Forecasting the Energy Consumption of an Actual Air Handling Unit and Absorption Chiller Using ANN Models

Abstract

Air conditioning in buildings accounts for 60% of the total energy consumption. Therefore, accurate predictions of energy consumption are needed to properly manage the energy consumption of buildings. For this purpose, many studies have been conducted recently on the prediction of energy consumption of buildings using machine learning techniques. The energy consumption of the air handling unit (AHU) and absorption chiller in an actual building’s air conditioning system is predicted in this paper using prediction models that are based on artificial neural networks (ANNs), which simply and accurately allow us to forecast energy consumption with limited variables. Using these ANN models, the energy usage of the AHU and chiller could be predicted by collecting a month’s worth of driving data during the summer cooling period. After the forecast models had been verified, the AHU prediction model showed performance in the ranges of 13.27% to 15.25% and 19.42% to 19.53% for the training period and testing period, respectively, and the mean bias error (MBE) ranges were 4.03% to 4.97% and 3.48% to 4.39% for the training period and testing period, respectively. The chiller prediction model satisfied the energy consumption forecast performance criteria presented by American Society of Heating, Refrigerating and Air-Conditioning Engineers (ASHRAE) guideline 14 (the measurement of energy and demand savings), with a performance of 24.64~25.58% and 7.12~29.39% in the training period and testing period, respectively, and MBE ranges of 2.59~3.40% and 1.35~2.87% in the training period and testing period, respectively. When the training period and testing period were combined for the AHU data, the actual energy usage forecast showed a lower error rate range of 0.22% to 1.11% for the training period and 0.17% to 2.44% for the testing period. For the chiller data, the error rate range was 0.22% to 2.12% for the entire training period, but was somewhat higher at 11.67% to 15.18% for the testing period. The study found that, even if the performance criteria were met, high accuracy results were not obtained, which was due to the poor data set quality. Although the forecast model based on artificial neural network can achieve relatively high-accuracy results with sufficient amounts of data, it is believed that this will require a thorough verification of the data used, as well as improvements in the predictive model to avoid overfitting and underfitting, to achieve such good results.

1. Introduction

Buildings use a substantial amount of energy to provide a pleasant environment for occupants and residents. Statistics show that buildings consume about 35% of the world’s final energy and cause 75% of the world’s greenhouse gas emissions. Most of the energy consumed in buildings is used for air conditioning; lighting; water supply; and transportation during the heating, ventilation, and air conditioning (HVAC) system’s life cycle, with air conditioning accounting for 60 percent [1,2,3]. Therefore, the proper operation of various facilities is important to reduce energy consumption in buildings. The concept of a Building Energy Management System (BEMS) is being implemented to support energy and management efficiency at the operational level. A BEMS is an integrated system of measurement, control, management, and operation that provides optimized building energy management measures by monitoring the energy usage not only for efficient energy management but also to maintain a pleasant indoor environment.

The ability to predict energy consumption and demand is essential for optimizing energy performance from the design stage to the operational stage. The system’s design and the selection of appropriate facility capacities are determined during the design stage, and an optimal control plan and proper operations plan are established in the operational stage to improve the energy performance. Researchers have employed many techniques to predict energy consumption and demand, including machine learning, which has been used to predict power demand since the 1990s. One major type of machine learning model is the artificial neural network (ANN), which researchers have used extensively to investigate building energy predictions in various ways.

