Demand-side adjustable resources unified data model study

Abstract

With the large-scale grid integration of new energy sources, the power relationship between source and demand side is becoming more and more complex, the flexibility requirements of the distribution network are increasing, and new requirements for system safety and reliability are put forward, and it can’t meet the peak regulating requirements of grid by relying solely on source side to cope with variable loads. In view of this, to make the grid operation flexibility improved by using an adjustable load capacity, the study first constructs a load classifying method on the foundation of fuzzy style K-plane clustering method to understand the interaction information of the load devices. Then the adjustable value of the load is analyzed from the demand response and standby perspectives, respectively, and an improved dynamic time-bending-based source-load similarity inscription method is proposed, which aims to unify the multiple load information obtained by clustering. The proposed clustering algorithm takes the highest value of 0.096 for the Davies-Bouldin index index, which is 0.012 and 0.014 higher than the K-plane clustering algorithm and the K-means algorithm, respectively. In addition, the load demand response regulation with the improved dynamic time bending method has a higher capacity for new energy consumption than the variance method, with a difference of 2.0 kW. This indicates that using load regulation to consume new energy to exploit the load curve will stimulate the active participation of customer-side load in maintaining power balance between the electricity consumption side and the demand side, and form an interest community between power supplying and demanding sides.

Keywords

Renewable energy demand side clustering median dynamic time warping standby capacity

Abbreviations

DTW	Dynamic Time Warping
S-KPC	fuzzy style K-plane clustering
M-DTW	Median Dynamic Time Warping
DBI	Davies-Bouldin index

Notations

$N$	the total samples number in dataset $X$
$k$	the total samples number in $X$

$\mu^{m}_{ij}$	$i$ th sample’s fuzzy affiliation in $X$ contained in the $j$ th cluster
$x_{i}^{s}$	the original sample $x_{i}$ generated after $s$ stepwise dimensional expansion
$A_{j}$	$j$ th cluster’s style matrix
$\{{\omega_{j},b_{j}}\}$	the jth cluster plane discriminant function parameter
$\lambda$	the regularization parameter, and $I$ is the unit matrix
$\xi_{j}$	the minimum eigenvalue
$D_{j}$	the K-plane clustering learning parameter
$U({:,j})$	the column vectors in fuzzy subordinate matrix $U$
$e$	the column vector with each element being 1
$\Lambda$	the diagonal matrix
$v_{i}$	the equivalence transformation matrix
$\overline{XY}$	the average difference between $X$ and $Y$
$C$	the similarity index
$\Delta XY$	the median distance between two time series curves’ corresponding points
$L({i,j})$	the construction matrix

1. Introduction

While the proportion of renewable energy sources increase, some of the problems brought about will gradually emerge [1]. The output condition of new energy sources will be affected by many uncertain factors, so it does not reflect the output stability of conventional units, for example, wind power and photoelectricity, which account for the most, will be directly affected by wind speed and light intensity, thus making their output less stable than that of generating units, with strong volatility and uncertainty [2]. With the continuous progress of intelligent technologies, the traditional power supply service is gradually transitioning to a multi-source convergence-based power supply service model [3, 4]. Due to the continuous change of load characteristics, the impact on flexible adjustable loads at the customer side is also increasing. Taking demand as the driving mode and the adjustable load as the entry point, it is an important part of the adjustable power system to give full play to the active function of the energy side in the new energy consumption scenario, to effectively suppress the volatility and intermittency of new energy, and to achieve the maximum consumption of new energy. Demand-side adjustable load can not only utilize its own elasticity, but also provide auxiliary services such as peaking, frequency regulation and standby for users, which has great application prospects [5]. In response to the challenge that it is difficult to achieve unified scheduling because of the huge scale and wide distribution of adjustable loads, load aggregators can gather adjustable loads together and contract with power companies to achieve unified scheduling of adjustable loads.

In view of this, this study will innovatively propose an analysis method of power consumption data of load equipment based on S-KPC algorithm, and propose an improved dynamic time bending method to describe the similarity between load and new energy output, so as to realize the regulation of demand-side adjustable resources.

