A multi-type load collaborative control method and system for industrial parks

Abstract

The present invention discloses a method and system for collaborative control of multiple types of loads for industrial parks, the method including data collection and processing, generating adversarial networks, enriching original data, building multi-type load prediction models and specific operations. The present invention belongs to the field of load collaborative control, specifically refers to a method and system for collaborative control of multiple types of loads for industrial parks. This scheme generates simulation data by constructing generator and discriminator models, designing regularized loss functions, and using back-propagation to update parameters to generate adversarial networks, thereby enriching the amount of original data and improving training effects; by constructing a multi-factor neural network, designing a gradient moment estimation function, designing a matrix offset correction term, designing a step-based parameter update method, and designing an optimal initial model parameter search method, the model accuracy is improved, the accuracy of the model parameter optimization direction is improved, the adaptive correction capability is improved, and the optimization time is shortened.

Description

Technical Field

[0001] The present invention relates to the field of load collaborative control, and specifically refers to a multi-type load collaborative control method and system for industrial parks. Background Art

[0002] A multi-type load collaborative control method and system for industrial parks refers to a method and system that uses machine learning to analyze and model multi-type load data in industrial parks to achieve collaborative control of loads. However, traditional collaborative control methods have problems such as insufficient original data volume, poor training effect and low generalization ability; single consideration of influencing factors, poor parameter optimization methods, low model accuracy, inaccurate parameter optimization direction, inability to adaptively correct and slow speed. Summary of the invention

[0003] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a multi-type load collaborative control method and system for industrial parks. In view of the problem that the traditional collaborative control method has insufficient original data volume, resulting in poor training effect and low generalization ability, this solution constructs a generator and a discriminator model, designs a regularized loss function, and uses back propagation to update the parameters to generate an adversarial network to generate simulation data, thereby enriching the original data volume and improving the training effect and generalization ability; in view of the problem that the influencing factors are single and the parameter optimization method is poor, resulting in low model accuracy, inaccurate parameter optimization direction, inability to adaptively correct and slow speed, this solution constructs a multi-factor neural network, designs a gradient moment estimation function, designs a matrix offset correction term, designs a step-based parameter update method and designs an optimal initial model parameter search method to improve the model accuracy, improve the accuracy of the model parameter optimization direction, improve the adaptive correction ability and shorten the optimization time.

[0004] The technical solution adopted by the present invention is as follows: The present invention provides a multi-type load coordinated control method for an industrial park, the method comprising the following steps:

[0005] Step S1: data collection and processing;

[0006] Step S2: Generate adversarial network;

[0007] Step S3: enriching the original data;

[0008] Step S4: construct a multi-type load forecasting model;

[0009] Step S5: specific operation.

[0010] Furthermore, in step S1, the data collection and processing specifically includes the following steps:

[0011] Step S11: Collect external environment data, including weather data, seasonal data, and holiday data; operation status data, including operation time period data, operation mode data, and operation plan data; equipment status, including equipment operation status data, equipment fault information data, equipment energy consumption, and operation efficiency; work status data, including work plan data and personnel quantity data; set these data as features; collect historical load data of the industrial park; set the load data as a label;

[0012] Step S12: Data preprocessing, cleaning the collected data, filling missing data, detecting and removing abnormal data; screening, selecting data with key features and removing noise data; data conversion, multi-type data aggregation, adjusting data of different scales to the same range.

[0013] Furthermore, in step S2, the generative adversarial network specifically includes the following steps:

[0014] Step S21: constructing a generator and a discriminator model. The constructed generator receives a random noise vector and a conditional vector as input, and converts them into a false sample similar to a real sample through a neural network. The discriminator receives a real sample or a false sample generated by the generator and a target label as input. The GAN model optimizes the entire model by alternately training the generator and the discriminator, minimizing the loss function of the generator to make the generated sample closer to the real sample, and maximizing the loss function of the discriminator to improve its accuracy.

