A Novel Fractional Hausdorff Discrete Grey Model for Forecasting the Renewable Energy Consumption
Reducing carbon dioxide emissions and using renewable energy to replace fossil fuels have become an essential trend in future energy development. Renewable energy consumption has a significant impact on energy security so accurate prediction of renewable energy consumption can help the energy department formulate relevant policies and adjust the energy structure. Based on this, a novel Fractional Hausdorff Discrete Grey Model, abbreviated FHDGM (1,1), is developed in this study. The paper investigates the model’s characteristics. The fractional-order of the FHDGM (1,1) model is optimized using particle swarm optimization. Subsequently, through two empirical analyses, the prediction accuracy of the FHDGM (1,1) model is proven to be higher than that of other models. Finally, the proposed model is applied with a view to forecasting the consumption of renewable energy for the years 2021 to 2023 in three different areas: the Asia Pacific region, Europe, and the world. The study’s findings will offer crucial forecasting data for worldwide energy conservation and emission reduction initiatives.
The world economy is expanding, and so is energy consumption. Renewable energy consumption, as a part of global energy consumption, has continuously increased, reaching 2.9 exajoules in 2020. Indeed, it has become an essential part of energy consumption. The International Energy Agency (IEA) states although oil and other liquid fuels will remain the world’s leading energy source in 2050, renewable energy consumption, including solar and wind energy, will increase by 50% . Spencerdale, the chief economist of the BP group, said renewable energy, as part of total 2020 global power generation, had exhibited the fastest growth in history. This increase was mainly due to the decline in the proportion of coal power generation. The strong growth of renewable energy and the gradual replacement of coal align with the world’s need to transition to net-zero emissions . The accurate prediction of renewable energy consumption is therefore of great significance for both the government and the energy sector in their quest to adjust current energy structures and deploy resources.
In recent years, more and more scholars have focused on the prediction method of renewable energy consumption. There are three commonly used renewable energy prediction models: the first is the statistical model [3, 4]; the second is the intelligent predict model [5–7]; the third is the grey prediction model [8, 9]. The establishment of statistical models and intelligent prediction models requires a lot of data, but renewable energy, as a newly emerging energy in recent years, has a limited amount of data. As the grey prediction model is famous for its characteristics of “less data and higher prediction accuracy,” this paper chooses the grey prediction model to forecast renewable energy consumption.
In the 1980s, in order to solve the problem of data with the characteristics of “ limited sample and poor information,” Prof. Deng first proposed the grey system theory. Subsequently, numerous fields, such as agriculture, the economy, the environment, and others, have extensively used this theory. As it was the first model, the GM (1,1) model is the basic model of the theory and, since it was put forward, many scholars have used it. For instance, Zeng et al. employed GM (1,1) to analyze the functional status of the human kidney. They found that, even though the renal function index (serum creatinine) was within the normal range, human renal function could nevertheless be abnormal . For practical applications, however, it was later determined that the GM (1,1) model’s prediction accuracy had to be increased. Hence, an increasing number of researchers have optimized the GM (1,1) model to improve its predictive accuracy [11–19].
In the early stages of the development of the grey system theory, first-order accumulation generation was used for all raw data preprocessing. However, the first-order accumulation generation operator cannot reflect the importance of new information in the original data. In order to solve this problem, scholars have proposed different accumulation generation operators. Wu et al. used the least-square method perturbation theory to demonstrate how the new information priority principle of the grey system theory was broken in this processing technique. Then, he proposed a new fractional-order accumulation and built a new grey prediction model FGM (1,1) . Ma et al. introduced the conformable fractional accumulation and difference based on the definition of conformable fractional derivative. Then, they came up with a new conformable fractional grey model . In 2020, Chen et al. created a brand-new Fractional Hausdorff Grey Model (FHGM (1,1)) by incorporating the fractional Hausdorff derivative into the grey model . Shen et al. proposed a grey prediction model based on weighted fractional order accumulation generation operator and applied it to the prediction and analysis of the total output value of China’s construction industry . With the introduction of a new accumulation generation operator, a series of new grey prediction models have been established and applied to practice, further enriching grey system theory [24–28].
