High-energy-consumption industrial user load classification method and system, electronic equipment and medium

Abstract

The invention provides a high-energy-consumption industrial user load classification method and system, electronic equipment and a medium, and is applied to the technical field of intelligent power utilization. The method comprises the steps of determining a multi-dimensional load characteristic index based on daily load data of each industrial user by considering a load change characteristic, a load utilization rate and electricity price sensitivity of the user; time feature extraction is performed on the daily load data of each industrial user based on a convolutional auto-encoder, and power consumption features are determined based on the extracted load time features of each industrial user and the multi-dimensional load feature indexes; inputting the power consumption characteristics into a self-organizing mapping neural network to determine output nodes to which the input data are mapped; screening out a target output node based on the quantity of the input data corresponding to each output node; and clustering the power consumption characteristics of the industrial users by taking the target output node as an initial clustering center. According to the method, the problems of poor clustering effect, large calculation amount and low processing efficiency in the industrial load clustering process are solved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent electricity consumption technology, and in particular to a method and system for classifying loads of high-energy-consuming industrial users, electronic equipment, and a medium. Background Art

[0002] With the increasing reliance on renewable energy, the problem of balancing power grid supply and demand is becoming increasingly serious. To solve this problem, it is essential to use demand-side resources to enhance the stability of the power grid. High-energy-consuming industrial users play a key role in demand response due to their remarkably stable and manageable power consumption. Understanding and exploring the load characteristics of high-energy-consuming industrial users is essential to optimizing demand response capabilities, ensuring safe operation of the power grid, and achieving lean management.

[0003] Most existing methods use clustering methods to classify industrial users with similar consumption behaviors into groups, thereby implementing targeted load management strategies. However, the industrial power consumption data brought by the existing advanced metering infrastructure has high dimensionality and nonlinear characteristics, which increases the complexity and difficulty of clustering; and the clustering process often ignores the temporal characteristics of load data, resulting in poor clustering effect, large amount of calculation, and low processing efficiency. Summary of the invention

[0004] In order to overcome the problems of poor clustering effect, large amount of calculation and low processing efficiency in the above industrial load clustering process, the present invention provides a method and system for classifying loads of high-energy-consuming industrial users.

[0005] In one aspect, the present invention provides a method for classifying loads of high-energy-consuming industrial users, the method comprising: Considering the load variation characteristics, load utilization and user's sensitivity to electricity prices, the multi-dimensional load characteristic indicators of each industrial user are determined based on the daily load data of each industrial user; The time characteristics of the daily load data of each industrial user are extracted based on the convolutional autoencoder, and the power consumption characteristics of each industrial user are determined based on the extracted load time characteristics and multi-dimensional load characteristic indicators of each industrial user; Input the power consumption characteristics of each industrial user into the self-organizing map neural network to determine the output node to which each input data is mapped; filter out the target output node based on the number of input data corresponding to each output node; The target output node is used as the initial clustering center to cluster the power consumption characteristics of each industrial user to obtain the load classification result of each industrial user.

[0006] Optionally, determining the power consumption characteristics of each industrial user based on the extracted load time characteristics and multi-dimensional load characteristic indicators of each industrial user includes: For each industrial user, the index weight of the load characteristic index of each industrial user in each dimension is determined based on the entropy weight method; Based on the weighted fusion of the load characteristic indicators of each dimension and the corresponding indicator weights, the key load characteristics of the corresponding industrial users are obtained; The power consumption characteristics of each industrial user are determined based on the extracted load time characteristics of each industrial user and the corresponding key load characteristics.

[0007] Optionally, determining the index weight of the load characteristic index of each industrial user in each dimension based on the entropy weight method includes: Based on the daily load data of each industrial user, the entropy of the corresponding load characteristic index of each dimension is calculated; Based on the entropy of the load characteristic index of each dimension, the index weight of the load characteristic index of each dimension is calculated using an exponential function.

[0008] Optionally, the entropy of the load characteristic index of each dimension is calculated as follows: ; in, is the entropy of the load characteristic index c, a represents the total number of load characteristic indicators, d represents the index of daily load data, is the value of the load characteristic index c corresponding to the d-th daily load data sample, and m0 represents the total number of daily load data; The calculation formula of the indicator weight of the load characteristic indicator of each dimension is: ; in, is the index weight of load characteristic index c, Indicates The entropy of a load characteristic index, e is the base of the natural logarithm.

[0009] Optionally, the load variation characteristics, load utilization rate and user's sensitivity to electricity prices are considered, and multi-dimensional load characteristic indicators of each industrial user are determined based on the daily load data of each industrial user, including: For each industrial user, the corresponding daily average load, daily total load, daily maximum load, daily minimum load, daily peak period average load and daily valley period average load are calculated based on the daily load data of the industrial user; Determine the daily load rate based on the daily average load and the daily maximum load, and determine the daily load peak-to-valley difference rate based on the daily maximum load and the daily minimum load; determine the daily maximum load utilization time based on the daily total load and the daily maximum load; determine the peak period load rate based on the daily peak period average load and the daily maximum load, and determine the valley period load rate based on the daily valley period average load and the daily maximum load; The daily load rate and the daily load peak-to-valley difference rate are used as load characteristic indicators of the corresponding dimension of load change characteristics, the daily maximum load utilization duration is used as the load characteristic indicator of the corresponding dimension of load utilization rate, and the peak period load rate and the valley period load rate are used as load characteristic indicators of the corresponding dimension of user's electricity price sensitivity.

[0010] Optionally, the convolutional autoencoder includes an encoder and a decoder, and the training process of the convolutional autoencoder includes: Obtain historical load data of each industrial user as training data; Input the training data into the convolutional autoencoder, and perform convolution and pooling processing on the training data through the encoder to obtain a corresponding hidden layer feature vector; Deconvolution and upsampling are performed on the hidden layer feature vector by the decoder to reconstruct output data with the same dimension as the input data; A loss function is calculated based on the output data and the corresponding input data, and model parameters of the encoder and the decoder are updated based on the loss function.

[0011] Optionally, the extracting time features of the daily load data of each industrial user based on the convolutional autoencoder includes: The daily load data of each industrial user is input into the trained convolutional autoencoder, and the hidden layer low-dimensional feature vector processed by the encoder is extracted as the corresponding load time feature.

[0012] Optionally, the self-organizing map neural network includes input layer and output layer grid nodes, and the step of inputting the power consumption characteristics of each industrial user into the self-organizing map neural network to determine the output node to which each input data is mapped includes: The power consumption characteristics corresponding to the historical load data of each industrial user are used as training samples, and the self-organizing map neural network is trained using the training samples to obtain the weight of each output node in the output layer grid node; Based on the weight of each output node of the output layer, the power consumption characteristics of each industrial user are mapped, and the output node to which each input data is mapped is determined; The input data is the power consumption characteristics of industrial users, and the output nodes are used to characterize the regional classification of the corresponding power consumption characteristics to achieve a specific response to the input data.

[0013] Optionally, the step of selecting a target output node based on the amount of input data corresponding to each output node includes: Determine the local density of each output node based on the amount of input data corresponding to each output node; For each mapped output node, if an output node is determined to be the node with the largest local density in the neighborhood of the current output node, the distance between the current output node and the farthest node in the neighborhood of the current output node is taken as the attribute distance of the current output node; If it is determined that an output node is not the node with the maximum local density in the neighborhood of the current output node, the minimum distance between the current output node and each node in the target node set is taken as the attribute distance of the current output node; Screening out a plurality of nodes having local density and attribute distance from the output layer grid nodes as the target output nodes; The target node set is a node set whose local density in the current output node domain is higher than that of the current output node.

[0014] Optionally, clustering the power consumption characteristics of each industrial user is performed with the target output node as the initial clustering center to obtain the load classification result of each industrial user, including: Determine the target output nodes corresponding to different numbers of cluster centers, and use the weights corresponding to the target output nodes as the initial cluster centers under the corresponding number of cluster centers; For each number of cluster centers, cluster the electricity consumption characteristics of each industrial user based on the corresponding initial cluster center and K-means clustering algorithm to obtain the clustering results under each cluster center data; Based on the clustering performance analysis of the clustering results under different numbers of cluster centers, the target number of cluster centers is determined; The clustering results corresponding to the target number of cluster centers are used as load classification results for each industrial user.

[0015] Optionally, before determining the multi-dimensional load characteristic index of each industrial user based on the daily load data of each industrial user, taking into account the load variation characteristics, load utilization rate and user's electricity price sensitivity, the method further includes: Pre-process the daily load data of each industrial user; Wherein, the preprocessing includes at least one of missing data filling, outlier detection and correction, and data normalization.

