Demand response capability evaluation method and system for multi-element power load group
Through the autoencoder and convergence cross-mapping algorithm combined with the LSTM neural network, the demand response capability of multiple power load groups is evaluated, and the problem of inaccurate evaluation of diversified load groups is solved, and efficient and accurate load curve generation and demand response capability evaluation is achieved.
Patent Information
- Application Number
- CN202311821623.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-27
- Publication Date
- 2025-07-25
AI Technical Summary
The existing load calculation methods cannot fully consider the changing trends and fluctuations of diversified load groups. Especially in different environments, the differences in power consumption characteristics and rules of different industries affect the calculation results of power loads, resulting in inaccurate evaluation.
The deep embedding clustering algorithm based on the autoencoder is used to cluster the power load data, and the causal relationship between multiple load groups and multiload types is evaluated in combination with the convergence cross-mapping algorithm, a mathematical representation model for the typical load characteristics of multiple user groups is established, and a load curve in different environments is generated using the LSTM neural network to calculate the real-time demand response capabilities of multiple user groups.
It realizes reasonable, objective and accurate evaluation of the demand response capabilities of multiple power load groups, and can efficiently generate high-precision load curves, solves the problem of difficult-to-describe complex load curves, and improves the accuracy and efficiency of evaluation.
Smart Images
Figure CN120373678A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of power demand response capability evaluation, and particularly relates to a method and system for evaluating the demand response capability of a diversified power load group. Background Art
[0002] With the continuous construction and development of the "smart grid", the power system has entered a new era of accelerated development. The inherent randomness, volatility and other problems of renewable energy have affected the grid connection and consumption of renewable energy power generation, bringing challenges to ensuring the real-time power dynamic balance and safe and stable operation of the power grid. At the same time, the load characteristics on the demand side are also changing. Accurately describing the load change law and the correlation is an important means for the future energy structure reform of the power grid and an inevitable requirement for the rapid development of the future power grid. Therefore, accurately describing the dynamic correlation between different loads, revealing the "flexible" law therein, and evaluating its demand response capability have become the key to renewable energy consumption and the stable operation of the power system.
[0003] Currently, the general calculation methods for load are static load calculation methods and dynamic calculation methods. Static load calculation refers to calculating the load demand of the power system according to the load curve and load characteristics within a certain period of time. Static load calculation usually adopts the load curve method and the load rate method. The load curve method means that according to historical load data, a load curve is drawn, and then according to the load curve and load characteristics, the load demand of the power system within a future period of time is calculated. The load rate refers to the ratio of the actual load of the power system to the rated load. Dynamic load calculation refers to considering the influence of time variation and random factors, and calculating the real-time load demand of the power system according to the actual operation situation of the power system. Dynamic load calculation methods include short-term load forecasting and medium- and long-term load forecasting. Short-term load forecasting refers to predicting the power load demand within a future period of time according to historical load data and real-time meteorological data, etc. Medium- and long-term load forecasting refers to predicting the power load demand within a relatively long future period according to factors such as economic development trends, industrial structure changes, and climate changes.
[0004] However, the above technologies mainly focus on the change of a single load or the overall load change trend in the power system, but there are still the following problems for the increasingly complex current load composition:
[0005] 1. The composition of the diversified load group leads to an increasingly complex load change trend. Traditional load calculation methods cannot comprehensively consider the load change and fluctuation law, as well as the correlation between load groups.
[0006] 2. In different environments, the electricity consumption characteristics and laws of different industries may vary, which will affect the power load calculation results. For a certain load group, it is necessary to select a suitable power load calculation method according to the actual situations of different regions and industries and the current environmental conditions. Summary of the Invention
[0007] In view of the technical problems existing in the prior art, the present invention provides a method and system for evaluating the demand response ability of a multi-source power load group. An in-depth embedded clustering algorithm based on an autoencoder is proposed to cluster power load data to obtain the change characteristics of various load types of different load groups. A causal association relationship of the load process of the multi-source load group - multi-load type is also evaluated based on the convergent cross-mapping algorithm, and a mathematical representation model of the typical load characteristics of a real-time multi-source user group considering the dynamic association relationship of the load group is established. Combining with the multi-dimensional evaluation method of the demand response ability of the multi-source user group, the demand response ability of users can be evaluated reasonably, objectively and accurately.
[0008] The technical solution adopted by the present invention is as follows: A method for evaluating the demand response ability of a multi-source power load group includes the following steps:
[0009] Obtain the load data of the multi-source load group;
[0010] According to the load data of the multi-source load group, use the in-depth embedded clustering algorithm based on the autoencoder to calculate various load types of each load group;
[0011] Based on the load data of various load types of the multi-source load group, evaluate the causal association relationship of the load process of the multi-source load group - multi-load type based on the convergent cross-mapping algorithm, and establish a mathematical representation model of the typical load characteristics of a real-time multi-source user group considering the dynamic association relationship of the load group;
[0012] According to the mathematical representation model of the typical load characteristics of the real-time multi-source user group considering the dynamic association relationship of the load group, generate the load curves of the multi-source user group time series in different environments;
[0013] According to the load curves of the multi-source user group time series in different environments, calculate the real-time demand response ability of the multi-source user group.
[0014] Further, the multi-source load group includes multiple of industrial load, residential load, commercial load, transportation load, and agricultural load.
[0015] Further, after obtaining the load data of the multi-source load group, process the load data, and the processing process is as follows:
[0016] According to the time series, type, and load value of the load data, identify the outliers in the load data and delete the outliers;
[0017] Calculate the data missing ratio D of the load data a ,
[0018]
[0019] Wherein, M is the number of measurement time nodes within the time range, and N is the total number of types of the load data. Indicates whether the load data of type n is missing at time m. If it is missing, its value is 1; otherwise, it is 0.
[0020] Calculate the missing value of the load data at each moment in the load data respectively.
