Innovative machine learning approaches for indoor air temperature forecasting in smart infrastructure.

Area of Science:

• Building Science
• Machine Learning
• Energy Management

Background:

• Efficient energy management and maintaining optimal indoor climate are crucial for modern buildings.
• Accurate prediction of indoor air temperature (IAT) is key to achieving these goals.
• Traditional methods often struggle with the dynamic nature of building environments.

Purpose of the Study:

• To present an innovative surrogate modeling approach for IAT prediction using machine learning.
• To enhance time-series modeling capabilities for dynamic building data.
• To improve the robustness and generalizability of temperature forecasts.

Main Methods:

• Application of Long Short-Term Memory (LSTM) networks for time-series analysis.
• Implementation of Rolling Window Cross-Validation (RWCV) to adapt to evolving data trends.
• Development of a comprehensive evaluation framework including MSE, R², and cumulative error analysis.

Main Results:

• The proposed LSTM with RWCV demonstrates robust generalization, with minimal loss difference between training and testing datasets.
• Loss values ranged from 0.0004709 to 0.02819861, indicating effective prediction across different building conditions.
• Comparative analysis showed Adaboost and Gradient Boosting outperforming linear regression for IAT prediction.

Conclusions:

• The developed LSTM with RWCV approach is effective for accurate IAT prediction in buildings.
• The method offers improved adaptability and robustness compared to traditional LSTM models for dynamic time-series data.
• Findings support enhanced building climate management, energy conservation, and suggest avenues for future research in model optimization.