[0016] Optionally, after obtaining the load classification results of each industrial user, the method further includes: Carry out load pattern analysis based on the load classification results of each industrial user to determine the load pattern of each industrial user; Predict the electricity demand of each industrial user based on their load patterns; Based on the prediction results, load equipment of each industrial user is adjusted and controlled.

[0017] Optionally, before determining the multi-dimensional load characteristic index of each industrial user based on the daily load data of each industrial user, the method further includes: Collect real-time operation data of internal load equipment of each industrial user during the day; Collect real-time measurement data of internal measurement equipment of each industrial user during the day; Collect daily control status data of internal switch equipment of each industrial user; Based on the load feature extraction of the daily real-time operation data, the daily real-time measurement data and the daily control status data, the daily load data of each industrial user is obtained.

[0018] On the other hand, the present invention also provides a high-energy-consuming industrial user load classification system, the system comprising: The load feature extraction module is used to determine the multi-dimensional load feature indicators of each industrial user based on the daily load data of each industrial user, taking into account the load variation characteristics, load utilization rate and user's sensitivity to electricity prices; extract the time feature of the daily load data of each industrial user based on the convolutional autoencoder, and determine the power consumption characteristics of each industrial user based on the extracted load time characteristics and multi-dimensional load feature indicators of each industrial user; A node screening module is used to input the power consumption characteristics of each industrial user into the self-organizing map neural network to determine the output node to which each input data is mapped; and to screen out the target output node based on the number of input data corresponding to each output node; The clustering module is used to cluster the power consumption characteristics of each industrial user with the target output node as the initial clustering center to obtain the load classification result of each industrial user.

[0019] On the other hand, the present invention also provides an electronic device, comprising: at least one processor and a memory; the memory and the processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, any one of the above-mentioned methods for classifying loads of high-energy-consuming industrial users is implemented.

[0020] On the other hand, the present invention also provides a readable storage medium having an execution program stored thereon, and when the execution program is executed, any one of the above-mentioned methods for classifying loads of high-energy-consuming industrial users is implemented.

[0021] Compared with the prior art, the present invention has the following beneficial effects: The present invention provides a load classification method and system for high-energy-consuming industrial users. The method extracts multi-dimensional load characteristic indicators by considering load change characteristics, load utilization and user's electricity price sensitivity, reflects load characteristics from multiple dimensions such as load change, load utilization and user's electricity price sensitivity, and uses a convolutional autoencoder model to extract time features to ensure the comprehensiveness and accuracy of the extracted load features, thereby realizing accurate mining of the comprehensive load characteristics of industrial users. The feature extraction process before clustering can greatly reduce the clustering complexity, reduce the calculation amount of the clustering process, and improve the clustering efficiency.

[0022] The present invention inputs the extracted power consumption characteristics of each industrial user into a self-organizing mapping neural network, mines the distribution of the input characteristics, and screens out the target output node based on the number of input data corresponding to each output node, and clusters the power consumption characteristics of each industrial user with the target output node as the initial clustering center. The clustering process can be guided by mining the distribution of the input characteristics, thereby improving the clustering effect and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is one of the flow charts of a method for classifying loads of high-energy-consuming industrial users of the present invention; Figure 2 A schematic diagram of the structure of a convolutional autoencoder according to an example of the present invention; Figure 3 The structure and mapping diagram of a self-organizing map neural network of an example of the present invention; Figure 4 The second flow chart of a method for classifying loads of high-energy-consuming industrial users of the present invention; Figure 5 A schematic diagram of the architecture of a high-energy-consuming industrial user load classification system of the present invention; Figure 6 It is a schematic structural diagram of the electronic device of the present invention. DETAILED DESCRIPTION

[0024] The specific implementation modes of the present invention are further described in detail below with reference to the accompanying drawings.

[0025] Embodiment 1: The present invention provides a method for classifying loads of high energy-consuming industrial users, as shown in the schematic diagram Figure 1 As shown, including: Step S110, considering load variation characteristics, load utilization rate and user's sensitivity to electricity prices, determining multi-dimensional load characteristic indicators of each industrial user based on daily load data of each industrial user; Step S120, extracting time features of the daily load data of each industrial user based on the convolutional autoencoder, and determining the power consumption features of each industrial user based on the extracted load time features and multi-dimensional load feature indicators of each industrial user; Step S130, inputting the power consumption characteristics of each industrial user into the self-organizing map neural network to determine the output node to which each input data is mapped; and filtering out the target output node based on the number of input data corresponding to each output node; Step S140 , clustering the power consumption characteristics of each industrial user with the target output node as the initial clustering center to obtain a load classification result of each industrial user.

[0026] In this example implementation, characteristic indicators that can characterize load change characteristics, load utilization and user's electricity price sensitivity are extracted from daily load data, and the convolutional autoencoder and the self-organizing map neural network are obtained after training with the historical load data of each industrial user. The extracted load time feature can be a low-dimensional feature vector of the hidden layer of the convolutional autoencoder. The self-organizing map neural network is used to establish a mapping relationship between the input data and the output node of the network. The self-organizing map neural network includes an input layer and an output layer, and the output layer is composed of grid nodes. K-means, fuzzy C-means and spectral clustering algorithms can be used for clustering. The present invention extracts multi-dimensional load characteristic indicators by considering load change characteristics, load utilization and user's electricity price sensitivity, reflects load characteristics from multiple dimensions such as load change, load utilization and user electricity price sensitivity, and uses the convolutional autoencoder model to extract time features to ensure the comprehensiveness and accuracy of the extracted load features, thereby realizing accurate mining of the comprehensive load characteristics of industrial users; the feature extraction process before clustering can greatly reduce the complexity of clustering, reduce the amount of calculation in the clustering process, and improve the clustering efficiency. By inputting the extracted power consumption characteristics of each industrial user into the self-organizing map neural network, mining the distribution of the input characteristics, and screening the target output node based on the number of input data corresponding to each output node, clustering the power consumption characteristics of each industrial user with the target output node as the initial clustering center, the clustering process can be guided by mining the distribution of the input characteristics, thereby improving the clustering effect and efficiency. The present invention uses a self-organizing map (SOM) network to determine the initial clustering center of the K-means algorithm, overcoming the local optimal problem that may be caused by the traditional K-means algorithm when randomly selecting the initial clustering center, thereby improving the stability and effect of clustering.

[0027] During the load data collection process, due to measurement anomalies, communication interruptions and human errors, the data may be abnormal or missing. Therefore, it is necessary to preprocess the original load data. In an example implementation, before S110 considers the load change characteristics, load utilization and user's electricity price sensitivity, and determines the multi-dimensional load characteristic indicators of each industrial user based on the historical load data of each industrial user, it also includes: Pre-process the daily load data of each industrial user; Wherein, the preprocessing includes at least one of missing data filling, outlier detection and correction, and data normalization.

[0028] In this example implementation, it is assumed that is the daily load data of industrial users, is the daily load data of the i-th industrial user, is the daily load data of the nth industrial user, n is the total number of industrial users, where ; represents the Mth sampling point data of the i-th industrial user, and M is the total number of sampling points of the i-th industrial user. In order to maintain the smoothness of the load curve, the missing data can be filled by the mean value based on the same time, specifically: Assume is the m1th sampling point data of the i-th industrial user on date k. For missing data values, the filling value can be calculated by the following formula: (1) in, is the date set with measured values ​​at the m1th sampling point, which can be a date set close to the date of the value to be filled. represents the measurement value of the m1th sampling point of the i-th industrial user on date k, N is The number of days in is the missing filling value of the m1-th sampling point of the ith industrial user on date k. In particular, when the proportion of missing data in the daily load data exceeds 20%, the daily load data is considered invalid and can be deleted directly.

[0029] In addition to the case of missing data, load data may also be abnormal due to measurement equipment failure. For this case, the horizontal similarity detection method can be used to identify abnormal data. Specifically, assuming is the load curve of an industrial user on day k, is the load data to be tested, that is, the data of the m1th sampling point in the load curve of the kth day. First, the load change rate of the sampling point m1 is calculated according to formula (2): , then specify the normal range of the load change rate of each sampling point, and finally determine whether the daily load change rate of each sampling point is abnormal. If there is an abnormal value, use the value of the adjacent sampling point to make an abnormal correction. For this industrial user, (3) can be used for abnormal correction.