[0021]
[0022] When is 1, delete the load data at this moment.
[0023] Furthermore, the calculation process of multiple load types of the load group is as follows:
[0024] Input the load data of multiple load groups into a pre-trained autoencoder model to obtain embedded features.
[0025] Input the embedded features into a clustering layer network, use KL divergence for clustering training, and then calculate the preliminary clustering results of multiple load types.
[0026] The clustering training using KL divergence is to optimize the encoder parameters in the autoencoder model through the following formula.
[0027]
[0028] And optimize the clustering layer network through the following formula.
[0029]
[0030]
[0031] Wherein, P is the original distribution, p ij is a certain clustering distribution, Q is the assumed distribution, q ij is the probability that sample i is assigned to cluster j; L = KL, and KL is the KL-divergence distance; α is the degree of freedom of the Student-t distribution, initially set to 1; u j is the clustering center of cluster j, x i is the input load value, z i is the latent feature space, which is a non-linear mapping of x - z, z i = f θ (x i );
[0032] Filter out small clustering clusters in the preliminary clustering results through a threshold to obtain multiple load types for each of the load groups.
[0033] Further, based on the load data of multiple load types of a multi-load group, the process of evaluating the causal association relationship of the load process of the multi-load group - multiple load types based on the convergence cross-mapping algorithm and establishing a mathematical representation model of the typical load characteristics of real-time multi-user groups considering the dynamic association relationship of the load group is as follows:
[0034] Taking the load data of multiple load types of a multi-load group as observations, reconstruct a shadow manifold according to a certain embedding dimension E, calculate the E + 1 neighbor points and Euclidean distance or Mahalanobis distance of the shadow manifold, obtain the weights of each neighbor point using the Euclidean distance or Mahalanobis distance, and obtain the estimated value of the observation value through weighted average, and calculate the error value between the observation value and the estimated value; repeat the above process for different embedding dimensions E respectively, and calculate the error values between the observation value and the estimated value; select the embedding dimension E corresponding to the smallest error value as the optimal embedding dimension;
[0035] Based on Takens' embedding theorem, reconstruct the shadow manifold from the load data of multiple load types of the multi-load group according to the optimal embedding dimension; apply dynamic system theory to the reconstructed shadow manifold, and analyze whether the points on the reconstructed shadow manifold approach or move away from each other over time. If there is the above trend, it is considered that there is a causal influence between the load data;
[0036] Iterate the above process multiple times, use different optimal embedding dimensions each time, and observe which optimal embedding dimension results in the most obvious causal relationship between the load data; establish a mathematical representation model L of the typical load characteristics of real-time multi-user groups considering the dynamic association relationship of the load group based on the causal relationship between the load data obtained under this optimal embedding dimension e
[0037]
[0038] In the model, E is the association relationship between different load types in different load groups, N' is different load types of different load groups, and L n is the load curve of the time series of different types in different load groups.
[0039] Further, the mathematical representation model of the typical load characteristics of real-time multi-user groups considering the dynamic association relationship of the load group inputs a trained LSTM neural network model to calculate the load curves of the time series of multi-user groups in different environments.
[0040] Further, the process of calculating the real-time demand response capability of the multi-user group according to the load curves of the multi-user group time series in different environments is as follows:
[0041] Calculate the demand response indicators of the load curves of the multi-user group time series in different environments. The demand response indicators include the interruptible load ratio X1, the shiftable load ratio X2, the curtailable load ratio X3, the load interruptible time X4, the load shiftable time X5, the load curtailable time X6, the proportion of electricity cost in the cost expenditure X7, the user comfort requirement X8, the user economic level X9, the power supply reliability requirement X 10 , the smart meter coverage rate X 11 , and the benefits (price-based and incentive-based) obtained by users participating in demand response X 12 ;
[0042] Calculate the real-time demand response capability of the multi-user group.
[0043]
[0044]
[0045] In the formula, K is the demand capacity level, x1...x 12 are the values of the demand response indicators, k1...k 12 are the objective correlation weights of the indicators, q1...q 12 are the subjective weights of the indicators, and DR is the demand response capability.
[0046] The technical solution adopted by the present invention is: a demand response capability evaluation system for a multi-power load group, including a data acquisition unit, a deep embedded clustering unit, a mathematical representation model generation unit, a load curve generation unit, and a demand response capability evaluation unit connected in sequence.
[0047] The data acquisition unit is used to obtain the load data of the multi-load group.
[0048] The deep embedded clustering unit is used to calculate various load types of each load group by using a deep embedded clustering algorithm based on an autoencoder according to the load data of the multi-load group.
[0049] The mathematical representation model generation unit is used to evaluate the causal association relationship of the load process of the multi-load group - multi-load types based on the load data of various load types of the multi-load group by using a convergence cross mapping algorithm, and establish a real-time multi-user group typical load characteristic mathematical representation model considering the dynamic association relationship of the load group.
[0050] The load curve generation unit is configured to generate load curves of a multi - user group time series in different environments according to the real - time multi - user group typical load characteristic mathematical representation model considering the dynamic correlation relationship of the load groups;
[0051] The demand response ability evaluation unit is configured to calculate the real - time demand response ability of the multi - user group according to the load curves of the multi - user group time series in different environments.
[0052] Furthermore, the deep embedding clustering unit inputs the load data of multiple load groups into a pre - trained auto - encoder model to obtain embedding features;
[0053] The embedding features are input into the clustering layer network, and KL - divergence is used for clustering training, and then preliminary clustering results of multiple load types are calculated;
[0054] The clustering training using KL - divergence is to optimize the encoder parameters in the auto - encoder model through the following formula,
[0055]
[0056] and optimize the clustering layer network through the following formula,
[0057]
[0058]
[0059] In the formula, P is the original distribution, p ij is a certain clustering distribution, Q is the assumed distribution, q ij is the probability that sample i is assigned to cluster j; L = KL, and KL is the KL - divergence distance; α is the degree of freedom of the Student - t distribution, initially set to 1; u j is the clustering center of cluster j, x i is the input load value, z i is the latent feature space, is the non - linear mapping of x - z, z i = f θ (x i );
[0060] Small clustering clusters in the preliminary clustering results are filtered out through a threshold to obtain multiple load types of each load group.