This study is divided into the following four parts, the second part is a literature review about load equipment power consumption data or load demand response, the third part is about the construction of a clustering model for load equipment and the analysis of a unified model for demand response data, the fourth part is a utility analysis of the proposed model, and finally a summary conclusion about the effectiveness of the method.

2. Related work

In the clustering of load equipment electricity consumption data, many scholars have carried out rich researches. Ge et al. [6] proposed a new electric load forecasting method which considered the load characteristic, industries and production patterns. The new forecasting algorithm is able to separate out production modes with high forecasting accuracy. Si et al. [7] summarized some measurements and classes in electric load clustering and their advantages and disadvantages, and discussed in depth important applications as well as future trends. Sharma et al. [8] proposed a particle swarm algorithm based on grey wolf optimization to balance the load as well as energy efficiency optimization and experimental results showed high optimization performance. Jeong et al. [9] used an autoregressive integrated moving average model and K-means clustering to predict the peak electrical loads of university buildings. The method can also be implemented in demand response to reduce electricity bills by avoiding electricity consumption during high tariff hours. Han et al. [10] proposed a forecasting method, which can effectively extract similarities in residential loads and accurately perform residential load forecasting at the individual level.

Many scholars have conducted detailed studies on load demand response analysis. Wang et al. [11] designed a model-free deep reinforcement learning method with a deep Q network structure for optimal smart grid disaster recovery management under time-of-use tariffs and variable electricity consumption patterns. It verified that this method deals with the noise and instability found in conventional algorithms and makes peak load and costs reduced while regulating voltage to safe limits. Waseem et al. [12] presented the development of air conditioning systems and their impact on electricity demand, outlining possible disaster recovery scenarios. Then, a review of control techniques and disaster recovery plans for air conditioning systems to manage electricity consumption in the residential and commercial energy sectors is presented. Bahrami et al. [13] proposed a demand response algorithm for residential customers, taking into account the uncertainty in load demand and electricity prices, privacy issues of customers, and tidal constraints imposed by the distribution network. The results show that the proposed demand response algorithm can reduce the peak load by 33% and the expected cost per customer by 13%. Singh et al. [14] investigated how the demand response schemes effect the optimal microgrids dispatch. An elasticity model was used to describe the actual behavior of customers for electricity price changes. Finally, various indices were evaluated and priorities were assigned based on multi-criteria evaluation techniques. Al Hadi et al. investigated a simple and effective demand response method to mitigate the intermittency problem and provide uninterrupted power supply to the customers. The renewable energy is used to maximum extent. Peak demand and costs can be reduced with little carbon emission [15].

In summary, scholars mainly focus on the design of load forecasting and demand response algorithms, but fewer scholars consider the establishment of a unified data model of demand-side resource devices. This may result in a very complex and inefficient demand response calculation process, which cannot meet the required response time requirements of the system. In view of this, this study will design a grid control system model with adjustable load clusters participating in demand response from the perspective of load demand response regulation. This work will effectively reduce the operating costs of the power system while effectively consuming new energy, and meet the demand for adjustable resources in the development of the power system.

3. Clustering of load equipment based on clustering algorithm and its demand unification model construction

To classify the load devices and to analyze the characteristics of power consumption habits of various load devices, a S-KPC algorithm is introduced to cluster the load devices’ power consumption data in this experiment. In addition, in order to evaluate the value of demand response of loads in the consumption of new energy output, the M-DTW method is proposed to characterize the effectiveness of load regulation in the consumption of new energy.

3.1 Clustering of load equipment power consumption data based on S-KPC

Flexible loads are flooding into power system, various loads exhibit multiple operating conditions due to various reasons such as intrinsic reasons and equipment characteristics, and mining and analyzing the power consumption of these loads helps to analyze and master different loads’ power consumption habits, and to get the adjustable load equipment information uploaded to the demand response terminal, this study proposes A clustering method based on S-KPC of load equipment electricity consumption data information.

For different levels of load information, the application scope varies, and the selection of load characteristic indexes will have a greater impact on it. When load information is collected, data distortion or loss is inevitable due to communication interruptions, signal interference and other reasons. In the case of missing data, the mean interpolation strategy can be used to unfold the data complement, i.e., the mean value of the data at the two neighboring positions in the missing data is used as a substitute value for the missing position.