[0030] (2) (3) in, is the load data of the m1+1th sampling point in the load curve of the kth day, is the load curve of the kth day. The load data of the sampling points, is the load curve of the kth day. The load data of the sampling points, , are the number of sampling points taken forward and backward respectively, and Represents the total number of sampling points taken forward and backward respectively. By taking points before and after the sampling point m1, the average change rate of the sampling point is calculated to eliminate the abnormal change rate. In this example, you can set .

[0031] Since the load power values ​​of different industrial users may vary greatly, in order to better quantify the similarity between data and improve the accuracy of subsequent multi-dimensional power feature acquisition, the load data can be normalized, such as minimum-maximum normalization processing. For a certain industrial user, specifically: (4) in, for The normalized result is The data of the m1th sampling point in the load curve of the kth day, is the total number of sampling points of the daily load data of industrial users, and max and min are the maximum and minimum values ​​respectively.

[0032] This example performs preprocessing on the sampled daily load data, such as missing data filling, outlier detection and correction, and data normalization, to improve the accuracy of subsequent feature extraction.

[0033] In an example implementation, the method of determining the power consumption characteristics of each industrial user based on the extracted load time characteristics and multi-dimensional load characteristic indicators of each industrial user includes: For each industrial user, the index weight of the load characteristic index of each industrial user in each dimension is determined based on the entropy weight method; Based on the weighted fusion of the load characteristic indicators of each dimension and the corresponding indicator weights, the key load characteristics of the corresponding industrial users are obtained; The power consumption characteristics of each industrial user are determined based on the extracted load time characteristics of each industrial user and the corresponding key load characteristics.

[0034] In this example implementation, the load characteristic index of each dimension describes the power consumption characteristics of industrial users to different degrees, and their importance and importance vary. Therefore, it is necessary to set the weight coefficient of the load characteristic index so that it can objectively describe the importance of each load characteristic index. This example uses the entropy weight method to determine the index weight of each load characteristic index, and obtains the key load characteristics by weighted fusion of each load characteristic index and the corresponding index weight. For example, if there are 5 load characteristic indicators, the corresponding key load characteristics after weight assignment can represent the industrial load characteristic set of industrial user power consumption characteristics. , is the index weight of the first load characteristic index, L 1 is the first load characteristic index, such as L 1 It can be a daily load rate. For example, the index weight of the load characteristic index of each industrial user in each dimension is determined based on the entropy weight method, including: Based on the daily load data of each industrial user, the entropy of the corresponding load characteristic index of each dimension is calculated; Based on the entropy of the load characteristic index of each dimension, the index weight of the load characteristic index of each dimension is calculated using an exponential function.

[0035] In this example implementation, the entropy of the load characteristic index may be calculated based on the definition of entropy, and the calculation formula of the entropy of the load characteristic index of each dimension is: ; (5) in, is the entropy of the load characteristic index c, a represents the total number of load characteristic indicators, d represents the index of daily load data, is the value of the load characteristic index c corresponding to the d-th daily load data sample, and m0 represents the total number of daily load data; The calculation formula of the indicator weight of the load characteristic indicator of each dimension is: ; (6) in, is the index weight of load characteristic index c, Indicates The entropy of each load characteristic indicator, e is the base of the natural logarithm. It should be noted that the use of exponential functions for weight calculation in this example can more scientifically measure the uncertainty and information content of the distribution of each indicator, so as to better reflect the multiplication and accumulation properties of probability. In addition, the convexity of the exponential function helps to capture the uniformity and uncertainty of data distribution, while dealing with zero value problems, and can fairly measure the contributions of low-probability and high-probability events, so that the final key load characteristics can more accurately describe the load characteristics.

[0036] In some example implementations, the consideration of load variation characteristics, load utilization, and user's sensitivity to electricity prices in S110 to determine multi-dimensional load characteristic indicators of each industrial user based on daily load data of each industrial user includes: For each industrial user, the corresponding daily average load, daily total load, daily maximum load, daily minimum load, daily peak period average load and daily valley period average load are calculated based on the daily load data of the industrial user; Determine the daily load rate based on the daily average load and the daily maximum load, and determine the daily load peak-to-valley difference rate based on the daily maximum load and the daily minimum load; determine the daily maximum load utilization time based on the daily total load and the daily maximum load; determine the peak period load rate based on the daily peak period average load and the daily maximum load, and determine the valley period load rate based on the daily valley period average load and the daily maximum load; The daily load rate and the daily load peak-to-valley difference rate are used as load characteristic indicators of the corresponding dimension of load change characteristics, the daily maximum load utilization duration is used as the load characteristic indicator of the corresponding dimension of load utilization rate, and the peak period load rate and the valley period load rate are used as load characteristic indicators of the corresponding dimension of user's electricity price sensitivity.

[0037] In this example implementation, basic statistics of the data can be calculated first through daily load data, such as daily average load, daily total load, daily maximum load, daily minimum load, daily peak period average load and daily valley period average load, where the peak period and valley period can adopt the current regulations or standards. Based on the above basic statistics, load characteristic indicators that can reflect the load change characteristics, load utilization and user's electricity price sensitivity are determined. For example, the first load characteristic indicator can be selected as the daily load rate L 1 The second load characteristic index is the daily peak-to-valley difference rate L 2 The third load characteristic index is the daily maximum load utilization hours L 3 To reflect the load change characteristics and time utilization efficiency of industrial users; the electricity prices of industrial users in different time periods are significantly different, and the fourth load characteristic indicator is selected as the peak load rate L 4 and the fifth load characteristic index is the valley load factor L5 To reflect the user's sensitivity to electricity prices. The calculation method of each load characteristic index is shown in Table 1, which shows the specific calculation method of the above five load characteristic indicators.

[0038] Table 1 Calculation method of industrial user load characteristic index In Table 1, represents the load data of industrial users at time t, , , , , They represent the daily average load, daily maximum load, daily minimum load, daily peak-time average load, and daily valley-time average load of industrial users, respectively. This example describes the power consumption characteristics of industrial users by selecting load characteristic indicators to index the characteristics and attributes of industrial users' power consumption habits, which is convenient for subsequent data processing.

[0039] Traditional feature extraction methods such as principal component analysis (PCA) or independent component analysis (ICA) are difficult to mine nonlinear relationships in data. Therefore, this example uses a convolutional autoencoder for feature extraction. Before feature extraction, the convolutional autoencoder needs to be trained. The training process of the convolutional autoencoder includes: Obtain historical load data of each industrial user as training data; Input the training data into the convolutional autoencoder, and perform convolution and pooling processing on the training data through the encoder to obtain a corresponding hidden layer feature vector; Deconvolution and upsampling are performed on the hidden layer feature vector by the decoder to reconstruct output data with the same dimension as the input data; A loss function is calculated based on the output data and the corresponding input data, and model parameters of the encoder and the decoder are updated based on the loss function.

[0040] In this example implementation, the convolutional autoencoder (CAE) is an unsupervised learning algorithm, which consists of an input layer, a hidden layer, and an output layer, and is a type of neural network. Specifically, the convolutional autoencoder includes an encoder and a decoder. The convolutional autoencoder replaces the fully connected layer in the traditional autoencoder with a convolutional layer and a pooling layer. Its structure is as follows: Figure 2 As shown in Figure 2. Due to the weight sharing in the convolutional layer, the number of training parameters in CAE is less than that in the fully connected layer, which reduces overfitting, speeds up training, and is suitable for processing more complex high-dimensional data. For example, suppose the input data is is the industrial user load data set, is the load data of the i-th industrial user, , the dimension of the load data at each sampling point is , CAE includes an encoder and a decoder. During the encoding process, the initial input data First, through the convolution layer and pooling layer in the encoder, the feature vector q of the original vector in the hidden layer is obtained. In the decoding process, q in the hidden layer is reconstructed through the deconvolution layer and upsampling pooling layer in the decoder to obtain the same feature vector as the input data. The output data has the same dimension. The encoding and decoding process is shown in Formula 7 and Formula 8: (7) (8) Among them, formula (7) represents the encoding process of CAE, represents the one-dimensional encoder convolution kernel; Indicates bias in the encoding process; represents the convolution operator; q is the low-dimensional feature vector in the hidden layer after encoding, , is the mth element of q; represents the activation function of the encoder, which can be a ReLU function. Formula (8) represents the decoding process of CAE, represents the one-dimensional decoder convolution kernel; Indicates the deviation in the decoding process; Represents the reconstructed input data; represents the activation function of the decoder, which can be a Sigmoid function. The loss function in the training process can be calculated using the mean square error (MSE). CAE learns the optimal network parameters by minimizing the loss function to reconstruct the data. As close as possible to the input data .

