A load response potential perception method and system for steel enterprises

Through time series decomposition and binary semantic representation, the neural network model is improved and the learning rate is dynamically adjusted, which solves the supply and demand balance problem caused by the complexity of electricity consumption structure of steel enterprises, real-time evaluation of grid side load and accurate perception of user characteristics is achieved. , improves the power consumption scheduling capability of the power system.

CN118211789BActive Publication Date: 2024-11-12STATE GRID JIANGSU ELECTRIC POWER CO LTD NANTONG POWER SUPPLY BRANCH +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410308828.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-03-19
Publication Date
2024-11-12
Estimated Expiration
2044-03-19

AI Technical Summary

Technical Problem

The existing technology cannot make timely and accurate assessments of the real-time operation characteristics of the grid-side load, nor can it achieve effective perception of the load-side user characteristics and accurate description of the load-side resource response potential, especially in steel enterprises that use electricity is much more complex. In the case of change, the supply and demand balance is difficult to match.

Method used

The time series decomposition method is used to decompose the load of steel enterprises, obtain trend quantity, period quantity and residual components, and analyze user behavior characteristics through binary semantic characterization, and combine with improved neural network model to dynamically adjust the learning rate to obtain load response potential .

Benefits of technology

It realizes timely and accurate evaluation of real-time operation characteristics and user characteristics of the grid-side load, accurately characterizes the load-side resource response potential, and improves the power consumption prediction and scheduling capabilities of the power system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118211789B_ABST
    Figure CN118211789B_ABST
Patent Text Reader

Abstract

The present invention discloses a load response potential perception method and system for steel enterprises, the method comprising the following steps: S1 uses a time series decomposition method to decompose the load of the steel enterprise, and obtains the trend quantity, period quantity and residual component of the load of the steel enterprise; S2 analyzes the user behavior characteristics of the steel enterprise based on a binary semantic representation method, and quantifies the behavior characteristics to form quantitative data for calculation; S3 inputs the quantitative data and the obtained trend quantity, period quantity and residual component into the trained improved neural network model, so as to obtain the effective response potential of the steel enterprise. The present invention designs an improved neural network method with an adaptive learning rate, fully considers the influence of the enterprise user behavior characteristics, obtains the effective response potential of the steel enterprise, and then makes a timely and accurate evaluation of the real-time operation characteristics of the load on the power grid side, so as to achieve effective perception of the user characteristics on the load side and accurate characterization of the resource response potential on the load side.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of power system demand side resource management, and in particular to a load response potential perception method and system for a steel enterprise. Background Art

[0002] In the era of big data, cloud computing, the Internet of Things, and Internet technologies have emerged and gradually become popularized. As an important industry that guarantees the rapid development of the national economy, the power industry has also taken advantage of the times to develop rapidly. In addition, with the continuous expansion of the scale of meticulous service-oriented electricity consumption, the traditional power customer service method can no longer meet the requirements of users to actively participate in the reform of the power market, the adjustment of the new energy structure, and the social development centered on "innovation and creation", which has added new challenges to the power grid's control capabilities.

[0003] At present, the complex and changeable electricity consumption structure of steel enterprises leads to insufficient user load perception ability and lack of technical means to effectively judge the production status and electricity consumption characteristics of steel enterprises. Therefore, in order to cope with the increasingly serious supply and demand balance problem, the load information on the demand side is becoming more and more important in the load adjustment of the power system. The traditional dispatching method has limited improvement potential at this stage. Unless there is a major technical breakthrough, it will be difficult to match the existing problems in a short time. In-depth research on demand response technology has become an important way to alleviate the electricity consumption problem of the power system. However, it is still impossible to make timely and accurate evaluations of the real-time operating characteristics of the load on the grid side, and it is impossible to effectively perceive the characteristics of users on the load side and accurately characterize the response potential of resources on the load side. Summary of the invention

[0004] Purpose of the invention: In order to overcome the deficiencies of the prior art, the present invention provides a load response potential perception method for steel enterprises, which solves the problem that the prior art cannot make timely and accurate evaluation of the real-time operating characteristics of the grid-side load, and cannot achieve effective perception of the load-side user characteristics and accurate characterization of the load-side resource response potential. The present invention also provides a load response potential perception system for steel enterprises.

[0005] Technical solution: According to a first aspect of the present invention, a load response potential perception method for a steel enterprise is provided, the method comprising the following steps:

[0006] S1 uses the time series decomposition method to decompose the load of steel enterprises and obtain the trend, cycle and residual components of the load of steel enterprises;

[0007] S2 analyzes the user behavior characteristics of steel enterprises based on the binary semantic representation method, and quantifies the behavior characteristics to form quantitative data for calculation;

[0008] S3 inputs the quantitative data and the obtained trend quantity, cycle quantity and residual component into the trained improved neural network model, so as to obtain the effective response potential of the steel enterprise.

