The decision variables of local control loops significantly influence the overall energy use of the heating ventilation and air-conditioning (HVAC) systems. Therefore, model-based real-time optimization (MRTO) of those decision variables has been widely studied in recent years. The decision making by MRTO relies on the accuracy of the performance model used to describe the relationship between cost function and decision variables. However, due to high diversity in ambient conditions and demand load, it is very difficult to develop an accurate model of these systems without a downside. This paper presents a dissimilarity-based ensemble model of the HVAC system, which systematically ensembles two powerful modelling methods from the literature: simplified or semi-physical model; and artificial neural network model. A semi-physical model of the system was used as a generalized model to provide reliability in performance. A batch trained nonlinear-autoregressive neural network with exogenous input (NARX) was used as an artificial neural network model to provide robust and accurate performance. The ensemble technique proposed in this study is a dissimilarity-based boosting technique, i.e. able to capture the performance of the NARX model and ensure accuracy and reliability in final ensemble prediction. To evaluate its performance, it was tested for the performance prediction of a complex HVAC system under actual ambient conditions over a week and its performance was compared with both semi-physical and NARX models. It was demonstrated that the proposed dissimilarity-based ensemble model could provide improved accuracy with robust operation and reliability while maintaining reasonable computational efficiency. When compared with the semi-physical model, the proposed dissimilarity-based ensemble model provided; 60–78% reduction in prediction error in terms of normalized root-mean-square error, improvement in terms of peak signal-to-noise ratio by 55–63%, and consistent performance in terms of reduction in mean absolute percentage deviation by 51–75%.

本地控制回路的决策变量显着影响采暖通风和空调 (HVAC) 系统的整体能源使用。因此，这些决策变量的基于模型的实时优化（MRTO）近年来得到了广泛的研究。MRTO 的决策依赖于用于描述成本函数和决策变量之间关系的性能模型的准确性。然而，由于环境条件和需求负载的高度多样性，很难在没有缺点的情况下开发这些系统的准确模型。本文提出了一种基于差异的 HVAC 系统集成模型，它系统地集成了文献中两种强大的建模方法：简化或半物理模型；和人工神经网络模型。系统的半物理模型被用作通用模型以提供性能可靠性。使用具有外源输入的批量训练非线性自回归神经网络 (NARX) 作为人工神经网络模型，以提供稳健和准确的性能。本研究提出的集成技术是一种基于差异的提升技术，即能够捕捉NARX模型的性能并确保最终集成预测的准确性和可靠性。为了评估其性能，在一周的实际环境条件下对复杂 HVAC 系统的性能预测进行了测试，并将其性能与半物理和 NARX 模型进行了比较。结果表明，所提出的基于相异性的集成模型可以在保持合理的计算效率的同时提供更高的准确性、稳健的操作和可靠性。与半物理模型相比，提出的基于相异性的集成模型提供了；就归一化均方根误差而言，预测误差减少 60-78%，峰值信噪比提高 55-63%，平均绝对百分比偏差降低 51，性能一致–75%。