[0061] Further, the mathematical representation model generation unit uses the load data of multiple load types of a multi-load population as observations, reconstructs a shadow manifold according to a certain embedding dimension E, calculates the E+1 neighbor points and Euclidean distance or Mahalanobis distance of the shadow manifold, obtains the weights of each neighbor point using the Euclidean distance or Mahalanobis distance, and obtains the estimated value of the observation through weighted average, and calculates the error value between the observation and the estimated value; the above process is repeated by trying different embedding dimensions E respectively, and the error value between the observation and the estimated value is calculated; the embedding dimension E corresponding to the smallest error value is selected as the optimal embedding dimension;
[0062] Based on Takens' embedding theorem, the shadow manifold is reconstructed from the load data of multiple load types of the multi-load population according to the optimal embedding dimension; the dynamic system theory is applied to the reconstructed shadow manifold to analyze whether the points on the reconstructed shadow manifold approach or move away from each other over time. If there is the above trend, it is considered that there is a causal influence between the load data;
[0063] The above process is iterated multiple times. Each iteration uses a different optimal embedding dimension, and it is observed which optimal embedding dimension results in the most obvious causal relationship between the load data; based on the causal relationship between the load data obtained under the optimal embedding dimension, a mathematical representation model L of the typical load characteristics of real-time multi-user groups considering the dynamic correlation relationship of the load population is established e
[0064]
[0065] In the model, E is the correlation relationship between different load types in different load populations, N' is different load types of different load populations, and L n is the load curve of time series of different types in different load populations.
[0066] Further, the load curve generation unit is a trained LSTM neural network model.
[0067] Further, the demand response ability evaluation unit calculates the demand response indicators of the load curves of the time series of multi-user groups in different environments. The demand response indicators include the interruptible load ratio X1, the shiftable load ratio X2, the reducible load ratio X3, the load interruptible time X4, the load shiftable time X5, the load reducible time X6, the proportion of electricity cost in the cost expenditure X7, the user comfort requirement X8, the user economic level X9, the power supply reliability requirement X 10 and the smart meter coverage rate X 11 and the benefits (price-based and incentive-based) obtained by users participating in demand response X 12 ;
[0068] Calculate the real-time demand response capabilities of multiple user groups
[0069]
[0070]
[0071] Where K is the demand capacity level, x1...x 12 are the values of the demand response indicators, k1...k 12 are the objective correlation weights of the indicators, q1...q 12 are the subjective weights of the indicators, and DR is the demand response capacity.
[0072] It further includes a data processing unit for processing the load data obtained by the data acquisition unit and then transmitting it to the deep embedding clustering unit.
[0073] The processing process of the load data is as follows:
[0074] Identify the outliers in the load data according to the time series, type, and load value of the load data, and delete the outliers;
[0075] Calculate the data missing ratio D a ,
[0076]
[0077] Where M is the number of measurement time nodes within the time range, N is the total number of types of the load data, indicates whether the load data of type n is missing at time m. If it is missing, its value is 1, otherwise it is 0;
[0078] Calculate the load data missing values at each moment in the load data respectively
[0079]
[0080] When is 1, delete the load data at that moment.
[0081] The technical solution adopted by the present invention is also: a computer-readable storage medium including a program that can be executed by a processor to implement the method for evaluating the demand response capabilities of a multiple electric load group.
[0082] Compared with the prior art, the beneficial effects of the present invention are:
[0083] 1. The present invention proposes a deep embedding clustering algorithm based on an autoencoder to cluster power load data, obtain the change characteristics of multiple load types of different load groups, and solve the problem of difficult description of complex load curves.
[0084] 2. The present invention proposes to evaluate the causal association relationship of the load process of a multi - load group - multi - load type based on the convergent cross - mapping algorithm, and establish a mathematical representation model of the typical load characteristics of real - time multi - user groups considering the dynamic association relationship of the load group, solving the problem of the representation of different types of power loads in different environments.
[0085] 3. The present invention is based on the LSTM neural network model to realize a mathematical representation model of the typical load characteristics of real - time multi - user groups considering the dynamic association relationship of the load group, and can efficiently generate load curves of high - precision multi - user groups in different environments.
[0086] 4. The present invention proposes a multi - dimensional evaluation method for the demand response ability of multi - user groups, which can reasonably, objectively and accurately evaluate the demand response ability of users. BRIEF DESCRIPTION OF THE DRAWINGS
[0087] Figure 1 is a schematic flowchart of the method of the embodiment of the present invention;
[0088] Figure 2 is a schematic structural diagram of the principle of the autoencoder of the embodiment of the present invention;
[0089] Figure 3 is a schematic structural diagram of the system of the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0090] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
[0091] An embodiment of the present invention provides a method for evaluating the demand response ability of a multi - power load group, as Figure 1 shown, which includes the following steps:
[0092] Step 1: Obtain the load data of the multi - load group. The types of the load group include industrial load, residential load, commercial load, transportation load, and agricultural load. The multi - load group is multiple of industrial load, residential load, commercial load, transportation load, and agricultural load. The load group in this embodiment includes industrial load, residential load, commercial load, transportation load, and agricultural load.
[0093] In order to solve the problem that the collected load data is affected by various factors and has inaccuracies, the emergence of these outliers may have a negative impact on the data analysis results, so outlier data processing is required. Through the detection and processing of outliers, we can better understand the distribution and rules of data, improve the quality and reliability of data, and provide more accurate and reliable support for decision-making. In order to solve the problem of abnormal and inaccurate load data, a load data processing link is designed to compare various types of load data horizontally between load groups and vertically with the data of the same historical date to identify outliers. If there are abnormal load values at many times, the abnormal data will be deleted.