[0041] Based on the trained convolutional autoencoder, the step of extracting time features from the daily load data of each industrial user based on the convolutional autoencoder in S120 includes: The daily load data of each industrial user is input into the trained convolutional autoencoder, and the hidden layer low-dimensional feature vector processed by the encoder is extracted as the corresponding load time feature.

[0042] In this example implementation, the daily load data of each industrial user to be processed can be input into the trained convolutional autoencoder to extract the hidden layer low-dimensional feature vector, that is, to obtain the low-dimensional feature vector q of the original input, where the dimension of q is less than That is to say, the dimension of the extracted load time feature is smaller than the input daily load data, thereby realizing the time feature extraction of industrial user load data and reducing the dimension of input data. Afterwards, the low-dimensional feature vector H in the CAE hidden layer is extracted as the time feature of industrial user load data, and combined with the industrial user load feature set L to form a multidimensional feature set C, completing the construction of multidimensional power consumption features. , They represent the first to fifth load time characteristics, namely the daily load rate, daily peak-to-valley difference rate, daily maximum load utilization hours, peak load rate, and valley load rate in the above example Table 1, q 1 Represents the first load time feature. The convolutional autoencoder can overcome the subjectivity of the artificially specified load characteristic indicators and the difficulty in representing the complex time relationship in the load curve; by considering the time series characteristics of the load data, the time characteristics of the industrial user load data are extracted to improve the industrial load clustering effect.

[0043] Next, it is necessary to cluster the loads of industrial users based on the Selective Kernel algorithm. When the traditional K-means method processes clustering tasks, it is difficult to determine the initial cluster center. In some example implementations, the step of inputting the power consumption characteristics of each industrial user into the self-organizing map neural network to determine the output node to which each input data is mapped includes: The power consumption characteristics corresponding to the historical load data of each industrial user are used as training samples, and the self-organizing map neural network is trained using the training samples to obtain the weight of each output node in the output layer grid node; Based on the weight of each output node in the output layer, the power consumption characteristics of each industrial user are mapped, and the output node to which each input data is mapped is determined.

[0044] In this example implementation, the input data is the electricity consumption characteristics of industrial users, and the output nodes are used to characterize the regional classification of the corresponding electricity consumption characteristics to achieve a specific response to the input data. The activation of the output nodes corresponding to the input data reflects the distribution of the input data in the feature space. The distribution of the input data in the feature space is mined by the self-organizing map neural network to guide the clustering process, which can improve the clustering effect. The self-organizing map neural network in this example includes input layer and output layer grid nodes. A SOM network (Self-Organizing Map, SOM) can be established first. The electricity consumption characteristics corresponding to the historical load data of each industrial user can be used to train the self-organizing map neural network to obtain the weight of each output layer node, that is, to obtain the mapping relationship between the daily load data and the output layer nodes. SOM is an important unsupervised learning method, which is often used in clustering, high-dimensional visualization, data compression, feature extraction, etc. SOM is essentially a two-layer neural network consisting of an input layer and an output layer. The relationship between the output layer nodes and the input data is established through a competitive learning process. Each node competes to become the closest match to a given input data point. The output layer nodes have the same dimension as the input data. Its network structure is as follows: Figure 3 As shown. The training process of SOM may include: 1) Establish a self-organizing map neural network The SOM network is constructed using the historical multidimensional feature set of industrial users (i.e., the multidimensional historical electricity consumption characteristics) as input. Assume that the number of samples is , the sample dimension is f, the number of neurons in the input layer is set to m2, and the number of nodes in the output layer is set to ,For example, .

[0045] Randomly initialize the weights of the output layer neurons ,in , Represents the output node The f-th dimension weight of .

[0046] 2) Traverse each node in the output layer and find the winning node Read the initial feature set from the multidimensional feature set of the industrial user As input, the Euclidean distance between the input data and each output node is calculated, and the output node with the closest distance to the input data is selected as the winning node.

[0047] 3) Calculate the weight update amplitude of each output node Assume that the weight update amplitude of the winning node is 1. Based on the Gaussian kernel function, the update amplitude of each node is calculated according to the distance between each output node and the winning node. The update amplitude calculation formula is as follows: (9) in, Output node The update range, , , Indicates the output node closest to the current input data. is the distance parameter, which indicates the sensitivity of the update amplitude to the distance between nodes. increase, Gradually getting smaller, The update formula is as follows: (10) in, is the initial distance parameter, which can be taken ; maxstep is the maximum number of iterations, , T is the total time of the time level.

[0048] 4) Update the weight of each output node After obtaining the weight update amplitude of each output node, update the weights of the winning node and its domain nodes. The node weight update formula is as follows: (11) in, Represents the output node In the +The weight at 1 iteration, Represents the output node In the The weight at the iteration, Indicates The learning rate of the iteration is The update formula is as follows: (12) in, is the initial learning rate, which can be taken .

[0049] Repeat the above steps 2) to 4) until the number of iterations reaches maxstep and the training is completed. After the training is completed, the weight matrix of each output node in the output layer is obtained. Based on the weights of each output node in the output layer, the power consumption characteristics of each industrial user are mapped to determine the output node to which each input data is mapped.

[0050] In some example implementations, the step of selecting a target output node based on the amount of input data corresponding to each output node in step S130 includes: Determine the local density of each output node based on the amount of input data corresponding to each output node; For each mapped output node, if an output node is determined to be the node with the largest local density in the neighborhood of the current output node, the distance between the current output node and the farthest node in the neighborhood of the current output node is taken as the attribute distance of the current output node; If it is determined that an output node is not the node with the maximum local density in the neighborhood of the current output node, the minimum distance between the current output node and each node in the target node set is taken as the attribute distance of the current output node; Screening out a plurality of nodes having local density and attribute distance from the output layer grid nodes as the target output nodes; The target node set is a node set whose local density in the current output node domain is higher than that of the current output node.

[0051] In this example implementation, the mapping relationship between input and output nodes is obtained according to the weight matrix of the output layer nodes, and the number of inputs corresponding to each output node is counted, and the number of input data is used as the local density of the output layer nodes. To facilitate calculation, different columns of the weight matrix can be concatenated and converted into column vectors. For example, the matrix Convert to column vector For each node in the transformed column vector, if the current node is not the node with the highest local density, its attribute distance is defined as the minimum distance between other nodes in the field with higher local density than the current node. The distance between two nodes and the node attribute distance are calculated as follows: (13) (14) in, is the distance between node r and node s, is the node r's The weight of the dimension, is the node s The weight of the dimension, is the attribute distance of node r, represents the local density of node r, represents the local density of node s.

[0052] If the current node is the node with the largest local density, the distance attribute It is defined as the Euclidean distance between the current node and the farthest node in the neighborhood, and its calculation formula is: (15) in, Indicates the domain scope of node r.

[0053] According to the above process, the local density and attribute distance of each output node in the column vector are calculated, and then K output nodes with local density and attribute distance are randomly selected as target output nodes, where K is the number of clusters.

[0054] This example is based on two basic ideas of the Density Peak Clustering (DPC) algorithm: the cluster center usually has a large local density, that is, there are a large number of sample points around it; the distance between different cluster centers is far; based on the above two basic ideas, the local density and attribute distance of each output node are calculated, and the target output nodes are screened according to the local density and attribute distance of each node in the output layer to guide the determination of the initial cluster center and improve the clustering performance of industrial user loads.

[0055] Traditional clustering methods encounter the problem of difficulty in determining the number of clusters when processing clustering tasks, and improper selection of the number of clusters will have an adverse effect on the clustering effect. Therefore, in some embodiments, S140 clusters the power consumption characteristics of each industrial user with the target output node as the initial clustering center to obtain the load classification results of each industrial user, including: Determine the target output nodes corresponding to different numbers of cluster centers, and use the weights corresponding to the target output nodes as the initial cluster centers under the corresponding number of cluster centers; For each number of cluster centers, cluster the electricity consumption characteristics of each industrial user based on the corresponding initial cluster center and K-means clustering algorithm to obtain the clustering results under each cluster center data; Based on the clustering performance analysis of the clustering results under different numbers of cluster centers, the target number of cluster centers is determined; The clustering results corresponding to the target number of cluster centers are used as load classification results for each industrial user.