[0015] Step S22: Design a consistency regularization adding strategy, the formula is as follows:

[0016] ;

[0017] ;

[0018] Among them, L G and L D Represent the consistency regularization additions of the generator and the discriminator, respectively. , and represents the weight factor, represents the enhanced least squares loss, represents the enhanced real data loss, represents the enhanced simulation data loss, represents enhanced mixing loss;

[0019] Step S23: Design a consistency regularization term, the formula is as follows:

[0020] ;

[0021] Among them, L represents the consistency regularization addition term, represents the gradient weight, represents the balanced gradient, D(·) represents the discriminator function, T(x) represents the data enhancement function, ||·|| 2 represents the square of the norm;

[0022] Step S24: Design the loss function of the generator, the formula is as follows:

[0023] ;

[0024] Among them, Loss G represents the loss function of the generator; E[·] represents the mean, G(·) represents the generator function, s represents the random noise vector of the generator input, c represents the conditional vector of the generator input, represents the target label, λ represents the weight factor, represents the characteristic distribution of generated samples, represents the mean of the generated samples, represents the standard deviation of the generated samples, N(0,1) represents the standard normal distribution, D KL (·) represents KL divergence, log represents logarithm, and M(·) represents least squares loss;

[0025] Step S25: Design the loss function of the discriminator, the formula is as follows:

[0026] ;

[0027] Among them, Loss D represents the loss function of the discriminator; r represents the real sample;

[0028] Step S26: Iterative training, generate a batch of false samples through the generator, and take the real samples and false samples as input, calculate the loss function of the discriminator, and update the parameters of the discriminator through back propagation, then generate a batch of false samples through the generator, and take the false samples as input, calculate the loss function of the generator, and update the parameters of the generator through back propagation, repeat the above steps, alternately train the generator and the discriminator until the predetermined number of training times is reached or the loss function converges.

[0029] Furthermore, in step S3, the enriching of the original data specifically includes the following steps:

[0030] Step S31: Generate simulation data. Use the trained generator to input the prepared input vector into the generator model to generate simulation data. The generator will generate simulation data similar to the real data based on the input vector and the learned parameters.

[0031] Step S32: after generating simulation data, remove abnormal values ​​and select high-quality data that meets the conditions;

[0032] Step S33: enriching the original data, adding the processed simulation data to the original data.

[0033] Furthermore, in step S4, the construction of the multi-type load forecasting model specifically includes the following steps:

[0034] Step S41: construct a multi-factor neural network and a feedforward neural network model; the hidden layer uses a ReLU activation function and the output layer uses a Sigmoid activation function;

[0035] Step S42: define the loss function, the formula is as follows:

[0036] ;

[0037] Among them, J(t) represents the loss value at time t, lor represents the loss weight, Q(t) represents the actual load value at time t, represents the predicted load value at time t, n represents the sample index, N represents the total number of samples, and lop represents the loss bias;

[0038] Step S43: Initialize model parameters;

[0039] Step S44: Design a gradient moment estimation function, the formula is as follows:

[0040] ;

[0041] ;

[0042] Where t represents time, g(t) represents the gradient value at time t, represents the square of the gradient norm, a is the step size, m t represents the first-order moment of the gradient at time t, m t-a represents the first-order moment of the gradient at time ta, v t represents the second-order moment of the gradient at time t, v t-a represents the second-order moment of the gradient at time ta, and represents the attenuation coefficient;

[0043] Step S45: Design the matrix offset correction term, the formula is as follows:

[0044] ;

[0045] ;

[0046] in, and Respectively represent m t and v t Matrix offset correction;

[0047] Step S46: Design a parameter update method based on step length, the formula is as follows:

[0048] ;

[0049] Among them, θ t represents the parameter at time t, θ t-a represents the parameter at time ta, ε represents a small positive number;

[0050] Step S47: Iterative update: pre-set the threshold of the loss function, calculate the loss function value under the current model parameters after each iterative update, if the current loss function value is less than the threshold, stop the iteration and output the current model parameters as the optimal model parameters; if the current loss function value is greater than the threshold, continue to iterate until the maximum number of iterations is reached, if the parameters less than the threshold are still not obtained, reinitialize the model parameters and continue iterative update;

[0051] Step S48: Design an optimal initial model parameter search method, the formula is as follows:

[0052] ;

[0053] Among them, Iter represents the number of iterations, θ * (Iter+1) represents the initial model parameters searched for the Iter+1th iteration, ω is the search weight, best(θ * ) represents the global optimal initial model parameters currently searched, ibest(θ * ) represents the local optimal initial model parameters currently searched, rand(0,1) represents a random number between 0 and 1, θ * (Iter) represents the initial model parameters searched at the Iter-th iteration;

[0054] Step S49: Search for the optimal initial model parameters, set the model performance as the fitness value, pre-set the fitness threshold, iteratively search the initial model parameters, calculate the fitness value after each iteration, and if there is an initial model parameter whose corresponding fitness value is greater than the threshold, stop the iteration and set the initial model parameters at this position as the optimal initial model parameters; if the maximum number of iterations is reached, reinitialize the parameter position for search; otherwise, continue the iterative search.