In the study of the above grey prediction model, the estimated value of the model parameters is calculated based on the basic form (difference equation), while the time response function of the model is derived based on the whitening equation (differential equation). The jump from the difference equation to the differential equation will inevitably lead to errors in the model, which reduces the accuracy of the model in some applications. To solve this problem, in 2005, Xie and Liu proposed the discrete grey prediction model (DGM (1,1) model), which formally unified the model parameter estimation and the time response function, fundamentally eliminating the jump between the difference equation and the differential equation of the grey prediction model . Inspired by their work, many scholars have further studied the model and put forward a series of new discrete grey prediction models on this basis, all of which have greatly complemented the research content and application scope of the discrete grey prediction model [30–32].
Based on the research results of the literature mentioned above, this paper proposes a new Fractional Hausdorff Discrete Grey Model and utilizes the model to study the consumption of renewable energy. The main contributions of this paper can be exacted below: (1)The FHGM (1,1) model and the discrete model were combined to reduce errors in the FHGM (1,1) model discretization process(2)We investigate the properties of the FHDGM (1,1) model(3)To find the final optimal fractional-order , we create an optimization problem and utilize the particle swarm optimization algorithm to determine it(4)Two examples demonstrate the strong predictive qualities of the FHDGM (1,1) model, with the suggested model used to forecast renewable energy consumption over the following three years
The rest of this paper is structured as follows: Section 2 discusses the modeling mechanism of the proposed model. Section 3 describes the properties of the FHDGM (1,1) model in detail. Section 4 introduces the process of using PSO to find the optimal order of the FHDGM (1,1) model. Two numerical examples are used to verify the performance of the FHDGM (1,1) model in Section 5. Section 6 uses the newly proposed model to simulate and forecast renewable energy consumption. The conclusions are given in Section 7.
2.1. The Definitions of the Hausdorff Fractional Accumulation and Difference
Before introducing the definitions of the Hausdorff fractional accumulation and difference, the Hausdorff fractional derivative is given as shown in Definition 1.
Definition 1 (see ). The Hausdorff fractional derivative is defined as Based on Definition 1, the definitions of the Hausdorff fractional accumulation and difference are given in Definition 2 and Definition 3, respectively.
Definition 2 (see ). The Hausdorff fractional accumulation (HFA) with order is
Definition 3 (see ). The Hausdorff fractional difference (HFD) with order is
2.2. Traditional FHGM (1,1) Model
Definition 4. Assume is a nonnegative sequence; its -order accumulative sequence is then given by , so the differential equation of the FHGM (1,1) model is where .
Definition 5. The basic form of the FHGM (1,1) model is where , and .
Theorem 7. Let the initial value be ; then, the time response function of the FHGM (1,1) model is and the restored values are , and .
Obviously, Eq. (4) is a continuous equation, and Eq. (5) is a discrete equation. The time response function of the model is obtained by directly substituting the parameters obtained from Eq. (5) into Eq. (4). The model’s performance in terms of prediction will be reduced since the prediction of the time response formula derived using this method will result in inaccuracies in the model’s predictive results. To solve the problem, this paper proves the main reasons for the unstable simulation and accuracy of the FHGM (1,1) model through model derivation. In doing so, this paper constructs the FHDGM (1,1) model, which solves the error caused by the FHGM (1,1) model changing from a difference equation to a differential equation.
2.3. Fractional Hausdorff Discrete Grey Model
Definition 8. Let and be the same as Definition 4; the equation is called the Fractional Hausdorff Discrete Grey Model (FHDGM (1,1) for short).
Theorem 9. Using the least-square method, the parameters are estimated by where
Proof. When we enter into Eq. (9), we get . After that, the error sequence is obtained with by changing to . Let The parameters that make the S minimum should satisfy Recalculating the equation system, we have where Owing to Substituting Eq. (16) into Eq. (14), we get Quod erat demonstrandum (Q.E.D).