[0056] In this example implementation, several different numbers of cluster centers can be pre-configured according to actual conditions, for example, the number of candidate cluster centers is set to 6, 8, and 9. Then, for each number of cluster centers, the corresponding number of target output nodes is screened out, and the weight corresponding to the target output node is used as the initial cluster center. The K-means clustering algorithm is used to cluster the power consumption characteristics of each industrial user to obtain the clustering results under each number of cluster centers. The target number of cluster centers with the best clustering performance is screened out by analyzing the clustering performance of the clustering results under each number of cluster centers, and the load classification results of each industrial user corresponding to the clustering results of the target number of cluster centers are obtained. In this example, the clustering performance can be evaluated by observing the Calinski-Harabasz index (CHI), Davies-Bouldin index (DBI) and silhouette coefficient (SC) of the clustering results under different numbers of cluster centers. CHI is the ratio of the inter-class dispersion to the intra-class dispersion. The larger the CHI value, the greater the intra-class similarity and the smaller the inter-class similarity, that is, the better the clustering performance. The calculation formula of CHI is as follows: (16) in, is the Calinski-Harabasz index, and Represent the between-class variance and within-class variance respectively; represents the number of clusters; Indicates the amount of data; between-class variance and the intra-class variance The calculation formula is as follows: (17) (18) in, Indicates The cluster center of the class; Represents the center of all data samples; Indicates Samples With Class membership, if Belong to Class, then Takes 1, otherwise takes 0.

[0057] DBI describes the compactness of data in the same cluster. The smaller the DBI value, the lower the dispersion and the better the clustering performance. The calculation formula of DBI is as follows: (19) in, is the Davies-Bouldin index, and Respectively represent The average distance from the samples in the class and the jth class to the cluster center; It is The distance between the jth cluster center and the

[0058] SC combines the two factors of cohesion and separation to describe the definition of each type of profile. The larger the SC value, the better the clustering performance. The calculation formula of SC is as follows: (20) in, is the silhouette coefficient, Represents the average distance between data point i1 and other points of the same type; Represents the minimum average distance between data point i1 and samples in other classes.

[0059] Clustering performance evaluation load clustering expects each curve to have high intra-class similarity and low inter-class similarity. Therefore, the number of cluster centers corresponding to a large CHI value, a small DBI value, and a large SC value, that is, the K value, can be selected as the target number of cluster centers. This example further ensures the final clustering effect by screening the number of clusters.

[0060] In some example implementations, after obtaining the load classification results of each industrial user, the method further includes: Carry out load pattern analysis based on the load classification results of each industrial user to determine the load pattern of each industrial user; Predict the electricity demand of each industrial user based on their load patterns; Based on the prediction results, load equipment of each industrial user is adjusted and controlled.

[0061] In this example implementation, the load mode corresponding to the daily load data of each industrial user can be analyzed by analyzing the load classification results of each industrial user. For example, several load modes can be determined based on a large amount of historical load data of each industrial user. Each load mode has a typical load curve within the corresponding load cycle. The load cycle can be set based on the production plan or actual production situation of the industrial user. Based on the typical load curve, the electricity demand of the corresponding industrial user in the future period can be predicted, and the power of the load equipment can be controlled based on the prediction results. Through deep mining and analysis of the electricity consumption data of industrial users, this example can accurately predict the electricity demand through accurate extraction and clustering of load characteristics, and adjust the power of equipment in real time to achieve intelligent load management, which not only helps to optimize the electricity consumption period and power, but also can significantly improve the response speed and flexibility of the power grid in application scenarios such as day-ahead and intraday demand response and power auxiliary service market. Through precise control, the efficient consumption of renewable energy is promoted, and the adaptability of the power grid to distributed photovoltaic and other new energy sources is greatly improved.

[0062] In some example implementations, before determining the multi-dimensional load characteristic index of each industrial user based on the daily load data of each industrial user, the method further includes: Collect real-time operation data of internal load equipment of each industrial user during the day; Collect real-time measurement data of internal measurement equipment of each industrial user during the day; Collect daily control status data of internal switch equipment of each industrial user; Based on the load feature extraction of the daily real-time operation data, the daily real-time measurement data and the daily control status data, the daily load data of each industrial user is obtained.

[0063] In this example implementation, it is necessary to collect the daily load data of each industrial user before extracting the characteristics of the daily load data. The load data can be fully acquired through multi-level data collection. First, intelligent load equipment, such as self-provided generator sets, distributed power inverters, arc furnaces and ladle refining furnace equipment electronic control units, etc., can collect real-time operation data of these devices through wired or wireless communication. If the equipment does not have an open agreement, the equipment manufacturer must cooperate to access and debug. By monitoring the operating status of these devices and extracting their load characteristic data, the use and operating efficiency of the equipment can be analyzed. Secondly, the measurement equipment inside each industrial user can include electrical measuring instruments for monitoring electrical parameters (such as smart meters installed by power grid companies or power demand response implementation agencies) or system measurement and control modules, etc., which can collect real-time measurement data of these devices through wired or wireless communication. By continuously monitoring the measurement data of these devices, their load characteristics are analyzed, including parameters such as voltage, current, and power factor, especially when the equipment participates in the demand response strategy, real-time monitoring of the corresponding equipment load is carried out in accordance with the established requirements. Finally, the control status data of non-intelligent devices such as load switches or circuit breakers can be collected. The control status data can be collected through wired or wireless communication, and direct on-off control of these devices can also be achieved. By extracting features from these devices, their load characteristics are evaluated and corresponding optimization suggestions are provided. This example obtains load data of various types of equipment through multi-level data collection, enriches data sources, and provides data support for subsequent feature extraction.

[0064] The following is a specific example to illustrate the specific steps of a method for classifying loads of high energy-consuming industrial users of the present invention. Figure 4 As shown, the following processes may be included: (1) Input the original load data.

[0065] (2) Check whether the original load data has data missing or exceeds the threshold. If so, delete the data; otherwise, supplement the missing load data.

[0066] (3) Check whether there are abnormal data in the original load data. If so, correct the abnormal values ​​in the data. Otherwise, obtain the processed load data (i.e., the supplemented load data).

[0067] (4) Perform maximum-minimum normalization on the processed load data.

[0068] (5) Establish a load feature set based on the standardized load data.

[0069] (6) Based on the convolutional autoencoder, a low-dimensional feature vector of the load data after standardization in (4) is extracted to obtain the temporal characteristics of the load data.

[0070] (7) Based on the load feature set in (5) and the time feature set in (6), a multi-dimensional electricity consumption feature set is constructed.

[0071] (8) Establish a SOM (self-organizing map) network and train the SOM network to obtain the mapping relationship between input data and output nodes.

[0072] (9) Input the multi-dimensional electricity consumption feature set in (7) into the trained SOM network and determine the output node corresponding to each input data.

[0073] (10) The local density and attribute distance of each output node are calculated based on the number of input data corresponding to each output node.

[0074] (11) Determine the initial cluster centers under different numbers of clusters based on local density and attribute distance.

[0075] (12) Based on the initial cluster center and K-means algorithm, the multi-dimensional electricity consumption feature set in (7) is clustered to obtain the clustering results under different cluster numbers.

[0076] (13) Evaluate the clustering performance of clustering results under different numbers of clusters, determine the target number of clusters and use the corresponding clustering results as the load classification results.

[0077] (14) Analyze the load patterns of industrial users and conduct electricity demand forecasting and equipment power regulation.

[0078] The present invention calculates five key load characteristic indicators and weights them using the improved entropy weight method to form an industrial load characteristic set. At the same time, a convolutional autoencoder (CAE) is used to extract the time characteristics of the load data and combine it with the load characteristic set to construct a multidimensional characteristic set. In order to overcome the local optimal problem that may be caused by the traditional K-means algorithm when randomly selecting the initial clustering center, the present invention uses a self-organizing map (SOM) network to determine the initial clustering center of the K-means algorithm, thereby improving the stability and effect of clustering. Experimental verification shows that the method proposed in the present invention has better stability and clustering effect in solving the local optimal problem.