[0009] Further, including:

[0010] In step S1, decomposing the load of the steel enterprise by using the time series decomposition method specifically includes:

[0011] S11 Assumption It is the trend and period after the K-1th calculation of the inner loop is completed. Initially, it is assumed that And define the following parameters: n (i) is the number of inner layers, n (o) is the number of outer layers, n (p) is the number of samples in one cycle, n (s) 、n (l) 、n (t) is the LOESS smoothing parameter;

[0012] S12 groups the samples at the same position in each cycle into a subsequence. There are n(p) subsequences in total. The inner loop includes the following steps:

[0013] S121 detrends, removing the influence of the previous round of trend volume.

[0014] S122 subsequence smoothing is done by using the LOESS function to perform regression smoothing on all subsequences, and at the same time extending them by a period T before and after them, where q = n in the LOESS function n(s) , d = 1; after smoothing, the output result is recorded as v=-n (p) +1,N+n (p) ;

[0015] The S123 subsequence is low-throughput filtered, and the result sequence C(k+1)v of the previous step is moved averaged in turn, and the final result sequence is obtained through LOESS function regression. v=1,…,N, where q=n in the LOESS function n(l) ,d=1;

[0016] S124 to smooth subsequence trends,

[0017] S125 de-cycles, minus the cycle amount,

[0018] S126 trend smoothing, for the sequence after removing the period, do LOESS function regression to get the trend value Among them, q = n in the LOESS function n(t) ,d=1, the residual component is the original data minus the trend value and then minus the period value.

[0019] Further, including:

[0020] The use of the time series decomposition method to decompose the load of the steel enterprise also includes:

[0021] The outer loop of S13 is used to adjust the robust weights. If there are outliers in the sequence, the residual term is large, and the residual component is defined as:

[0022] h=6×median(|R v |) (1)

[0023] Among them, R v is the periodic component S v The maximum and minimum values ​​of R are a set, and median represents v The middle value part of

[0024] For a data point at position v, its robust weight is specified as:

[0025] ρ v =B(|R v | / h) (2)

[0026] Among them, the B function is a two-way function:

[0027]

[0028] Among them, u is the uncertainty of the periodic component, which is a subjective value;

[0029] In each subsequent iteration, after completing the Loess regression in steps S122 and S126, the neighborhood weight is multiplied by a specific parameter ρ v To reduce the influence of the remainder, the condition for judging the convergence of the inner loop is that n(i) reaches the set value.

[0030] Further, including:

[0031] The step S2 specifically includes:

[0032] The user behavior characteristics include production level, electricity price, production cost and enterprise electricity satisfaction. Each user behavior characteristic is represented by binary semantics, and the binary semantic representation is divided into high, medium and low. The three elements of triangular fuzzy numbers corresponding to the binary semantic representation are: high: (4,5,6), medium: (2,3,4), low: (1,2,3).

[0033] Further, including:

[0034] The quantification of the behavior characteristics to form quantitative data for calculation includes clarifying the triangular fuzzy number, and the formula used is:

[0035]

[0036] in, is a triangular fuzzy number, (a, b, c) are the three elements of the triangular fuzzy number, This is the result of clarifying the triangular fuzzy number.

[0037] Further, including:

[0038] In step S3, the improved neural network model is an improved BP neural network model, wherein the learning rate is improved to dynamically adjust the learning rate according to the error, specifically:

[0039] By comparing the size of the current error and the previous error in the iteration process, the learning rate is adaptively adjusted. Here, a factor λ that affects the learning rate is introduced. When the current error E(k) is greater than the error E(k-1) of the previous iteration, the learning rate η(k) of this iteration is reduced to (1-λ) times the learning rate η(k-1) of the previous iteration.

[0040] When the error E(k) of this iteration is less than the error E(k-1) of the previous iteration, the learning rate η(k) of this iteration is increased to (1+λ) times the learning rate η(k-1) of the previous iteration to get closer to the optimal weight. The specific adjustment format is as follows:

[0041]

[0042] Further, including:

[0043] The improved BP neural network model specifically includes:

[0044] S31 will be the trend quantity T v , periodic quantity S v and the residual component matrix C v and quantitative information D t With v 0,p The improved BP neural network is input in the form of forward calculation using the following formula, where t = 1, 2, 3, p = 1, 2, 3, 4, v 0,1 represents the trend quantity, v 0,2 represents the periodic quantity, v 0,3 represents the residual component, v 0,4 It represents the value after the quantitative information is clarified;

[0045]

[0046] In the formula, vk,j represents the output of the jth neuron in the kth layer, w k,ji represents the weight transmitted from the i-th neuron in the k-1th layer to the j-th neuron in the k-th layer;

[0047] The 0th layer of the neural network, that is, the input layer, has four neurons. The 1st, 2nd, and 3rd layers are hidden layers, and the 4th layer is the output layer, that is, k = 5, v k,j represents the jth neuron in layer k, i represents the node of the previous layer of neural network, N k-1 is the number of neurons in the k-1 layer, 4 in the 0th layer, 4 in the 1st, 2nd, and 3rd layers, and 1 in the 4th layer;

[0048] S32 calculates the error and updates the learning rate:

[0049]

[0050] Among them, E is the error of the current node, Y n is the expected output value, β o (n) is the output value of the nth neuron node on the output layer of a network, and c is the total number of nodes. Compare the error between this iteration and the previous iteration, and use formula (8) to update the learning rate.