Theorem 10. By assuming and as shown in Theorem 9 and letting the initial condition be . Then, the time response sequence of the FHDGM (1,1) model is given as
The restored value of is
Proof. If , .
If , .
3. Properties of the FHDGM (1,1) Model
Property 11. Assume is the original nonnegative sequence, is the multiple transformation sequence of , then, , where is a multiplied amount.
Property 12. Let and be the same as Property 11. Assuming the parameters in FHDGM (1,1) modeled by are and that are parameters in FHDGM (1,1) modeled by , then, .
Proof. Parameters satisfy , where Parameters satisfy , where Therefore, Q.E.D.
Property 13. If , , respectively, formed the FHDGM (1,1) model, and if are the restored values of those models, then, .
Proof. When , we have When , we have Q.E.D.
Property 14. Consider that are, respectively, the relative errors of the FHDGM (1,1) model built by , , then, .
From Properties 11 to 12, we see that the original data sequence can affect both the structural parameter and the restored value by multiplying it by a certain quantity, but it cannot affect the structural parameter and the relative error . As a result, when modeling with data of a large order of magnitude, we can perform data multiplication transformation on the original data to reduce the order of magnitude of the data for modeling without affecting the model’s ability to predict outcomes accurately and to streamline the modeling procedure.
Property 15. The initial value of the FHDGM (1,1) model is invalid.
Proof. If is defined as the disturbance of the initial value , then, , . If set are the parameters of the FHDGM (1,1) model constructed by , then, , where
Therefore, if , then
If , then
The fact that the restored value of the FHDGM (1,1) model is unaffected by the initial value change, as shown by Property 15, suggests that the initial value of the FHDGM (1,1) model is invalid.
Lemma 16 (see ). Suppose , , and are the generalized inverse of matrix , the column vectors of are linearly independent, and . Solutions to equations min and min are defined as and , respectively. If , then
Property 17. Assuming to be perturbation bound when is the disturbance of . If , , and , in which represents the generalized inverse of matrix , then, we have
Proof. When and are defined as, respectively, the disturbance of and its perturbation bound of the solution, then Therefore,
Based on Lemma 16, we have
When and are defined as, respectively, the disturbance of and its perturbation bound of the solution, then
When and are defined as, respectively, the disturbance of and its perturbation bound of the solution, then
According to Property 17, the disturbance bound is an increasing function depending on the sample . As increases, gradually increases, which leads to system instability. In other words, similar to the GM (1, 1) model and the FGM (1,1) model, the newly proposed FHDGM (1,1) model is also suitable for predicting small sample problems.
4. Seeking the Optimal Fractional-Order for FHDGM (1,1)
For a new model, it is very significant not only to explore its modeling mechanism but to also seek its optimal fractional-order. Referring to previous papers in the literature, searching for the optimal fractional-order is commonly carried out by constructing a simple constrained optimization problem; its formula is
It should be pointed out that since Eq. (46) has nonlinear characteristics, it is difficult to solve directly by ordinary methods. PSO is used to determine the optimal fractional-order . The concrete flowchart of PSO to seek the optimal fractional-order of the FHDGM (1,1) model is shown in Figure 1. The specific values of PSO are listed in Table 1.
5. Validation of the FHDGM (1,1) Model
5.1. Model Evaluation Metrics
Several judgment methods are used to evaluate the suitability of the novel model; the absolute percentage error (APE) and the mean absolute percentage error (MAPE) are chosen.
By defining as Eq. (47), APE is used to illustrate the absolute percentage error between the real value and the predicted value. The smaller the APE values, the better the predicted accuracy.
MAPE is another widely used metric to measure the overall difference between the real value and the predicted value. By defining Eq. (48), the smaller the MAPE values, the better the predicted accuracy.