[0079] In the prior art, the load feature extraction of high-energy-consuming industrial users mainly relies on traditional statistical methods and simple models based on experience. These methods are usually based on historical data for static analysis, lack of real-time response and prediction capabilities for dynamic load changes, and are difficult to meet the actual needs of modern smart grids and demand response. Among high-energy-consuming industrial users, the load characteristics are complex and changeable. Affected by various factors such as production processes, seasonal changes, and equipment status, traditional methods are difficult to capture the subtle characteristics and potential laws of load changes, and cannot achieve accurate management and optimization of loads. The existing processing of high-dimensional load data still has problems of high complexity, large amount of calculation, and low efficiency. For example, the Chinese patent with publication number CN103390117A, named a method for extracting features of dynamic characteristics of power loads, first pre-processes the three-phase transient current data reflecting the dynamic characteristics of power loads, and selects the sampling data of one cycle before the fault and two cycles after the fault starts as samples. Next, a wavelet basis is selected and its lifting format is determined. This wavelet has superior performance in terms of spectrum leakage and resolution. After selecting the wavelet basis, the number of decomposition layers is determined, and the three-phase transient current data of each sample is decomposed at each layer using the lifting format wavelet packet transform. Next, the original signal is reconstructed, and the feature vector is constructed by extracting the corresponding reconstruction coefficients and calculating the energy value. This method depends on the selection of the wavelet basis. If the wavelet basis is not selected properly, the accuracy of feature extraction may be affected, and the process of wavelet transform and inverse transform is relatively complicated, which has certain requirements for computing resources. In addition, this method requires data of one cycle before the fault and two cycles after the fault, which has high requirements for data acquisition and storage, and increases the complexity of data processing. For example, the Chinese patent with the publication number CN116451052A, named a system for extracting characteristics of industrial power load, consists of four main modules: external analysis module, confidence configuration module, caliber grouping module and feature extraction module. The external analysis module is responsible for connecting different external data platforms, scanning and generating sensitive data indexes in real time, so as to obtain industry confidence information. This module ensures the reliability and accuracy of data through confidence analysis strategies and confidence analysis models. Each industry confidence information includes several industry segments and their corresponding benchmark confidence values. The confidence configuration module uses dynamic configuration strategies to adjust the benchmark confidence value to generate a dynamic confidence value in combination with the electricity address and historical electricity usage information of each electricity user. This module selects the industry segment with the highest dynamic confidence value for each user as the industry segmentation feature of the user, thereby improving the accuracy and precision of the analysis. The caliber grouping module uses the caliber grouping strategy to generate a scale grouping benchmark based on feature retrieval requirements, and divides electricity users with the same industry segmentation characteristics into different scale groups. This grouping strategy can improve the efficiency and accuracy of data analysis and ensure that the extracted power load characteristics are more in line with actual needs.The feature extraction module is equipped with a variety of feature extraction algorithms, which can call the corresponding algorithms according to different needs to extract the corresponding power load characteristics from the target power consumption data of the scale group. This technology is highly dependent on the accuracy of the external data platform and internal power consumption data, and has high computational complexity and resource requirements.

[0080] The present invention aims at the problems existing in the existing methods for extracting load characteristics of high-energy-consuming industrial users, such as reliance on traditional statistical methods, lack of intelligent analysis means, and inability to accurately reflect the load change law and response characteristics. A method for extracting load characteristics of high-energy-consuming industrial users based on data and model driving is proposed. By introducing big data analysis and machine learning models, the method aims to accurately extract the load characteristics of high-energy-consuming industrial users and improve the scientificity and effectiveness of load management and optimization. With the rapid development of big data technology and machine learning algorithms, load characteristic extraction methods based on data driving and model driving have gradually become a research hotspot. By introducing big data analysis and machine learning models to conduct in-depth mining and analysis of massive electricity consumption data, the present invention can extract the time series characteristics and nonlinear characteristics of the load of high-energy-consuming industrial users, establish an accurate load prediction model, and provide real-time response capabilities to load changes; breaking through the limitations of traditional methods and improving the accuracy and efficiency of load characteristic extraction.

[0081] The present invention calculates key load characteristic indicators and weights them using an improved entropy weight method to form an industrial load characteristic set; uses a convolutional autoencoder to extract the time characteristics of the load data and combines it with the load characteristic set to construct a multidimensional characteristic set. The initial clustering center of the K-means algorithm is determined by a self-organizing mapping network, which overcomes the local optimal problem that may be caused by the traditional K-means algorithm when randomly selecting the initial clustering center, and improves the stability and effect of clustering.

[0082] The industrial user load classification method proposed in the present invention, which is based on multi-dimensional electricity consumption characteristics, has a significant positive impact on the power system. By accurately mining and classifying the load characteristics of industrial users, this method improves the stability and operation efficiency of the power grid, helps to reduce the imbalance between supply and demand, optimize power dispatching and improve load management. It supports the construction of smart grids, provides data-driven decision support, and enhances the flexibility of demand-side response. The method also improves the security of the power grid, responds to emergencies in a timely manner through anomaly detection and early warning, and promotes the process of electricity marketization, supports market pricing mechanisms and user interactions, promotes the overall optimization and intelligent development of the power system, and improves the safety and stability of the power system.

[0083] The present invention can realize the refined management of the internal equipment level of high-efficiency energy-consuming industrial users. The method helps users identify energy consumption peaks and optimize equipment operation strategies through real-time analysis of equipment operation data, thereby significantly improving energy utilization efficiency and reducing energy consumption costs. For market-oriented users, participating in the power demand response and ancillary service markets, through accurate load feature extraction, corresponding market incentives can be obtained, further reducing production costs and enhancing corporate competitiveness.

[0084] The present invention can provide protection for the environmental protection and sustainable development of industrial users. By optimizing the allocation of power resources, promoting load transfer and peak shaving and valley filling by extracting the characteristics of high-energy-consuming industrial users, this method improves energy utilization efficiency and reduces carbon emissions and pollution. It supports the access of renewable energy, promotes the development of the green power market, enhances the flexibility and reliability of the power grid, and helps achieve sustainable development goals. Using data provided by advanced metering infrastructure, this method promotes smart grids and digital transformation, optimizes industrial electricity consumption behavior, and enhances users' environmental awareness and social responsibility, which overall promotes the use of green energy and environmental protection.

[0085] The present invention can enhance the market competitiveness of industrial users: by implementing the load feature extraction method based on data and models, industrial users can not only achieve cost control and energy efficiency improvement, but also obtain additional benefits by participating in the electricity market. The realization of this dual benefit enables enterprises to occupy an advantageous position in the fierce market competition and promote industrial upgrading and technological progress.

[0086] The present invention accurately extracts and depicts the load characteristics of high-energy-consuming industrial users, depicts the operating characteristics of high-energy-consuming industrial loads, realizes load refinement modeling, supports accurate evaluation and refined management of various types of industrial load resources, helps improve the quality and efficiency of on-site investigation of high-energy-consuming industrial users, and provides scientific, reasonable, efficient and accurate theoretical support for power load management services such as power demand response and orderly power consumption. At the same time, the application of the present invention can support the optimization of power consumption plans for high-energy-consuming industrial users, reduce electricity costs, and improve the stability and reliability of power grids. The method can also promote the consumption of renewable energy and provide support for achieving green and low-carbon development. The method has broad application prospects in load management, smart grids, demand response and energy big data. The high-energy-consuming industrial load characteristic analysis system will support the construction of a new power load management system, realize accurate load control and normalized and refined energy management of users under the requirements of power safety and supply, help build an adjustable load resource library with a certain proportion (such as 5%) of maximum load, and delay power grid construction investment.

[0087] Example 2 Based on the same inventive concept, the present invention also provides a high-energy-consuming industrial user load classification system, the system comprising: The load feature extraction module is used to determine the multi-dimensional load feature indicators of each industrial user based on the daily load data of each industrial user, taking into account the load variation characteristics, load utilization rate and user's sensitivity to electricity prices; extract the time feature of the daily load data of each industrial user based on the convolutional autoencoder, and determine the power consumption characteristics of each industrial user based on the extracted load time characteristics and multi-dimensional load feature indicators of each industrial user; A node screening module is used to input the power consumption characteristics of each industrial user into the self-organizing map neural network to determine the output node to which each input data is mapped; and to screen out the target output node based on the number of input data corresponding to each output node; The clustering module is used to cluster the power consumption characteristics of each industrial user with the target output node as the initial clustering center to obtain the load classification result of each industrial user.

[0088] In a possible implementation, the load feature extraction module includes: A weight calculation submodule is used to determine the index weight of the load characteristic index of each industrial user in each dimension based on the entropy weight method for each industrial user; A fusion submodule is used to obtain the key load characteristics of the corresponding industrial users based on the weighted fusion of the load characteristic indicators of each dimension and the corresponding indicator weights; The feature determination submodule is used to determine the power consumption feature of each industrial user based on the extracted load time feature of each industrial user and the corresponding key load feature.

[0089] In a possible implementation manner, the weight calculation submodule is specifically used for: Based on the daily load data of each industrial user, the entropy of the corresponding load characteristic index of each dimension is calculated; Based on the entropy of the load characteristic index of each dimension, the index weight of the load characteristic index of each dimension is calculated using an exponential function.