[0051] Adjustment of S33 output layer weights and thresholds

[0052] The adjustment formula for the output layer weight is:

[0053]

[0054] Among them, η is the learning rate of this iteration, v bn Indicates the weight value of the bth node in the hidden layer of the network connected to the nth neuron node in the output layer of the network, β i (n) represents the input of the a-th network output layer;

[0055] The adjustment formula of the output layer threshold is:

[0056]

[0057] Among them, γ n Represents the output threshold of each neuron node on the network output layer;

[0058] S33 hidden layer weight and threshold adjustment

[0059] The adjustment formula for the hidden layer weight is:

[0060]

[0061] Among them, ω cdRepresents the weight value of the connection between the cth node of the network input layer and the dth neuron node of the hidden layer; α c (d) represents the input value of the dth neuron node in the hidden layer of the network; α o (d) represents the output value of the dth neuron node in the hidden layer;

[0062] The adjustment formula of the hidden layer threshold is:

[0063]

[0064] Among them, θ d Represents the output threshold of the dth neuron node in the hidden layer of the network;

[0065] S34 updates the weights and thresholds of each layer:

[0066] v' bn =v bn +Δv bn (14)

[0067] ω' cd =ω cd +Δω cd (15)

[0068] θ' d =θ d +Δθ d (16)

[0069] γ' n =γ n +Δγ n (17);

[0070] S35 uses the updated weights and thresholds to further train the samples and calculate the error. If the error still exceeds the set value, the gradient descent method is continued to be used for error back propagation, and the weights and thresholds are repeatedly updated until the final output error is within an acceptable range. The cycle is terminated and the model training is completed.

[0071] On the other hand, the present invention also provides a load response potential perception system for steel enterprises, the system comprising:

[0072] A decomposition module is used to decompose the load of the steel enterprise by using the time series decomposition method to obtain the trend quantity, cycle quantity and residual component of the load of the steel enterprise;

[0073] The data processing module is used to analyze the user behavior characteristics of the steel enterprise based on the binary semantic representation method, and quantify the behavior characteristics to form quantitative data for calculation;

[0074] The data output module is used to input the quantitative data and the acquired trend quantity, cycle quantity and residual component into the trained improved neural network model, so as to obtain the effective response potential of the steel enterprise.

[0075] Further, including:

[0076] In the data output module, the improved neural network model is an improved BP neural network model, wherein the learning rate is improved to dynamically adjust the learning rate according to the error, specifically:

[0077] By comparing the size of the current error and the previous error in the iteration process, the learning rate is adaptively adjusted. Here, a factor λ that affects the learning rate is introduced. When the current error E(k) is greater than the error E(k-1) of the previous iteration, the learning rate η(k) of this iteration is reduced to (1-λ) times the learning rate η(k-1) of the previous iteration.

[0078] When the error E(k) of this iteration is less than the error E(k-1) of the previous iteration, the learning rate η(k) of this iteration is increased to (1+λ) times the learning rate η(k-1) of the previous iteration to get closer to the optimal weight. The specific adjustment format is as follows:

[0079]

[0080] Finally, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to implement the above-mentioned load response potential perception method for steel enterprises when executed.

[0081] Beneficial effects: Compared with the prior art, the present invention has the following advantages: (1) The present invention proposes a steel enterprise load decomposition technology based on a time series decomposition method, which converts disordered enterprise load into an ordered load that can be quantitatively analyzed; (2) The present invention integrates binary semantic technology to construct an indicator system that can reflect the user behavior characteristics of steel enterprises, and uses triangular fuzzy numbers to represent it, effectively integrating multidimensional data and enriching the input dimension of the steel enterprise load response capability prediction model; (3) The present invention designs an improved neural network method with adaptive learning rate, fully considering the influence of enterprise user behavior characteristics, and analyzes the trend quantity, periodic quantity and residual component of the steel enterprise load to obtain the effective response potential of the steel enterprise. Therefore, the present invention can realize timely and accurate evaluation of the real-time operating characteristics of the grid side load, realize effective perception of the load side user characteristics, and accurately characterize the load side resource response potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 This is a flow chart of the time series decomposition method of the present invention;

[0083] Figure 2 A flow chart for generating load response potential of a steel enterprise for the improved neural network method based on adaptive learning rate according to the present invention;

[0084] Figure 3 It is a schematic diagram of decomposition of trend quantity, period quantity and residual component of the present invention;