If the neighboring positions are also missing values, then they are found, calculated and filled in a sequential forward or backward progression. If the feature vector boundary has been found, but it is still a null value, then 0 is used to fill it. Equation (1) is the expression of the objective function of S-KPC.

$\displaystyle\left\{{\begin{array}[]{l}\mathop{\min}\limits_{\mu_{ij},A_{j},% \omega_{j},b_{j}}J_{\textit{S-KPC}}=\sum\limits_{i=1}^{N}\sum\limits_{j=1}^{k}% \mu_{ij}^{m}({({x_{i}^{s}})^{T}A_{j}\omega_{j}+b_{j}})^{2}+\lambda\sum\limits_% {j=1}^{k}{\|{A_{j}-I}\|_{F}^{2}}\\ s.t.\sum\limits_{j=1}^{k}{\mu_{ij}=1,\|{\omega_{j}}\|^{2}=1}\\ \end{array}}\right.$ (1)

In Eq. (1), $N$ refers to the total samples number in dataset $X$ , $k$ refers to the total samples number in $X$ , $\mu^{m}_{ij}$ means $i$ th sample’s fuzzy affiliation in $X$ contained in the $j$ th cluster. $x_{i}^{s}$ refers to the original sample $x_{i}$ generated after $s$ stepwise dimensional expansion, $A_{j}$ refers to $j$ th cluster’s style matrix, $\{{\omega_{j},b_{j}}\}$ is the jth cluster plane discriminant function parameter. $\lambda$ represents the regularization parameter, and $I$ is the unit matrix.

We can get the solution from the perspective of the eigenvalue problem. As such, the Lagrangian of (1) can be expressed as:

$\displaystyle{L}^{\prime}(\omega_{j},b_{j},\xi_{j})=\sum\limits_{i=1}^{N}\sum% \limits_{j=1}^{k}\mu_{ij}^{m}({({x_{i}^{s}})^{T}A_{j}\omega_{j}+b_{j}})^{2}-% \sum\limits_{j=1}^{k}{\xi_{j}({\omega_{j}^{T}\omega_{j}-1})}$ (2)

Compared with traditional clustering methods such as distance and density, S-KPC is a clustering method for mining type data from an unsupervised perspective, which can learn not only the physical attributes of sample information, but also the style types. An objective function is established based on the dual-knowledge representation of the data, i.e., a step-by-step dimensional expansion of the characteristics of the data samples using a style matrix, and the algorithm is able to finely discern the subtle differences within data styles.

After setting the partial derivatives of the Lagrangian of (2) to be zero, namely letting $\partial$ L// $\partial\omega j=$ 0 and $\partial$ L’/ $\partial bj=$ 0, we have:

$\displaystyle D_{j}\omega_{j}=\xi_{j}\omega_{j}$ (3)

In Eq. (3), $\xi_{j}$ denotes the minimum eigenvalue. Equation (4) is the iterative rule expression.

$\displaystyle b_{j}=\frac{U({:,j})^{T}({X^{s}})^{T}A_{j}\omega_{j}}{\sum% \limits_{i=1}^{N}{\mu_{ij}^{m}}}$ (4)

Where $D_{j}$ is the K-plane clustering learning parameter, and the calculation expression is as follows:

$\displaystyle D_{j}=\frac{A_{j}^{T}({X^{s}})^{T}U({:,j})U({:,j})^{T}X^{s}A_{j}% }{e^{T}U({:,j})}-A_{j}^{T}({X^{s}})^{T}\Lambda X^{s}A_{j}$ (5)

In Eq. (5), $U({:,j})$ is all the column vectors in fuzzy subordinate matrix $U$ , $e$ refers to the column vector with each element being 1, and $\Lambda$ refers to the diagonal matrix.

The update rule can be obtained in Eq. (6).