[0055] Furthermore, in step S5, the specific operation is to collect the load data of the industrial park, and predict the load status of the industrial park based on the constructed multi-type load prediction model, that is, by collecting the load data of the industrial park in real time, inputting it into the established multi-type load prediction model, predicting the load of the industrial park, and setting a threshold to monitor whether an overload state will occur. If the predicted load does not exceed the threshold, the system remains in a normal state. If the predicted load exceeds the threshold, the system enters an alarm state and issues an overload alarm message. According to various load demands in the park, the system is optimized and scheduled to achieve coordinated control of multiple types of loads.

[0056] The present invention provides a multi-type load collaborative control system for industrial parks, including a data acquisition and processing module, a generative adversarial network module, an original data enrichment module, a multi-type load prediction model construction module and a specific operation module;

[0057] The data acquisition and processing module collects historical load data, external environment data, operation status, equipment status, and working status data of the industrial park and performs cleaning, screening, data conversion, and multi-type data aggregation, and sends the processed data to the generative adversarial network module and the enriched raw data module;

[0058] The generative adversarial network module receives the data sent by the data acquisition and processing module, generates an adversarial network by constructing generator and discriminator models, designing consistency regularization adding strategies and regularization terms, designing loss functions of the generator and the discriminator, and sends the data to the enriched original data module;

[0059] The enriched raw data module receives data sent by the generated adversarial network module and the data acquisition and processing module, generates simulation data through the adversarial network, enriches the raw data after removing abnormalities, and sends the data to the module for building a multi-type load prediction model;

[0060] The multi-type load forecasting model module is constructed by constructing a multi-factor neural network, defining a loss function, initializing model parameters, designing a gradient moment estimation function, designing a matrix offset correction term, designing a step-based parameter update method, designing an optimal initial model parameter search method, iteratively updating and constructing an optimal parameter model, and sending data to a specific operation module;

[0061] The specific operation module receives data sent by the module for constructing a multi-type load prediction model, performs load prediction based on the historical load information of the industrial park, and optimizes and schedules the system based on the prediction results to achieve coordinated control of multi-type loads.

[0062] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0063] (1) In order to solve the problem that the traditional collaborative control method has insufficient original data volume, resulting in poor training effect and low generalization ability, this scheme constructs a generator and discriminator model, designs a regularized loss function, and uses back-propagation to update the parameters of the generative adversarial network to generate simulation data, thereby enriching the original data volume and improving the training effect and generalization ability.

[0064] (2) In order to address the problems of low model accuracy, inaccurate parameter optimization direction, inability to adaptively correct, and slow speed due to single consideration of influencing factors and poor parameter optimization methods, this solution improves model accuracy, improves the accuracy of model parameter optimization direction, improves adaptive correction capabilities, and shortens optimization time by constructing a multi-factor neural network, designing a gradient moment estimation function, designing a matrix offset correction term, designing a step-based parameter update method, and designing an optimal initial model parameter search method. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 A schematic diagram of a flow chart of a multi-type load collaborative control method for an industrial park provided by the present invention;

[0066] Figure 2 A schematic diagram of a multi-type load collaborative control system for an industrial park provided by the present invention;

[0067] Figure 3 is a schematic flow chart of step S2;

[0068] Figure 4 It is a schematic diagram of the process of step S4.

[0069] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0070] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0071] In the description of the present invention, it should be understood that terms such as “upper”, “lower”, “front”, “back”, “left”, “right”, “top”, “bottom”, “inside” and “outside” indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as limiting the present invention.

[0072] Example 1, see Figure 1 The present invention provides a smart book management method for reading promotion, which includes the following steps:

[0073] Step S1: Data collection and processing, collecting historical load data, external environment data, operation status, equipment status, working status data of the industrial park and cleaning, screening, data conversion, and multi-type data aggregation;

[0074] Step S2: Generate an adversarial network by constructing generator and discriminator models, designing consistency regularization adding strategies and regularization terms, and designing loss functions of the generator and discriminator.