[0094] After obtaining the load data of the multivariate load group, the load data is processed, and the processing process is as follows:
[0095] Step 1.1: The load data is a time series load curve of multiple types of loads. According to the time series, type and load value of the load data, outliers in the load data are identified based on the k-means method and deleted; the specific process is:
[0096] Step 1.11: Select k initialized samples as initial cluster centers a=a1,a2,...,a k ;
[0097] Step 1.12: For each sample X in the data set, calculate its distance to the k cluster centers and divide it into the class corresponding to the cluster center with the smallest distance;
[0098] Step 1.13: For each category c i , recalculate its cluster center (μ i That is, the centroid of all samples belonging to this class);
[0099] Step 1.14: Repeat the above steps 1.12 and 1.13 until the minimum error change is achieved. The minimum stops the cycle;
[0100] Step 1.15: Check the distance of each data point from the centroid of the cluster to which it belongs to determine the outliers and remove the outliers. If the distance of a data point from the centroid of the cluster to which it belongs exceeds the threshold (the average distance of the centroids of all clusters), then this data point can be considered an outlier.
[0101] Step 1.2: Calculate the missing data ratio D of the load data a ,
[0102]
[0103] In the formula, M is the number of measurement time nodes within the time range, and N is the total number of types of the load data. indicates whether the load data of type n is missing at time m. If it is missing, its value is 1; otherwise, it is 0.
[0104] Step 1.3: According to the data missing ratio D a judge the missing risk of the load values. If the missing value is greater than the data missing ratio D a , delete the missing load curves. Calculate the missing values of the load data at each moment in the load data respectively
[0105]
[0106] When is 1, delete the load data at this moment.
[0107] The load data processing method adopted by the present invention deletes the identified outliers, converts them into missing values, and together with the original missing values of the load data, calculates the data missing ratio, judges the missing risk of the load values, and deletes the load data with more outliers and missing values, improving the accuracy of the later data analysis results.
[0108] Step 2: According to the load data of the multi-load population, use the deep embedding clustering algorithm based on the autoencoder to calculate various load types of each load population.
[0109] The autoencoder has advantages in processing various data types, data with noise, and data compression, etc. At the same time, its unsupervised learning feature also makes it perform well in processing unlabeled data. Combining with the power load change rules of the multi-load population to further identify the load change characteristics can provide a basis for load characteristic representation.
[0110] The autoencoder model is an unsupervised learning model. Based on the backpropagation algorithm and optimization method, it uses the input data, that is, the load data itself, as supervision to guide the neural network to learn the fluctuation characteristics of each type of load under the time series. The role of the encoder is to input and encode the power load data of different high-dimensional load populations into low-dimensional latent variables, so that the neural network learns the most informative features. The role of the decoder is to restore the latent variables in the hidden layer to the output data X of the initial dimension R , and the best state is that the output of the decoder can perfectly or approximately recover the original input, that is, X R ≈X. The principle structure diagram of the autoencoder is as shown in Figure 2 shown, and the calculation methods from the input layer of the model to the hidden layer and from the hidden layer to the output layer are as follows:
[0111] h = σ(W1x + b1) (3)
[0112]
[0113] MiniLoss = dist(X, X R ) (5)
[0114] where x is the input layer data, h is the hidden layer, is the output layer data (i.e., the reconstructed input data), σ is the activation function, W1 and W2 are the neural network weights, b1 and b2 are the bias terms, and dist is the mean square variance of the distance metric function between the two.
[0115] The deep embedded clustering algorithm can make full use of the powerful feature extraction ability of the neural network, use more discriminative features to find a more accurate subspace. At the same time, during the clustering process, the deep embedded clustering algorithm considers not only the local features of the data but also the global features of the data. Therefore, the deep embedded clustering algorithm has a good feature extraction effect for processing high-dimensional data. Based on the deep embedded clustering algorithm, clustering is performed on the power load to obtain the change laws of multiple types of loads in different load groups. As Figure 1 shown, the deep embedded clustering algorithm based on the autoencoder consists of two parts. The first part pre-trains an autoencoder model; the second part selects the encoder part in the autoencoder model, adds a clustering layer, and uses the KL divergence for training and clustering.
[0116] The calculation process of multiple load types of the load group is as follows:
[0117] Step 2.1: Input the load data of multiple load groups into the pre-trained autoencoder model to obtain the embedded features.
[0118] Step 2.2: Input the embedded features into the clustering layer network, use the KL divergence for clustering training, and then calculate the preliminary clustering results of multiple load types.
[0119] The clustering training using the KL divergence is to optimize the encoder parameters in the autoencoder model through formula (6), and optimize the clustering layer network through formula (7) and formula (8), which can simultaneously achieve the effect of optimizing the relevant parameters in the clustering model and the encoder network.
[0120]
[0121]
[0122]
[0123] where P is the original distribution, p ijis a certain clustering distribution, Q is the hypothesized distribution, and q ij is the probability that sample i is assigned to cluster j; L = KL, where KL is the KL-divergence distance; α is the degree of freedom of the Student-t distribution, initially set to 1; u j is the cluster center of cluster j, and x i is the input load value, and z i is the latent feature space, a non-linear mapping of x - z, and z i = f θ (x i ).
[0124] Step 2.3: Filter out small clusters in the preliminary clustering results through a threshold to obtain multiple load types for each load group. In the process of setting the threshold, first analyze the size distribution of the clusters, and then select a suitable quantile as the threshold. For example, if the data distribution is relatively uniform, the median or average can be selected as the threshold; if the data distribution is highly skewed, a relatively high quantile can be selected as the threshold.
