Comparative Analyses of Energy Efficiency between on ‐ Demand and Predictive Controls for Buildings’ Indoor Thermal Environment

: Advanced thermal control technologies have been continuously developed to complement conventional models and algorithms to improve their performance regarding control accuracy and energy efficiency. This study analyses the strengths and weaknesses of simultaneous controls for the amount of air and its temperature by use of on ‐ demand and predictive control strategies responding to two different outdoor conditions. The framework performs the comparative analyses of an on ‐ demand model, which reacts immediately to indoor conditions, and a predictive model, which provides reference signals derived from data learned. Two models are combined to make a comparison of how much more efficient the combined model operates than each model when abnormal situations occur. As a result, when the two models are combined, its efficiency improves from 20.0% to 33.6% for indoor thermal dissatisfaction and from 13.0% to 44.5% for energy use, respectively. This result implies that in addition to creating new algorithms to cope with any abnormal situation, combining existing models can also be a resource ‐ saving approach.


Building Thermal Controls
In order to improve the performance of building energy supply methods, the efficiency of the Heating, Ventilating, and Air-Conditioning (HVAC) systems commonly has been investigated by use of controlling fuel amount into boilers, fan motor speed, and distribution networks as control targets. Complementing the control strategies and tuning rules have helped to upgrade the inner algorithms of Proportional-Integral-Derivative (PID) control systems such as optimization of fuel amount or speed of turbines. In such models, the focus on the analyses was placed on reducing absolute errors by analyzing the differences between control volume and actual demand. As a metric, the root mean square error was frequently used to explain the model's efficiency by analyzing the differences between existing and new regression models. In these cases, some parametric simulation processes were employed to supplement the existing regression models, to compensate for the weaknesses of statistical approaches [1][2][3][4].
The rapid development of computing and statistical technologies made the models improve their efficiency against complex calculations reflecting huge amount of data. Specifically, the Fuzzy Inference System (FIS) and Artificial Neural Network (ANN), which were preferred to deal with the large amount of data, were gradually developed to improve control strategies in sensitive and adaptive control areas [5,6]. For the processes, developing data-driven genetic algorithms for linguistic models to solve ambiguous situations in thermal networks was an effective way to define their inner relationship of numerous events. Finding statistical patterns at controlling signals using the FIS model was frequently adopted to compare the efficiency of conventional PID rules for specific situations [5,7,8]. The energy use efficiency from novel models adopting the FIS algorithm was frequently utilized, and the FIS systems were developed to increase the efficiency of fuel injection in boilers as well as fuel distribution and heat gain in networks of spaces, buildings, and districts. Plantscaled control systems have also been preferred to increase energy savings in distribution systems because of the immediate and noticeable economic benefits due to the reduction of fuel costs [9][10][11]. The linguistic logics, which consist of fuzzy membership functions reflecting ambiguous expressions, effectively helped to solve some difficult problems including high complex parametric approaches or conventional control methods based on time-consuming experimental studies [5,7,12]. By effectively combining generic PID and FIS systems, several variations of controlling were investigated to increase the control efficiency at the stages of fuel use for combustion systems or energy distribution networks in a single building or building districts [9,13].
Furthermore, such advances in computing technology have allowed many researchers to develop another very useful way of dealing with a very large amount of data. In several studies, many complex problems, because the number of combinations increases exponentially with just a few variables, can be investigated using the Artificial Neural Network (ANN). Its algorithm helped to solve complex calculations for analyzing the effects of day-lighting, infiltration and ventilation using small gaps, and latent heat of living, which require time-consuming processes to organize data and find their hidden correlations [14,15]. In addition to the control algorithms mixing the PID, FIS, and ANN models, specific differential functions dealing with the amount of opening for valves or dampers were tested by combining experimental models and ANN algorithms, which react to different demands at specific geometries in buildings or weather conditions [16][17][18].
