Development of knowledge sharing AI-based prediction framework for optimal HVAC control in building cluster

Abstract

With the explosive proliferation of population and urbanization, buildings are one of the largest subjects to consume a variety of energy resources. Particularly, it is accounting for approximately 40% of the final energy in Europe and more than 30% of the greenhouse gas emissions. Therefore, buildings have to be considered as one of the huge single domains for improving energy efficiency and issues that conventional cities should fundamentally solve to be transformed into smart cities. In general, efforts that aim to fulfill requirements to make the buildings cost- and energy-efficient can be largely classified into two categories: (i) passive building systems and (ii) active building systems, respectively. The passive system is to be helpful for reducing energy consumption and maximize the occupant comforts in indoor built environments by acclimating and optimizing the design and construction materials (e.g. orientation, types and ratio of windows in facade, etc.) of buildings to the natural environment without consumption of any other energy resources. In contrast, the active system can be referred as to leveraging all energy systems in the building to preserve high sustainable environments based on optimal and efficient control. The most representative active system is related to heating, ventilation, and air conditioning (HVAC) to regulate the indoor thermal condition and supply the fresh air in the operating phase. Particularly, energy consumed from operating HVAC in terms of heating and cooling systems has been largely consumed by more than approximately 43% and 59% in both residential and commercial sectors of South Korea, respectively. Thus, the active system should be considered as one of the most potential energy-saving factors in the operational phase during the entire life cycle of buildings as well as the advanced management and control systems for these systems should need to be explored and designed additionally. Diverse control models have been studied for the efficient operation of HVAC in the building. Compared with conventional control systems known as rule-based control (RBC) such as PID control, model predictive control (MPC) has been now in spotlight and been considered as one of the promising alternatives for the establishment of efficient control in the HVAC systems. Moreover, MPC is capable of adapting different system dynamics and disturbances, enhancing the thermal comfort condition and energy performances, simultaneously. In other words, MPC can be more prominent when the control tasks go beyond setpoint management including occupant-centric control or peak demand applications by predicting and reflecting the variations of future states in a variety of control parameters (temperature, relative humidity, etc.). To date, MPC based on data-driven approach has attracted the great attentions from many researchers over the years due to its flexibility and easy-to-use. In other words, data-driven approaches have accelerated comprehensive development of state-of-the-art predictive models by learning non-linear relationships latent in collected training samples without the need to deeper understand detailed physical expertise. Moreover, those approaches in the fields for optimal control of HVAC systems have demonstrated the superior capabilities than other conventional control models. However, the performance of data-driven MPCs would vary significantly depending on the accuracy of the prediction model. Thus, the development of a more efficient and accurate predictive model plays a crucial role in deriving the operational strategy of HVAC systems to ensure high energy performances and an acceptable indoor environment. The emerging issue in data-driven MPCs has gone beyond the application of individual buildings to multiple buildings where all characteristics are fully different. But, previous research in HVAC control of buildings based on data-driven MPC has been focused on and applied to individual buildings due to a lack of collected data and generalization performances of predictive models. Namely, even when prediction model is applied to individual or multiple buildings, reducing the modeling effort (not time-consuming) and enhancing the model reliability or fidelity are still essential problems to be tacked in the control of HVAC system based on data-driven MPCs. In addition, there has been a lack of practical implementation frameworks providing holistic prediction tasks (three in this dissertation) for obtaining control parameters of the HVAC systems in building grid scenarios. To sum up, the current data-driven MPCs for optimal control of the building HVAC in an operational phase have three representative issues: (i) limitations for dealing with multiple building scenarios, only applying to a single building, (ii) lack of generalization performances in the prediction model guaranteeing more accurate results even if insufficient datasets exist in the buildings, and (iii) lack of practical implementation framework for predicting control parameters used for MPCs applicable to multiple buildings. For addressing these challenges, we aim to develop the knowledge sharing AI-based prediction framework for providing accurate results when applied to other different buildings through more generalized performances despite the lack of an insufficient dataset and the different characteristics of buildings and each individual. The proposed prediction framework deals with three types of prediction tasks widely used for the establishment of optimal control strategies of HVAC systems in buildings: short-term energy consumption, individual occupant thermal comfort, and natural ventilation rate. Those factors should be necessarily considered in order to develop the MPC model for providing reasonable energy efficiency and an acceptable indoor environment for occupants simultaneously. Moreover, develop framework leverages the transfer learning (TL) in which knowledge learned from pre-trained model with usage of datasets in particular building by using deep learning (DL) algorithms is reused for application of target buildings in order to satisfy multiple building scenarios as well as each model used for TL is integrated with different types of ensemble strategies to improve generalization performances when performing model transfer from the source domain to the target domain. Lastly, pre-trained models developed by DLs are interpreted by shapley additive explanation (SHAP), which is one of the explainable artificial intelligence (XAI) methods widely utilized for the investigation of input variables affecting the prediction results. The proposed framework is composed of two main steps: (i) development of the pre-trained model as a feature extractor by using state-of-the-art deep learning models and interpretation of the model, and (ii) application of various ensemble strategies according to characteristics of different tasks when transferring the features previously learned from the pre-trained model. Through the results of the proposed prediction framework, the optimal operational strategies for energy systems of multiple buildings can be established and the cost generated from data acquisition and construction of prediction model for each individual building would be drastically reduced (not time-consuming and reliable prediction results, simultaneously). In addition, adequate interpretation of prediction models enables the decision-makers and operators of building energy systems to obtain comprehensive deeper insights of which features are affected to the prediction results and can provide latent clues for updating the prediction models.