Topic 14 – Electricity Load Diagrams (Energy Consumption, Library Dataset)

Dataset (Example Source)

• Source: Electricity load diagrams dataset (via libraries like GluonTS/HuggingFace, etc.)
• Example Link: https://huggingface.co/datasets/tulipa762/electricity_load_diagrams

Download (Simple Manual Approach)

1. Open the link above.
2. Download the CSV file containing the load data (if provided).
3. Save to data/electricity/.

Data Loading (If CSV is Available)

import pandas as pd
df = pd.read_csv("data/electricity/electricity.csv")  # adjust filename
print(df.head())
Example: parse datetime column like "timestamp"
df["timestamp"] = pd.to_datetime(df["timestamp"])
df = df.set_index("timestamp").sort_index()

Implementation Steps

1. Data Loading and Exploration

• Load electricity load data
• Understand data structure (multiple consumers, time series)
• Inspect data frequency (hourly, daily)
• Examine data quality and missing values
• Identify available variables

2. Exploratory Data Analysis (EDA)

• Plot load time series (overall and by consumer if multiple)
• Identify strong seasonality patterns:
• Daily patterns (hourly data)
• Weekly patterns (weekday vs weekend)
• Yearly patterns (seasonal variations)
• Analyze load distributions
• Calculate basic statistics
• Perform time series decomposition

3. Data Preprocessing

• Handle missing values
• Handle outliers (may be real events)
• Create time-based features (hour, day of week, month, season)
• Normalize or scale if needed
• Prepare data for modeling

4. Feature Engineering

• Time Features:
• Hour of day, day of week, month, season
• Holiday indicators
• Weekend indicators
• Lag Features:
• Previous hour, day, week loads
• Same hour previous day, week
• Rolling Statistics:
• Rolling mean, std (daily, weekly windows)
• External Features (if available):
• Temperature, weather data
• Economic indicators

5. Model Building

• Univariate Models (per consumer):
• ARIMA/SARIMA (strong seasonality - multiple seasonal patterns)
• Exponential smoothing (Holt-Winters with multiple seasonality)
• Multivariate Models:
• VAR models for multiple consumers
• Hierarchical models
• Machine Learning:
• Feature-based models (XGBoost, LightGBM)
• Handle multiple seasonality with features
• Advanced (if using libraries):
• GluonTS, darts, or similar time series libraries
• Deep learning models (LSTM, Transformer)

6. Model Evaluation

• Use time series cross-validation
• Split data temporally
• Calculate metrics (MAE, RMSE, MAPE)
• Evaluate at different time horizons (1-hour, 1-day, 1-week ahead)
• Compare multiple approaches

7. Forecasting

• Generate forecasts for different horizons
• Include prediction intervals
• Aggregate forecasts if multiple consumers
• Visualize forecasts with actual values
• Analyze forecast accuracy by time of day/day of week

8. Advanced Analysis (Optional)

• Analyze load patterns by consumer type
• Identify peak demand periods
• Analyze seasonal variations
• Compare different modeling approaches
• Analyze feature importance

Expected Deliverables

1. EDA Report:

• Load time series plots
• Seasonality analysis (daily, weekly, yearly)
• Pattern identification
• Statistical summaries

1. Model Results:

• Selected model(s) with parameters
• Performance metrics (by horizon)
• Forecast plots
• Comparison of approaches

1. Code:

• Complete Python notebook
• Data processing functions
• Feature engineering utilities
• Modeling pipeline

Tips

• Electricity load has very strong and complex seasonality (multiple patterns)
• Daily patterns are crucial (peak hours, off-peak hours)
• Weekly patterns (weekday vs weekend) are important
• Yearly patterns (summer vs winter) affect demand
• Multiple seasonality requires SARIMA with multiple seasonal components or feature engineering
• Consider using specialized time series libraries (GluonTS, darts) for complex seasonality
• Feature engineering is essential (time features, lags, rolling stats)
• External features (temperature) can significantly improve forecasts
• Handle multiple consumers appropriately (panel data structure)
• Use appropriate evaluation metrics and cross-validation
• Document all preprocessing and feature engineering steps
• Compare simple models (ARIMA) with complex models (XGBoost, LSTM)
• Consider ensemble methods for better performance
• This is a challenging problem - expect to try multiple approaches