[0125] Each load group in this embodiment obtains 3 load types by setting a reasonable threshold. For example, industrial load calculates industrial load type A, industrial load type B, and industrial load type C; similarly, commercial load calculates commercial load type A, commercial load type B, and commercial load type C. Residential load, transportation load, and agricultural load are similar.
[0126] Step 3: Based on the load data of multiple load types of the multi-load group, evaluate the causal association relationship of the load process of the multi-load group - multi-load type based on the convergent cross-mapping algorithm, and establish a mathematical representation model of the typical load characteristics of real-time multi-user groups considering the dynamic association relationship of the load group. The specific process is as follows:
[0127] Step 3.1: Using the load data of multiple load types of the multi-load group as observations, reconstruct the shadow manifold according to a certain embedding dimension E, calculate the E + 1 neighbor points of the shadow manifold and the Euclidean distance or Mahalanobis distance, obtain the weights of each neighbor point using the Euclidean distance or Mahalanobis distance, and obtain the estimated value of this observation through weighted average, and calculate the error value between the observation and the estimated value; repeat the above process by trying different embedding dimensions E respectively, and calculate the error values between the observations and the estimated values; select the embedding dimension E corresponding to the smallest error value as the optimal embedding dimension; this step is carried out in the way of cross-validation, that is, by reconstructing the input data in a specific dimension E, and then calculating the error between the reconstructed data and the original data to determine the optimal embedding dimension.
[0128] Step 3.2: Based on Takens' embedding theorem, reconstruct the shadow manifold from the load data of various load types in the multi-load group of the load group according to the optimal embedding dimension; apply the dynamic system theory to the reconstructed shadow manifold, and analyze whether the points on the reconstructed shadow manifold approach or move away from each other over time. If there is such a trend, it is considered that there is a causal influence between the load data.
[0129] Step 3.3: Iterate Steps 3.1 and 3.2 multiple times, each time using a different optimal embedding dimension, and observe which optimal embedding dimension results in the most obvious causal relationship between the load data; based on the causal relationship between the load data obtained under this optimal embedding dimension, establish a mathematical representation model L of the typical load characteristics of real-time multi-user groups considering the dynamic association relationship of the load group. e
[0130]
[0131] In the model, E represents different electricity consumption environment conditions, that is, the association relationship between different load types in different load groups. N' represents different load types of different load groups, and L n is the load curve of the time series of different types in different load groups. Under different electricity consumption environment conditions, different load groups have different dynamic relationships. Therefore, formula (9) is used to express the set of load groups with the strongest causal relationship to improve the accuracy of the model's demand response.
[0132] In this embodiment, under different electricity consumption environment conditions, based on the convergent cross-mapping algorithm, combinations of 3 load types for each of 5 load groups are formed to form a combination of characteristic loads of dynamic multi-load groups - multi-load types. For example, if the causal association relationship between residential load type A and commercial load type B is the closest, then it corresponds to a certain electricity consumption environment condition.
[0133] Step 4: Generate the load curves of the time series of multi-user groups in different environments according to the mathematical representation model of the typical load characteristics of real-time multi-user groups considering the dynamic association relationship of the load group.
[0134] The mathematical representation model of the typical load characteristics of real-time multi-user groups considering the dynamic association relationship of the load group inputs the trained LSTM neural network model to calculate the load curves of the time series of multi-user groups in different environments.
[0135] By introducing structures such as forget gates, LSTM selectively stores information and extends the time series information learning range of the recurrent neural network.
[0136] A single-layer LSTM neural network consists of a set of recurrent structures, and each recurrent structure includes: an input gate, an output gate, a forget gate, and a cell state module.
[0137] The input gate determines how much of the data currently input to the neural network can be saved to the cell state. By inputting all the input parameters x at the current time t and the output parameter h of the hidden state at the previous time t-1 , the calculation of the input gate is realized:
[0138] i t = σ(W i [h t-1 , x t + b i ) (10)
[0139] The forget gate controls the amount of data from the previous time saved to the data at this time. Using the tanh activation function, the calculation formula is as follows:
[0140] f t = σ(W f [h t-1 , x t + b f ) (11)
[0141] C′ t = tanh(W c [h t-1 , x t + b c ) (12)
[0142] The output gate controls how much of the current cell state will be output to the current data. The calculation formula of the output gate O t is as follows:
[0143] o t = σ(W o [h t-1 , x t + b o ) (13)
[0144] The LSTM neural network includes the transfer of long-term memory and short-term memory. The transfer of long-term memory is realized by calculating the input gate, the output gate, and the memory information at the previous time. The transfer of short-term memory is realized by calculating the output gate and activating the information in the long-term memory. The formulas for realizing the transfer of long-term memory and short-term memory are as follows:
[0145] C t = f t C t-1 + i t C′ t (14)
[0146] h t = o t tanh(C t ) (15)
[0147] In the formula, W i is the weight matrix of the input gate; W f and W C are the weight matrices of the forgetting gate; W o is the weight matrix of the output gate; x t is the input parameter; h t is the output parameter in the hidden state; b is the bias.
[0148] Step 5: Calculate the real-time demand response capabilities of the multi-user group according to the load curves of the multi-user group time series in different environments. The specific process is as follows:
[0149] Step 5.1: Calculate the demand response indicators of the load curves of the multi-user group time series in different environments. The demand response indicators include the interruptible load ratio X1, the shiftable load ratio X2, the curtailable load ratio X3, the load interruptible time X4, the load shiftable time X5, the load curtailable time X6, the proportion of electricity cost in the cost expenditure X7, the user comfort requirement X8, the user economic level X9, the power supply reliability requirement X 10 , the smart meter coverage rate X 11 and the benefits (price-based and incentive-based) obtained by users participating in demand response X 12 .
[0150] Step 5.2: Calculate the real-time demand response capabilities of the multi-user group,
[0151]
[0152]
[0153] In the formula, K is the demand capacity level, x1...x 12 are the values of the demand response indicators, k1...k 12 are the objective correlation weights of the indicators, q1...q 12 are the subjective weights of the indicators, and DR is the demand response capability.
