Intelligent Prediction of Air Conditioning System Performance using Artificial Neural Network
概要
Energy provision and the control of the emissions related to energy consumption have arisen as major global issues in the last decade. The largest share of the global energy provision still relies on depletable fossil fuels, such as coal, oil, and gas, which are non-renewable and becoming very limited. The large-scale fossil fuels usage leads to the raising of potential supply difficulties and critical environmental impacts such as greenhouse gas emission, climate change, ozone layer depletion, etc. The energy consumption reduction and environmental safety of energy conversion systems are issues currently targeted by many researchers in the world, especially in building sector. The statistical data show that residential buildings account for the second largest share of energy consumption after the industrial sector. The energy consumption in buildings is steadily increasing. The main reasons contributing to this fact include the population growth, larger demand for building services, more advanced thermal comfort standards, and longer permanence of the occupants inside buildings. These circumstances forewarn the rising of the related energy demand in the future. Accordingly, improving energy efficiency in the building sector has recently become the main concern for policy makers and stakeholders. Today, the air conditioning systems have become a necessity in buildings. The recent change of lifestyle, which has brought many people to spend most of their time inside the building, pushes AC to work for long time, in different conditions, and different climates. In these modern days, people are estimated to spend approximately 80-90% of their time in conditioned indoor spaces. The International Energy Agency (IEA) estimates that 1.9 billion units of air conditioning (AC) systems are operating all over the world in 2020. This number is projected to experience an increase of as much as 50% by 2030. A more extensive use of air conditioners is caused by stricter standard requirements of thermal comfort, the established development of this technology (large number of experienced manufactures providing systems at affordable price) and economic welfare in developing countries. Moreover, the systems used by consumers are most likely to demonstrate less than half of best available efficiency of this technology. As a result, the growth in energy demand for AC systems tends to increase continuously and, if not operated properly, the related environmental footprint could become a substantial cause for unsustainable greenhouse emissions. As the AC systems are responsible for the largest energy demand in the building sector, there are large opportunities to reduce the energy consumption in the building sector through efficiency improvement of these systems and their operation strategy.
A vapor compression technology is most commonly applied for AC systems in residential building applications. In order to increase the systems efficiency several works have been conducted to develop advanced control strategies and improve component efficiency. However, these efforts improved the system performance only slightly due to technical limitations of each system component. On the other hand, the optimal operation management can be an effective measure to substantially reduce the energy consumption of AC systems. The system cooling capacity represents the main extensive indicator of the system performance, response to the external disturbances and internal load. This quantity is generally related to the input driving energy to evaluate the system performance in terms of first and second law thermodynamic efficiency. Ideally, the system should be working to deliver appropriate cooling capacity in response to the instantaneous cooling load at the potentially maximum efficiency. According to the investigation presented in previous research, the optimal operation of AC systems can be achieved in range of 50-70% of the nominal capacity. To this purpose, a reliable prediction method for actual system performance is necessary for defining advanced operation strategies, adapting the available capacity with the load of a facility, and establishing system efficiency monitoring techniques. As billions of these systems have already been installed in different types of buildings, capacity sizes, grids, and climates and encountered performance degradation over time due to fouling of the heat exchanger surface, pipe leakage, improper refrigerant mass charge, etc., the actual system performance during operation is generally unknown. This suggests the fundamental necessity of a cost-effective, accurate, and non-intrusive method for predicting the AC system performance for better system operation management.
Due to the several issues that should be considered in a performance prediction method for air conditioners, such as accessibility for input measurement, cost, reliability and applicability, an effective method is presently not available. In practice, the systems operate in various dynamic conditions and there are only few parameters that can be measured using inexpensive and non- intrusive sensor as input of prediction. Moreover, installed systems have different size, manufactures, configurations, and characteristics. The above mentioned issues have so far represented overwhelming obstacles to the development of a physical model, to the measurement of the required input parameters, and to the implementation of the method in field tests. Among the several approaches to estimate system performances, direct measurement, statistics, physical model, and machine learning are mostly adopted.
The direct measurement method cannot be extensively applied as the measurement of mass flow rate and pressure of the refrigerant is intrusive and requires high investment cost. Physical models have a good approximation capability and can be reliable to extrapolate the system behavior outside the validation range. Unfortunately, to the accurate knowledge of the numerous input parameters required for calculation, such as component geometry, refrigerant properties, system specifications, etc. are commonly not available far installed systems. On the other hand, statistics method can be proposed evaluate system performance on the basis of a set of recorded data. However, these data are not presently able to capture the extensive variability of the system performances, dynamic behavior, and operating conditions. Alternatively, machine learning technique can be used to predict the system performance without requiring the complex mathematical functions as used in physical model. This method has been proven to be able to approximate the input output data accurately when the sufficient training data are available. Regarding to the complexity of AC systems phenomena, this method offers the possibility for simplification of cooling capacity estimation. In physical model, the relationship between the physical phenomena and system performance is established using the complex mathematical functions based on first principle theory. Instead of using the mathematical model, an ANN black box model can be developed to learn the physical phenomena using few, indirectly related input parameters (non-intrusive and inexpensive) to predict the system performance. This method can be developed using relevant input parameters that are non-intrusive, inexpensive to measure and generally represent system performance behavior.
