Accurately predicting heat transfer performance of ground-coupled heat pump system using improved autoregressive model

Related work

With the flourish of machine learning and big data technology, some data-mining based methods, such as ANN and SVM have been applied to predict the energy transfer performance. Zhao, Zhong & Zhang (2016) applied ANN and SVM models to do energy consumption predicting of variable refrigerant volume (VRV) system in office buildings. Le Cam, Daoud & Zmeureanu (2016) presented an application of the process of knowledge discovery in databases (KDD) using nonlinear ANN for the forecasting of the electrical power demand of a supply fan of an AHU (air handling unit). Ceci, Corizzo & Fumarola (2017) tackled the problem of power prediction of several photovoltaic (PV) plants and indicated that regression trees provided better models than artificial neural networks on two PV power plant datasets. Ceci et al. (2020) also proposed an unsupervised change detection method which is able to analyze streaming data generated by sensors located in smart grid. Yan, Hu & Li (2016) studied the performance prediction of ground source heat pump (GSHP) systems by real-time monitoring data and data-driven models. They used back-propagation neural network (BPNN) algorithm to establish the data-driven models. Candanedo, Feldheim & Deramaix (2017) presented and discussed data-driven predictive models for the energy use of appliances. They trained four statistical models: multiple linear regression, support vector machine with radial kernel, random forest and gradient boosting machines (GBM) to predict energy used in a low-energy house. Xiao, Wang & Cheng (2017) explored the potential application of advanced data mining techniques for effective utilization of big building operational data. Deep learning-based prediction techniques, decision tree and association rule mining were adopted to analyze the operational data. Bedi & Toshniwal (2018) proposed an empirical mode decomposition method combined with the long short-term memory network to estimate electricity demand for the given season, day, and time interval of a day. Mathioulakis, Panaras & Belessiotis (2018) introduced artificial neural networks for the performance prediction of heat pump hot water heaters, which verified that a trained ANN could represent an effective tool for the prediction of the air-to-water heat pump (AWHP) performance in various operation conditions and the parametrical investigation of their behavior. Xia, Ma & Kokogiannakis (2018) presented a model-based design optimization strategy for ground source heat pump systems with integrated solar photovoltaic thermal collectors (GSHP-PVT). The highlight of this paper was an artificial neural network (ANN) model was used for performance prediction and a genetic algorithm (GA) was implemented as the optimization technique.

Some data mining methods were also applied to forecast the performance of ground-coupled heat pump system. Esen, Inalli & Sengur (2008b) tried to improve the performance of an artificial neural network with a statistical weighted pre-processing method to learn to predict ground source heat pump systems with the minimum data set. They used three work conditions as input layer, while the output was coefficient of performance of system. Besides, they reported on a modeling study of ground coupled heat pump system performance by using a support vector machine method (Esen, Inalli & Sengur, 2008d). Wang et al. (2013) proposed an application of artificial neural networks based on improved Radial Basis Function (NNCA-RBF) to predict performance of a horizontal ground-coupled heat pump system. Roberto et al. (2021) propose a Tucker tensor decomposition, capable of extracting a new feature space for forecasting the multiaspect renewable energy.

These mentioned methods only use either time series information to make a time series prediction or external inputs to simulate a regression function. This paper proposes an autoregressive model with working condition inputs to accurately predicting heat transfer performance of ground-coupled heat pump system.

Data

A quantitative validation of heat transfer performance data analysis has been obtained according to GCHP software, which is developed by ZhongRui New Energy Science & Technology Co., Ltd. 5,000 working conditions of solar-assisted GCHP system with vertical GHE during 30 years (360 months) is simulated, and the detailed description of the input data (working conditions variables) can be found in our previous work (Zhuang, Ben & Yan, 2017).

Working conditions variables

The 30 years’ heat transfer performance data including heat exchange ability, temperature of circulating liquid, inlet and outlet temperature of heat pump are simulated. Table 1 shows the detailed characteristics of input parameters. In Table 1, numerical types (Type for short) of the data include real and nominal numbers. The minimum (Min. for short) and maximum (Max. for short) are the smallest and largest values of a set of an input variable in 5,000 working conditions, respectively. The average (Avrg. for short) is the mean value of a set of an input variable in 5,000 working conditions, denoting a measure of central tendency. Standard deviation (Std. for short) is a measure used to quantify the amount of variation or dispersion of a set of data values. Range measures the difference between the Max. and Min. in a set. The median (Med. for short) can be found by arranging all the observations from lowest value to highest value and picking the middle one.

Heat pump performance output data

The solar-assisted GCHP system for 360 months are simulated and the heat transfer performance Y₁–Y₄ are recorded. Y₁ denotes the monthly power consumption of heat pump. Y₂ denotes the monthly power consumption of whole system, including heat pump and water pump. Y₃ denotes the ratio of the heating capacity or the refrigerating capacity of heat pump to the operating power of heat pump. Y₄ denotes the ratio of the heating capacity or the refrigerating capacity of heat pump to the total operating power of system. The overall tendencies of Y₁–Y₄ are given in Fig. 1.