[0154] An embodiment of the present invention also provides a demand response capability evaluation system for a multi-power load group, as Figure 3 shown, including a data acquisition unit 1, a data processing unit 2, a deep embedding clustering unit 3, a mathematical representation model generation unit 4, a load curve generation unit 5, and a demand response capability evaluation unit 6 that are connected in sequence.
[0155] The data acquisition unit 1 is used to obtain the load data of multiple load groups. The types of the load groups include industrial load, residential load, commercial load, transportation load, and agricultural load. The load data is the time-series load curves of multiple types of loads.
[0156] The data processing unit 2 is used to process the load data obtained by the data acquisition unit 1 and then transmit it to the deep embedding clustering unit 3.
[0157] The processing process of the load data is as follows:
[0158] According to the time series, type, and load value of the load data, identify the outliers in the load data and delete the outliers;
[0159] Calculate the data missing ratio D of the load data a ,
[0160]
[0161] where M is the number of measurement time nodes within the time range, N is the total number of types of the load data, indicates whether the load data of type n is missing at time m. If it is missing, its value is 1; otherwise, it is 0;
[0162] Calculate the load data missing values at each moment in the load data respectively
[0163]
[0164] When is 1, delete the load data at this moment.
[0165] The deep embedding clustering unit 3 is used to calculate multiple load types of each load group by using a deep embedding clustering algorithm based on an autoencoder according to the load data of multiple load groups, and then transmit the calculation result to the mathematical representation model generation unit 4.
[0166] The deep embedding clustering unit 3 receives the load data processed by the data processing unit, inputs the load data of multiple processed load groups into a pre-trained autoencoder model to obtain embedding features;
[0167] Input the embedding features into the clustering layer network, use KL divergence for clustering training, and then calculate the preliminary clustering results of multiple load types;
[0168] The clustering training using KL divergence is to optimize the encoder parameters in the autoencoder model through the following formula
[0169]
[0170] And optimize the clustering layer network through the following formula,
[0171]
[0172]
[0173] In the formula, P is the original distribution, p ij is a certain clustering distribution, Q is the assumed distribution, q ij is the probability that sample i is assigned to cluster j; L = KL, where KL is the KL-divergence distance; α is the degree of freedom of the Student-t distribution, initially set to 1; u j is the cluster center of cluster j, x i is the input load value, z i is the latent feature space, which is a non-linear mapping of x - z, z i = f θ (x i );
[0174] Filter out small clusters in the preliminary clustering results through a threshold to obtain multiple load types for each of the load groups.
[0175] The mathematical representation model generation unit 4 is used to evaluate the causal association relationship of the load process of the multi-load group - multi-load type based on the load data of multiple load types of the multi-load group according to the convergent cross-mapping algorithm, establish a mathematical representation model of the typical load characteristics of the real-time multi-user group considering the dynamic association relationship of the load group, and then transmit it to the load curve generation unit 5.
[0176] The mathematical representation model generation unit 4 takes the load data of multiple load types of the multi-load group as the observed value, reconstructs the shadow manifold according to a certain embedding dimension E, calculates the E + 1 neighbor points of the shadow manifold and the Euclidean distance or Mahalanobis distance, obtains the weights of each neighbor point using the Euclidean distance or Mahalanobis distance, and gets the estimated value of the observed value through weighted average, and calculates the error value between the observed value and the estimated value; repeat the above process by trying different embedding dimensions E respectively, calculate the error values between the observed value and the estimated value; select the embedding dimension E corresponding to the minimum error value as the optimal embedding dimension;
[0177] Based on Takens' embedding theorem, the shadow manifold is reconstructed from the load data of multiple load types of the multi-load group of the load group according to the optimal embedding dimension; applying the dynamic system theory to the reconstructed shadow manifold, analyzing whether the points on the reconstructed shadow manifold approach or move away from each other over time, and if there is such a trend, it is considered that there is a causal influence between the load data;
[0178] The above process is iterated multiple times, and each iteration uses a different optimal embedding dimension, and observes which optimal embedding dimension results in the most obvious causal relationship between the load data; based on the causal relationship between the load data obtained under this optimal embedding dimension, a mathematical representation model L of the typical load characteristics of real-time multi-user groups considering the dynamic association relationship of the load group is established e
[0179]
[0180] In the model, E is the association relationship between different load types in different load groups, N' is different load types of different load groups, and L n is the load curve of the time series of different types in different load groups.
[0181] The load curve generation unit 5 uses a trained LSTM neural network model to generate the load curves of the time series of multi-user groups in different environments according to the mathematical representation model of the typical load characteristics of real-time multi-user groups considering the dynamic association relationship of the load group; and then transmits them to the demand response capacity evaluation unit 6.
[0182] The demand response capacity evaluation unit 6 is used to calculate the real-time demand response capacity of the multi-user group according to the load curves of the time series of multi-user groups in different environments, and output the calculation result.
[0183] The demand response capacity evaluation unit 6 calculates the demand response indicators of the load curves of the time series of multi-user groups in different environments, and the demand response indicators include the interruptible load ratio X1, the shiftable load ratio X2, the reducible load ratio X3, the load interruptible time X4, the load shiftable time X5, the load reducible time X6, the proportion of electricity cost in the cost expenditure X7, the user comfort requirement X8, the user economic level X9, the power supply reliability requirement X 10 and the smart meter coverage rate X 11 and the benefits (price-based and incentive-based) obtained by users participating in demand response X 12 ;
[0184] Calculate the real-time demand response capacity of the multi-user group,
[0185]
[0186]
[0187] In the formula, K is the demand capacity level, x1...x 12 is the value of the demand response index, k1...k 12 is the objective correlation weight of the index, q1...q 12 is the subjective weight of the index, and DR is the demand response capacity.
