This paper introduces an innovative approach to forecasting indoor air temperature (IAT) in smart buildings, using machine learning techniques, particularly Long Short-Term Memory (LSTM) networks. By incorporating Rolling Window Cross-Validation (RWCV), the model adapts to dynamic changes in building conditions, improving the accuracy and generalizability of temperature predictions over traditional LSTM models. The study emphasizes the need for efficient energy management and optimal indoor climates, particularly in the context of smart buildings. It compares the performance of LSTM models with other machine learning models like Adaboost and Gradient Boosting, which also outperform linear regression in terms of energy efficiency and comfort. The study further introduces a novel cumulative error analysis method for real-time model adjustment and accuracy maintenance.
Takeaways:
- LSTM with RWCV for Dynamic Environments: The introduction of RWCV to LSTM networks enhances their ability to adapt to new trends in data, making the model robust in dynamic building conditions.
- Real-time Monitoring and Adjustment: A novel cumulative error analysis method ensures that model performance remains accurate over time, even in fluctuating building conditions.
- Comparative Performance: The study shows that machine learning models like Adaboost and Gradient Boosting provide superior performance over linear regression, improving energy management and comfort in smart buildings.
- Surrogate Modeling for Building Energy Efficiency: Surrogate models, such as LSTM, serve as simplified representations of complex systems, improving the efficiency of building climate control while reducing computational costs.
- Energy Savings and Comfort: Accurate IAT predictions lead to better HVAC optimization, reducing energy costs, enhancing occupant comfort, and supporting energy conservation efforts in smart buildings.