[0085] Figure 4 It is a schematic diagram of the trend component and load curve of the present invention;

[0086] Figure 5 It is a schematic diagram of periodic components and load steps participating in demand response according to the present invention;

[0087] Figure 6 It is a comparison diagram of the load corresponding potential prediction value and the actual value according to the present invention;

[0088] Figure 7 This is the user behavior characteristic evaluation information representation intention described in the present invention. DETAILED DESCRIPTION

[0089] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0090] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0091] The first aspect of the present invention provides a method for sensing load response potential for a steel enterprise, such as Figure 1 and 2 As shown, the method comprises the following steps:

[0092] S1 uses the time series decomposition method to decompose the load of steel enterprises and obtain the trend quantity, cycle quantity and residual component of the load of steel enterprises.

[0093] In step 1, the time series decomposition is divided into two cycles, inner and outer, and the former completes trend fitting and periodic quantity calculation. It is the trend and period component after the K-1th pass of the inner loop is completed. Initially, it is assumed that And define the following parameters: n (i) Number of inner layers, n (o) Number of outer layers, n (p) The number of samples in one cycle, n (s) 、n(l) 、n (t) is the LOESS smoothing parameter in Step 2, 3, and 6. The samples at the same position in each cycle are grouped into a subsequence. There are n(p) subsequences in total, called cycle-subseries. Overall, the inner loop has the following 6 steps:

[0094] (1) Detrending, removing the influence of the previous round of trend volume,

[0095] (2) Subsequence smoothing is performed by using LOESS (q = n n(s) , d = 1) function performs regression smoothing for all subsequences, and at the same time extends a period T before and after them; after smoothing, the output result is recorded as v=-n (p) +1,N+n (p) ;

[0096] (3) Subsequence low-throughput filtering, the result sequence C(k+1)v of the previous step is moved averaged in turn, and finally LOESS (q=n n(l) ,d=1) regression, and obtain the final result sequence v=1,…,N; the purpose is to propose the low throughput of subsequence;

[0097] (4) To smooth the subsequence trend, Same effect as 1;

[0098] (5) De-periodize, subtract the periodic component,

[0099] (6) Trend smoothing: LOESS (q = n n(t) ,d=1) regression to obtain the trend component

[0100] The role of the outer loop is to adjust the robust weights. If there are outliers in the sequence, the remainder will be larger. Definition

[0101] h=6×median(|R v |)(1)

[0102] For a data point at position v, its robust weight is specified as

[0103] ρ v =B(|R v | / h)(2)

[0104] Among them, the B function is a two-way function:

[0105]

[0106] In each subsequent iteration, after completing the Loess regression in steps 2 and 6, the neighborhood weight needs to be multiplied by a specific parameter ρ v , to reduce the influence of the remainder. In particular, the condition for judging the inner loop convergence is that n(i) reaches the set value; if the remainder in the sequence is not obvious, n(0) can be set to 0.

[0107] S2 analyzes the user behavior characteristics of steel enterprises based on the binary semantic representation method, and quantifies the behavior characteristics to form quantitative data for calculation.

[0108] The binary semantic representation method is used to analyze the user behavior characteristics of steel enterprises, forming quantitative data that can be used for quantitative calculation, providing support for the calculation of the load response potential of steel enterprises. The semantic representation method for constructing the user behavior characteristics of steel enterprises is shown in Table 1:

[0109] Table 1 Semantic representation method of user behavior characteristics in steel enterprises

[0110]

[0111] The corresponding relationship between the semantics and triangular fuzzy numbers is shown in Table 2:

[0112] Table 2 Corresponding semantics of triangular fuzzy numbers

[0113]

[0114] Furthermore, in step 2, based on the binary semantic representation method, the triangular fuzzy number formed is clarified according to the following formula (4) to form quantitative data that can be used for calculation, providing support for the calculation of the load response potential of steel enterprises

[0115]

[0116] in, is a triangular fuzzy number, (a, b, c) are the three elements of the triangular fuzzy number, This is the result of clarifying the triangular fuzzy number.

[0117] S3 inputs the quantitative data and the obtained trend quantity, cycle quantity and residual component into the trained improved neural network model, so as to obtain the effective response potential of the steel enterprise.

[0118] First, the traditional BP neural network training process is introduced as follows:

[0119] 1) Determine relevant parameters and activation functions

[0120] Determine the number of nodes of each layer of neurons in the input layer, hidden layer and output layer, set the learning rate η of the neural network, the number of iterations epochs, and the activation function f(x). The formula of the S-type activation function is:

[0121]

[0122] 2) Forward calculation

[0123] After determining the relevant parameters and activation functions, perform the forward calculation of the BP neural network:

[0124]

[0125] Where v k,j represents the output of the jth neuron in the kth layer, w k,ji Represents the weight transmitted from the i-th neuron in the k-1th layer to the j-th neuron in the kth layer.