[0090] In a possible implementation, the calculation formula of the entropy of the load characteristic index of each dimension is: ; in, is the entropy of the load characteristic index c, a represents the total number of load characteristic indicators, d represents the index of daily load data, and m0 represents the total number of daily load data; The calculation formula of the indicator weight of the load characteristic indicator of each dimension is: ; in, is the index weight of load characteristic index c, Indicates The entropy of a load characteristic index, e is the base of the natural logarithm.

[0091] In a possible implementation, the load feature extraction module includes an index calculation submodule, and the index calculation submodule is used to: For each industrial user, the corresponding daily average load, daily total load, daily maximum load, daily minimum load, daily peak period average load and daily valley period average load are calculated based on the daily load data of the industrial user; Determine the daily load rate based on the daily average load and the daily maximum load, and determine the daily load peak-to-valley difference rate based on the daily maximum load and the daily minimum load; determine the daily maximum load utilization time based on the daily total load and the daily maximum load; determine the peak period load rate based on the daily peak period average load and the daily maximum load, and determine the valley period load rate based on the daily valley period average load and the daily maximum load; The daily load rate and the daily load peak-to-valley difference rate are used as load characteristic indicators of the corresponding dimension of load change characteristics, the daily maximum load utilization duration is used as the load characteristic indicator of the corresponding dimension of load utilization rate, and the peak period load rate and the valley period load rate are used as load characteristic indicators of the corresponding dimension of user's electricity price sensitivity.

[0092] In a possible implementation, the convolutional autoencoder includes an encoder and a decoder, and the system further includes an encoder training module, where the encoder training module is used to: Obtain historical load data of each industrial user as training data; Input the training data into the convolutional autoencoder, and perform convolution and pooling processing on the training data through the encoder to obtain a corresponding hidden layer feature vector; Deconvolution and upsampling are performed on the hidden layer feature vector by the decoder to reconstruct output data with the same dimension as the input data; A loss function is calculated based on the output data and the corresponding input data, and model parameters of the encoder and the decoder are updated based on the loss function.

[0093] In a possible implementation manner, the load feature extraction module includes a time feature extraction submodule, and the time feature extraction submodule is used to: The daily load data of each industrial user is input into the trained convolutional autoencoder, and the hidden layer low-dimensional feature vector processed by the encoder is extracted as the corresponding load time feature.

[0094] In a possible implementation, the self-organizing map neural network includes input layer and output layer grid nodes, and the node screening module includes a mapping submodule, and the mapping submodule is used to: The power consumption characteristics corresponding to the historical load data of each industrial user are used as training samples, and the self-organizing map neural network is trained using the training samples to obtain the weight of each output node in the output layer grid node; Based on the weight of each output node of the output layer, the power consumption characteristics of each industrial user are mapped, and the output node to which each input data is mapped is determined; The input data is the power consumption characteristics of industrial users, and the output nodes are used to characterize the regional classification of the corresponding power consumption characteristics to achieve a specific response to the input data.

[0095] In a possible implementation, the node screening module includes a target node screening submodule, and the target node screening submodule is used to: Determine the local density of each output node based on the amount of input data corresponding to each output node; For each mapped output node, if an output node is determined to be the node with the largest local density in the neighborhood of the current output node, the distance between the current output node and the farthest node in the neighborhood of the current output node is taken as the attribute distance of the current output node; If it is determined that an output node is not the node with the maximum local density in the neighborhood of the current output node, the minimum distance between the current output node and each node in the target node set is taken as the attribute distance of the current output node; Screening out a plurality of nodes having local density and attribute distance from the output layer grid nodes as the target output nodes; The target node set is a node set whose local density in the current output node domain is higher than that of the current output node.

[0096] In a possible implementation, the clustering module includes: An initial cluster center determination submodule is used to determine target output nodes corresponding to different numbers of cluster centers, and use the weights corresponding to the target output nodes as initial cluster centers under the corresponding number of cluster centers; The clustering submodule is used to cluster the power consumption characteristics of each industrial user based on the corresponding initial clustering center and the K-means clustering algorithm for each number of clustering centers, and obtain the clustering results under each clustering center data; A performance analysis submodule, used to determine the target number of cluster centers based on clustering performance analysis of clustering results under different numbers of cluster centers; The classification submodule is used to use the clustering results corresponding to the target number of cluster centers as the load classification results of each industrial user.

[0097] In a possible implementation, the load feature extraction module further includes a preprocessing submodule, and the preprocessing submodule is used to: Pre-process the daily load data of each industrial user; Wherein, the preprocessing includes at least one of missing data filling, outlier detection and correction, and data normalization.

[0098] In a possible implementation, the system further includes a prediction and control module, which is used to: Carry out load pattern analysis based on the load classification results of each industrial user to determine the load pattern of each industrial user; Predict the electricity demand of each industrial user based on their load patterns; Based on the prediction results, load equipment of each industrial user is adjusted and controlled.

[0099] In a possible implementation, the system further includes a data acquisition module, and the data acquisition module is used to: Collect real-time operation data of internal load equipment of each industrial user during the day; Collect real-time measurement data of internal measurement equipment of each industrial user during the day; Collect daily control status data of internal switch equipment of each industrial user; Based on the load feature extraction of the daily real-time operation data, the daily real-time measurement data and the daily control status data, the daily load data of each industrial user is obtained.

[0100] The high-energy-consuming industrial user load classification system of the present invention extracts the load characteristics of high-energy-consuming industrial users through the above modules, and realizes the collection of electric quantity and non-electric quantity parameters through the intensive sensing platform and the existing platform integration or direct technical transformation. It combines the key parameters of the main equipment of the enterprise production to analyze the load characteristics of high-energy-consuming industrial users, supports the formulation of load participation in the power grid interaction strategy, and realizes the operation requirements of the power grid such as peak shaving and valley filling and new energy consumption.

[0101] The high energy consumption industrial user load classification system architecture of the present invention is as follows Figure 5As shown in the figure, it can be deployed in factories and enterprises in industrial parks, including power stations, distribution rooms and other areas with good public wireless communication network coverage and less electromagnetic interference, as well as areas where operators in the factory are on duty daily. The system can connect to the upper master station through the vertical security authentication platform to ensure the security of the system and the effective transmission of data. For downlink connections, the system adopts a hybrid networking method of wired and wireless to connect various types of equipment, such as load edge terminals and load monitoring terminals, to obtain electrical parameters, state parameters and environmental parameters of various types of equipment. The system extracts load characteristic data of various types of equipment through multi-level data collection and analysis, and conducts detailed analysis to optimize the use efficiency and regulation capacity of the equipment. When executing the demand response strategy, the system monitors the response of the equipment in real time to ensure that it is executed according to the predetermined plan. The entire system achieves efficient communication through hybrid networking of wired and wireless, and ensures overall security through the vertical security authentication platform. This deployment and connection method of the load characteristic extraction system for high-energy-consuming industrial users not only improves the intelligent management level of the power system and improves energy utilization efficiency, but also ensures the safety and reliability of the system.

[0102] Example 3 like Figure 6 As shown, the present invention also provides an electronic device, which may be a computer device, a single-chip device, an intelligent mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, the processor, and the transceiver component are connected via a bus; the memory may be used to store an execution program, and an exemplary execution program may include instructions; the processor is used to execute the instructions stored in the memory. The memory may also be used to store data, which may be called and / or modified when the instructions are executed.

[0103] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in a storage medium to implement the corresponding method flow or corresponding function, so as to implement the steps of a high-energy-consuming industrial user load classification method in the above embodiment.

[0104] Example 4 Based on the same inventive concept, the present invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory), which is a memory device in an electronic device for storing programs and data. It can be understood that the storage medium here can include both built-in storage media in electronic devices and, of course, extended storage media supported by electronic devices. The storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by a processor are also stored in the storage space, and these instructions can be one or more execution programs (including program codes). It should be noted that the storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. The processor loads and executes one or more instructions stored in the storage medium, which can implement the steps of a high-energy-consuming industrial user load classification method in the above embodiment.

[0105] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0106] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0107] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0108] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit its protection scope. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that after reading the present invention, those skilled in the art can still make various changes, modifications or equivalent substitutions to the specific implementation methods of the application, but these changes, modifications or equivalent substitutions are all within the protection scope of the claims to be approved.