For example, Peng et al. proposed a combined model of two ANN models (Box and Jenkins) to predict cooling loads, with a less than 2.1% mean absolute percentage error [4]. Cheng et al. used an ANN model with input variables such as building envelope performance, parameters, and the heating degree day and cooling degree day to predict building energy consumption quickly and effectively, resulting in a forecast that was 96% more successful than the existing method [5]. Roldán et al. proposed a method for predicting short-term building energy consumption using an ANN-based time–temperature curve prediction model [6]. The input variables included temperature, type of day, etc., which can affect energy consumption. Testing in real buildings for over a year resulted in a high predictive accuracy [6]. Turhan et al. used an ANN model to predict the thermal load of an existing building based on the width/length ratio, wall overall heat transfer coefficient, area/volume ratio, total external surface area, total window area/total external surface area ratio of the building, etc., and compared the results with those obtained from a building energy simulation tool [7]. They observed a good correlation between the ANN model results and the building energy simulation tool results, with an average absolute percentage of 5.06% and a success forecast of 0.977 [7]. Ferrito et al. developed an ANN model that uses monthly building electrical energy consumption data [8]. The predicted accuracy of the developed model indicated that a root mean square percentage error of 15.7% to 17.97% could be obtained [8]. Ahmed et al. used weather data to compare and analyze the predicted power load performance of ANN and random forest (RF) models for a single building [9]. To improve the predictions, the ANN models achieved an average coefficient of variance of the root mean square error (CV(RMSE)) of 4.91% by extracting and removing normalized techniques and input variables, whereas the RF models produced an average CV(RMSE) of 6.10% by changing the tree depth [9]. Li et al. proposed ANN-based forecasting methods that could quicken the prediction of the energy consumption of complex types of buildings in the initial design phase [10]. They proposed a method to simplify a complex type of building into several blocks. As a result, the relative deviation of heating and cooling energy consumption was within ±10%, and the relative deviation of total energy consumption was within 10 percent [10]. Ding et al. used ANN models and a support vector machine (SVM) in a study of prediction accuracy by combining eight input variables [11]. When K-means and hypersarchical clustering methods were used to obtain a combination of optimized variables, the accuracy was better than otherwise, and the historical cooling capacity data had the greatest impact on the predicted accuracy [11].

Koschwitz et al. predicted data-driven thermal loading using two nonlinear autoregressive exogenous recurrent neural networks (NARX RNN) of different depths and ε-SVM region models [12]. They used historical data from non-residential regions in Germany for training and testing to predict the monthly load. The evaluation results show that the NARX RNN was more accurate than the ε-SVM region model [12].

In short, various studies of machine learning methods, including ANN models, have been conducted in the field of building energy. Predictions of the energy consumption and cooling loads of buildings show a high prediction accuracy.

In a study on predicting air handling unit (AHU) energy consumption, Niu et al. conducted thermal energy consumption forecasts for AHU using AutoRegressive with eXternal inputs (ARX), State Space (SS), Subspace state space (N4S), and Bayesian Network (BN). All four models satisfied the criteria in the American Society of Heating, Refrigerating and Air-Conditioning Engineers (ASHRAE) guideline 14, and Bayesian Network’s forecast results were the most accurate [13]. Le Cam et al. used a closed-loop nonlinear ANN model to predict the supply fan power demand of an AHU. During the test period, the Root Mean Squared Error (RMSE) was predicted at 5.5% and the CV (RMSE) at 17.6%. Fan electricity demand was predicted to be 1.4 kW RMSE, CV(RMSE) 30%, over 6 h. A sensitivity analysis shows that reducing the size of the training data set from 23 days to 4 or 8 days does not adversely affect the RMSE values [14].

In addition, a study of chiller plants with an absorption chiller showed that Adnan et al. used ANN to model the baseline electrical energy use of chiller system, providing a higher predictive performance when ANN is over 93% R, and showing small error values for mean square error (MSE) and mean absolute percentage error (MAPE) for the selected ANN structure [15]. Lazrak et al. used ANN to forecast energy consumption under the weather conditions of the solar combisystem combined with an absorption chiller, with annual energy forecast errors mostly showing less than 5%. As such, it can be seen that ANN techniques are valid for predicting energy consumption for AHUs and chillers. However, there were not many studies of individual components of HVAC systems for energy consumption forecasts in buildings [16].

This research team is developing a centralized air conditioning system and energy management technique for BEMS applications and has researched energy consumption and load predictions based on ANN (Artificial Neural Networks). Using ANN models, the team conducted a prediction study of chiller energy consumption and achieved results that satisfy the American Society of Heating, Refrigerating, and Air-Conditioning Engineers (ASHRAE) criteria, with an average CV(RMSE) of 19.49% in the training period and an average CV(RMSE) of 22.83% in the testing period [17]. In addition, the team conducted studies to optimize the cooling load prediction model based on MATLAB’s NARX (with eXigenous) Feedforward Neural Networks model to confirm that a forecast accuracy of less than 7% CV(RMSE) can be obtained depending on the conditions [18]. The accuracy of the ANN prediction model was verified through prior research, whose results were based on the data generated by simulation programs. In this study, the prediction accuracy of the ANN prediction model was investigated using both the air handling unit (AHU) and absorption chiller operations data from an actual building.