[0188] An embodiment of the present invention also provides a computer-readable storage medium, including a program that can be executed by a processor to implement the above-mentioned method for evaluating the demand response capacity of a multi-source power load group.
[0189] The above has described the present invention in detail through embodiments, but the above content is only an exemplary embodiment of the present invention and cannot be considered as limiting the implementation scope of the present invention. The protection scope of the present invention is defined by the claims. Any use of the technical solutions described in the present invention, or those designed by those skilled in the art under the inspiration of the technical solutions of the present invention, within the essence and protection scope of the present invention, to achieve the above technical effects by designing similar technical solutions, or any equivalent changes and improvements made to the application scope, shall still fall within the patent coverage protection scope of the present invention.
Claims
1. A method for evaluating the demand response capacity of a multi-source power load group, characterized in that Including the following steps: Obtain the load data of the multi-load population; According to the load data of the multi-load population, adopt a deep embedding clustering algorithm based on an autoencoder to calculate various load types of each load population; Based on the load data of various load types of the multi-load population, evaluate the causal association relationship of the load process of the multi-load population - multi-load types based on the convergent cross mapping algorithm, and establish a mathematical representation model of the typical load characteristics of real-time multi-user populations considering the dynamic association relationship of the load population; Generate the load curves of the time series of the multi-user population in different environments according to the mathematical representation model of the typical load characteristics of the real-time multi-user population considering the dynamic association relationship of the load population; Calculate the real-time demand response ability of the multi-user population according to the load curves of the time series of the multi-user population in different environments; 2. The demand response capacity evaluation method for a multi-source power load group according to claim 1, characterized in that The multi-load population includes multiple of industrial load, residential load, commercial load, transportation load, and agricultural load; 3. The demand response ability evaluation method for a multi-source power load group according to claim 1 or 2, characterized in that, After obtaining the load data of the multi-load population, process the load data, and the processing process is as follows: Identify the outliers in the load data according to the time series, type, and load value of the load data, and delete the outliers; Calculate the data missing ratio D of the load data a , Where M is the number of measurement time nodes within the time range, and N is the total number of types of the load data, indicates whether the load data of type n is missing at time m. If it is missing, its value is 1; otherwise, it is 0. Calculate the missing values of the load data at each moment in the load data respectively When is 1, the load data at that moment is deleted.
4. A method for evaluating the demand response ability of a multi-source power load group according to claim 1 or 2, characterized in that, The calculation process of various load types of the load population is as follows: Input the load data of multiple load populations into a pre-trained autoencoder model to obtain embedding features; Input the embedding features into the clustering layer network, use KL divergence for clustering training, and then calculate the preliminary clustering results of various load types; The use of KL divergence for clustering training is to optimize the encoder parameters in the autoencoder model through the following formula, and optimize the clustering layer network through the following formula, where P is the original distribution, p ij is a certain cluster distribution, Q is the hypothesized distribution, q ij is the probability that sample i is assigned to cluster j; L = KL, where KL is the KL-divergence distance; α is the degree of freedom of the Student-t distribution, initially set to 1; u j is the cluster center of cluster j, x i is the input load value, z i is the latent feature space, a non-linear mapping of x - z, z i = f θ (x i ); Filter out small clustering clusters in the preliminary clustering results through a threshold to obtain various load types of each load population; 5. The demand response ability evaluation method for a multi-source power load group according to claim 1 or 2, characterized in that The process of evaluating the causal association relationship of the load process of the multi-load population - multi-load types based on the convergent cross mapping algorithm and establishing a mathematical representation model of the typical load characteristics of real-time multi-user populations considering the dynamic association relationship of the load population is as follows: Taking the load data of various load types of the multi-load population as observations, reconstruct the shadow manifold according to a certain embedding dimension E, calculate the E + 1 neighbor points and Euclidean distance or Mahalanobis distance of the shadow manifold, obtain the weights of each neighbor point using the Euclidean distance or Mahalanobis distance, and obtain the estimated value of the observation through weighted average, and calculate the error value between the observation and the estimated value; Try different embedding dimensions E respectively and repeat the above process to calculate the error values between the observations and the estimated values; select the embedding dimension E corresponding to the smallest error value as the optimal embedding dimension; Based on Takens' embedding theorem, the shadow manifold is reconstructed from the load data of multiple load types of the multi-load population of the load population according to the optimal embedding dimension; applying dynamic system theory to the reconstructed shadow manifold, analyzing whether the points on the reconstructed shadow manifold approach or move away from each other over time, and if there is such a trend, it is considered that there is a causal influence between the load data; The above process is iterated multiple times, and each iteration uses a different optimal embedding dimension, and observes which optimal embedding dimension results in the most obvious causal relationship between the load data; Based on the causal relationship between the load data obtained under the optimal embedding dimension, a mathematical representation model L of the typical load characteristics of real-time multi-user groups considering the dynamic correlation relationship of the load group is established. e In the model, E is the association relationship between different load types in different load groups, N’ is the different load types of different load groups, and L n is the load curve of the time series of different types in different load groups.
6. The demand response capacity evaluation method for a multi-source power load group according to claim 1 or 2, characterized in that The mathematical characterization model of the typical load characteristics of the real-time multi-user population considering the dynamic correlation relationship of the load population inputs the trained LSTM neural network model to calculate the load curves of the multi-user population time series under different environments.