[0126] 3) Error back propagation

[0127] The objective function of the neural network is expressed as the sum of squares of the total error:

[0128]

[0129] Where the number of iterations is k, y represents the expected value of the output, z represents the actual value of the output, and the tolerable total error ε is determined. If E(k)>ε, the error is back-propagated, the weights and thresholds are updated, and retraining is performed.

[0130] 4) Repeat the training until the network converges

[0131] Use the updated weights and thresholds to further train the samples and calculate the error. If the error still exceeds the set value, continue to use the gradient descent method to backpropagate the error, repeatedly update the weights and thresholds, and terminate the loop until the final output error is within an acceptable range.

[0132] Furthermore, in step S3 of the present invention, an improved neural network method for designing an adaptive learning rate is used, which fully considers the influence of enterprise user behavior characteristics, and analyzes the trend quantity, cycle quantity and residual component of the steel enterprise load to obtain the effective response potential of the steel enterprise.

[0133] The model can dynamically adjust the learning rate according to the error during the training process, and adapt the learning rate by comparing the size of the current error and the previous error during the iteration process. A factor λ that affects the learning rate is introduced. When the current error E(k) is greater than the error E(k-1) of the previous iteration, the learning rate η(k) of this iteration is reduced to (1-λ) times the learning rate η(k-1) of the previous iteration; when the current error E(k) is less than the error E(k-1) of the previous iteration, the learning rate η(k) of this iteration is expanded to (1+λ) times the learning rate η(k-1) of the previous iteration to get closer to the optimal weight. The specific adjustment format is as follows:

[0134]

[0135] Furthermore, the step 3 designs an improved neural network method with adaptive learning rate, fully considers the influence of enterprise user behavior characteristics, and analyzes the trend, cycle and residual components of the steel enterprise load to obtain the effective response potential of the steel enterprise. The steps of the proposed improved neural network model are as follows:

[0136] 1) Calculation error

[0137]

[0138] Y is the expected output value, β o (n) is the output value of the nth node in the network output layer.

[0139] 2) Update the learning rate

[0140] Compare the error of this iteration with that of the previous iteration, and use formula (8) to update the learning rate.

[0141] 3) Adjustment of output layer weights and thresholds

[0142] The adjustment formula for the output layer weight is:

[0143]

[0144] Among them, v jn represents the weight value of the jth node in the hidden layer of the network connected to the nth neuron in the output layer, β i (n) represents the input of the nth neuron node in the output layer.

[0145] The adjustment formula of the output layer threshold is:

[0146]

[0147] Among them, γ n Represents the output threshold of each neuron node in the output layer.

[0148] 4) Adjustment of hidden layer weights and thresholds

[0149] The adjustment formula for the hidden layer weight is:

[0150]

[0151] Among them, ω ij Represents the weight value of the connection between the i-th node of the network input layer and the j-th neuron node of the hidden layer; α i (j) represents the input value of the jth node in the hidden layer of the network; α o (j) represents the output value of the jth neuron node in the hidden layer.

[0152] The adjustment formula of the hidden layer threshold is:

[0153]

[0154] Among them, θ j Represents the output threshold of each neuron node in the hidden layer of the network.

[0155] 5) Update the weights and thresholds of each layer

[0156] v' jn =v jn +Δv jn (14)

[0157] ω i ' j =ω ij +Δω ij (15)

[0158] θ j '=θ j +Δθ j (16)

[0159] γ' n =γ n +Δγ n (17)

[0160] Furthermore, the step 3 designs an improved neural network method with adaptive learning rate, fully considers the influence of enterprise user behavior characteristics, and analyzes the trend quantity, cycle quantity and residual component of the steel enterprise load to obtain the effective response potential of the steel enterprise.

[0161] According to the specific power system and the method mentioned above, the corresponding specific implementation is:

[0162] Step 1: Use the time series decomposition method to analyze the load of steel enterprises. It can carefully extract the important components of the load, including trend quantity, cycle quantity and residual component, so as to more comprehensively understand the law and trend of load change. The trend quantity reflects the overall trend of the load in the long term, the cycle quantity reveals the law of the load in periodic changes, and the residual component captures the random fluctuations outside the trend and cycle effects. By decomposing and more accurately identifying the fundamental characteristics of load changes, more accurate data support can be provided for the energy management and production plans of steel enterprises. This method of in-depth analysis of load data not only helps to predict future load trends, but also provides an important basis for formulating more effective energy utilization and production strategies. Decomposition quantities such as Figure 3 shown.

[0163] Step 2: By using binary semantic representation to capture the key features of user behavior and converting them into quantifiable data, a more powerful foundation is provided for data analysis and mining, thereby providing more detailed and actionable information support for corporate decision-making. Based on the key features of the captured user behavior, the triangular fuzzy numbers are clarified according to the following formula (4) to form quantitative data that can be used for calculation, such as Figure 7 shown.

