A novel twin time series network for building energy consumption predicting

Area of Science:

• Building energy management
• Artificial intelligence in energy systems
• Time series forecasting

Background:

• Accurate building energy consumption prediction is vital for efficient energy management.
• Current methods struggle with temporal correlation, prediction accuracy, timestamp embedding, and multi-source data fusion.
• Existing models often fail to fully leverage the potential of time series decomposition and adaptive data integration.

Purpose of the Study:

• To address the limitations in current building energy prediction models.
• To propose an advanced model that enhances temporal correlation extraction and prediction accuracy.
• To effectively integrate multi-source data, including energy consumption and meteorological information.

Main Methods:

• Development of Twin Time-Series Networks (T2SNET).
• Integration of a time-embedding layer and Temporal Convolutional Network (TCN) for pattern extraction from Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN).
• Implementation of an adaptive fusion gate for combining energy consumption and meteorological data.

Main Results:

• T2SNET demonstrated significant improvements over baseline methods across diverse building types (dormitories, offices, classrooms).
• On a university classroom dataset, T2SNET achieved a 4.56% reduction in Mean Absolute Error (MAE), 9.45% in Root Mean Square Error (RMSE), and 3.16% in Mean Absolute Percentage Error (MAPE) compared to CEEMDAN-RF-LSTM.
• The model effectively extracts temporal patterns and fuses multi-source data, leading to superior prediction performance.

Conclusions:

• T2SNET provides a robust and effective solution for building energy consumption prediction.
• The proposed method overcomes key challenges in temporal correlation and data fusion.
• The findings support the adoption of T2SNET for advanced energy management systems.