Seasonal factor decomposition

The relevant temperature data, strongly influenced by the weather, is a seasonal periodical time series. Generally speaking, seasonal time series have the following components: tendency (T), cycle (C), seasonality (S), irregular variations (I). Additive seasonal decomposition model can be expressed as: Y_t = T_t + S_t + C_t + I_t and a term-by-term separation is as follows:

Autoregressive model with working condition inputs

For an output variable Y, a time serie can be defined as $\left\{y_t,t=1,2,...,T\right\}$, under a working condition parameter x. T denotes the length of time series, and x denotes a d-dimensional vector of working condition inputs. The aim is to construct the appropriate model and to predict its future values by using historical data of Y as well as these working condition parameters (X₁–X₁₂).

The predictive value ${\hat{y}}_t$ for the moment t can be obtained by general form of autoregressive (AR) model:

(4) $${\hat{y}}_t=f(y_{t-1},y_{t-2},...,y_{t-M}),$$

where M is a lagged term. Therefore (4) indicates that the historical data of Y can predict its future values. Furthermore, these exogenous variables such as working condition parameters to the autoregressive (AR) model is introduced, therefore

(5) $${\hat{y}}_t=f(\mathbf{x},y_{t-1},y_{t-2},...,y_{t-M}).$$

Formula (5) indicates that the future values ${\hat{y}}_t$ are predicted by not only the historical data of Y but also the working condition parameters and this is autoregressive model with working condition inputs.

The next step is to construct Random Forests according to the data of the working conditions and time series. Random Forests are constituted by Classification And Regression Tree (CART). Firstly, m attributes from all (d + 1) × M + d attributes are randomly extracted and the optimal attributes should be selected from m attributes to be split. Then the subtrees are generated iteratively until attributes cannot be split. Ultimately, a regression tree is formed. The optimal split property selection is based on the minimum variance of the two sub-nodes that the property is divided into. In terms of attribute set A of a note, one of the attributes (a) is regarded as split attribute to split into two nodes D₁, D₂, and the optimal split attribute is as follows:

(6) $$a^{\ast }=\arg \underset{a\in A}{min}[V(D_1)+V(D_2)],$$

where V(D₁), V(D₂) are variances of D₁, D₂ respectively.

By using the method described above, K classification and regression trees are generated, then the random forest is constructed. The average of regression K results is the predicted value for the moment t.

(7) $${\hat{y}}_t={\displaystyle \frac{1}{K}\sum _{k=1}^K{f}_k(\mathbf{x},y_{t-1},y_{t-2},...,y_{t-M}),}$$

where f_k(·) denotes the regression result of a regression tree k.

Error measurements

Three error indicators including Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE) are used in this paper to measure the prediction accuracy. N donates the number of time series and T denotes the length of a time series predicted. ${\hat{y}}_t^{(i)}$ and $y_t^{(i)}$ donates the predicted value and the observed value for the moment t of the i-th time series, respectively.

(8) $$\mathrm{M}\mathrm{A}\mathrm{E}={\displaystyle \frac{1}{\mathrm{T}\mathrm{N}}\sum _{t=1}^T\sum _{i=1}^N|{\hat{y}}_t^{(i)}-y_t^{(i)}|}$$

(9) $$\mathrm{M}\mathrm{A}\mathrm{P}\mathrm{E}={\displaystyle \frac{100}{\mathrm{T}\mathrm{N}}\sum _{t=1}^T\sum _{i=1}^N\left|{\displaystyle \frac{{\hat{y}}_t^{(i)}-y_t^{(i)}}{y_t^{(i)}}}\right|}$$

(10) $$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{1}{\mathrm{T}\mathrm{N}}\sum _{t=1}^T\sum _{i=1}^N{({\hat{y}}_t^{(i)}-y_t^{(i)})}^2}}$$

MAE, which is the rate of absolute value of difference between predicted value and observed value and the number of predicted data, reflects the average distance of predicted value’s deviation. MAPE is another accuracy measurement in statistics and indicates the proportion of predicted value in observed value, reflecting the relative error. RMSE is square root of ratio, which can be calculated by square sum of difference between predicted value and observed value and the number of predicted data. RMSE indicates the extent of deviation and is sensitive to particularly large or particularly small errors in predicted values.

Seasonal factor decomposition results and analysis

The trend item, seasonal factor itemcycle item and stochastic error item for Y₁–Y₄ are given in Figs. 2–5. As for Y₁, the trend item shows a considerably downward trend during the period. From seasonal items it has obvious seasonal characteristic. In practical terms, the power consumption of heat pump drops year by year, while the working state of it changes with different seasons. The cycle item of implies that the time series is periodical and it can be seen that its value is small and its influence is weak. Stochastic error item reflects some uncertain factors of series. It has a small amount of fluctuation with 0, and the effect is very small. Y₂, Y₃ and Y₄ have similar features with Y₁. The periods of Y₁, Y₂, Y₃ and Y₄ are all 12 months.