Claims

1. A method for classifying loads of high energy-consuming industrial users, characterized in that: include: Considering the load variation characteristics, load utilization and user's sensitivity to electricity prices, the multi-dimensional load characteristic indicators of each industrial user are determined based on the daily load data of each industrial user; The time characteristics of the daily load data of each industrial user are extracted based on the convolutional autoencoder, and the power consumption characteristics of each industrial user are determined based on the extracted load time characteristics and multi-dimensional load characteristic indicators of each industrial user; Input the power consumption characteristics of each industrial user into the self-organizing map neural network to determine the output node to which each input data is mapped; filter out the target output node based on the number of input data corresponding to each output node; The target output node is used as the initial clustering center to cluster the power consumption characteristics of each industrial user to obtain the load classification result of each industrial user.

2. The method according to claim 1, characterized in that The method of determining the power consumption characteristics of each industrial user based on the extracted load time characteristics and multi-dimensional load characteristic indicators of each industrial user includes: For each industrial user, the index weight of the load characteristic index of each industrial user in each dimension is determined based on the entropy weight method; Based on the weighted fusion of the load characteristic indicators of each dimension and the corresponding indicator weights, the key load characteristics of the corresponding industrial users are obtained; The power consumption characteristics of each industrial user are determined based on the extracted load time characteristics of each industrial user and the corresponding key load characteristics.

3. The method according to claim 2, characterized in that The method of determining the index weight of the load characteristic index of each industrial user in each dimension based on the entropy weight method includes: Based on the daily load data of each industrial user, the entropy of the corresponding load characteristic index of each dimension is calculated; Based on the entropy of the load characteristic index of each dimension, the index weight of the load characteristic index of each dimension is calculated using an exponential function.

4. The method according to claim 3, characterized in that: The calculation formula of the entropy of the load characteristic index of each dimension is: ； in, is the entropy of the load characteristic index c, a represents the total number of load characteristic indicators, d represents the index of daily load data, is the value of the load characteristic index c corresponding to the d-th daily load data sample, and m0 represents the total number of daily load data; The calculation formula of the indicator weight of the load characteristic indicator of each dimension is: ； in, is the index weight of load characteristic index c, Indicates The entropy of a load characteristic index, e is the base of the natural logarithm.

5. The method according to claim 1, characterized in that The load characteristic indexes of each industrial user in multiple dimensions are determined based on the daily load data of each industrial user, taking into account the load variation characteristics, load utilization rate and user's sensitivity to electricity prices, including: For each industrial user, the corresponding daily average load, daily total load, daily maximum load, daily minimum load, daily peak period average load and daily valley period average load are calculated based on the daily load data of the industrial user; Determine the daily load rate based on the daily average load and the daily maximum load, and determine the daily load peak-to-valley difference rate based on the daily maximum load and the daily minimum load; determine the daily maximum load utilization time based on the daily total load and the daily maximum load; determine the peak period load rate based on the daily peak period average load and the daily maximum load, and determine the valley period load rate based on the daily valley period average load and the daily maximum load; The daily load rate and the daily load peak-to-valley difference rate are used as load characteristic indicators of the corresponding dimension of load change characteristics, the daily maximum load utilization duration is used as the load characteristic indicator of the corresponding dimension of load utilization rate, and the peak period load rate and the valley period load rate are used as load characteristic indicators of the corresponding dimension of user's electricity price sensitivity.

6. The method according to claim 1, characterized in that The convolutional autoencoder includes an encoder and a decoder, and the training process of the convolutional autoencoder includes: Obtain historical load data of each industrial user as training data; Input the training data into the convolutional autoencoder, and perform convolution and pooling processing on the training data through the encoder to obtain a corresponding hidden layer feature vector; Deconvolution and upsampling are performed on the hidden layer feature vector by the decoder to reconstruct output data with the same dimension as the input data; A loss function is calculated based on the output data and the corresponding input data, and model parameters of the encoder and the decoder are updated based on the loss function.

7. The method according to claim 6, characterized in that The time feature extraction of the daily load data of each industrial user based on the convolutional autoencoder includes: The daily load data of each industrial user is input into the trained convolutional autoencoder, and the hidden layer low-dimensional feature vector processed by the encoder is extracted as the corresponding load time feature.

8. The method according to claim 1, characterized in that The self-organizing map neural network includes input layer and output layer grid nodes, and the step of inputting the power consumption characteristics of each industrial user into the self-organizing map neural network to determine the output node to which each input data is mapped includes: The power consumption characteristics corresponding to the historical load data of each industrial user are used as training samples, and the self-organizing map neural network is trained using the training samples to obtain the weight of each output node in the output layer grid node; Based on the weight of each output node of the output layer, the power consumption characteristics of each industrial user are mapped, and the output node to which each input data is mapped is determined; The input data is the power consumption characteristics of industrial users, and the output nodes are used to characterize the regional classification of the corresponding power consumption characteristics to achieve a specific response to the input data.

9. The method according to claim 8, characterized in that The step of selecting a target output node based on the amount of input data corresponding to each output node includes: Determine the local density of each output node based on the amount of input data corresponding to each output node; For each mapped output node, if an output node is determined to be the node with the largest local density in the neighborhood of the current output node, the distance between the current output node and the farthest node in the neighborhood of the current output node is taken as the attribute distance of the current output node; If it is determined that an output node is not the node with the maximum local density in the neighborhood of the current output node, the minimum distance between the current output node and each node in the target node set is taken as the attribute distance of the current output node; Screening out a plurality of nodes having local density and attribute distance from the output layer grid nodes as the target output nodes; The target node set is a node set whose local density in the current output node domain is higher than that of the current output node.

10. The method according to claim 1, characterized in that The target output node is used as the initial clustering center to cluster the power consumption characteristics of each industrial user, and the load classification results of each industrial user are obtained, including: Determine the target output nodes corresponding to different numbers of cluster centers, and use the weights corresponding to the target output nodes as the initial cluster centers under the corresponding number of cluster centers; For each number of cluster centers, cluster the electricity consumption characteristics of each industrial user based on the corresponding initial cluster center and K-means clustering algorithm to obtain the clustering results under each cluster center data; Based on the clustering performance analysis of the clustering results under different numbers of cluster centers, the target number of cluster centers is determined; The clustering results corresponding to the target number of cluster centers are used as load classification results for each industrial user.

11. The method according to any one of claims 1 to 10, characterized in that: Before determining the multi-dimensional load characteristic indicators of each industrial user based on the daily load data of each industrial user, considering the load variation characteristics, load utilization rate and user's electricity price sensitivity, it also includes: Pre-process the daily load data of each industrial user; Wherein, the preprocessing includes at least one of missing data filling, outlier detection and correction, and data normalization.

12. The method according to claim 1, characterized in that After obtaining the load classification results of each industrial user, it also includes: Carry out load pattern analysis based on the load classification results of each industrial user to determine the load pattern of each industrial user; Predict the electricity demand of each industrial user based on their load patterns; Based on the prediction results, load equipment of each industrial user is adjusted and controlled.

13. The method according to claim 1, characterized in that Before determining the multi-dimensional load characteristic indicators of each industrial user based on the daily load data of each industrial user, it also includes: Collect real-time operation data of internal load equipment of each industrial user during the day; Collect real-time measurement data of internal measurement equipment of each industrial user during the day; Collect daily control status data of internal switch equipment of each industrial user; Based on the load feature extraction of the daily real-time operation data, the daily real-time measurement data and the daily control status data, the daily load data of each industrial user is obtained.

14. A load classification system for high energy consumption industrial users, characterized in that: include: The load feature extraction module is used to determine the multi-dimensional load feature indicators of each industrial user based on the daily load data of each industrial user, taking into account the load variation characteristics, load utilization rate and user's sensitivity to electricity prices; extract the time feature of the daily load data of each industrial user based on the convolutional autoencoder, and determine the power consumption characteristics of each industrial user based on the extracted load time characteristics and multi-dimensional load feature indicators of each industrial user; A node screening module is used to input the power consumption characteristics of each industrial user into the self-organizing map neural network to determine the output node to which each input data is mapped; and to screen out the target output node based on the number of input data corresponding to each output node; The clustering module is used to cluster the power consumption characteristics of each industrial user with the target output node as the initial clustering center to obtain the load classification result of each industrial user.

15. An electronic device, characterized in that: include: at least one processor and memory; The memory and the processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, the method according to any one of claims 1 to 13 is implemented.

16. A readable storage medium, characterized in that: An execution program is stored thereon, and when the execution program is executed, the method according to any one of claims 1 to 13 is implemented.

Images