In addition to the improvement of machine's performance, there have been various approaches to investigate thermal comfort levels in buildings including questionnaire-based studies based on users' actual responses and simulation studies based on some systematic assumptions [19][20][21]. Unlike some qualitative approaches, the Predicted Mean Vote (PMV) and Predicted Percentage of Dissatisfied (PPD) indices were preferred to define the mechanical and mathematical thermal sensation [22][23][24][25]. In order to improve thermal comfort without compromising thermal regulations, various types of energy conservation measures were adopted. The measures include two different types of factors such as wall, roof, penetration, electric equipment, and HVAC system, and such as mean radiant temperature, air speed, and relative humidity [26][27][28][29]. At phases for the improvement, system models to immediately consider hidden relationships between the factors were refined with dynamic variables from co-simulation programs. In addition, users' responses obtained using systematic survey formats helped to recognize occupant behavior affected by control rules for widening energy saving strategies [30][31][32][33].
By means of the advanced strategies, several studies have been conducted from the various viewpoints to build more appropriate indoor thermal conditions in buildings. The existing space planning and thermal control methods may not operate properly, and excessive energy use can be derived from energy intensive control methods. Therefore, further discussion is required on the issue of developing a completely new model to improve the thermal control systems currently in use. This can also be addressed to be linked to the economic comparison between continuously creating new models for a more effective control approach and properly combining properly existing models. This study presents a combined model for controlling supply air mass and its temperature utilizing two existing algorithms. The Methodology section describes the basic structures of the simulation models, equations, and algorithms used. The Results and Discussion section indicates the advantages and disadvantages of single-algorithm control models and a combined model, and the characteristics of the models' signals are discussed regarding how the models work to improve thermal comfort and energy use. In the final section, Conclusion and some possible follow-up studies are addressed to improve the model's performance. Figure 1 and Table 1 describe a simplified building model and the design parameters for the simulations. This room is a single module equipped with one HVAC system. The pressure variations of air speed in the building were not considered. In addition, air leakage within the building's duct and air flow in the building space were neglected in this modeling. Other geometries of the building and simulation parameters to investigate indoor thermal transfer and comfort levels, except for some fixed values such as external work rate (0 W*m −2 ) and air velocity (0.1 m*s −1 ), were adopted from the templates in some frequently referred to sources [34][35][36]. In addition, one of the main objectives of this study is how much efficiency these models show when abnormal climate conditions are inputted that are hard to predict. Recently, the daily temperature gap has widened due to abnormal temperature drops, which led to the choice of the date around mid-spring or fall, assuming that heating and cooling are simultaneously required during one day. For this reason, in the climate condition of Table 1, the condition of April 25th in Xuzhou was used as the control group of normal condition, and the temperature and relative humidity of May 7th were used as abnormal condition. The weather file was obtained from the website of EnergyPlus Weather Data (https://energyplus.net/weatherlocation/asia_wmo_region_2/CHN//CHN_Jiangsu.Xuzhou.580270_CSWD). Except for the conditions addressed, other simulation parameters, such as territorial conditions, construction sets, schedules, and simulation algorithms, followed the template of ASHRAE9012016_SchoolSecondary in the EnergyPlus simulation program. In the process of setting parameters, some important values related to the building geometries and weather conditions were set, as presented in Table 1 and Figures 2 and 3. Wall thickness (Dwall) 15 cm 4

Research Framework
Wall thermal conductivity (kwall) 0.038 W*m −1 *K −1 5 Window area (Awindow) 12 m 2 6 Window thickness (Dwindow) 2 cm 7 Window thermal conductivity (kwindow) 0.780 W*m −1 *K −1 8 Mass flow rate into room (ṁ) 3600 kg*h −1 9 Weather condition Xuzhou, Jiangsu in China  From the thermodynamics, the energy transfer of heating and cooling in a room is given as Equation (1). In this equation, it was assumed that there was no work in the system from the mass flow rate and enthalpy, there was no change in the flow rate from the law of conservation of mass flow rate, and there was thermal energy loss in the room, which only occurred from conduction through the building's envelopes.