[0164] Step 3: Using the improved neural network method based on adaptive learning rate, the behavioral characteristics of enterprise users are comprehensively considered, and the load of steel enterprises is deeply analyzed to extract trend quantity, cycle quantity and residual components to obtain the effective response potential of steel enterprises.

[0165] First, by introducing adaptive learning rate, the improved neural network can dynamically adjust the learning rate during the training process, making it more adaptable to changes in different periods and load data, which helps to improve the stability and convergence speed of the neural network. Secondly, the deep learning structure of the improved neural network enables it to better understand and extract the complex and nonlinear relationships in the load data. Through multi-level feature extraction and learning, the neural network can analyze the data more comprehensively and capture potential trend changes, thereby improving the accuracy of the trend quantity. In addition, the improved neural network method can also prevent the model from overfitting and enhance its generalization ability through effective regularization technology. This is particularly critical when processing long-term time series data, which helps to improve the prediction accuracy of the trend quantity.

[0166] (1) Trend load potential assessment

[0167] The trend load is Tv, which indirectly represents the production scale of the enterprise. It is generated by cluster equipment working continuously. Its load curve changes synchronously with the change of production scale and is consistent. This type of component is generally invited by the enterprise and the power grid company to participate in the regulation of the power grid, and its adjustability is determined along with the formulation of production calculations. When steel enterprises and power grid companies invite each other to participate in load response demand regulation, it is generally necessary to re-formulate the production plan to ensure that it meets the needs of power grid regulation. At the same time, this load response potential is generally related to the production equipment of the steel enterprise, and its adjustability is generally easy to be restricted, so it is relatively fixed. Therefore, the improved neural network method based on adaptive learning rate is used to analyze the load curve of the enterprise in the state of participating in load regulation in the past time period, so as to evaluate the trend load potential of the enterprise. The resulting response potential and trend component curves are shown in the figure below. Figure 4 shown.

[0168] (2) Evaluation of periodic component load potential

[0169] Periodic load component S v , represents the regular electricity consumption behavior of the electricity user, such as the regular start and stop of production equipment, which is specifically manifested as a step change in the load curve. Therefore, the periodic component S v The step part of the curve represents the start and stop of some large-scale electrical equipment in the production process or represents the change of the production process; while the smoother part of the curve is usually identified as the period of stable load output in the corresponding production process, which can be recorded as a load step. At present, most of the DR plans implemented in China are peak-shaving plans. At this time, when the demand response begins, the sudden drop in load before and after the response can be regarded as the user's adjustable load and regarded as load potential.

[0170] The improved neural network method based on adaptive learning rate is used to analyze the demand response cycle quantity, and the response load step curve is obtained as follows: Figure 5 shown.

[0171] The periodic load component of this user includes four load steps. Therefore, one of the user load characteristics can be identified as P p The difference between the load steps of L1, L2 and L3. The periodic component is directly generated by the STL algorithm. Due to the principle of the algorithm, the decomposed waveform cannot be very smooth, and even some burrs are generated. In order to accurately obtain the curve characteristics, the invention adopts the SG filtering method to remove the interference of the burr processing and retain the original characteristics.

[0172] When the enterprise participates in load response regulation under the invitation state, the mean value of the load regulation potential of its periodic component is 38.64kW and the standard deviation is 2.83kW, that is, under this working state, the periodic component DR load will fall between 29.98 and 44.30kW with a mean value of 38.64kW and a probability of 95%.

[0173] (3) Assessment of residual load potential

[0174] After investigation and analysis, it is known that the residual amount obtained by decomposition is the interruptible load fluctuation caused by irregular or unnecessary operations of the enterprise in the production process. These loads can be adjusted during the operation process by standardizing the production process, improving the operator's proficiency, and saving unnecessary electricity costs (such as lighting). This patent regards it as a responsive potential load.

[0175] To further verify the effectiveness of this method, an improved neural network method based on adaptive learning rate is used to predict the load demand response potential in the first week of July using the data from May to June as training sample input. The results are compared with the actual load response demand potential value. Figure 6 As shown, it can be seen that the demand response prediction value obtained by the proposed scheme can better predict the load response potential in the future, and the difference with the true value can be controlled below 5%. Table 4 shows the predicted value of the load response potential of a certain day under the improved neural network method with adaptive learning rate.

[0176] Table 4 Load response potential prediction values

[0177]

[0178] On the other hand, the present invention also provides a load response potential perception system for steel enterprises, the system comprising:

[0179] A decomposition module is used to decompose the load of the steel enterprise by using the time series decomposition method to obtain the trend quantity, cycle quantity and residual component of the load of the steel enterprise;

[0180] The data processing module is used to analyze the user behavior characteristics of the steel enterprise based on the binary semantic representation method, and quantify the behavior characteristics to form quantitative data for calculation;

[0181] The data output module is used to input the quantitative data and the acquired trend quantity, cycle quantity and residual component into the trained improved neural network model, so as to obtain the effective response potential of the steel enterprise.

