Electricity Load Forecasting: Statistical vs. ML vs. DL Approaches

Project Overview

This project focuses on predicting electricity load consumption using three different approaches: Statistical (VAR), Machine Learning (XGBoost), and Deep Learning (LSTM). The goal is to identify the most accurate model for short-term energy demand forecasting, which is crucial for grid stability and energy management.

Dataset Information

• Source: Electricity Load Diagrams 2011-2014 (UCI Machine Learning Repository).
• Format: Hourly resampled time-series data.
• Features: Load (Target), Temperature, Humidity, Wind speed, and Hour of the day.

Key Data Insights (EDA)

Before modeling, I performed Exploratory Data Analysis to understand the data patterns:

• Seasonality: High demand during specific hours of the day (Business hours).
• Correlation: Analyzed the relationship between weather variables (Temperature, Wind) and electricity load.

Methodology

1. Data Preprocessing: Handled missing values (zeros), resampled to hourly frequency, and performed feature scaling using StandardScaler.
2. Feature Engineering: Created lag features (1-24h), rolling means, and rolling standard deviations to capture temporal trends.
3. Modeling:
   • VAR (Vector Autoregression): A traditional statistical model for multivariate time series.
   • XGBoost: A powerful gradient boosting algorithm with custom time-series features.
   • LSTM (Long Short-Term Memory): A Recurrent Neural Network (RNN) designed to learn long-term dependencies.

Results & Comparison

Model	RMSE	MAE	MAPE (%)	Accuracy (%)
VAR	19.94	18.28	71.29%	28.71%
XGBoost	0.47	0.14	0.74%	99.26%
LSTM	3.63	2.87	12.01%	87.99%

Backtesting Visualization

The following chart compares the predictions of all three models against the actual load for the last 100 hours of the test set:

Business Insights for Data Analyst Position

• Model Selection: XGBoost significantly outperformed other models in this specific dataset, making it the most reliable for short-term forecasting.
• Actionable Advice: Energy providers can use the XGBoost model to predict peak loads with >99% accuracy, allowing for better resource allocation and cost reduction.

How to Run

1. Clone the repository:

   git clone https://github.com/diphacf/TSD_models.git
   

2. Install dependencies:

   pip install -r requirements.txt
   

3. Open TSD_models.ipynb and run all cells.