Shakhovska Nataliya, Mochurad Lesia, Caro Rosana, Argyroudis Sotirios
Artificial Intelligence Department, Lviv Polytechnic National University, 12 S. Bandery St, Lviv, 79013, Ukraine.
Brunel University of London, Uxbridge, UB8 3PH, UK.
Sci Rep. 2025 Jan 2;15(1):47. doi: 10.1038/s41598-024-85026-3.
Efficient energy management and maintaining an optimal indoor climate in buildings are critical tasks in today's world. This paper presents an innovative approach to surrogate modeling for predicting indoor air temperature (IAT) in buildings, leveraging advanced machine learning techniques. At the core of this study is the application of Long Short-Term Memory (LSTM) networks for time-series modeling, which significantly enhances the capture of temporal dependencies in temperature predictions. The proposed LSTM with RWCV (Rolling Window Cross-Validation) offers significant advantages over a usual LSTM in time-series tasks, particularly due to its ability to adapt to new data trends through the rolling window mechanism. It provides more robust and generalizable forecasts in dynamic environments, prevents overfitting through dropout and cross-validation, and improves model evaluation with temporal integrity. In contrast, traditional LSTM models are better suited for static, non-evolving datasets and may not handle dynamic time-series data effectively. To rigorously assess model performance, a comprehensive evaluation framework is developed, incorporating metrics such as mean square error (MSE) and the coefficient of determination (R²). Additionally, a novel cumulative error analysis method is introduced enabling real-time monitoring and model adjustment to maintain predictive accuracy over time. Test results demonstrate that model losses on the test dataset are only marginally higher than those on the training dataset, indicating robust generalization capabilities. Loss values range from 0.0004709 to 0.02819861, depending on building operating conditions. A comparative analysis reveals that Adaboost and Gradient Boosting models outperform linear regression, highlighting their potential for achieving energy-efficient and comfortable indoor climate management in buildings. The findings underscore the efficacy of the proposed approach for IAT prediction and point towards further research possibilities in dataset expansion and model optimization to enhance building climate management and energy conservation.
在当今世界，高效的能源管理和维持建筑物内的最佳室内气候是至关重要的任务。本文提出了一种创新的代理建模方法，用于预测建筑物内的室内空气温度（IAT），该方法利用了先进的机器学习技术。本研究的核心是应用长短期记忆（LSTM）网络进行时间序列建模，这显著增强了对温度预测中时间依赖性的捕捉。所提出的带有滚动窗口交叉验证（RWCV）的LSTM在时间序列任务中比普通LSTM具有显著优势，特别是由于其能够通过滚动窗口机制适应新的数据趋势。它在动态环境中提供更稳健和可推广的预测，通过随机失活和交叉验证防止过拟合，并通过时间完整性改进模型评估。相比之下，传统的LSTM模型更适合静态、非演化的数据集，可能无法有效地处理动态时间序列数据。为了严格评估模型性能，开发了一个综合评估框架，纳入了均方误差（MSE）和决定系数（R²）等指标。此外，还引入了一种新颖的累积误差分析方法，能够进行实时监测和模型调整，以长期保持预测准确性。测试结果表明，测试数据集上的模型损失仅略高于训练数据集上的损失，表明具有强大的泛化能力。损失值范围从0.0004709到0.02819861，具体取决于建筑物的运行条件。对比分析表明，Adaboost和梯度提升模型优于线性回归，突出了它们在实现建筑物内节能和舒适室内气候管理方面的潜力。研究结果强调了所提出的IAT预测方法的有效性，并指出了在数据集扩展和模型优化方面进一步研究的可能性，以加强建筑物气候管理和节能。
