Industrial user interaction behavior label analysis method under spot market condition

Abstract

The invention discloses an industrial user interaction behavior label analysis method under spot market conditions, and belongs to the technical field of power system demand side management, and the method comprises the following steps: constructing a load model; screening the power utilization indexes according to the load model, and recombining the screened indexes to form a recombined data set; inputting the data set and the power generation data of the self-contained power plant into an adjustable deduction model to obtain equipment regulation and control measurement and calculation results in different time periods; and carrying out industrial user interaction behavior label attribute training by using the adjustable deduction model, and identifying the interaction behavior characteristics of the industrial user. According to the industrial user interaction behavior label analysis method under the spot market condition, model training is optimized, the model generalization ability is enhanced, and the efficiency and accuracy of power consumption behavior analysis of industrial large users are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of demand - side management of power systems, and particularly relates to a method for analyzing industrial user interaction behavior tags under the conditions of the spot market. Background Art

[0002] With the increasing global attention to environmental protection and sustainable development, many countries have put forward the goals of carbon peak and carbon neutrality, and are committed to building a new power system. These goals have promoted the transformation of the energy system from traditional fossil energy to renewable energy and new energy. In the field of traditional loads, due to the contraction of conventional power generation construction and the continuous growth of electricity demand, there is a shortage in the peak period in a short time. Among them, the load data of large industrial users is huge and complex, and their electricity consumption behaviors have significant periodicity and randomness. How to effectively manage and optimize these loads has become an important issue for improving the flexibility and stability of the power system.

[0003] Traditional load forecasting and regulation models often face challenges such as inaccurate fitting and insufficient generalization ability when dealing with complex actual application scenarios. Therefore, a new method is urgently needed. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for analyzing industrial user interaction behavior tags under the conditions of the spot market. This method significantly improves the efficiency and accuracy of analyzing the electricity consumption behaviors of large industrial users by comprehensively applying various technical means to improve the fitting accuracy, generalization ability, and short - time correction ability of the model.

[0005] To achieve the above - mentioned purpose, the present invention provides a method for analyzing industrial user interaction behavior tags under the conditions of the spot market, including the following steps: constructing a load model, where the load model includes a continuous impact load equipment model, an indirect impact load equipment model, and an auxiliary production load model; screening electricity consumption indicators according to the load model, recombining the screened indicators to form a recombined data set; inputting the data set and the self - owned power plant generation data into an adjustable deduction model to obtain the equipment regulation calculation results for each time period, where the adjustable deduction model includes a leaky rectified linear unit, a Gaussian error linear unit, and an improved random forest algorithm; setting the initial weight of the sample, defining the cumulative error amount, and using the adjustable deduction model to train the industrial user interaction behavior tag attributes to identify the interaction behavior characteristics of industrial users.

[0006] Preferably, the calculation method of the continuous impact load equipment model is as follows:

[0007]

[0008] Among them, P ss (τ) is the equipment power; τ represents the time period; τoff is the shutdown time period; τ on is the startup time period; P ess is the rated power; Δτ ss is the time period required for the device to reach its rated power from startup;

[0009] The calculation method of the indirect impact load device model is as follows:

[0010]

[0011] Among them, P is (t) represents the device power; t ison is the device power-on time; t isof is the device power-off moment; P eis is the device rated power; Δt up is the time required for the device to reach the rated power from power-on; Δt down is the time for the device to reach zero power from power-off; α(t) is the power fluctuation of the device during steady-state operation.

[0012] Preferably, when training the load model, the training process of the model is optimized through an adaptive learning rate training method and an early termination mechanism. The adaptive learning rate training method is to set the initial learning rate as α0, then the learning rate α stu is calculated as follows:

[0013]

[0014] Among them, t is the current calculation time; T max is the complete model training time; α stu,min is the minimum value of the tracking learning rate for calculating the electricity load of industrial large users; α stu,max is the maximum value of the tracking learning rate for calculating the electricity load of industrial large users; the minimum and maximum values of the learning rate are related to the maximum, minimum, and average values of the load of industrial large users within the statistical time range, and the calculation method is as follows:

[0015]

[0016] Among them, P min is the minimum load of industrial large users within the statistical time range; P max is the maximum load of industrial large users within the statistical time range; is the average load of industrial large users within the statistical time range; is calculated as follows:

[0017]

[0018] For the industrial large user load sequence with a statistical point number of N, P min = min{P(1), P(2), …, P(N)}; P max = max{P(1), P(2), …, P(N)}; where P(1), P(2), …, P(N) represent the 1st, 2nd, …, Nth sampling points respectively.