2. ANN-Based Air Handling Unit and Absorption Chiller Prediction Models

In this study, the energy consumption of AHUs and absorption chiller during the hvac system was predicted using ANN-based forecasting models and the operation data of the hvac system collected from office buildings.

2.1. ANN Model for Predicting Energy Consumption

In this study, the NARX Feedforward Neural Networks model included in MATLAB’s Natural Networks Toolbox (R2018b) was used to predict energy consumption. This NARX model has been validated to show accurate results in predicting time series values [19,20,21]. The NARX Feedforward Neural Networks model consists of an input layer, a hidden layer, and an output layer. Outside conditions, the operation conditions of equipment, seasonality data, and historical energy input data are used as input values for neural network learning in the input layer. The input signal of the input layer is received in the hidden layer, and the neural network operation is performed through the internal neuron. In this study, the number of hidden layers and neurons was handled differently; the output layer releases the predicted energy consumption results. A schematic of the ANN model is shown in Figure 1.

2.2. Collection of AHU and Absorption Chiller Operational Data

The data used to forecast results via ANN models are real-world data obtained from an office building with 18 floors and a total floor area of 41,005.32 square meters. The refrigeration equipment is equipped with two absorption chiller units with a capacity of 600 USRT (U.S. Refrigeration Ton), but only one chiller unit is operational and, therefore, only one chiller is used in this study. Eight constant air volume air conditioners are used in the business facilities.

The research team predicted the energy consumption of one absorption chiller and one air conditioner. Only one air conditioning unit could be used to measure the amount of electricity, because power meters were not installed in all the air-conditioning systems at the time of data collection. In August 2019, data collected for 15 min were used, and the energy consumption predictions were carried out using 860 data points that eliminated the missing values of the air-conditioning facilities.

3. Prediction Condition for ANN Model

3.1. Input Values

For the purpose of predicting energy consumption, the input values of the absorption chiller were the ambient air conditions, dry bulb temperature, relative humidity, cold water supply temperature, and inlet coolant temperature. The AHU’s input values were similarly selected as the ambient air conditions, dry bulb temperature, relative humidity, and supply air temperature. Seasonality data for both the absorption chiller and the air conditioner used the year, month, day, and time.

3.2. Structural and Learning Parameters for ANN Models

The structural and learning parameters must be set for accurate prediction results. The structural parameters are the hidden layer that forms the structure of the ANN and the neuron that exists inside it where the actual learning takes place; the learning capacity is determined by the number of neurons. The learning parameter—i.e., learning rate—determines the amount of learning and ensures that weighting parameters are updated after learning with a single amount of learning. Epoch, which is a unit of learning, indicates that one time passes the entire data set completely once.

Table 1. Structural and learning parameters for ANN models.

Division	Condition
Number of hidden layers	1
Number of neurons	5
Learning rate	0.0001
Epochs	50

presents details of the conditions.

3.3. Preprocessing and Training Size

In this study, missing values were removed from the data set and the data used as input values were normalized to obtain more accurate forecasting results. Non-operating hours, weekends, and holidays resulted in missing values for these times when the HVAC system was inoperable. The energy consumption was predicted using 860 data points per month in August. All the learning data were then normalized to have values within the same range and converted to values between 0 and 1. Of the 860 preprocessed data points, the training data size changed to 50 (430 points), 60 (516 points), 70 (602 points), 80 (688 points), and 90 (774 points)%, to a comparative evaluation of the results of the forecast of energy consumption due to changes in the amount of training.

3.4. Performance Evaluation Indicators for Predictive Models

The results of the energy consumption predictions include the American Society of Heating, Refrigerating and Air-Conditioning Engineers’ (ASHRAE) data, U.S. Department of Energy Federal Energy Management Program (FEMP) data, and International Performance Measurement and Verification Protocol (IPMVP) data.

Table 2. Acceptable calibration tolerances.

