24-Hour Electricity Demand Forecasting

A fast-track electricity demand forecasting solution for Indian cities using smart meter data and weather information.

Overview

This project forecasts the next 24 hours of electricity demand for Bareilly using historical smart meter readings and weather data. It processes 3-minute smart meter data into hourly forecasts using Ridge Regression with engineered features.

Features

• Data Processing: Converts 3-minute smart meter readings to clean hourly data
• Weather Integration: Fetches weather forecasts from Open-Meteo API
• Multiple Models: Seasonal naive baseline + Ridge Regression with engineered features
• Uncertainty Quantification: Provides 10th, 50th, and 90th percentile forecasts
• Daily Calibration: Aligns forecasts with recent consumption patterns
• Automated Reporting: Generates PDF reports with plots and metrics

Requirements

Install dependencies using:

pip install -r requirements.txt

Main dependencies:

• pandas
• numpy
• scikit-learn
• requests
• matplotlib
• fpdf

Quick Start

Run the complete forecast pipeline with a single command:

python run_forecast.py --city Bareilly --history_window days:7 --with_weather true --make_plots true --save_report true

Command-line Arguments

• --city: City name (Bareilly or Mathura)
• --history_window: Training data window (format: days:7)
• --with_weather: Include weather data (true/false)
• --make_plots: Generate visualization plots (true/false)
• --save_report: Create PDF report (true/false)

Project Structure

├── src/
│   ├── run_forecast.py
│   ├── utils.py
│   ├── features.py
│   ├── models.py
│   ├── evaluation.py
│   ├── plot.py
│   └── report.py
├── data/
│   ├── bareilly.csv
│   └── mathura.csv
├── artifacts/
│   └── fast_track/
│       ├── forecast_T_plus_24.csv
│       ├── metrics.csv
│       └── plots/
│           ├── actual_vs_forecast.png
│           └── horizon_mae.png
├── reports/
│   └── fast_track_report.pdf
├── .gitignore
├── Pipfile
├── Pipfile.lock
├── requirements.txt
└── README.md

Data Sources

• Smart Meter Data: Kaggle - Smart Meter Data (Mathura and Bareilly)
• Weather Data: Open-Meteo API

Methodology

Data Preparation

1. Aggregates 3-minute readings to hourly consumption (kWh)
2. Handles missing values using forward fill
3. Removes outliers using 99th percentile capping
4. Merges weather forecasts by timestamp

Feature Engineering

• Temporal features: Hour-of-day (sine/cosine encoding), day-of-week
• Lag features: 1-hour, 2-hour, 3-hour, 24-hour lags
• Rolling statistics: 24-hour rolling mean
• Weather features: Temperature (when available)

Models

1. Baseline: Seasonal naive (previous day, same hour)
2. Ridge Regression: Engineered features with L2 regularization
3. Quantile Forecasts: 10th, 50th, 90th percentiles using residual scaling

Evaluation Metrics

• MAE (Mean Absolute Error)
• WAPE (Weighted Absolute Percentage Error)
• sMAPE (Symmetric Mean Absolute Percentage Error)

Output Files

metrics.csv

Contains model performance metrics for baseline and Ridge Regression models.

forecast_T_plus_24.csv

24-hour forecast with columns:

• timestamp: Hourly timestamps for next 24 hours
• yhat: Point forecast
• y_p10, y_p50, y_p90: Quantile forecasts

Plots

• Forecast Overlay: Last 3 days of actual demand + 24-hour forecast
• Horizon MAE: Error analysis across all 24 forecast horizons

Results Summary

The model achieves consistent performance with strong day-to-day demand patterns. The Ridge Regression model provides reliable forecasts suitable for grid operations and capacity planning.

Reproducibility

The entire pipeline is fully reproducible with a single command. All data processing, model training, forecasting, and report generation happen automatically.

Author

Prantik Mukhopadhyay

• Email: prantik25m@gmail.com
• LinkedIn: Prantik Mukhopadhyay