The quantification of thermal comfort is preferably measured using Predicted Mean Vote (PMV), developed by Profs. P.O. Fanger at the International Centre for Indoor Environment and Energy at the Technical University of Denmark, and Predicted Percentage of Dissatisfied (PPD). From an exponential of human factors and thermal loads, a function of the PMV is developed as follows [35][36][37].
where M is Metabolic Rate, L is Thermal Load, Tcl = average surface temperature of clothed body, fcl = ratio of clothed surface area to DuBois surface area (Acl/AD), Rcl = effective thermal resistance (Rvalue) of clothing, Ta = air temperature, hc = convection heat transfer coefficient , Tr = mean radiant temperature, hr = radiative heat transfer coefficient, Wa = air humidity ratio, and Wsk = saturated humidity ratio at the skin temperature. An exponential for the PPD from the empirical relationship between the metrics is defined as follows.
. (4) There are three different ways to process signals for efficient control of the building thermal and PPD models above: Thermostat On/Off (On/Off), Fuzzy Inference System (FIS), and Artificial Neural Network (ANN). The On/Off model is a simple and conventional model most commonly used in real life, switched on and off according to the previously set up heating and cooling temperature.
The purpose of the FIS model is to determine the optimized values of supply air as the amount of mass and the temperature, which are directly linked to the difference between Tset and Troom. Equations (5)-(7) describe a membership function for two input variables in the FIS model, wherein the differences between the set-point and room temperature (E) are derivative of the temperature difference (ΔE) [38]: In two outputs for supply air as the amount of mass and its temperature, the FIS model uses five membership functions for each input variable with discourse 0(0%) to opening 1(100%) for Mass (amount of air) and −10 to 10 for Temp (air temperature). It was assumed that the range of controlling Mass in the output was 0 (0%) to 1 (100%). The first layer consists of input variables (MFs), input 1 and input 2. This layer just supplies the input values to the next layer. In the first layer every node is an adaptive node. This study utilizes triangle MFs with maximum equal to 1 and minimum equal to 0.
; , , Like the above functions and sets for each range, it was assumed that the intersections of each expression, such as Large, Medium, Low, and Zero, occur at an interval. For instance, when E and ΔE are estimated as 0.1 and −10, the model converts the indoor temperature as Positive-Very-Small and Negative-Big, respectively. Then, the model interprets them as OFF for Mass and Medium-Large for Temp, respectively, from the membership rule matrix. Then, the system converts the linguistic values into numerical values as around 0% for Mass and around −4~−2 °C for Temp, respectively, using the fuzzy membership functions. This process is called fuzzification and defuzzification, but the parts related to simple arithmetic operations are recommended to see other references.
The ANN structure includes a large class of several nodes and flows, and the selections of a nonlinear mapping function x with a network are needed [39]. This algorithm, used in function approximation, is the multilayer perception, which consists of two inputs, ten neurons in one hidden layer, and one output. Figure 4 indicates a node and flow diagram in the ANN used [39]. The inputs of x1, …, xk to the neuron are multiplied by the weights w k i and summed up with the constant bias term of θi. The resulting ni is the input to the activation function g(n k i) below [39].
The ANN models were executed by using the two inputs of E and ΔE from each output of the FIS model with the threshold setup from −1 to +1. In addition, it had ten single nodes in the multilayer perception network as hidden layers within inputs and an output.