$\displaystyle\mu_{ij}=\frac{1}{\sum\limits_{{j}^{\prime}=1}^{k}{\left({\frac{(% {X^{s}})^{T}A_{j}\omega_{j}-b_{j}}{({X_{i}^{s}})^{T}A_{j}^{\prime}\omega_{j}^{% \prime}-b_{j}^{\prime}}}\right)}^{\frac{2}{m-1}}}$ (6)

Once the sum fuzzy affiliation matrix U is determined, the style matrix’s iterative formula can be obtained in Eq. (7).

$\displaystyle A_{j}=I-\frac{1}{\lambda}\sum\limits_{i=1}^{N}{\mu_{ij}^{m}v_{i}% x_{i}^{s}\omega_{j}}$ (7)

In Eq. (7), $v_{i}$ denotes the equivalence transformation matrix. Figure 1 shows the flow chart of the algorithm execution.

Figure 1.

Algorithm execution flowchart.

The input processed load data set, i.e., typical daily load consumption curve, the clusters $k$ , the largest dimensional expansions of the augmentation nodes $S$ , the most cycles for S-KPC algorithm learning and style matrix learning $H$ , and the iterating terminal threshold $\varepsilon$ . The parameters initialized include $U$ , $A_{j}$ , $\omega_{j}$ , and $b_{j}$ , as well as the style matrix $\lambda$ , and Gaussian kernel function’s kernel width to generate the augmentation nodes $\sigma$ . Figure 2 shows the block diagram of the process of clustering analysis of load device data.

Figure 2.

Flow chart of cluster analysis of load equipment.

First, the raw information is preprocessed, which includes the analysis of anomalous information and the analysis of typical electricity consumption curves of load devices, and then the feature vectors on the time scale are obtained using averaging. On this basis, the number of categories k is selected based on the Calinski-Harabaz principle, and these numbers of categories are input into the classification algorithm to classify them. Finally, a comparison with several other methods is made.

3.2 Uniform model construction of adjustable load data based on source-load similarity

In order to enhance the aggregation value of adjustable load clusters, we take “source-load-business” fusion as the entry point to build a unified data model for different devices, so as to build a fused grid-connected dispatching model. To enhance the aggregation value of adjustable load clusters. Figure 3 is a schematic diagram of an aggregation architecture in which demand-side adjustable resource clusters participate in demand response.

Figure 3.

Schematic diagram of aggregation architecture of demand-side adjustable resource cluster participating in demand response.

On the generation side, the main body of this system is new energy generation and conventional unit generation. Most of the electricity it generates is used to meet the daily electricity demand of local customers, while the rest, which is used as an adjustable resource, is sent out to the grid company. However, unlike the conventional grid model, in the real grid model in the region, the load agent first signs a power purchase contract with the power trading market, then signs a contract with the grid dispatch and control center, and finally forms an orderly power consumption model with the customer as the core, and signs a contract with the customer to achieve centralized control of the customer. This approach can increase customers’ use of new energy sources, thus increasing the equipment’s ability to regulate the power system. On this basis, the power consumption in the contract is optimized by optimizing several aspects such as regulation periods, demand response volume, and tariff incentives. Figure 4 shows the optimization process of the load cluster standby aggregation model.

Figure 4.

The optimization process of load cluster backup aggregation model.

In practice, to ensure the effectiveness and reliability of system operation, the allocation of adjustable loads in the power system is usually done in an agency manner. In the power market, in order to ensure the control of the power market, load agents sign contracts with customers and power companies in the power market and gain control of the power market. Figure 2 shows that in the electricity market, in order to maximize their revenues, load agent aggregators sign contracts with both parties for electricity consumption/electricity consumption, thus exerting some influence on electricity consumption and electricity consumption. The price of electricity purchased in the power trading market, the dispatch strategy developed with the dispatch center, and the orderly consumption strategy developed with the customers are important influences on their interests. The game between grid companies and grid companies focuses on whether to purchase low-priced electricity and obtain high-priced electricity; the demand for information exchange is mainly reflected in the game with users, which mainly revolves around the adjustable amount of user load curve, adjustable time period, and the price of electricity for users. In addition, it is also possible to purchase low-priced electricity from the grid company at night and discount the electricity consumption of customers who participate in night dispatch, so as to attract more customer loads to participate in operational backup [16].

