2025 Competition on Electric Energy Consumption Forecasting - Smart Building Dataset

Overview

The 2025 Competition on Electric Energy Consumption Forecasting dataset published on Zenodo in July 2025 was released as part of an international challenge to compile and evaluate the latest advances in building energy forecasting methods.

Dataset Versions and Structures

• v1.0 (Long-term high-resolution): One full year of data from a smart building with readings taken every 5 minutes, totaling over 105,000 data points. Suitable for evaluating long-term forecasting models that must handle seasonal patterns, weekday/weekend differences, and holiday effects.
• v2.0 (Medium-term high-resolution): 40 consecutive days of 5-minute interval data (over 11,500 points), designed for models that focus on weekly and daily cycles without the complexity of full seasonal variations.
• v3.x (Short-term low-resolution): A single day of hourly energy consumption data (24 points), targeting simple forecasting tasks and providing a baseline for rapid algorithm prototyping.

Measured Variables

• Total building energy consumption (kWh): Whole-building aggregate electricity usage at each time step.
• Timestamps: Date and time markers in standard format for temporal analysis.
• Contextual metadata (where available): Building type (office, mixed-use), floor area, occupancy indicators, and HVAC operational status.
• Weather data (for some versions): Outdoor temperature, humidity, solar irradiance to support exogenous variable modeling.

Competition Objectives

• Benchmark the performance of classical statistical models (ARIMA, exponential smoothing) against modern machine learning techniques (LSTM, GRU, Transformer, ensemble methods).
• Evaluate generalization across different time horizons (intraday, next-day, next-week forecasts).
• Assess the impact of data granularity (5-min vs. hourly) on forecasting accuracy and computational requirements.
• Encourage innovation in hybrid models, transfer learning, and interpretable forecasting approaches.

Use Cases

• Energy forecasting research: Developing and testing novel forecasting algorithms with publicly available benchmark data.
• Building management systems: Training models for real-time energy prediction to optimize HVAC scheduling, energy procurement, and demand response participation.
• Educational purposes: Providing students and practitioners with realistic time-series datasets for learning forecasting techniques.
• Smart grid integration: Supporting grid operators in predicting building-level demand for aggregated load forecasting and grid balancing.