5.2. Case 1: Prediction of Coal Consumption in the United States
Identical information can be found in this example from the literature . Different grey models are constructed using the data from 2010 to 2015 (in-sample data), and these are used to forecast the values from 2016 to 2018 (out-of-sample). Four models were examined, with the actual and forecast values shown in Table 2.
Table 2 shows that FHDGM (1,1) produced the lowest MAPE in-sample data and out-of-sample, which are 3.884% and 4.029%, respectively. The MAPE of the total data from 2010 to 2018 is 3.938%, which is also the smallest MAPE of the four models. This indicates that the FHDGM (1,1) model is better at predicting coal consumption in the United States.
5.3. Case 2: Prediction of the Natural Gas Consumption in North America
We look at a case in the literature  that offers sample data. Different grey models are constructed using the data from 2010 to 2015 (in-sample data), and the values from 2016 to 2018 (out-of-sample) were forecast. The actual and forecast values of the four models were then compared.
From Table 3, FHDGM (1,1) has the lowest MAPE in-sample data, out-of-sample, and total data, 0.220%, 2.732%, and 0.974%, respectively. This indicates that FHDGM (1,1) is better at predicting natural gas consumption in North America.
5.4. Case 3: Prediction of Natural Gas Consumption in Thailand
We look at another case in the literature  that offers sample data. Different grey models are constructed using the data from 2008 to 2017 (in-sample data), and the value of 2018 (out-of-sample) was forecast. Table 4 displays the actual and forecast values of the four models that were compared.
From Table 4, we see that FHDGM (1,1) has the lowest MAPE in-sample data, out-of-sample, and total data, 0.994%, 1.367%, and 1.031%, respectively. This indicates that FHDGM (1,1) is better at predicting natural gas consumption in Thailand.
6. Application of the FHDGM (1,1) Model
In this section, the new model that we propose is utilized to forecast renewable energy consumption in the Asia Pacific region, in all of Europe and globally to evaluate its applicability to other competing models such as the FDGM (1,1) model , the CFGM (1,1) model , and the DGM (1,1) model . The data came from the BP Statistical Review of World Energy 2020, which can be downloaded at https://www.bp.com/en/global/corporate/energy-economics/statistical-review-of-world-energy/renewable-energy.html. Different grey models are constructed using the data from 2008 to 2017, with the accuracy of the models tested using data from 2018 to 2020. Based on this, the structural chart of forecasting renewable energy consumption can be drawn in Figure 2.
6.1. Renewable Energy Consumption of the Asia Pacific Region
In order to fit and forecast the consumption of renewable energy in the Asia Pacific region, the suggested FHDGM (1,1) model is compared with three conventional grey models, DGM (1,1), FDGM (1,1), and CFGM (1,1). The whole data set is split into two sections: data from 2008 to 2017 (in-sample periods), which are used to evaluate grey model accuracy at fitting data, and data from 2018 to 2020 (out-of-sample periods), which are used to assess their accuracy at forecasting data. Second, PSO determines the parameters for FHDGM (1,1), FDGM (1,1), and CFGM (1,1). The minimum MAPE and corresponding parameters are listed in Table 5, with both the track searching for the minimum MAPE, as well as the optimal parameters, graphed in Figure 3.
From Table 8 and Figure 5, the four grey model MAPE values in the in-sample period are 1.68%, 1.78%, 1.87%, and 1.97%, respectively, while the competing MAPE values in the out-of-sample period are 3.39%, 3.53%, 4.21%, and 3.58%. From a comparison of these MAPE values, it is apparent that FHDGM (1,1) has the smallest MAPE value in fitting and predicting the renewable energy consumption of the Asia Pacific region. Therefore, the new model performs well with respect to fitting and forecasting the Asia Pacific region’s renewable energy consumption.
