on electricity, Verizon seeks a robust forecasting solution to enhance budgeting, improve decision-making, and reduce operational costs.

By building a time-series model, we aimed to:

• Forecast electricity prices by state and sector through 2030.
• Provide actionable insights into pricing trends to support strategic planning.
• Develop an interactive dashboard for easy visualization of predictions and trends.

Successful implementation will empower Verizon to anticipate price fluctuations, optimize energy procurement, and enhance operational efficiency.

Objectives and Goals

• Primary Objective: Develop a predictive model for electricity prices that can help optimize energy strategies and inform policy decisions.
• Goals:
  • Analyze historical energy pricing data through comprehensive exploratory data analysis (EDA).
  • Engineer features that capture key patterns, such as seasonal trends and sector-based distinctions.
  • Build and evaluate machine learning models to predict electricity prices accurately.

---

Exploratory Data Analysis (EDA)

• Identified correlations among key variables (sales, revenue, customers, price).
• Visualized distributions and trends across states and regions using tools like SweetViz, KLib, and Dabl.
• Highlighted invalid data points (e.g., zero customers with non-zero revenue) and addressed them.

---

Data Preparation and Validation

Dataset Description

• Source: Public U.S. electricity market data (2001–2024).
• Features: Prices (cents/kWh), sales, revenue, customers, states, and sectors (residential, commercial, industrial).

Data Preprocessing Steps

1. Handling Missing Values:
   • Imputed missing data in the customers column using mean imputation to preserve correlations.
2. Addressing Invalid Data Points:
   • Removed records with contradictory zero-value combinations (e.g., customers = 0 but revenue ≠ 0).
3. Standardization:
   • Normalized numerical features (price, sales, revenue, customers) using Min-Max scaling.
4. Feature Leakage Mitigation:
   • Excluded dependent variables (sales, revenue, customers) from model training to avoid leakage.
5. Feature Engineering:
   • Created a Season feature to capture temporal trends in energy consumption.

---

Modeling Approach

Selected Models

1. ARIMA:
   • Chosen for its simplicity and effectiveness in time-series forecasting.
   • RMSE: 11.5%.
2. SARIMA:
   • Tested but showed higher RMSE due to weak seasonality in the dataset.
3. Exploratory Models:
   • VAR and LSTM models were explored but deprioritized due to computational inefficiency and dataset limitations.

Validation Strategy

• Split the dataset into 200 subsets (50 states × 4 sectors) to ensure accurate predictions for regional and sector-specific data.
• Used train-test splits and cross-validation to assess model performance.

---

Key Findings and Insights

Model Performance

• ARIMA Model:
  • RMSE: 11.5%, making it the most reliable predictor.

Visualizations

• Heatmaps of correlations.
• Bar charts showing state-wise distributions of price, revenue, and sales.
• Forecasting plots comparing predicted prices to actual values.

Dashboard Highlights

• Interactive Visualization: Developed a Tableau dashboard with features such as:
  • Energy Price Mapping: State-wise electricity prices in cents/kWh.
  • Price Trends: Time-series analysis by sector (residential, commercial, industrial).
  • Price Rankings: Top states based on electricity costs.

Insights

• Linear Price Growth: Electricity prices are increasing linearly, driven by factors like infrastructure investments and market demand.
• Sector Variability: Residential sectors show higher pricing trends compared to commercial and industrial sectors.
• Regional Disparities: States like California and Alaska demonstrate significantly higher prices due to unique market conditions.

Potential Next Steps

• Incorporate external factors such as weather data or fuel prices to improve predictions.
• Experiment with deep learning models (e.g., LSTMs) for better time-series forecasting.
• Deploy the model as an API for real-time electricity price prediction.

---

Credits & Acknowledgments

We extend our gratitude to:

• Verizon AI Studio: For providing this challenge and the opportunity to contribute meaningful solutions.
• Team Members: Sneha Nangunoori, Pooja Ginjupalli, Kelly Chan, Emily-Ann Willix, and Annie Zhang, for their dedication and collaboration.
• Advisors: Nandini Proothi, Christian Winter, and Frankie Delgado for their guidance and support.
• Break Through Tech AI Program: For fostering this learning experience and connecting us with industry leaders.