In order to unify the clustering information of multiple load devices, this study proposes a unified model of adjustable load demand data based on the similarity of source loads. Similarity can be used to describe how close the load curve is to the reference line pattern. Euclidean distance, general normative distance, etc. can be used as an important means to calculate the similarity. In general, variance can weigh the dispersion of random variable or data sets. If the variance is low, the dispersion of two data sets’ difference is low, which indicates that the fluctuation scenarios of two data sets are similar. Based on this, a population-based synergistic difference matrix is proposed to quantify the distribution characteristics of the groups between tasks, and the degree of similarity between the groups is described by the difference of the distribution characteristics of the groups and the distance of the distribution characteristics of the groups. When variance weighing the similarity, the first step is to make two time series curves normalized, so $X=({x_{1},x_{2},\ldots,x_{n}})$ , $Y=({y_{1},y_{2},\ldots,y_{n}})$ can be obtained. The second step is to calculate the variance difference between these curves after normalizing in Eq. (8).

$\displaystyle C=\frac{1}{n}\sum\limits_{i=1}^{n}{({x_{i}-y_{i}-\overline{XY}})% ^{2}}$ (8)

In Eq. (8), $\overline{XY}$ refers to the average difference between $X$ and $Y$ . $C$ refers to the similarity index. The $C$ is smaller, these curves are more similar. The classical Euclidean distance is limited to the magnitude order of two-time series and cannot precisely characterize the morphological similarity. DTW uses the time domain features of curvature to precisely characterize the degree of similarity between time domain features. However, the DTW method requires the Euclidean distance to construct the distance matrix, and there is a great difference in order of magnitude between the two compared with new generating energy, so researchers put forward an improved M-DTW. The new energy daily output and load curves are normalized separately, and a matrix consisting of two time series curves has been obtained $D$ , as shown in Eq. (9).

$\displaystyle D({i,j})=\sqrt{({x_{i}-y_{i}-\Delta XY})^{2}}$ (9)

In Eq. (9), $\Delta XY$ is the median distance between two time series curves’ corresponding points. Equation (10) is the expression of cumulative distance calculation.

$\displaystyle L({i,j})=D({i,j})+\min\left\{{\begin{array}[]{l}L({i-1,j-1})\\ L({i-1,j})\\ L({i,j-1})\\ \end{array}}\right.$ (10)

In Eq. (10), $L({i,j})$ represents the construction matrix. Find the best fold path that minimizes the cumulative distance from the start point to the end point, and construct the matrix reflecting the cumulative distance by iteration $L({i,j})$ .

4. Clustering model of demand-side adjustable resource data and utility analysis of unified model of load regulation

The unique advantages of the S-KPC method in terms of clustering characteristics are examined by comparing the S-KPC method with other methods. On this basis, the effectiveness of the method in the application of new energy consumption is examined by comparing the variance method with the M-DTW method, with a view to promoting efficient synergy between the source and the demand side and ensure the reliable power supply of the power system during peak hours.

4.1 Performance analysis of S-KPC clustering algorithm

The load equipment used in the study is the power consumption curve of industrial and agricultural load equipment at the minute level for 30 days. The curves with zero operating power for whole day are excluded from 30 curves. Then the remaining curve data are averaged, and what is obtained is load equipment’s typical power consumption curve at minute. Then, the average of the remaining curves is averaged to obtain the hourly typical power consumption curves, as shown in Fig. 5.

Figure 5.

Electricity consumption curve.

In Fig. 5, the aggregated curve tends to be more linear in shape, with an average power per hour between (12, 16) kW and a high power period from 13:00 to 14:00. The clustering algorithms selected for the study were accuracy, $F$ -value, and DBI. The comparison algorithms are KPC and K-means. Figure 6 shows the accuracy curves of the different algorithms during training and testing.

Figure 6.

Accuracy curves of different algorithms during training and testing.