[0019] Preferably, the full convolutional network model is further used to select the power consumption data of industrial large users. First, the input data is formatted. Second, the load curve is converted into a picture form. Next, convolution processing operations are performed on the matrix formed by the load pictures, and the calculation method is as follows:

[0020]

[0021] where P(x, y) represents the load picture matrix; x, y represent the xth row and the yth column of the load picture; G(x, y, σ) is the convolution processing kernel of the load picture; represents convolution; L g (x, y, σ) represents the matrix after convolution processing.

[0022] Preferably, the original load picture is selected as the 0th layer, and the difference layer is generated by calculating the difference between the load images of two adjacent time sections. The calculation method is as follows:

[0023]

[0024] where represents the difference layer data of the ith user at the jth time section; x (i,j) and y (i,j) respectively represent the abscissa and ordinate of a certain point in the difference layer; σ (i,j) determines the width of the convolution kernel and the sensitivity to different scale features; represents the matrix after convolution processing, corresponding to the data of the ith user at the jth time section; represents the matrix after convolution processing.

[0025] Preferably, the calculation formula of the Gaussian error linear unit is as follows:

[0026]

[0027] where is the input of the ith node in the nth layer of the neural network, is the output value after being processed by the GeLU activation function.

[0028] Preferably, the steps of training the industrial user interaction behavior label attributes are as follows:

[0029] Set the initial weights, and set the initial sample weight set D1 for the data set with the total number of sample attributes being N. The calculation formula is as follows:

[0030]

[0031] where, w 1n represents the weight of the i-th sample in the first iteration;

[0032] The number of samples in the data set is M, and D m (m = 1, 2,..., M) represents the sample weight set at the m-th iteration; Define the user recognizer G m (x), G m (x): x → {-1, +1} is the m-th weak classifier;

[0033] Define the cumulative error amount e m , and the calculation formula is as follows:

[0034]

[0035] where, N is the total number of training samples; x n is the feature vector of the n-th training sample; y n is the true label of the n-th training sample; G m (x n ) is the predicted output of the m-th weak classifier for the n-th sample; w mn is the weight of the n-th sample at the m-th iteration; I(·) is the indicator function;

[0036] Calculate the weight coefficient α m of the user label recognizer, and the calculation formula is:

[0037]

[0038] where, α m is the weight coefficient of the m-th weak classifier G m ; e m is the cumulative error amount at the m-th iteration, and update the corresponding weight vector:

[0039] D m+1 =(w m+1,1 , w m+1,2 … w m+1,i …, w m+1,N );

[0040] w m+1,n = w m,n exp(-α m y i G m (xi ));

[0041] Among them, D m+1 is the set of sample weights at the (m + 1)-th iteration; w m+1,n is the weight of the n-th sample at the (m + 1)-th iteration;

[0042] Normalize the weights, and the normalization factor Z m is calculated as follows:

[0043]

[0044] The formula for normalization is as follows:

[0045]

[0046] Among them, w' m+1,n represents the normalized sample weights.

[0047] Preferably, the industrial user interaction behavior label attributes include response capacity, response time interval, maximum response duration, response rate, ramp rate, and unit response cost.

[0048] An industrial user interaction behavior label analysis device under spot market conditions includes:

[0049] A load model construction module for constructing the load model;

[0050] A data processing module, connected to the load model construction module, for screening power consumption indicators according to the load model and recombining the screened indicators to form a recombined data set;

[0051] An adjustable deduction module, connected to the data processing module, for receiving the recombined data set and self-owned power plant generation data, and using an adjustable deduction model including a leaky rectified linear unit, a Gaussian error linear unit, and an improved random forest algorithm to obtain equipment regulation calculation results in sub-periods;

[0052] A label attribute training module, connected to the adjustable deduction module, for setting initial sample weights, defining the cumulative error amount, and using the adjustable deduction model to train the industrial user interaction behavior label attributes to identify the interaction behavior characteristics of industrial users;

[0053] The load model construction module further includes a parameter adjustment sub-module for dynamically adjusting the parameter values in the load model training according to historical data;

[0054] The data processing module further includes an outlier detection unit for detecting outliers and removing abnormal data before screening the power consumption indicators;

[0055] The adjustable deduction module further includes a feedback correction unit for correcting the results of the adjustable deduction model according to the actual operation data to improve the prediction accuracy.

[0056] Preferably, when the computer program is executed by a processor, it implements the steps of any one of the above methods.