Simulation Model
The main approach in this study is to test the model of finding the optimized patterns more efficiently by combining the control characteristics and advantages of the FIS and ANN models described in Section 2.1. This is to analyze the signals generated by the two models with specific intervals, keeping the indoor thermal comfort level at given conditions, but choosing the one with less energy consumption. If the amount of error locates in the lower 25%, which means "inefficiency", in 4 times tracking by the 15 min interval, the signal is sent into the ANN model. If not, the controller sends the signal into the FIS model to modify the signal. In Figure 5, this flow is indicated in the simulation blocks of signal merging. In this model, there are some assumptions to facilitate the entire process. The room model is equipped with one central HVAC system using a dual duct. The air pressure variations of indoor air speed and the air leakage and airflow between the walls, roof, floor, and duct systems in the zone are neglected. Figures 1-3 and Table 1 describe the design framework and the model's geometry used for the entire simulation. In addition, except for some fixed values such as external work (0 W*m −2 ) and air velocity (0.1 m*s −1 ), the values changed as users' characteristics and built environments were derived from some preferred references [34][35][36].

Thermal Comfort
Through the simulations, controlling heating and cooling air supply were investigated in accordance with human factors, which caused some unexpected changes in the PPD results. As indicated in Figures 6 and 7, there were quite different characteristics of control results. In both normal and abnormal conditions for the On/Off model, except for the daytime when the thermal system is completely turned off, the model has consistent patterns in PPD because it simply repeated the task of turning off when it met model's Tset and turning on until it reached the Tset in most of the time range. However, in other models, such as FIS, ANN, and Comb, it can be seen that a steady flow of sensitive controls was taking place to meet the Tset.  Unlike the FIS and ANN, the Comb associated with the PPD model was kept below the set value of 5% of the PPD model. This advantage is more visible under the abnormal condition of Figure 7. Although it has failed to lower the level of PPD to 5% or less due to some fixed internal thermal conditions and human factors, it can be found that the average degree of PPD was maintained as low as possible compared to the FIS and ANN models. It can also be seen that the signal combinations were performed to find the optimized points of signals from the two models, as shown in Figure 7, at 14:00 and 19:00. Thus, the patterns of the two figures below is strong evidence that the simple structure of the Comb model, which compares optimized control points derived from the two models' algorithms and selects better results, can achieve high efficiency without having to build a separate algorithm. Figures 8 and 9 describe the typical control patterns of the On/Off model, whereas the FIS model in Figures 10 and 11, which has to determine the optimized heating and cooling supply for each situation in all simulated time intervals, shows a very complex and sensitive control patterns. Therefore, it would be advantageous to keep the Troom to a Tset, but there may be a disadvantage in maintaining a constant thermal comfort level during times when the heating and cooling air supply volume changes. Unlike these two models, the ANN model in Figures 12 and 13 smoothly controls the heating and cooling supply. In particular, in situations where the precision of the control sensitivity is significantly increased, the outdoor temperature is accurately predicted, even in the abnormal condition, to send an effective signal. This pattern may have quite huge advantages in maintaining indoor thermal comfort, as shown in Figures 6 and 7. The analogy in this situation may be that the on demand model is disadvantageous to maintaining the constancy of comfort, but it may have the advantage of mitigating the increase in energy consumption On the other hand, the ANN, as a predictive model, is advantageous for maintaining a constant indoor thermal comfort level, but it can be inferred that the inevitable increase in energy consumption will not be avoided.        Therefore, a model or algorithm that compensates for the weaknesses of the theses two models is necessary. In Figures 14 and 15, it was confirmed that the Comb model, which determined the adequacy of the signals at 15 min intervals, indicated a relatively low value of heating and cooling energy usage compared to the FIS model and indicated a complex graph pattern of the ANN model. It can be inferred that the optimized supply of heating and cooling energy required over time is working properly in a way that minimizes energy consumption and indoor thermal dissatisfaction. However, from the pattern of these graphs, it is difficult to ascertain how the actual energy consumption and the level of indoor comfort have changed. Therefore, the advantages and disadvantages of each model are compared through accurate numerical values.