6.2. Renewable Energy Consumption for All of Europe
The total of European renewable energy data, as it pertains to consumption, is broken down into two categories: in-sample period and out-of-sample period. The raw data from 2008 to 2017 is utilized to test how well the FHDGM (1,1) model fits; out-of-sample period: the raw data from 2018 to 2020 is also applied to verify the prediction accuracy of the FHDGM (1,1) model. A comparison is made between the model and the three other models, FHDGM (1,1), FDGM (1,1), and CFGM (1,1). Similarly, the parameters of FHDGM (1,1), FDGM (1,1), and CFGM (1,1) are decided by PSO, with the minimum MAPE and corresponding parameters listed in Table 9. The tracks of both searching for the minimum MAPE and the optimal parameters are given in Figure 6.
Table 12 and Figure 8 show that the MAPE values of the four grey models are 1.90%, 1.74%, 1.79%, and 4.92% in describing total European renewable energy consumption; the MAPE values of the competitors are 1.24%, 2.99%, 1.70%, and 8.61%, respectively, for predicting renewable energy consumption in Europe. It can be seen that, although the MAPE of the FHDGM (1,1) model is not the smallest in the in-sample period, it is very close to the minimum MAPE 1.74%. The MAPE of the FHDGM (1,1) model is the smallest in the out-of-sample period, which is lower than that of the other three grey models. In general, the FHDGM (1,1) model is better able to predict when compared to the other three grey models.
6.3. Renewable Energy Consumption of the World
Four prediction models are developed using data on global renewable energy consumption from 2008 to 2017; the performance of the models is tested using data from 2018 to 2020. Similar comparisons are made between the model and the FHDGM (1,1), the FDGM (1,1), and the CFGM (1,1). PSO then determines the parameters of FHDGM (1,1), FDGM (1,1), and CFGM (1,1). The minimum MAPE and corresponding parameters are listed in Table 13; the track of searching for the minimum MAPE and the optimal parameters are shown in Figure 9.
From Table 16 and Figure 11, we see that the MAPE values of the four grey models are 0.46%, 0.51%, 0.52%, and 1.42% in the “in-sample” period. The competitor MAPEs are 0.93%, 0.94%, 0.98%, and 3.05% in the “out-of-sample” period. The MAPE value of FHDGM (1,1) is the lowest among these grey models, which also verifies that the predictive ability of the FHDGM (1,1) model is higher than that of the other models. Moreover, the maximum APE of the FHDGM (1,1) model is 1.26%, which is less than the 1.32% of the FDGM model, the 1.56% of the CFGM (1,1) model, and the 5.53% of the DGM model. That is to say, the proposed new model, FHDGM (1,1), outperforms all the other competitors because of its lowest fitting and the predictive ability of the MAPE values.
6.4. Forecasting Renewable Energy Consumption from 2021 to 2023
Since the novel FHDGM (1,1) model proposed in this paper performs well in predicting renewable energy consumption, we employ the FHDGM (1,1) model to predict renewable energy consumption from 2021 to 2023. Considering the principle of new information priority, we use the data on renewable energy consumption from 2008 to 2020 to establish the FHDGM (1,1) model and make a three-step prediction to obtain the prediction data from 2021 to 2023. The values of the optimal parameters are shown in Table 17, and the prediction results are shown in Table 18 and Figure 12.
As can be seen from the above analysis, the Asia Pacific, European, and global consumption of renewable energy will continue to grow over the next three years, with annual increase rates shown in Table 19. Apparently, the annual growth rate of renewable energy consumption in the Asia Pacific region and the world in the next three years will exceed 10%, while the annual growth rate in Europe will exceed 5%. This demonstrates that, as global energy growth moves forward, the coal-based energy system will continue to change and renewable energy sources will take on a more significant role. This is because both the government and the energy industry have pointed towards boosting the use of clean and renewable energy to decrease carbon dioxide emissions, lessen the burden on the environment, and hasten the transition to a low-carbon and green economy.