[0057] Therefore, by adopting the above method for analyzing industrial user interaction behavior labels under spot market conditions, compared with the prior art, the present invention has the following remarkable beneficial effects:

[0058] (1) By using a leaky rectified linear unit to simulate the power generation data of self-provided power plants and introducing a Gaussian error linear unit to improve the generalization ability of the model, the model can more accurately adapt to diverse actual situations;

[0059] (2) Based on the Pearson-mutual information index selection model to reconstruct the energy consumption data, screening out the data sets with high adjustable potential, and further using the fully convolutional network model to select the electricity consumption data of large industrial users, more accurately analyzing the electricity consumption behavior of industrial users and accurately determining the user labels;

[0060] (3) By conducting labeled analysis on the interaction behavior of industrial users, the enterprise can more accurately identify the needs and preferences of users, thereby formulating personalized marketing strategies. At the same time, these analysis results provide data support for industrial enterprises to participate in the spot market;

[0061] (4) The present invention designs a complete label iterative training model, which can effectively identify the characteristics of each type of user, and then realize the automatic query of the adjustable capacity of demand-side resources, improving the load management ability of the new power system, promoting the construction of the adjustable resource pool, and contributing to the stable operation of the power grid.

[0062] The technical solution of the present invention will be further described in detail below with reference to the drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a sliding window selection diagram under different power consumption periods of the method for analyzing industrial user interaction behavior labels under spot market conditions of the present invention, where a represents the low valley period, b represents the peak period, and c represents the normal period;

[0064] Figure 2 It is a user label recognition flow chart of the method for analyzing industrial user interaction behavior labels under spot market conditions of the present invention;

[0065] Figure 3Random forest optimization algorithm diagram for time domain feature expansion of an industrial user interaction behavior label analysis method under spot market conditions according to the present invention;

[0066] Figure 4 Diagram for label application description of an industrial user interaction behavior label analysis method under spot market conditions according to the present invention. Detailed implementation manners

[0067] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention. Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the ordinary meanings understood by those of ordinary skill in the art to which the present invention pertains.

[0068] Embodiment 1

[0069] As Figure 1 shown, an industrial user interaction behavior label analysis method under spot market conditions of the present invention includes the following steps:

[0070] The present invention constructs load models for different types of industrial equipment to effectively utilize the regulation potential of these equipment. For continuous impact load equipment (such as motors), its power changes extremely violently, showing a short fluctuation period and a large fluctuation amplitude. By recording the start-stop time period and power fluctuation conditions, a continuous impact load equipment model is established to evaluate the regulation potential. The calculation formula is as follows:

[0071]

[0072] Wherein, P ss (τ) is the equipment power; τ represents the time period; τ off is the shutdown time period; τ on is the startup time period; P ess is the rated power; Δτ ss is the time period required for the equipment to reach its rated power from startup, which is an important indicator for evaluating its performance. When the demand response mechanism operates, according to the adjustment requirements of different loads, by utilizing the inherent volatility of continuous impact loads, the operation or suspension of the equipment in a specific time period can be planned to achieve accurate load adjustment goals.

[0073] For indirect impact load equipment (such as electric arc furnaces), this type of load exhibits significant indirect fluctuation characteristics. Based on its specific working characteristics and operation process, an indirect impact load equipment model is constructed to deeply understand and optimize its application. The calculation formula is as follows:

[0074]

[0075] Where, P is (t) represents the equipment power; t ison is the equipment power-on time; t isof is the equipment power-off moment; P eis is the rated power of the equipment; Δt up is the time required for the equipment to reach the rated power from power-on; Δt down is the time for the equipment to reach zero power from power-off; α(t) is the power fluctuation of the equipment during steady-state operation. When this type of equipment completes the established task, its power supply will be temporarily cut off. The adjustment space of this part of adjustable load is mainly affected by the feeding speed of the previous process and the production progress of the subsequent process. Therefore, the production plan of this type of production equipment can be formulated in advance at different times before and after. During the demand response period, by flexibly adjusting the production rhythm before and after it, the effective regulation of electrical equipment can be achieved.

[0076] For auxiliary production equipment (such as raw material handling, ventilation and exhaust, etc.), these equipment account for about 10% of the factory's electricity consumption. By modeling and analyzing the working modes of these equipment, recording their running time and power consumption, an auxiliary production load model is established to describe their electricity consumption behavior, and these equipment are given priority in adjusting their operation when urgently adjusting the electricity load. The calculation formula is as follows:

[0077]

[0078] Where, P s (t) represents the auxiliary equipment power; P sn is the rated power value of the auxiliary production equipment; P s1 is the adjustment power value; k1 and k2 are the ramp-up speeds of the equipment power reduction and increase respectively; t on and t of are the start and stop times of the auxiliary equipment respectively; t1 and t2 are the equipment power-down and ramp-up times respectively.