Calibration Type	Index	ASHRAE Guidelines 14 [22]	FEMP [23]	IPMVP [24]
Monthly	MBE monthly	±5%	±5%	±20%
	CV(RMSE) monthly	15%	15%	-
Hourly	MBE hourly	±10%	±10%	±5%
	CV(RMSE) hourly	30%	30%	20%

shows the ASHRAE, FEMP, and IPMVP measurement and verification protocols for the energy management of buildings that set the criteria for the prediction accuracy of a BEMS to predict building energy performance. In this study, the coefficient of variation of the root mean square error (CV(RMSE)) and mean bias error (MBE) were used as the performance indicators. The CV(RMSE) represents the fragmentation of estimates in consideration of the variance. The MBE is an error analysis index that identifies errors by tracking how closely the estimates cluster to the target through the bias of the data.

4. Results and Discussion

4.1. Accuracy of Prediction Models

Figure 2 presents the CV(RMSE) and MBE data for each training data size of the AHU energy consumption forecast results. The CV(RMSE) is a minimum of 19.42% and a maximum of 19.53% for the training period; the predicted accuracy meets the ASHRAE criterion of 30% for all conditions. The MBE of the AHU energy consumption forecast is shown to be at least 4.03% for the training period and 4.97% for 90% of the training period. The MBE for the testing period is the lowest at 3.48% when the training size is 90% and the highest at 60% at 4.56% in the testing period. The MBE data also show a predictive performance that meets the ASHRAE criteria under all conditions.

Figure 3 presents data regarding the prediction accuracy of the absorption chiller model. The CV(RMSE) of the training period is the lowest at 24.64% for the training size of 50% and the highest at 25.58% for the training size of 60 percent. For the testing period, the lowest CV(RMSE) is 27.12% with the training size of 80% and the highest is 29.39% with the training size of 50 percent. The ASHRAE criterion was satisfied for all conditions. The MBE results likewise satisfy the ASHRAE criteria with 2.59% to 3.40% for the training period and 1.35% to 2.87% for the testing period.

The accuracy of the AHU and absorption chiller energy consumption prediction models used in this study was found to be somewhat high in terms of the CV(RMSE) and low in high variance-low bias form. High variance-low bias means that an error occurred in the testing data due to over-fitting the training data [25]. However, both models satisfy the ASHRAE guidelines and were considered suitable for use in the study.

4.2. Energy Consumption Forecast Results

Figure 4 and

Table 3. Air handling unit energy consumption forecasting results: error rate and standard deviation.

Training Size (%)	Error Rate (%)
	Average		Standard Deviation
	Training	Testing	Training	Testing
50	6.03	11.24	28.82	38.96
60	8.46	15.20	30.60	46.47
70	8.92	10.87	30.31	25.33
80	9.65	14.76	30.03	30.12
90	8.08	12.82	26.53	22.53

provide a times series summary of the results of the energy consumption forecast for the AHU. The AHU energy consumption forecasts under all conditions show similar trends. The average error rate ranges of 6.03% to 9.65% in the training period and 10.87% to 15.20% in the testing period result from organizing the predicted data and error rate by time steps of 15 min. The standard deviations show relatively high values of 26.53~30.06 for the training period and 22.53~46.47 for the testing period. The predicted results by time step show an error rate of less than 1% in most sections, but the average of the error rate is somewhat higher due to the point at which the error rate was sometimes more than 100%, thereby increasing the standard deviation.

Figure 5 and

Table 4. Error rate of the total air handling unit energy consumption forecasting results.

Training Size (%)	Error Rate (%)
	Training	Testing	Total
50	1.11	0.17	0.47
60	0.67	1.36	0.95
70	0.22	0.42	0.28
80	0.85	0.68	0.82
90	0.85	2.44	1.01

present the results of the total energy consumption forecast with the training period and testing period combined. The error rate is 0.22~1.11% in the training period and 0.17~2.44% in the testing period. The error rate is 0.28% to 1.01% for the total energy consumption when the training period and testing period are combined. Unlike the error rates shown in the individual forecast results, the error rates across both the training period and testing period are shown to be significantly lower. The ANN model used in this study is considered to be suitable for predicting data over a certain period of time, but it is disadvantageous for predicting small intervals.

Figure 6 and

Table 5. Absorption chiller energy consumption forecasting results: error rate and standard deviation.