Discussion
Tables 2 and 3 summarize the results of control and energy efficiencies using the Integral of Absolute Error (IAE), average of the PPD, and the Energy Use Intensity (EUI).  As indicated in Table 2, the FIS and ANN models show quite high efficiency in the control accuracy for the normal situation. However, in the case of abnormal situation, the FIS model worked inefficiently as compared to the On/Off model. Although the difference is not quite large, it can be inferred that the FIS, as an on-demand model in an unpredictable situation, shows some inefficiency. The ANN model as a predictive model shows relatively small errors even in abnormal situations. Therefore, it is confirmed that the predictive model has strengths in accurately responding to the Tout to maintain Tset. However, with the combination of the FIS and the ANN, it can be seen that the accuracy is greatly improved. The strength of this combined model is also found in the PPD, where unlike the accuracy shown by FIS and ANN models in control errors, there is no obvious advantage in keeping PPD low, whereas the combined model significantly lowers PPD levels by over 29%.
This aspect is more pronounced in the case of energy consumption. As indicated in Table 3, in normal situations where cooling is not required, both the FIS and ANN models used more energy than the On/Off model, the FIS model even consumed more than 50%. However, the combined model was able to achieve about 7.5% energy efficiency as it chose optimized control values meeting the minimum set conditions for the signals from the two models. This study tested whether both the control accuracy and energy efficiency could be maintained even in abnormal situations which were not previously learned.
As indicated in Table 3, in the normal situation, the FIS as an on-demand model consumed about 50% more energy, and the ANN as a predictive model consumed about 1% more energy than the On/Off model. However, it was confirmed that the combined model could lower about 13.0% of EUI. Even if the amount of heating energy decreased significantly, and the overall energy consumption of all the control models reduced, the energy savings of the combined model were confirmed to be more noticeable. In the abnormal situation, the results of the FIS and the ANN models indicated an increase in energy consumption of about 44.5% and 4.7%, respectively, in comparison with the On/Off model. However, the combined model indicated about 18.5% energy savings. As mentioned just before, given that 13.0% of EUI was saved under the normal situation, the combined model is found to have more efficient energy savings of about 5.5% points even under the abnormal situation.
Overall, when the two models were combined, it reduced the thermal dissatisfaction by min. 20.0% and max. 33.6% and improved the energy savings by min. 13.0% and max. 44.5% as compared to each independent model. The thermostat on/off controller can still be effective if productivity issues related to indoor control accuracy are combined with real energy savings, including the initial installation and maintenance costs. Therefore, a follow-up study will be conducted to validate the practical performance of the model used in this study in a more synthesized way reflecting control accuracy, indoor comfort, energy use, and productivity in built environment. Regarding this, the combined mode can be used in certain buildings or spaces, and large resources may be consumed to supplement or replace existing indoor thermal control systems such as intensive care units, biology labs, data and distribution centers.

Conclusions
This study investigated the performance of combining on-demand and predictive models as compared to each model, reflecting deterministic algorithms and learning results derived from past data. Each model has been validated in its own area, and advanced techniques in various fields have been adopted to improve its efficiency. However, there has been no clear discussion of the economics of using these models independently or creating new algorithms for better performance. In order to verify this performance, the control results of each model and a combined model were compared assuming abnormal conditions with data in quite a different pattern from those already used in learning. As a result, it was confirmed that the combined model of on-demand and predictive algorithms improved the control efficiency from min. 20.0% to max. 33.6% and energy efficiency from min. 13.1% to max. 44.4% as compared to each independent model. From this result, it can be confirmed that when inefficiency is verified in certain models, attempts to combine them in parallel and properly should precede to compensate for the weaknesses before attempting to improve their internal algorithms or structures. However, specific economic considerations should be made that the organization or design for combining existing models consume minimal resources. These viewpoints prescribe omissions for the various unpredictable physical conditions that may occur in the actual experimental model. As a follow-up study, a lab-scaled or a plug-in model which can be equipped on real buildings will be tested to define more accurate and reliable patterns of energy consumption and thermal indoor comfort to validate the actual effectiveness of the reference model used in this study.  Acknowledgments: This work was supported by the Dong-A University research fund.

Conflicts of Interest:
The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.