[0079] In order to improve the matching degree between the regulation cost and the load regulation, the present invention introduces a sliding window corresponding to the regulation period to process the load data. Using the sliding window technology to segment the load data can closely combine the equipment regulation cost of industrial users with their actual electricity consumption situation, thereby effectively reducing the amount of calculation data. The calculation method is as follows:

[0080]

[0081] The above formulas are for selecting inputs in the sliding window during the low - valley period, peak period, and normal period respectively. The length selection is based on the number of data points within 15 days in the corresponding period, corresponding to the cycle of producing and storing a sufficient quantity of products in the warehouse. For example Figure 1 As shown, different window lengths are selected for different electricity - consumption periods (e.g., different window lengths are used for the low - valley period, peak period, and normal period respectively).

[0082] Among them, y 1 、y 3 and y 5 are the equipment power consumption corresponding to the low - valley period, peak period, and normal period respectively. y 2 and y 6 are the upward regulation costs for the corresponding periods respectively, and y 4 is the downward regulation cost for the corresponding period.

[0083] To solve the problems existing in the traditional fixed learning rate when dealing with rapid load fluctuations, the present invention adopts an adaptive learning rate training method. Assuming the initial learning rate is α0, the specific learning rate α stu is calculated as follows:

[0084]

[0085] Among them, t is the current calculation time; T max is the complete model training time; α stu,min is the minimum value of the tracking learning rate for calculating the electricity load of large industrial users; α stu,max is the maximum value of the tracking learning rate for calculating the electricity load of large industrial users; the minimum and maximum values of the learning rate are related to the maximum, minimum, and average values of the load of large industrial users within the statistical time range, and the calculation methods are as follows:

[0086]

[0087] Among them, P min is the minimum value of the load of large industrial users within the statistical time range; P max is the maximum value of the load of large industrial users within the statistical time range; is the average value of the load of large industrial users within the statistical time range; The calculation method of

[0088]

[0089] For the load sequence of large industrial users with N statistical points, P min =min{P(1),P(2),…,P(N)}; Pmax = max{P(1), P(2), …, P(N)}; where P(1), P(2), …, P(N) represent the 1st, 2nd, …, Nth sampling points respectively.

[0090] In addition, when the load model training reaches the optimal performance, it is difficult to further improve its accuracy. At this time, continuing to run the model will not only consume time in vain, but also occupy a large amount of precious computing resources. Therefore, the present invention sets up an early termination mechanism for training. The mechanism is as follows: when the prediction accuracy of the adjustable capacity of industrial large users does not exceed the maximum accuracy value within these N iterations for N consecutive iterations, the training program will automatically stop. Through this mechanism, it can be ensured that once the model reaches the optimal training state, the training process will be terminated immediately, thereby significantly shortening the training cycle, avoiding the unnecessary consumption of computing resources, and realizing the efficient, reasonable configuration and utilization of computing resources. The present invention selects N to be set as 100 times, and the specific number of iteration cut-offs can be adjusted independently according to the scale of the data set and the size of the operating parameters. Different parameter setting methods do not constitute a limitation to the present invention.

[0091] The load model can more accurately guide which data need to be focused on and optimized.

[0092] In order to further screen the electricity consumption indicators, the present invention uses a Pearson-mutual information index selection model. This model can identify the indicators highly correlated with the load regulation potential, so as to screen out the most representative data. The calculation method of the Pearson-mutual information index selection model is as follows:

[0093]

[0094] where r is the Pearson mutual information value; and are the mean values of index X and index Y respectively; X i is the real-time value of index X; Y i is the real-time value of index Y.

[0095] When r > 0, it indicates that there is a positive correlation between the two indicators; when r < 0, it means that there is a negative correlation between the two indicators; and when r = 0, it means that there is no correlation between the two indicators. Based on the similarity principle, for all indicators showing positive or negative correlation, the absolute value of their correlation coefficients is uniformly taken, and they are sorted in descending order according to the size of the absolute value. On this basis, the indicators with high regulation potential are screened and extracted to focus on the key factors and improve the effectiveness and pertinence of the analysis.