From Fig. 6a, the proposed clustering algorithm achieves the optimal accuracy curve on the training samples. In terms of the accuracy starting value, the clustering algorithm takes a value of 95.2%, which is higher than 90%; from the curve fluctuation, the clustering algorithm only fluctuates before the number of iterations is 3, and the fluctuation is small, less than 2%; from the curve convergence performance in the middle and late stages, the curve is in the horizontal convergence state in the middle and late stages, and there is no fluctuation, and the final convergence value is 96%. In contrast, the starting values of the accuracy curves of the training samples corresponding to the KPC and K-means algorithms are lower than 90%, and there is a significant decreasing trend in the middle and late stages, and the final accuracy convergence values are 92.8% and 94.5%, respectively. In comparison, the proposed algorithms are improved by 3.2% and 1.5% in terms of accuracy.

From Fig. 6b, the proposed clustering algorithm achieved the optimal accuracy curve on the test sample. In terms of the starting accuracy value, the clustering algorithm takes a value of 86.8%, which is higher than 85%; from the curve fluctuation, the clustering algorithm only has obvious ups and downs before the number of iterations is 2, and the magnitude is less than 8%; from the convergence performance in the middle and late stages of the curve, the curve is in a horizontal convergence state with almost no fluctuation, and the final convergence value is 92.4%. In contrast, the accuracy curves of KPC and K-means algorithms for the test samples start at less than 85%, and there is a significant downward trend in the middle and late stages, with a decline of more than 3% and 8%, and the final accuracy convergence values are 85.1% and 72.4%, respectively. In comparison, the proposed algorithms are improved by 7.3% and 20% in terms of accuracy. Figure 7 shows the F-value curves of different algorithms during training and testing. Figure 8 shows the DBI curves of different algorithms during training and testing.

Figure 7.

$F$ -value curves of different algorithms during training and testing.

It can be seen from Fig. 7a. For the training samples, the $F$ -value curve of the proposed clustering algorithm shows a trend of increasing in the first period, decreasing in the middle period, and increasing slightly in the later period. When the number of iterations is 3, the curve obtains the highest $F$ -value of 96.3% for the whole segment, and then the curve drops to 95.9%, and when the number of iterations is 10, the curve starts to rise again and maintains at about 96.2%. The starting value of the K-means algorithm is higher than 80%, but there is an obvious downward trend in its middle period, and the final convergence value is 80%. In comparison, it can be seen that the study mentions that the clustering algorithm improves 8.2% and 16.2% over the latter two. It can be seen from Fig. 7b. For the test sample, the $F$ -value curve of the proposed clustering algorithm shows a gradient upward trend. Before the number of iterations is 6, the curve $F$ -value is maintained near 92.3%; when the number of iterations is 12, the curve $F$ -value is maintained near 95.2%, which increases by 2.9%; finally, the convergence value of this curve is 97.5%. In contrast, the starting values of the curves of the KPC algorithm and the K-means algorithm are lower, both below 80%. And both of them show a slow incremental change trend, and when the number of iterations is 7, the two $F$ -value curves increase to 97.6% and 82.8%, respectively. After that, the KPC algorithm achieves a final convergence value of 88.7% with small fluctuations, while the K-means algorithm achieves a smoother convergence value of 85%. Figure 8 shows the DBI curves of the different algorithms during training and testing.

Figure 8.

DBI curves of different algorithms during training and testing.

As can be seen from Fig. 8a, for the training samples, the DBI curve of the proposed algorithm is studied to show a trend of increasing first and then level. The curve increases from the starting value of 0.086 to the inflection point value of 0.096 with an increase of 0.01 when the number of iterations is between 1 and 6; after that, the curve tends to converge with almost no fluctuation and finally achieves a DBI convergence value of 0.096, which is higher than 0.09. While the DBI curve of the KPC algorithm shows large fluctuations in the early iterations and the DBI approaches 0.088 in the middle; when the number of iterations The DBI curve of the K-means algorithm showed less fluctuation in general, but showed a certain degree of decreasing trend in the middle period, and the final convergence value was 0.087. From Fig. 8b, it can be seen that for the test samples, the DBI curve of the proposed algorithm still showed a trend of increasing first and then level. The DBI curve of the KPC algorithm and the K-means algorithm also shows a trend of increasing first and then level change in general, but the DBI of the proposed algorithm increases by 0.012 and 0.014 in terms of the final convergence value.