[0075] Step S3: enriching the original data, generating simulation data through the adversarial network, and enriching the original data after removing abnormalities;

[0076] Step S4: construct a multi-type load forecasting model by constructing a multi-factor neural network, defining a loss function, initializing model parameters, designing a gradient moment estimation function, designing a matrix offset correction term, designing a step-based parameter update method, designing an optimal initial model parameter search method, and iteratively updating to construct an optimal parameter model;

[0077] Step S5: Specific operation: load forecasting is performed based on the historical load information of the industrial park, and the system is optimized and scheduled based on the forecast results to achieve coordinated control of multiple types of loads.

[0078] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, the data collection and processing specifically includes the following steps:

[0079] Step S11: Collect external environment data, including weather data, seasonal data, and holiday data; operation status data, including operation time period data, operation mode data, and operation plan data; equipment status, including equipment operation status data, equipment fault information data, equipment energy consumption, and operation efficiency; work status data, including work plan data and personnel quantity data; set these data as features; collect historical load data of the industrial park; set the load data as a label;

[0080] Step S12: Data preprocessing, cleaning the collected data, filling missing data, detecting and removing abnormal data; screening, selecting data with key features and removing noise data; data conversion, multi-type data aggregation, adjusting data of different scales to the same range.

[0081] Example 3, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In step S2, the generative adversarial network specifically includes the following steps:

[0082] Step S21: constructing a generator and a discriminator model. The constructed generator receives a random noise vector and a conditional vector as input, and converts them into a false sample similar to a real sample through a neural network. The discriminator receives a real sample or a false sample generated by the generator and a target label as input. The GAN model optimizes the entire model by alternately training the generator and the discriminator, minimizing the loss function of the generator to make the generated sample closer to the real sample, and maximizing the loss function of the discriminator to improve its accuracy.

[0083] Step S22: Design a consistency regularization adding strategy, the formula is as follows:

[0084] ;

[0085] ;

[0086] Among them, L G and L D Represent the consistency regularization additions of the generator and the discriminator, respectively. , and represents the weight factor, represents the enhanced least squares loss, represents the enhanced real data loss, represents the enhanced simulation data loss, represents enhanced mixing loss;

[0087] Step S23: Design a consistency regularization term, the formula is as follows:

[0088] ;

[0089] Among them, L represents the consistency regularization addition term, represents the gradient weight, represents the balanced gradient, D(·) represents the discriminator function, T(x) represents the data enhancement function, ||·|| 2 represents the square of the norm;

[0090] Step S24: Design the loss function of the generator, the formula is as follows:

[0091] ;

[0092] Among them, Loss G represents the loss function of the generator; E[·] represents the mean, G(·) represents the generator function, s represents the random noise vector of the generator input, c represents the conditional vector of the generator input, represents the target label, λ represents the weight factor, represents the characteristic distribution of generated samples, represents the mean of the generated samples, represents the standard deviation of the generated samples, N(0,1) represents the standard normal distribution, D KL (·) represents KL divergence, log represents logarithm, and M(·) represents least squares loss;

[0093] Step S25: Design the loss function of the discriminator, the formula is as follows:

[0094] ;

[0095] Among them, Loss D represents the loss function of the discriminator; r represents the real sample;

[0096] Step S26: Iterative training, generate a batch of false samples through the generator, and take the real samples and false samples as input, calculate the loss function of the discriminator, and update the parameters of the discriminator through back propagation, then generate a batch of false samples through the generator, and take the false samples as input, calculate the loss function of the generator, and update the parameters of the generator through back propagation, repeat the above steps, alternately train the generator and the discriminator until the predetermined number of training times is reached or the loss function converges.

[0097] By performing the above operations, in order to address the problem that the traditional collaborative control method has insufficient original data volume, resulting in poor training effect and low generalization ability, this solution constructs generator and discriminator models, designs regularized loss functions, and uses back-propagation to update parameters to generate adversarial networks to generate simulation data, thereby enriching the original data volume and improving training effect and generalization ability.