[0096] Further screen the electricity consumption data of large industrial users using a fully convolutional network model. First, format the input data. Second, convert the load curve into a picture form, and the picture is represented by a two-dimensional matrix. In this two-dimensional matrix, the rows of the matrix correspond to the load curve data at 96 time points in a day, and the columns of the matrix represent the statistical time range. For example, if the load curve data for one year is selected, then the number of columns of the matrix is 365. After completing the conversion of the load curve to a picture, the matrix formed by the load pictures will be subjected to a convolution operation next, and the calculation method is as follows:

[0097]

[0098] Among them, P(x, y) represents the load picture matrix; x, y represent the x-th row and the y-th column of the load picture; G(x, y, σ) is the convolution processing kernel of the load picture; represents convolution, and L g (x, y, σ) represents the matrix after convolution processing, and this matrix contains the information obtained after convolution processing, which is helpful for subsequent analysis tasks such as identifying the electricity consumption behavior labels of users.

[0099] To effectively reduce the computational amount and fully consider the correlation characteristics presented by the load of large industrial users during the day, a difference calculation is specifically performed on the load pictures, and the specific operation is as follows: Select the original load picture as the 0th layer, and use this as the starting benchmark to further derive multiple differential layers. Then, generate the differential layers by calculating the difference between the load images at two adjacent time sections, and the calculation method is as follows:

[0100]

[0101] Among them, represents the differential layer data of the i-th user at the j-th time section; x (i,j) and y (i,j) respectively represent the abscissa and ordinate of a certain point in this differential layer; σ (i,j) determines the width of the convolution kernel and the sensitivity to different scale features; represents the matrix (or image) after convolution processing, corresponding to the data of the i-th user at the j-th time section; represents the matrix (or image) after convolution processing, but corresponding to the data of the i-th user at the previous time section.

[0102] This method of generating differential layers helps to capture the dynamic change characteristics of the load curve, especially in the case of large load fluctuations in a short period of time. Thus, capture the load disturbance variables from industrial users. In this way, the electricity consumption behavior of industrial users can be analyzed more accurately, and the labels of users can be determined accurately.

[0103] As shown Figure 2 in the figure, a method for identifying user behavior tags based on TCN. This method introduces many factors such as the regulation power of industrial users, the time sequence relationship of the technological process, meteorological parameters, economic conditions, etc., and adds the upward adjustment cost and the downward adjustment cost on the basis of time-of-use electricity price. Then, through the time convolutional network stacked residual link structure, the time sequence causal characteristics of industrial users are simulated and trained, and at the same time, the neural network is used to train the historical data of industrial load. For the differential image, the corresponding column-row vector y is selected and input into the TCN model. Then, calculation is carried out by using processing methods such as normalized weight, activation function, regularization, etc., so as to extract the influence generated by the part with a long interval in the load sequence and form a reorganized data set.

[0104] The reorganized data set and the power generation data of the self-provided power plant are input into the adjustable deduction model, and the adjustable deduction model includes a leaky rectified linear unit, a Gaussian error linear unit and an improved random forest algorithm. Thus, the equipment regulation measurement results under different time periods are obtained.

[0105] In order to distinguish the data, the power generation data of the self-provided power plant is input in the form of negative numbers. However, the TCN model will automatically ignore the power generation data in the form of negative numbers to achieve fast convergence. Therefore, a leaky rectified linear unit (LeakyRectified Linear Unit) is introduced to assign parameters to the power generation data. The leaky rectified linear unit is used to simulate the power generation data of the self-provided power plant, and the formula of the leaky rectified linear unit is as follows:

[0106]

[0107] Among them, is the corresponding input data; a i is a fixed parameter; is the power generation data. The TCN network is composed of multiple layers of residual links. Therefore, the leaky rectified linear unit is introduced corresponding to the residual connection, and after the power generation data is assigned parameters, it is calculated separately, and the power generation results are obtained in different time periods. The model enables the power generation data of the self-provided power plant to be calculated simultaneously with the power consumption data of industrial equipment through the leaky rectified linear unit.

[0108] In order to enhance the generalization ability of the model and make it adapt to diverse actual situations. A Gaussian error linear unit (Gaussian Error Linear Unit, GeLU) is introduced to enhance the generalization ability of the model. GeLU can attach a mask to the device input data, and the generation of the mask is randomly dependent on the input according to probability, so that the TCN model can improve the generalization ability and thus be more sensitive to changes in the input data. The calculation formula of GeLU is as follows:

[0109]

[0110] Among them, is the input of the i-th node in the n-th layer of the neural network; is the output value after being processed by the GeLU activation function.

