Transformers for Energy Forecast.

Area of Science:

• Building Science
• Artificial Intelligence
• Energy Systems

Background:

• Growing energy demand and climate change necessitate enhanced energy efficiency in buildings.
• Accurate energy consumption forecasting is key to optimizing building performance and operations.
• Identifying energy efficiency upgrades relies on reliable consumption prediction.

Purpose of the Study:

• To develop and evaluate an advanced forecasting model for building energy consumption.
• To address the challenge of multi-variable time series forecasting in energy usage.
• To improve the accuracy and robustness of energy consumption predictions for optimized building management.

Main Methods:

• A modified multi-head transformer model was proposed for multivariate time series analysis.
• A learnable weighting feature attention matrix was introduced to combine input variables.
• The model's performance was benchmarked against recurrent neural network (RNN) models.

Main Results:

• The proposed transformer-based model demonstrated robust performance in forecasting energy consumption.
• The model achieved a lower mean absolute percentage error compared to RNN models.
• The multivariate approach effectively integrated various input factors for prediction.

Conclusions:

• The modified transformer model offers superior performance for multivariate energy consumption forecasting.
• This advanced model can be integrated into future systems for tracing energy scenarios.
• The findings contribute to creating more sustainable and energy-efficient building usage patterns.