[0098] Example 4, see Figure 1 This embodiment is based on the above embodiment. In step S3, the enriching of original data specifically includes the following steps:

[0099] Step S31: Generate simulation data. Use the trained generator to input the prepared input vector into the generator model to generate simulation data. The generator will generate simulation data similar to the real data based on the input vector and the learned parameters.

[0100] Step S32: after generating simulation data, remove abnormal values ​​and select high-quality data that meets the conditions;

[0101] Step S33: enriching the original data, adding the processed simulation data to the original data.

[0102] Example 5, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In step S4, the construction of a multi-type load forecasting model specifically includes the following steps:

[0103] Step S41: construct a multi-factor neural network and a feedforward neural network model; the hidden layer uses a ReLU activation function and the output layer uses a Sigmoid activation function;

[0104] Step S42: define the loss function, the formula is as follows:

[0105] ;

[0106] Among them, J(t) represents the loss value at time t, lor represents the loss weight, Q(t) represents the actual load value at time t, represents the predicted load value at time t, n represents the sample index, N represents the total number of samples, and lop represents the loss bias;

[0107] Step S43: Initialize model parameters;

[0108] Step S44: Design a gradient moment estimation function, the formula is as follows:

[0109] ;

[0110] ;

[0111] Where t represents time, g(t) represents the gradient value at time t, represents the square of the gradient norm, a is the step size, m t represents the first-order moment of the gradient at time t, m t-a represents the first-order moment of the gradient at time ta, v t represents the second-order moment of the gradient at time t, v t-a represents the second-order moment of the gradient at time ta, and represents the attenuation coefficient;

[0112] Step S45: Design the matrix offset correction term, the formula is as follows:

[0113] ;

[0114] ;

[0115] in, and Respectively represent m t and v t Matrix offset correction;

[0116] Step S46: Design a parameter update method based on step length, the formula is as follows:

[0117] ;

[0118] Among them, θ t represents the parameter at time t, θ t-a represents the parameter at time ta, ε represents a small positive number;

[0119] Step S47: Iterative update: pre-set the threshold of the loss function, calculate the loss function value under the current model parameters after each iterative update, if the current loss function value is less than the threshold, stop the iteration and output the current model parameters as the optimal model parameters; if the current loss function value is greater than the threshold, continue to iterate until the maximum number of iterations is reached, if the parameters less than the threshold are still not obtained, reinitialize the model parameters and continue iterative update;

[0120] Step S48: Design an optimal initial model parameter search method, the formula is as follows:

[0121] ;

[0122] Among them, Iter represents the number of iterations, θ * (Iter+1) represents the initial model parameters searched for the Iter+1th iteration, ω is the search weight, best(θ * ) represents the global optimal initial model parameters currently searched, ibest(θ * ) represents the local optimal initial model parameters currently searched, rand(0,1) represents a random number between 0 and 1, θ * (Iter) represents the initial model parameters searched at the Iter-th iteration;

[0123] Step S49: Search for the optimal initial model parameters, set the model performance as the fitness value, pre-set the fitness threshold, iteratively search the initial model parameters, calculate the fitness value after each iteration, and if there is an initial model parameter whose corresponding fitness value is greater than the threshold, stop the iteration and set the initial model parameters at this position as the optimal initial model parameters; if the maximum number of iterations is reached, reinitialize the parameter position for search; otherwise, continue the iterative search.

[0124] By performing the above operations, in order to solve the problems of low model accuracy, inaccurate parameter optimization direction, inability to adaptively correct and slow speed due to single consideration of influencing factors and poor parameter optimization methods, this solution improves model accuracy, improves the accuracy of model parameter optimization direction, improves adaptive correction capabilities and shortens optimization time by constructing a multi-factor neural network, designing gradient moment estimation functions, designing matrix offset correction terms, designing step-based parameter update methods and designing optimal initial model parameter search methods.

[0125] Example 6, see Figure 1 This embodiment is based on the above embodiment. In step S5, the specific operation is to collect the load data of the industrial park, and realize the prediction of the load state of the industrial park based on the constructed multi-type load prediction model, that is, by collecting the load data of the industrial park in real time, inputting it into the established multi-type load prediction model, predicting the load of the industrial park, and setting a threshold to monitor whether an overload state will occur. If the predicted load does not exceed the threshold, the system remains in a normal state. If the predicted load exceeds the threshold, the system enters an alarm state and issues an overload alarm message. According to various load demands in the park, the system is optimized and scheduled to realize coordinated control of multiple types of loads.