[0111] As Figure 3 shown, in order to better capture the time-domain dynamic characteristics of the adjustable load of industrial large users, the present invention provides an improved random forest (RF) optimization algorithm. This method enhances the model's ability to express the correlation of time series by expanding traditional time-domain independent features into time-domain vectors. The steps are as follows:

[0112] First, construct a feature vector. For any spatial set S i , let p i be the power of the i-th industrial adjustable flexible load object. The feature vector of each time section is expressed as [D t1 , D t2 , …, D tm . These feature vectors are used to describe the correlation between this load object and other devices, as follows:

[0113]

[0114] Then, add a time vector. Add the features of l time sections of each load object to the original matrix elements in the form of an l-dimensional time vector, as follows:

[0115]

[0116] Secondly, considering the computational efficiency and model complexity, it is necessary to perform dimensionality reduction on the extended feature matrix, as follows:

[0117]

[0118] After that, based on the extended feature matrix, sampling with replacement is performed to generate a training set. This sampling method ensures that each generated training set has a certain randomness, thereby improving the generalization ability of the model. And attribute reduction is performed according to the importance matrix generated from the training set. The importance matrix reflects the influence degree of each feature on the model prediction result. By selecting the most important features, the model structure can be simplified and the prediction performance can be improved.

[0119] Finally, after completing the attribute reduction, reset the feature parameters and continue with the splitting, growth, and statistical prediction processes of the random forest. Splitting refers to performing node splitting operations based on the selected features to build a decision tree. Growth means gradually increasing the depth of the tree until the stopping condition is met. Statistical prediction refers to obtaining the final prediction result by comprehensively statistically analyzing the results of multiple decision trees.

[0120] Based on the industrial large-user high-adjustability potential dataset, with the help of the improved TCN model, fit the electricity consumption data of industrial large users and the power generation data of self-owned power plants. After short-term correction by the neural network and power constraint processing, output the results of the adjustable potential of industrial large users, and finally obtain the adjustable potential range of industrial large users by time period and direction.

[0121] By conducting labeled analysis on the interaction behaviors of industrial users, enterprises can more accurately understand the actual needs and preference tendencies of users, and then formulate personalized marketing strategies that suit the characteristics of users. This measure not only significantly improves the efficiency of marketing work, but also effectively enhances the user experience in the process of interacting with the enterprise, and strongly promotes the improvement of user loyalty.

[0122] Figure 4 Gives the application description of labels in the process of industrial user groups participating in market-oriented operation interactions. In the process of industrial users participating in the electricity demand response, they can participate through direct power participation or through the agency of load aggregators. These users will follow the price or incentive guidance signals issued by the superior management system and actively change or adjust their electricity consumption demands to respond to the power grid's supply and demand balance requirements. For industrial users who do not have the conditions for direct participation, they can participate in the demand response projects organized by the power grid through the agency of load aggregators. Such users usually have relatively small production volumes and more random load fluctuations. Therefore, through the design of the label system for industrial users, the energy consumption portraits of industrial users can be realized more accurately. At the same time, these users can also make specific response behaviors according to the price signals.

[0123] In the process of participating in the optimization project, first of all, industrial users need to register on the provincial-level smart energy service platform where they are located and bind their power marketing accounts. According to their own energy consumption patterns, industrial users need to submit the information of the equipment resources they manage, including: parameters such as equipment name, inspection type, and rated power, and report their adjustable response capacity and time period. For industrial users who have been labeled, this part of the information can be calculated by the local intelligent control device and reported to the cloud platform on the power grid side for dynamic analysis. It can also be reported to the cloud platform for calculation and analysis by summarizing the local original data. It should be noted that the specific label calculation and application methods do not constitute a limitation to the technology described in this specification.

[0124] For the characteristics that industrial users care about during the dynamic response process, relevant attributes of the impact load are selected for reporting, specifically including: user number, response capacity, response credibility, maximum response, duration, response rate, recovery rate, unit response cost, etc., as shown in Table 1 below:

[0125] Table 1 Impact Load Attribute Table

[0126]

[0127] Classification labels are set according to the corresponding attributes, as shown in Table 2:

[0128] Table 2 Impact Load Attribute Classification Label Table

[0129]

[0130] Among them, the attribute labels are represented by the set A = {A1, A2,..., A8}, and the user types are the label contents that the main station on the grid side cares about, namely economic type, stable type, and frequent response type. The specific classification method can be set according to the interactive service requirements of the power grid.

[0131] The steps of the training method for the interactive behavior label attributes of industrial users are as follows:

[0132] Set the initial weight, and set the initial sample weight set D1 for the data set with the total number of sample attributes being N. The calculation formula is as follows:

[0133]

[0134] Among them, w 1n represents the weight of the i-th sample in the first iteration;

[0135] The number of samples in the data set is M, and D m (m = 1, 2,..., M) represents the sample weight set at the m-th iteration; Define the user identifier G m (x), G m (x): x → {-1, +1} is the m-th weak classifier, which is used to map the input feature vector x to the binary classification label -1 or +1.