Training Size (%)	Error Rate (%)
	Average		Standard Deviation
	Training	Testing	Training	Testing
50	19.73	27.38	27.30	15.02
60	22.08	27.41	26.55	24.38
70	23.45	26.49	28.62	14.97
80	21.29	25.95	18.31	15.18
90	24.35	28.54	20.90	12.76

present a time series summary of the results of the energy consumption predictions for the absorption chiller. The trends of the results are similar under all conditions. An assessment of the predicted results for each 15 min time step and error with the measured data indicate an average error rate range of 19.73% to 24.35% in the training period. The standard deviations range from 18.31 to 28.62. For the testing period, the average error rates range from 25.95% to 28.54% and the standard deviations from 12.76 to 24.38. The error in the results for the predictions of the micro-sections is greater for the chiller than for the AHU.

Figure 7 and

Table 6. Error rate of the total absorption chiller energy consumption forecast results.

Training Size (%)	Error Rate (%)
	Training	Testing	Total
50	0.78	15.18	6.39
60	2.12	12.69	3.18
70	0.22	14.17	3.58
80	0.92	11.67	1.29
90	2.07	14.69	0.60

present the results for the total energy consumption forecasts for the training period and testing period. The error rate range for the total energy use in the training period is 0.22~2.12%, indicating the results of the forecast. However, the predicted results of the testing period show high error rates of 11.67% to 15.18 percent. The error rate range is 0.60~6.39% when the training period and testing period results are combined. The performance of the absorption chiller energy consumption prediction model satisfies the performance indicators, CV(RMSE) and MBE, in accordance with the ASHRAE guidelines. However, the error rate appears to be as high as the variance in the prediction results.

5. Conclusions

In this study, the energy consumption of the AHU and absorption chiller in an actual building was forecast using prediction models that are based on ANNs. The performance of the ANN prediction models that are based on training size was evaluated by collecting a month’s worth of driving data during the cooling period.

The performance of both the AHU model and the absorption chiller model was verified in this study. The results for the AHU prediction model show that the CV(RMSE) ranged from 13.27% to 15.25% for the training period and 19.42% to 19.53% for the testing period. The MBE ranged from 4.03% to 4.97% for the training period and 3.48% to 4.39% for the testing period. The results for the absorption chiller prediction model show that the CV(RMSE) ranged from 24.64% to 25.58% for the training period and 27.12% to 29.39% for the testing period. The MBE ranged from 2.59% to 3.40% for the training period and 1.35% to 2.87% for the testing period. These results satisfy ASHRAE guidelines.

As a result of predicting energy consumption using actual data, the AHU model showed an error rate of more than 10% and a relatively high standard deviation in the testing period by time step. However, when the training period and testing period results were combined, the error rate ranged from 0.22 to 1.11% in the training period and 0.17~2.44% in the testing period. The absorption chiller model showed a high error rate range of 19.73% to 28.54% per time step. The total energy consumption of the training period showed an error rate ranging from 0.22 to 2.12%, but the error rate was somewhat higher at 11.67% to 15.18% in the estimation of the total consumption of the testing period.

When the energy consumptions of both the AHU and absorption chiller were predicted together, the predicted results by time step showed a high error rate, whereas the predicted results for the training and testing periods combined showed a low error rate. The prediction models used in this study were more effective in forecasting over a certain time interval than for smaller sections. The predicted performance indicators, CV(RMSE) and MBE, satisfied the performance criteria of ASHRAE and showed a modest error rate in predicting energy usage over the entire period, indicating that ANN-based prediction models had adequate predictive performance.

However, it seems that errors occurred because the models used in this study had high variance in the form of high variance-low bias, which was somewhat higher than the CV(RMSE) at the test period and the error rate at maximum of 15.18% at the test period when predicting the absorption chiller. In addition, the data used in the study were collected shortly after the BEMS was installed in the target building, and the accuracy was insufficient. Therefore, a high error is thought to have occurred due to the inaccuracy of the data itself.

However, despite the poor quality of the data, some predictive accuracy has been confirmed, and the performance of the ANN-based predictive model has been verified. In order to achieve better results, it is deemed necessary to improve the predictive model and verify the data set to avoid overfitting and underfitting. To improve the predictive model, studies will be conducted such as optimizing the parameters of the NARX feedforward neural network model for improved accuracy when there is not enough data, and comparing them with other machine learning models. To improve the quality of the data set, more HVAC operation data will be collected and the correlation with the energy consumption is analyzed to verify the data set through various methods. such as the analysis of the results of the correlation of the input value and cross validation methods.