[0182] Further, including:

[0183] In the data output module, the improved neural network model is an improved BP neural network model, wherein the learning rate is improved to dynamically adjust the learning rate according to the error, specifically:

[0184] By comparing the size of the current error and the previous error in the iteration process, the learning rate is adaptively adjusted. Here, a factor λ that affects the learning rate is introduced. When the current error E(k) is greater than the error E(k-1) of the previous iteration, the learning rate η(k) of this iteration is reduced to (1-λ) times the learning rate η(k-1) of the previous iteration.

[0185] When the error E(k) of this iteration is less than the error E(k-1) of the previous iteration, the learning rate η(k) of this iteration is increased to (1+λ) times the learning rate η(k-1) of the previous iteration to get closer to the optimal weight. The specific adjustment format is as follows:

[0186]

[0187] The other technical features of the load response potential perception system for steel enterprises in the present invention are similar to those of the load response potential perception method for steel enterprises, and will not be repeated in this application.

[0188] Finally, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to implement the above-mentioned load response potential perception method for steel enterprises when executed.

[0189] In addition, it can be understood by a person skilled in the art that all or part of the processes in the method for implementing the above embodiment can be completed by instructing the relevant hardware through a computer program. The computer program includes program instructions, and the computer program can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the control terminal to implement the process steps of the embodiment of the above method.

[0190] Therefore, the present invention also provides a computer-readable storage medium, which stores a job risk identification program. When the job risk identification program is executed by a processor, the various steps of the job risk identification method described in the above embodiment are implemented.

[0191] It should be noted that since the storage medium provided in the embodiment of the present application is the storage medium used to implement the method of the embodiment of the present application, based on the method introduced in the embodiment of the present application, the person skilled in the art can understand the specific structure and deformation of the storage medium, so it is not repeated here. All storage media used in the method of the embodiment of the present application belong to the scope of protection of this application.

[0192] 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.

[0193] 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.

[0194] 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.

[0195] 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 The steps for the functions specified in one or more boxes.

[0196] It should be noted that in the claims, any reference signs placed between parentheses shall not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claim. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention may be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by one and the same item of hardware. The use of the words first, second, and third etc. does not indicate any order. These words may be interpreted as names.

[0197] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0198] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