4.2 Uniform model utility analysis of adjustable load demand data based on source-load similarity

In this study, a local area is selected as the research object, where the new energy output is mainly composed of distributed photovoltaic, and the shape of the new energy output curve tends to be normally distributed, and a number of customer load curves in the area are randomly selected as the curve to be regulated. The response step is 1 kW, and the variance and M-DTW optimal are used as the target for stepwise regulation. Figure 9 shows the variation curve of the relationship between the increment of new energy consumption and the response of load demand.

Figure 9.

Change curve of relationship between new energy consumption increment and load demand response.

Figure 9 shows the consumption process of new energy power by load device 1 as well as load device 2. It is easy to find that the M-DTW method and the variance method of the study both show a stepwise climbing change as the response step gradually increases. For load device 1, the M-DTW method climbs from 0 to 12.8 kW, while the variance method climbs to 10.9 kW, with a difference of 1.9 kW. For load device 2, the difference between the final values of the two methods is 2.0 kW. Therefore, this indicates that the M-DTW method is able to consume more new energy output as the response gradually increases, which indicates that the unified model with the proposed data is more advantageous in This indicates that the unified model with the proposed data is more advantageous for demand dispatch. Figure 10 shows the load curve before and after the load demand response.

Figure 10.

Load curve before and after demand response.

Figure 11.

Real time electricity price and upper and lower reserve capacity price.

As can be seen from Fig. 10, for either load device, the variance and M-DTW methods can adjust the response downward when the PV generation is lower than the initial load; while when the PV generation is higher than the initial load, both methods can adjust the response upward to accommodate the energy change. For example, for load device 2, when the time period is between 1 h and 12 h, the initial load is greater than the new generating energy and the response needs to be adjusted downward, so the adjusted load curve has a significant decrease compared with the initial load curve, with the average value of the decrease being 0.25; while when the time period is between 12 h and 17 h, the new generating energy is higher and the response needs to be adjusted upward, so the adjusted The variance method has a mean value of 0.11, while the M-DTW method has a mean value of 0.15, with a difference of 0.01. This indicates that the unified model of the proposed data can better adapt to the needs of source load changes.

To verify the ability of agricultural irrigation load to participate in operating standby, 500 mu of winter wheat is planted in an agricultural area is selected for the study. The minimum daily winter wheat water demand is 2000 m ${}^{3}$ and the maximum demand is 2500 m ${}^{3}$ . The maximum pumping capacity of a single irrigation pump is 360 m ${}^{3}$ and the maximum operating power is 98.1 kW. Figure 11 shows the real-time power price and the upper and lower standby capacity prices.

As seen in Fig. 11, the upper reserve capacity price curve shows a decreasing trend between 0:00 and 3:00, between 10:00 and 14:00, and between 20:00 and 24:00, while it shows an increasing trend in the rest of the time period. The lowest value of the upper standby price is $0.01 and the highest value is $0.20. The lower standby capacity price shows a decreasing trend between 0:00 and 3:00 and between 19:00 and 24:00 with a minimum value of $0.01 and a maximum value of $0.08. Figure 12 shows the real-time price lower reserve capacity curve.

Figure 12.

Reserve capacity curve under real-time pricing.

It can be seen from Fig. 12, when the time period is between 0:00 and 6:00, the upper standby capacity is activated because the upper standby capacity is less expensive, and the size of the standby capacity is 9.8 kW at this time, while when the time period is between 6:00 and 24:00, the lower standby capacity is less expensive, so the lower standby capacity is activated, and the size is 9.8 kW. It can be seen that the pump load is concentrated in the time interval when the price of electricity is lower, which is beneficial to achieve economic efficiency.