[0126] Embodiment 7, see Figure 2 , this embodiment is based on the above embodiment, and the present invention provides a multi-type load collaborative control system for industrial parks, including a data acquisition and processing module, a generative adversarial network module, an enriched original data module, a multi-type load prediction model building module and a specific operation module;

[0127] The data acquisition and processing module collects historical load data, external environment data, operation status, equipment status, and working status data of the industrial park and performs cleaning, screening, data conversion, and multi-type data aggregation, and sends the processed data to the generative adversarial network module and the enriched raw data module;

[0128] The generative adversarial network module receives the data sent by the data acquisition and processing module, generates an adversarial network by constructing generator and discriminator models, designing consistency regularization adding strategies and regularization terms, designing loss functions of the generator and the discriminator, and sends the data to the enriched original data module;

[0129] The enriched raw data module receives data sent by the generated adversarial network module and the data acquisition and processing module, generates simulation data through the adversarial network, enriches the raw data after removing abnormalities, and sends the data to the module for building a multi-type load prediction model;

[0130] The multi-type load forecasting model module is constructed by constructing a multi-factor neural network, defining a loss function, initializing model parameters, designing a gradient moment estimation function, designing a matrix offset correction term, designing a step-based parameter update method, designing an optimal initial model parameter search method, iteratively updating and constructing an optimal parameter model, and sending data to a specific operation module;

[0131] The specific operation module receives data sent by the module for constructing a multi-type load prediction model, performs load prediction based on the historical load information of the industrial park, and optimizes and schedules the system based on the prediction results to achieve coordinated control of multi-type loads.

[0132] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0133] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