The application of ANN to predict AC system performances have been investigated in previous works. It is generally shown that the performance of a given AC system can be well predicted by ANN model within the training range of the representative data used for training. However, ANN models were mostly developed using input parameters that are intrusive to measure (refrigerant mass charge, compressor speed, valve opening) or affected by high location variance (air-side temperatures and velocity), which could hinder their actual implementation in operative systems. Moreover, the predictions were carried out using the training and testing data generated from the same system, in the same structural condition. Therefore, the developed ANN model in previous studies cannot be applied in various systems with different rated capacity.
The present study aims to develop reliable ANN model that can be generally used to predict cooling capacity of different AC systems, in diverse operating conditions, while using non-intrusive and cost effective input parameters. The research is conducted under the hypothesis that system performance can be predicted by properly teaching ANN the physics of air conditioning cycle. The method adopted involves few input parameters representing the air conditioning cycle and easily measurable in actual operative conditions. Specifically, the method is demonstrated by utilizing an ANN model exclusively based on four input refrigerant temperatures that are easily accessible from the outdoor unit. A scaling method with data normalization is introduced to predict the system with different nominal capacity and from different manufacturers.
The ANN model is developed using a multilayer perceptron model to predict the cooling capacity of the systems. It is trained and tested with different data and optimized by a varying number of neurons and hidden layers. The training and testing data for the prediction are generated by five different systems using AC simulator and experimental facility. The simulator has been validated and designed to meet the system characteristics in close resemblance of the actual system performance in either steady or steady conditions. The three systems developed by a numerical simulator include 50 kW, 7.1 kW, and 2.5 kW nominal capacity, respectively. While the two actual machines used in experimental facility have nominal capacity of 33.5 kW and 28 kW, respectively. In order to comprehensively cover the realistic range of operation of AC systems, the performance behaviors are characterized in various operating conditions with varying cooling load, indoor and outdoor temperatures.
The applicability of different sets of inputs, including air temperatures, controllable parameters, and refrigerant temperatures, representing different categories of the influent parameters for system performance prediction, is investigated. The data generated from the numerical models of two AC systems with nominal capacity of 7.1 kW and 2.5 kW are used to train and test the ANN models using different input parameters for performance prediction. The investigation is conducted by testing the developed ANN model in different possible case: on data simulated within the range of training scenarios, on data outside the range of training, and on data obtained from the simulation of a different system with a different nominal capacity. The investigation indicate that the refrigerant temperatures at the inlet and outlet of evaporator and condenser are selected as the best option for an intelligent prediction method of AC systems as they meet the following criteria: inexpensive to measure, non-intrusive, relatively low uncertainty, represent air-conditioning cycle, scalable for different systems with different rated capacity and manufacturer, and sensitive to cooling capacity and external disturbances.
Consequently, the performance prediction of actual AC systems, represented by the experimental data collected in actual machines of 33.5 kW and 28 kW, is attempted. The training data are generated through the simulation of the 50 kW multi-evaporator systems. This investigation aims to verify the generalization capability of the fundamental hypothesis underlying this method for actual implementation. The hypothesis guiding this research effort is that, although AC systems have different configuration and size, and the specific data characteristics of simulation and experiments differ in the specific response determined by the control package, the training data obtained via a reliable simulator still provide a fundamental representation of the refrigerant cycle realized in actual systems. If such hypothesis is proved true and the simulator could reliably approximate the data behavior of actual AC systems, the ANN model can learn the system cycle from simulation data with a much broader variability of conditions, configurations and climates, which could be impossible (in terms of time, cost, and measuring method) to collect experimentally. This will reduce the time, cost, and complexity of data generation. Additionally, the simulator can be flexible and could provide the data that are difficult to generate in experimental facility. The results revealed that the ANN model has successfully predicted on smooth data characteristics. However, larger deviations appear when applied on high fluctuation testing data characteristics (due to the intermittent operation occurring at low partial loads) since this kind of data behavior are not presently available in the simulation training data. Therefore, due to the lack of information provided by the simulation data during such circumstances, the ANN model trained with simulation data could not well recognize the intermittent (on/off) or strongly dynamic experimental data. .
In order to demonstrate the possibility of higher accuracy prediction results by relying on the suggested method, the prediction of air conditioners’ performance on different systems is investigated by using experimental data for both training and testing. This aims to teach ANN a higher data variability especially during intermittent operation of the actual machine. The training and testing data are generated from experimental facility with the system of 33.5 kW and 28 kW, respectively. The ANN prediction includes the inputs from the previous time steps to capture the system dynamic behavior. The prediction results yield good agreement between predicted and experimental values. The inclusion of intermittent data characteristics in the training improves the data variability and generalization ability of the ANN model. This result suggests that the ANN model could be applied to predict the performance of different systems using sufficiently representative training data from a reference system.