5. Conclusion

With the increasing complexity of China’s power system and the increasing prominence of problems such as conventional resource scarcity, how to effectively reduce the operating costs of the power system while effectively consuming new energy is an important way to enhance the safety, stability, stability and efficiency of the power system. In this study, an S-KPC-based clustering method for load equipment power consumption data and an adjustable load demand response analysis method based on source-load similarity are proposed, and the accuracy curve of this cluster method is in a state of horizontal convergence with a final convergence value of 92.4%, which is 7.3% and 20% higher than KPC and K-means algorithms. In addition, the M-DTW method can dissipate more new energy output. When the time period is between 12 h and 17 h, the mean value of the upper body amplitude of the M-DTW method is 0.15, which is higher than that of the variance method. It can be seen that the S-KPC method can effectively use the power consumption data in the power system and can obtain the power consumption cycles of various power systems, so as to lay a solid foundation for further development of power aggregation and regulation policies of the power system. And the customer curve-based electricity demand response regulation can provide the optimal electricity market incentive path, which can improve the efficiency of new energy consumption while reducing the incentive cost of electricity market demand response. This article mainly focuses on the research of a unified data model for industrial and agricultural loads. In the future, it is necessary to further expand load types and build a demand response oriented load data model library to promote the further development of demand response business and assist in the consumption of renewable energy.

Funding

The work was financially supported by the Science and Technology Project of the State Grid Corporation of China (Research and verification of precise prediction and aggregation regulation technology for demand side flexible resource regulation ability to adapt to renewable energy consumption: 5108-202218280A-2-242-XG).

References

Kebede

Kalogiannis

Van Mierlo

Berecibar

. A comprehensive review of stationary energy storage devices for large scale renewable energy sources grid integration. Renew Sust Energ Rev. 2022; 159: 112213.

Qazi

Bhowmik

Hussain

Yang

Naseem

Adebayo

Gumaei

Al-Rakhami

. Analyzing the public opinion as a guide for renewable-energy status in malaysia: a case study. IEEE T Eng Manage. 2021; 40(2): 371-385.

Zhu

Mortazavi

Maleki

Aslani

Yousefi

. Analysis of the robustness of energy supply in Japan: Role of renewable energy. Energy Rep. 2020; 6: 378-391.

Tian

Huang

Liu

Sun

Deng

Zhao

. Application of photovoltaic power generation in rail transit power supply system under the background of energy low carbon transformation. Alex Eng J. 2021; 60(6): 5167-5174.

. An edge computing framework for powertrain control system optimization of intelligent and connected vehicles based on curiosity-driven deep reinforcement learning. IEEE T Ind Electron. 2021; 68(8): 7652-7661.

Guo

Jiang

Yao

Zhang

Hua

. Industrial power load forecasting method based on reinforcement learning and PSO-LSSVM. IEEE T Cybernetics. 2020; 52(2): 1112-1124.

CMZ

Wan

Chen

Cui

Zhao

. Electric load clustering in smart grid: Methodologies, applications, and future trends. J Mod Power Syst Cle. 2021; 9(2): 237-252.

Sharma

Vashisht

Singh

. WOATCA: A secure and energy aware scheme based on whale optimisation in clustered wireless sensor networks. IET Commun. 2020; 14(8): 1199-1208.

Jeong

Park

. Short-term electric load forecasting for buildings using logistic mixture vector autoregressive model with curve registration. Appl Energ. 2021; 282: 116249.

10.

Han

Taylor

. Short-term forecasting of individual residential load based on deep learning and K-means clustering. CSEE J Power Energy. 2020; 7(2): 261-269.

11.

Wang

Ming

Wang

SHR

. Deep reinforcement learning method for demand response management of interruptible load. IEEE T Smart Grid. 2020; 11(4): 3146-3155.

12.

Waseem

Lin

Ding

Wen

Liu

Palu

. Technologies and practical implementations of air-conditioner based demand response. J Mod Power Syst Cle. 2020; 9(6): 1395-1413.

13.

Bahrami

Chen

Wong

VWS

. Deep reinforcement learning for demand response in distribution networks. IEEE T Smart Grid. 2020; 12(2): 1496-1506.

14.

Singh

Ding

Raju

Raghav

Kumar

. A swarm intelligence approach for energy management of grid-connected microgrids with flexible load demand response. Int J Energ Res. 2022; 46(4): 4301-4319.

15.

Hadi

Silva

CAS

Hossain

Challoo

. Algorithm for demand response to maximize the penetration of renewable energy. IEEE Access. 2020; 8: 55279-55288.

16.

Namalomba

Shi

. Agent based simulation of centralized electricity transaction market using bi-level and Q-learning algorithm approach. Int J Elec Power. 2022; 134: 107415.