Time series analysis results and analysis

In time series analysis, there are 5,000 series in Y₁, Y₂, Y₃ and Y₄ respectively and each series consists of 360-month data, which means 5,000 × 360 dimensions. For the prediction of the value in future 60 months, the model is trained by the historical data of the first 300 months. We trained our model i.e. autoregressive model with working condition inputs (ARX for short) at each time step by using random-forest library in Matlab. It is note that we utilize the earlier predicting results as known data to predict later values at test stage. Combined with the results of seasonal decomposition of output data in “Seasonal Factor Decomposition Results and Analysis”, the lagged term in regression model should be set to an integral multiple of one period. If the setting is too small (i.e., one period (12 month)), the regular pattern of time series is not clearly caught. On the contrary, if the setting is too large, the number of random trees participating in the calculation will increase, which makes the model more complex. Therefore, the value of the lagged term M should be more than one period, and we tested two smaller values (M = 24, and M =3 6) in the ARX model. Table 2 lists the MAE results. The lower MAEs are obtains under M = 24. So the lagged term in regression model is set as 24 months. The sliding window is same as lagged term which is 24. We compare several time series analysis methods, including Exponential Smoothing (ES), Autoregressive Model (AR), Autoregressive Moving Average Model (ARMA) and Auto-regressive Integrated Moving Average Model (ARIMA). Exponential Smoothing (ES) is one of moving average methods which is characterized by applying different weights to the observed values of the past. The weight of short-term observed value is greater than the long-term. The basic idea is that predicted value is a weighted sum of the observed values with different weights (new data with greater weight). ARIMA regards the data series formed by forecasting object with the change of time as the random series and sets certain model to approximately describe this series. Once the model is recognized, the time series’ past value and present value can be applied to forecast its future value. Similarly, AR and ARMA can be set according to ARIMA. In detail, the values of both difference and moving average are 0 in AR, while only the value of difference is 0 in ARMA. The smoothing factor of ES is 0.1. For ARIMA, the auto-regressive p, the integrated term (differential order) d and the moving average lag q are set as 12, 1 and 12, respectively. And the parameters of the others methods is same with ARIMA, i.e., AR (p = 12) and ARMA (p = 12, q = 12).

In order to measure overview errors of the predicted 60-month value, we average 60-month errors of Y₁, Y₂, Y₃ and Y₄ respectively (Table 3). Compared with ES, AR, ARMA, ARIMA, Autoregressive model with working condition inputs (ARX) has prominent advantage with smallest mean absolute error, mean absolute percentage error and root mean square error. In “Seasonal Factor Decomposition Results and Analysis”, according to decomposition of seasonality, four outputs are stationary time series. Therefore, the large error can be witnessed in moving average method, which is the reason of low precision in ES. ARX added 12 working condition parameters to predict values, which lack in traditional time series analysis. The results of experiment show that adding working parameters can improve the predicted precision of time series. This is because that four output variables including the power consumption of heat pump, the power consumption of system, the ratios of the heating capacity (or the refrigerating capacity) of heat pump to the operating powers of heat pump and to the total system, respectively show different tendencies under various working parameters. In other words, it is limited to simply use historical data to predict the future values and ignore the influence of system parameters, such as hole, u-tube and collector. In practical working situation, working states of the system are definitely different with different parameters. The prediction of ignoring these parameters is only general and does not take the particularity of different systems into account. In terms of model, the introduction of these additional parameters has a similar effect on the prediction model with a window limit. Prediction result cannot deviate from its output restrictions of particular parameters. Ultimately, the prediction curve will be adjusted in a better direction.

In order to prove the positive effects of our prediction model, we draw the error graphs of Y₁, Y₂, Y₃ and Y₄ in 60 months with using MAE as an example respectively (see Fig. 6). First of all, obvious periodicity can be witnessed in MAE, periodic value of which is 12 (1 year), meeting the periodical requirement of practical system. In addition to the ES method, prediction accuracy shows a downward trend because the model predictions of heat pump and system’s electricity consumption become inaccurate as time goes on. As for this problem, the Autoregressive model with working condition inputs model well embodies the adjustment function of the external parameters, though not completely eliminate the accuracy decreasing trend, but the performance is much better. Taking seasonality into consideration, we find that the working state of heat pump in every month is regular. Therefore, between-year is used to predict, which means the prediction of a month is based on the data of the same month over the past years. In practice, the system is closed in April and October without electricity consumption. The predictions of these 2 months are not 0 in ARIMA, so the model is unrealistic. Nevertheless, Autoregressive model with working condition inputs has no prediction error in these months is in accordance with the practical circumstances. In spite of introducing extra parameters, the between-year-based model avoids deviating prediction from practical situation.