[0136] Define the cumulative error amount e m , and the calculation formula is as follows:

[0137]

[0138] Among them, N is the total number of training samples; x n is the feature vector of the n-th training sample; y n is the true label of the n-th training sample; G m (xn ) is the prediction output of the m-th weak classifier for the n-th sample; w mn is the weight of the n-th sample at the m-th iteration; I(·) is the indicator function.

[0139] Calculate the weight coefficient α of the user label recognizer m , and the calculation formula is:

[0140]

[0141] where α m is the weight coefficient of the m-th weak classifier G m ; e m is the cumulative error amount at the m-th iteration, and update the corresponding weight vector:

[0142] D m+1 =(w m+1,1 , w m+1,2 …w m+1,i …, w m+1,N );

[0143] w m+1,n =w m,n exp(-α m y i G m (x i ));

[0144] where D m+1 is the sample weight set at the (m + 1)-th iteration; w m+1,n is the weight of the n-th sample at the (m + 1)-th iteration; adjust the sample weights according to the prediction result and the true label of the current weak classifier G m .

[0145] Normalize the weights, and the calculation formula of the normalization factor Z m is as follows:

[0146]

[0147] The calculation formula for normalization is as follows:

[0148]

[0149] where w′ m+1,n represents the normalized sample weights. Combine multiple weak classifiers into a strong classifier and finally output the result for identifying the interaction behavior characteristics of industrial users.

[0150] Note that different methods of setting tag classifications may be related to specific industrial user owners. This invention only assumes the most common situation. Different classification calibration methods are related to the data annotation preferences of industrial users, and different annotation methods do not constitute a limitation to this invention.

[0151] Therefore, this invention adopts the above-mentioned method for analyzing industrial user interaction behavior tags under spot market conditions. By comprehensively applying various technical means, this method improves the fitting accuracy, generalization ability, and short-term correction ability of the model, significantly enhancing the efficiency and accuracy of analyzing the electricity consumption behavior of large industrial users.

[0152] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this invention and not to limit them. Although this invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions of this invention or make equivalent replacements, and these modifications or equivalent replacements do not enable the modified technical solutions to deviate from the spirit and scope of the technical solutions of this invention.

Claims

1. An analysis method for industrial user interaction behavior tags under spot market conditions, characterized in that, It includes the following steps: constructing a load model, which includes a continuous impact load equipment model, an indirect impact load equipment model, and an auxiliary production load model; screening power consumption indicators according to the load model, recombining the screened indicators to form a recombined data set; inputting the data set and the self - contained power plant power generation data into an adjustable deduction model to obtain the equipment regulation calculation results under different time periods, where the adjustable deduction model includes a leaky rectified linear unit, a Gaussian error linear unit, and an improved random forest algorithm; Setting the initial weight of the sample, defining the cumulative error amount, and using the adjustable deduction model to train the industrial user interaction behavior label attributes to identify the interaction behavior characteristics of industrial users.

2. The industrial user interaction behavior label analysis method under the spot market condition according to claim 1, wherein The calculation method of the continuous impact load equipment model is as follows: Among them, P ss (τ) is the device power; τ represents a time period; τ off is the shutdown time period; τ on is the startup time period; P ess is the rated power; Δτ ss The time period required for the device to reach its rated power from startup The calculation method of the indirect impact load equipment model is as follows: Among them, P is (t) represents the device power; t ison is the device power-on time; t isof is the device power-off moment; P eis is the rated power of the device; Δt up is the time required for the device to reach the rated power from power-on; Δt down is the time for the device to reach zero power from power-off; α(t) is the power fluctuation of the device during steady-state operation.

3. The industrial user interaction behavior label analysis method under spot market conditions according to claim 1, characterized in that, When training the load model, the training process of the model is optimized through an adaptive learning rate training method and an early termination mechanism. For the adaptive learning rate training method, assume the initial learning rate is α0, then the learning rate α stu is calculated as follows: where t is the current calculation time; T max is the complete model training time; α stu,min is the minimum value of the tracking learning rate for calculating the electricity load of industrial large users; α stu,max is the maximum value of the tracking learning rate for calculating the electricity load of industrial large users; the minimum and maximum values of the learning rate are related to the maximum, minimum, and average values of the load of industrial large users within the statistical time range, and the calculation method is as follows: Among them, P min is the minimum load of large industrial users within the statistical time range; P max is the maximum load of large industrial users within the statistical time range; is the average load of large industrial users within the statistical time range; The calculation method is as follows: For the industrial large customer load sequence with a statistical point number of N, P min = min{P(1), P(2), …, P(N)}; P max = max{P(1), P(2), …, P(N)}; where P(1), P(2), …, P(N) represent the 1st, 2nd, …, Nth sampling points respectively.