Accurate estimates of renewable energy consumption can be used to inform stakeholders and decision-makers about the worldwide renewable energy supply. This has significant practical implications and will assist in adjusting and enhancing the energy structure. Based on this, Hausdorff fractional-order accumulation is introduced into the discrete model, and a new model FHDGM (1,1) is proposed to predict renewable energy consumption. We then not only examine the features of the FHDGM (1,1) model parameters but also assess their impact on modeling precision using a multiplication transformation. We discover that the prediction performance of the FHDGM (1,1) model is unaffected by the multiplication transformation of modeling data. Subsequently, the optimal order of the FHDGM (1,1) model is sought by PSO. Finally, two numerical examples show that the FHDGM (1,1) model has good predictive performance. Using the FHDGM (1,1) model to forecast renewable energy consumption shows that renewable energy consumption will continue to increase over the next three years and that the annual growth rate will be more than 5%. With the increasing use of renewable energy, carbon emissions pressure will be reduced.
Although the FHDGM (1,1) model has decent prediction performance, there is still some opportunity for improvement. This paper does not consider some important factors that affect renewable energy consumption, which may reduce the predictive accuracy. In future research, we will add influencing factors to construct a multivariable grey prediction model. We will also consider combining the model with different intelligent optimization methods, including the ant lion algorithm , the whale optimizer , the moth-flame optimization algorithm , and other intelligent optimization algorithms, to improve the prediction accuracy of the FHDGM (1,1) model and enrich grey system theory.
The data used to support the findings of this study have been deposited in https://www.bp.com/en/global/corporate/energy-economics/statistical-review-of-world-energy/renewable-energy.html.
Conflicts of Interest
The authors declare that they have no conflicts of interest.
This research was supported by the Projects of General Program of Humanities and Social Sciences from Ministry of Education (20YJC630083) and Scientific and Technological Project in Henan Province of China (no. 222102110104).
L. Liu, “EIA predicts that world energy consumption will increase by nearly 50% in 2050,” Oil and Gas and New Energy, vol. 5, p. 57, 2021.View at: Google Scholar
Y.-W. Liu, “BP world energy statistical yearbook, 2021 edition: the energy market has suffered a huge impact,” China Petroleum and chemical industry, vol. 8, pp. 32-33, 2021.View at: Google Scholar
J. Cui, Y.-G. Dang, and S.-F. Liu, “Novel grey forecasting model and its modeling mechanism,” Control and Decision, vol. 24, no. 11, pp. 1702–1706, 2009.View at: Google Scholar
M.-Y. Tong, X.-H. Zhou, and B. Zeng, “Optimization of background value in grey NGM(1, 1, k) model,” Control and Decision, vol. 32, no. 3, pp. 507–514, 2017.View at: Google Scholar
Z. Liu, Y.-G. Dang, and L. Wei, “Optimization of background value and time response function in NGM(1,1,k),” Control and Decision, vol. 31, no. 12, pp. 2225–2231, 2016.View at: Google Scholar
J. Cui, Y.-G. Dang, and S.-F. Liu, “Study on morbidity of NGM (1, 1, k) model based on conditions of matrix,” Control and Decision, vol. 25, no. 7, pp. 1050–1054, 2010.View at: Google Scholar
D.-W. Li, M.-L. Qiu, J.-M. Jiang, S.-P. Yang, School of Mathematics and Statistics, Huizhou University, Huizhou 516007, China, and School of Mathematics and Statistics, Baise University, Baise 533000, China, “The application of an optimized fractional order accumulated grey model with variable parameters in the total energy consumption of Jiangsu Province and the consumption level of Chinese residents,” Electronic Research Archive, vol. 30, no. 3, pp. 798–812, 2022.View at: Publisher Site | Google Scholar
N.-M. Xie and S.-F. Liu, “Discrete GM(1,1) and mechanism of grey forecasting model,” Systems Engineering -- Theory & Practice, vol. 25, no. 1, pp. 93–99, 2005.View at: Google Scholar
L.-F. Wu, S.-F. Liu, and L.-G. Yao, “Discrete grey model based on fractional order accumulate,” Systems Engineering -- Theory & Practice, vol. 34, no. 7, pp. 1822–1827, 2014.View at: Google Scholar
G. H. Golub and C. Loan, Matrix Computations, Johns Hopkins University Press, 1996.