7. The demand response ability evaluation method for a multi-source power load group according to claim 1 or 2, characterized in that, The process of calculating the real-time demand response ability of the multi-user population according to the load curves of the multi-user population time series under different environments is as follows: Calculate the demand response indicators of the load curves of multi - user groups in different environments. The demand response indicators include the proportion of interruptible load X1, the proportion of shiftable load X2, the proportion of curtailable load X3, the interruptible time of load X4, the shiftable time of load X5, the curtailable time of load X6, the proportion of electricity cost in the cost expenditure X7, the user comfort requirement X8, the user economic level X9, the power supply reliability requirement X 10 , the smart meter coverage rate X 11 and the benefits obtained by users participating in demand response (price - based and incentive - based) X 12 ; Calculate the real-time demand response ability of the multi-user population, Where K is the demand capacity level, x1...x 12 is the value of the demand response index, k1...k 12 is the objective correlation weight of the index, q1...q 12 is the subjective weight of the index, and DR is the demand response capacity.
8. A demand response capacity evaluation system for a multi-source power load group, characterized in that, including a data acquisition unit, a deep embedding clustering unit, a mathematical characterization model generation unit, a load curve generation unit, and a demand response ability evaluation unit connected in sequence, The data acquisition unit is used to obtain the load data of the multi-load population; The deep embedding clustering unit is used to calculate multiple load types of each load population by using a deep embedding clustering algorithm based on an autoencoder according to the load data of the multi-load population; The mathematical characterization model generation unit is used to evaluate the causal association relationship of the load process of the multi-load population - multiple load types based on the convergence cross-mapping algorithm according to the load data of multiple load types of the multi-load population, and establish a mathematical characterization model of the typical load characteristics of the real-time multi-user population considering the dynamic correlation relationship of the load population; The load curve generation unit is used to generate the load curves of the multi-user population time series under different environments according to the mathematical characterization model of the typical load characteristics of the real-time multi-user population considering the dynamic correlation relationship of the load population; The demand response ability evaluation unit is used to calculate the real-time demand response ability of the multi-user population according to the load curves of the multi-user population time series under different environments.
9. The demand response capacity evaluation system for a multi-source power load group according to claim 8, characterized in that The deep embedding clustering unit inputs the load data of multiple load populations into a pre-trained autoencoder model to obtain embedding features; The embedding features are input into the clustering layer network, and KL divergence is used for clustering training, and then a preliminary clustering result of multiple load types is calculated; The use of KL divergence for clustering training is to optimize the encoder parameters in the autoencoder model through the following formula, and optimize the clustering layer network through the following formula, where P is the original distribution, p ij is a certain cluster distribution, Q is the hypothesized distribution, q ij is the probability that sample i is assigned to cluster j; L = KL, where KL is the KL-divergence distance; α is the degree of freedom of the Student-t distribution, initially set to 1; u j is the cluster center of cluster j, x i is the input load value, z i is the latent feature space, which is a non-linear mapping of x - z, z i = f θ (x i ); Filter out small clustering clusters in the preliminary clustering result through a threshold to obtain multiple load types of each load population.
10. The demand response capacity evaluation system for a multi-source power load group as described in claim 8, wherein, The mathematical representation model generation unit uses the load data of multiple load types of a multi-load population as observed values, reconstructs a shadow manifold according to a certain embedding dimension E, calculates the E+1 neighbor points and Euclidean distance or Mahalanobis distance of the shadow manifold, obtains the weights of each neighbor point using the Euclidean distance or Mahalanobis distance, and obtains an estimated value of the observed value through weighted average, and calculates the error value between the observed value and the estimated value; Repeat the above process by trying different embedding dimensions E respectively, and calculate the error values between the observed value and the estimated value; select the embedding dimension E corresponding to the smallest error value as the optimal embedding dimension; Based on Takens' embedding theorem, reconstruct the shadow manifold from the load data of multiple load types of the multi-load population according to the optimal embedding dimension; apply the dynamic system theory to the reconstructed shadow manifold, and analyze whether the points on the reconstructed shadow manifold approach or move away from each other over time. If there is the above trend, it is considered that there is a causal influence between the load data; Perform multiple iterations on the above process, use different optimal embedding dimensions in each iteration, and observe which optimal embedding dimension results in the most obvious causal relationship between the load data; Based on the causal relationship between the load data obtained under the optimal embedding dimension, a mathematical representation model L of the typical load characteristics of real-time multi-user groups considering the dynamic correlation relationship of the load group is established e In the model, E is the association relationship between different load types in different load groups, N’ is different load types of different load groups, and L n is the load curve of time series of different types in different load groups.
11. The demand response capability evaluation system for a multi-source power load group according to claim 8, characterized in that, The demand response capacity evaluation unit calculates the demand response indicators of the load curves of multi - user groups in different environments over time series. The demand response indicators include the interruptible load ratio X1, the shiftable load ratio X2, the curtailable load ratio X3, the interruptible load time X4, the shiftable load time X5, the curtailable load time X6, the proportion of electricity cost in the cost expenditure X7, the user comfort requirement X8, the user economic level X9, the power supply reliability requirement X 10 , the smart meter coverage rate X 11 and the benefits obtained by users participating in demand response (price - based and incentive - based) X 12 ; Calculate the real-time demand response ability of the multi-user population, where K is the demand capacity level, x1...x 12 are the values of the demand response indicators, k1...k 12 are the objective correlation weights of the indicators, q1...q 12 are the subjective weights of the indicators, and DR is the demand response capacity.
12. A demand response capacity evaluation system for a multi-source power load group according to any one of claims 8-11, characterized in that, It further includes a data processing unit for processing the load data acquired by the data acquisition unit and then transmitting it to the deep embedding clustering unit, The processing process of the load data is as follows: According to the time series, type and load value of the load data, identify the outliers in the load data and delete the outliers; Calculate the data missing ratio D of the load data a , Where M is the number of measurement time nodes within the time range, and N is the total number of types of the load data, indicates whether the load data of type n is missing at time m. If it is missing, its value is 1; otherwise, it is 0. Calculate the missing values of the load data at each moment in the load data respectively When is 1, the load data at that moment is deleted.
13. A computer-readable storage medium, characterized in that, It includes a program that can be executed by a processor to implement a method for evaluating the demand response ability of a multi-source power load population according to any one of claims 1-7.