4. The industrial user interaction behavior label analysis method under spot market conditions according to claim 1, characterized in that Using a fully convolutional network model to select the power consumption data of large industrial users. First, format the input data. Second, convert the load curve into a picture form. Next, perform a convolution operation on the matrix formed by the load pictures. The calculation method is as follows: Among them, P(x, y) represents the load picture matrix; x and y represent the x-th row and the y-th column of the load picture; G(x, y, σ) is the convolution processing kernel of the load picture; represents convolution; L g (x, y, σ) represents the matrix after convolution processing.

5. The industrial user interaction behavior label analysis method under spot market conditions according to claim 4, characterized in that Select the original load picture as the 0th layer, and generate a differential layer by calculating the difference between the load images of two adjacent time sections. The calculation method is as follows: Among them, represents the differential layer data of the i-th user at the j-th time section; x (i,j) and y (i,j) respectively represent the abscissa and ordinate of a certain point in the differential layer; σ (i,j) determines the width of the convolutional kernel and the sensitivity to features of different scales; represents the matrix after convolutional processing, corresponding to the data of the i-th user at the j-th time section; represents the matrix after convolutional processing.

6. The industrial user interaction behavior label analysis method under the spot market conditions according to claim 1, characterized in that The calculation formula of the Gaussian error linear unit is as follows:: Among them, is the input of the i-th node in the n-th layer of the neural network, is the output value after being processed by the GeLU activation function.

7. The industrial user interaction behavior label analysis method under spot market conditions according to claim 1, wherein The steps of the industrial user interaction behavior label attribute training are as follows: Set the initial weight, and set the initial sample weight set D1 for the data set with the total number of sample attributes N. The calculation formula is as follows: Among them, w 1n represents the weight of the i-th sample in the first iteration; The number of dataset samples is M, D m (m = 1, 2, ..., M) represents the set of sample weights at the m-th iteration; define the user recognizer G m (x), G m (x): x → {-1, +1} is the m-th weak classifier; Define the cumulative error amount e m , and the calculation formula is as follows: where N is the total number of training samples; x n is the feature vector of the n-th training sample; y n is the true label of the n-th training sample; G m (x n ) is the predicted output of the m-th weak classifier for the n-th sample; w mn is the weight of the n-th sample at the m-th iteration; I(·) is the indicator function; Calculate the weight coefficient α of the user tag recognizer m , and the calculation formula is as follows: Among them, α m is the weight coefficient of the m-th weak classifier G m ; e m is the cumulative error amount at the m-th iteration, and update the corresponding weight vector: D m+1 = (w m+1,1 , w m+1,2 … w m+1,i …, w m+1,N ); w m+1,n = w m,n exp(-α m y i G m (x i )); Among them, D m+1 is the set of sample weights at the (m + 1)-th iteration; w m+1,n is the weight of the n-th sample at the (m + 1)-th iteration. Normalize the weights, and the normalization factor Z m is calculated as follows: The calculation formula of normalization is as follows: Among them, w′ m+1,n represents the normalized sample weight.

8. The industrial user interaction behavior label analysis method under the spot market condition according to claim 7, characterized in that The industrial user interaction behavior label attributes include response capacity, response time interval, maximum response duration, response rate, ramp rate, and unit response cost.

9. An industrial user interaction behavior label analysis device under spot market conditions, characterized in that It includes: A load model construction module for constructing the load model; A data processing module connected to the load model construction module, which is used to screen power consumption indicators according to the load model and recombine the screened indicators to form a recombined data set; An adjustable deduction module connected to the data processing module, which is used to receive the recombined data set and the self - contained power plant power generation data, and use an adjustable deduction model including a leaky rectified linear unit, a Gaussian error linear unit, and an improved random forest algorithm to obtain the equipment regulation calculation results under different time periods; A label attribute training module connected to the adjustable deduction module, which is used to set the initial weight of the sample, define the cumulative error amount, and use the adjustable deduction model to train the industrial user interaction behavior label attributes to identify the interaction behavior characteristics of industrial users; The load model construction module further includes a parameter adjustment sub - module for dynamically adjusting the parameter values in the load model training according to historical data; The data processing module further includes an outlier detection unit for detecting outliers and removing abnormal data before screening power consumption indicators; The adjustable deduction module further includes a feedback correction unit for correcting the results of the adjustable deduction model according to the actual operation data to improve the prediction accuracy.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-8.

Images