[0199] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A load response potential perception method for steel enterprises, characterized in that: The method comprises the following steps: S1 uses the time series decomposition method to decompose the load of steel enterprises and obtain the trend, cycle and residual components of the load of steel enterprises; S2 analyzes the user behavior characteristics of steel enterprises based on the binary semantic representation method, and quantifies the behavior characteristics to form quantitative data for calculation; S3 inputs the quantitative data and the obtained trend quantity, cycle quantity and residual component into the trained improved neural network model, so as to obtain the effective response potential of the steel enterprise; In step S1, decomposing the load of the steel enterprise by using the time series decomposition method specifically includes: S11 Assumption It is the trend and period after the K-1th calculation of the inner loop is completed. Initially, it is assumed that And define the following parameters: n (i) is the number of inner layers, n (o) is the number of outer layers, n (p) is the number of samples in one cycle, n (s) 、n (l) 、n (t) is the LOESS smoothing parameter; S12 groups the samples at the same position in each cycle into a subsequence. There are n(p) subsequences in total. The inner loop includes the following steps: S121 detrends, removing the influence of the previous round of trend volume. S122 subsequence smoothing is done by using the LOESS function to perform regression smoothing on all subsequences, and at the same time extending them by a period T before and after them, where q = n in the LOESS function n(s) , d = 1; after smoothing, the output result is recorded as The S123 subsequence is low-throughput filtered, and the result sequence C(k+1)v of the previous step is moved averaged in turn, and the final result sequence is obtained through LOESS function regression. Among them, q = n in the LOESS function n(l) ,d=1; S124 to smooth subsequence trends, S125 de-cycles, minus the cycle amount, S126 trend smoothing, LOESS function regression is performed on the sequence after removing the period to obtain the trend value Among them, q = n in the LOESS function n(t) ,d=1, the residual component is the original data minus the trend value and then minus the period value; The use of the time series decomposition method to decompose the load of the steel enterprise also includes: The outer loop of S13 is used to adjust the robust weights. If there are outliers in the sequence, the residual term is large, and the residual component is defined as: h=6×median(|R v |) (1) Among them, R v is the periodic component S v The maximum and minimum values ​​of R are a set, and median represents v The middle value part of For a data point at position v, its robust weight is specified as: r v =B(|R v | / h) (2) Among them, the B function is a two-way function: Among them, u is the uncertainty of the periodic component, which is a subjective value; In each subsequent iteration, after completing the Loess regression in steps S122 and S126, the neighborhood weight is multiplied by a specific parameter ρ v To reduce the influence of the remainder, the condition for judging the convergence of the inner loop is that n(i) reaches the set value; The step S2 specifically includes: The user behavior characteristics include production level, electricity fee, production cost and enterprise electricity satisfaction. Each user behavior characteristic is represented by binary semantics. The binary semantic representation is divided into high, medium and low. The three triangular fuzzy number elements corresponding to the binary semantic representation are: high: (4, 5, 6), medium: (2, 3, 4), low: (1, 2, 3); In step S3, the improved neural network model is an improved BP neural network model, wherein the learning rate is improved to dynamically adjust the learning rate according to the error, specifically: By comparing the size of the current error and the previous error in the iteration process, the learning rate is adaptively adjusted. Here, a factor λ that affects the learning rate is introduced. When the current error E(k) is greater than the error E(k-1) of the previous iteration, the learning rate η(k) of this iteration is reduced to (1-λ) times the learning rate η(k-1) of the previous iteration. When the error E(k) of this iteration is less than the error E(k-1) of the previous iteration, the learning rate η(k) of this iteration is increased to (1+λ) times the learning rate η(k-1) of the previous iteration to get closer to the optimal weight. The specific adjustment format is as follows: The improved BP neural network model specifically includes: S31 will be the trend quantity T v , periodic quantity S v and the residual component matrix C v and quantitative information D t With v 0,p The improved BP neural network is input in the form of forward calculation using the following formula, where t = 1, 2, 3, p = 1, 2, 3, 4, v 0,1 represents the trend quantity, v 0,2 represents the periodic quantity, v 0,3 represents the residual component, v 0,4 It represents the value after the quantitative information is clarified; In the formula, v k,j represents the output of the jth neuron in the kth layer, w k,ji represents the weight transmitted from the i-th neuron in the k-1th layer to the j-th neuron in the k-th layer; The 0th layer of the neural network, that is, the input layer, has four neurons. The 1st, 2nd, and 3rd layers are hidden layers, and the 4th layer is the output layer, that is, k = 5, v k,j represents the jth neuron in layer k, i represents the node of the previous layer of neural network, N k-1 is the number of neurons in the k-1 layer, 4 in the 0th layer, 4 in the 1st, 2nd, and 3rd layers, and 1 in the 4th layer; S32 calculates the error and updates the learning rate: Among them, E is the error of the current node, Y n is the expected output value, β o (n) is the output value of the nth neuron node on the output layer of a network, and c is the total number of nodes. Compare the error between this iteration and the previous iteration, and use formula (8) to update the learning rate. Adjustment of S33 output layer weights and thresholds The adjustment formula for the output layer weight is: Among them, η is the learning rate of this iteration, v bn Indicates the weight value of the bth node in the hidden layer of the network connected to the nth neuron node in the output layer of the network, β i (n) represents the input of the a-th network output layer; The adjustment formula of the output layer threshold is: Among them, γ n Represents the output threshold of each neuron node on the network output layer; S33 hidden layer weight and threshold adjustment The adjustment formula for the hidden layer weight is: Among them, ω cd Indicates the weight value of the connection between the cth node of the network input layer and the dth neuron node of the hidden layer; α c (d) represents the input value of the dth neuron node in the hidden layer of the network; α o (d) represents the output value of the dth neuron node in the hidden layer; The adjustment formula of the hidden layer threshold is: Among them, θ d Represents the output threshold of the dth neuron node in the hidden layer of the network; S34 updates the weights and thresholds of each layer: v' bn =v bn +Δv bn (14) oh cd =ω cd +See cd (15) I will d =θ d +Δθ d (16) c' n =c n +Δγ n (17); S35 uses the updated weights and thresholds to further train the samples and calculate the error. If the error still exceeds the set value, the gradient descent method is continued to be used for error back propagation, and the weights and thresholds are repeatedly updated until the final output error is within an acceptable range. The cycle is terminated and the model training is completed.

2. The load response potential perception method for steel enterprises according to claim 1 is characterized in that: The quantification of the behavior characteristics to form quantitative data for calculation includes clarifying the triangular fuzzy number, and the formula used is: in, is a triangular fuzzy number, (a, b, c) are the three elements of the triangular fuzzy number, This is the result of clarifying the triangular fuzzy number.

3. A load response potential perception system for steel enterprises, characterized in that: The system is used to implement the perception method described in claim 1, and specifically includes: A decomposition module is used to decompose the load of the steel enterprise by using the time series decomposition method to obtain the trend quantity, cycle quantity and residual component of the load of the steel enterprise; The data processing module is used to analyze the user behavior characteristics of the steel enterprise based on the binary semantic representation method, and quantify the behavior characteristics to form quantitative data for calculation; The data output module is used to input the quantitative data and the acquired trend quantity, cycle quantity and residual component into the trained improved neural network model, so as to obtain the effective response potential of the steel enterprise.

4. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to implement the load response potential perception method for steel enterprises as described in claim 1 or 2 when executed.