[0134] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A multi-type load collaborative control method for industrial parks, characterized by: The method comprises the following steps: Step S1: data collection and processing; Step S2: Generate adversarial network; Step S3: enriching the original data; Step S4: construct a multi-type load forecasting model; Step S5: specific operation; In step S4, the construction of a multi-type load forecasting model specifically includes the following steps: Step S41: construct a multi-factor neural network and a feedforward neural network model; the hidden layer uses a ReLU activation function and the output layer uses a Sigmoid activation function; Step S42: define the loss function, the formula is as follows: ； Among them, J(t) represents the loss value at time t, lor represents the loss weight, Q(t) represents the actual load value at time t, represents the predicted load value at time t, n represents the sample index, N represents the total number of samples, and lop represents the loss bias; Step S43: Initialize model parameters; Step S44: Design a gradient moment estimation function, the formula is as follows: ； ； Where t represents time, g(t) represents the gradient value at time t, represents the square of the gradient norm, a is the step size, m t represents the first-order moment of the gradient at time t, m t-a represents the first-order moment of the gradient at time ta, v t represents the second-order moment of the gradient at time t, v t-a represents the second-order moment of the gradient at time ta, and represents the attenuation coefficient; Step S45: Design the matrix offset correction term, the formula is as follows: ； ； in, and Respectively represent m t and v t Matrix offset correction; Step S46: Design a parameter update method based on step length, the formula is as follows: ； Among them, θ t represents the parameter at time t, θ t-a represents the parameter at time ta, ε represents a small positive number; Step S47: Iterative update: pre-set the threshold of the loss function, calculate the loss function value under the current model parameters after each iterative update, if the current loss function value is less than the threshold, stop the iteration and output the current model parameters as the optimal model parameters; if the current loss function value is greater than the threshold, continue to iterate until the maximum number of iterations is reached, if the parameters less than the threshold are still not obtained, reinitialize the model parameters and continue iterative update; Step S48: Design an optimal initial model parameter search method, the formula is as follows: ； Among them, Iter represents the number of iterations, θ * (Iter+1) represents the initial model parameters searched for the Iter+1th iteration, ω is the search weight, best(θ * ) represents the global optimal initial model parameters currently searched, ibest(θ * ) represents the local optimal initial model parameters currently searched, rand(0,1) represents a random number between 0 and 1, θ * (Iter) represents the initial model parameters searched at the Iter-th iteration; Step S49: Search for the optimal initial model parameters, set the model performance as the fitness value, pre-set the fitness threshold, iteratively search the initial model parameters, calculate the fitness value after each iteration, and if there is an initial model parameter whose corresponding fitness value is greater than the threshold, stop the iteration and set the initial model parameter at this position as the optimal initial model parameter; if the maximum number of iterations is reached, re-initialize the parameter position for searching; otherwise, continue the iterative search; In step S2, the generative adversarial network specifically includes the following steps: Step S21: constructing a generator and a discriminator model. The constructed generator receives a random noise vector and a conditional vector as input, and converts them into a false sample similar to a real sample through a neural network. The discriminator receives a real sample or a false sample generated by the generator and a target label as input. The GAN model optimizes the entire model by alternately training the generator and the discriminator, minimizing the loss function of the generator to make the generated sample closer to the real sample, and maximizing the loss function of the discriminator to improve its accuracy. Step S22: Design a consistency regularization adding strategy, the formula is as follows: ； ； Among them, L G and L D Represent the consistency regularization additions of the generator and the discriminator, respectively. , and represents the weight factor, represents the enhanced least squares loss, represents the enhanced real data loss, represents the enhanced simulation data loss, represents enhanced mixing loss; Step S23: Design a consistency regularization term, the formula is as follows: ； Among them, L represents the consistency regularization addition term, represents the gradient weight, represents the balanced gradient, D(·) represents the discriminator function, T(x) represents the data enhancement function, ||·|| 2 represents the square of the norm; Step S24: Design the loss function of the generator, the formula is as follows: ； Among them, Loss G represents the loss function of the generator; E[·] represents the mean, G(·) represents the generator function, s represents the random noise vector of the generator input, c represents the conditional vector of the generator input, represents the target label, λ represents the weight factor, represents the characteristic distribution of generated samples, represents the mean of the generated samples, represents the standard deviation of the generated samples, N(0,1) represents the standard normal distribution, D KL (·) represents KL divergence, log represents logarithm, and M(·) represents least squares loss; Step S25: Design the loss function of the discriminator, the formula is as follows: ； Among them, Loss D represents the loss function of the discriminator; r represents the real sample; Step S26: iterative training, generating a batch of false samples through the generator, and taking the real samples and the false samples as input, calculating the loss function of the discriminator, and updating the parameters of the discriminator through back propagation, then generating a batch of false samples through the generator, and taking the false samples as input, calculating the loss function of the generator, and updating the parameters of the generator through back propagation, repeating the above steps, alternately training the generator and the discriminator until a predetermined number of training times is reached or the loss function converges; In step S3, the enrichment of raw data specifically includes the following steps: Step S31: Generate simulation data. Use the trained generator to input the prepared input vector into the generator model to generate simulation data. The generator will generate simulation data similar to the real data based on the input vector and the learned parameters. Step S32: after generating simulation data, remove abnormal values ​​and select high-quality data that meets the conditions; Step S33: enriching the original data, adding the processed simulation data to the original data; In step S1, the data collection and processing specifically includes the following steps: Step S11: Collect external environment data, including weather data, seasonal data, and holiday data; operation status data, including operation time period data, operation mode data, and operation plan data; equipment status, including equipment operation status data, equipment fault information data, equipment energy consumption, and operation efficiency; work status data, including work plan data and personnel quantity data; set these data as features; collect historical load data of the industrial park; set the load data as a label; Step S12: Data preprocessing, cleaning the collected data, filling missing data, detecting and removing abnormal data; screening, selecting data with key features and removing noise data; data conversion, multi-type data aggregation, adjusting data of different scales to the same range; In step S5, the specific operation is to collect the load data of the industrial park, and predict the load status of the industrial park based on the constructed multi-type load prediction model, that is, by collecting the load data of the industrial park in real time, inputting it into the established multi-type load prediction model, predicting the load of the industrial park, and setting a threshold to monitor whether it exceeds the preset value. If the predicted load does not exceed the threshold, the system remains in a normal state. If the predicted load exceeds the threshold, the system enters an alarm state and issues an alarm message. According to various load demands in the park, the system is optimized and scheduled to achieve coordinated control of multiple types of loads.

2. A multi-type load collaborative control system for industrial parks, used to implement a multi-type load collaborative control method for industrial parks as claimed in claim 1, characterized in that: It includes data collection and processing module, generative adversarial network module, raw data enrichment module, multi-type load forecasting model construction module and specific operation module; The data acquisition and processing module collects historical load data, external environment data, operation status, equipment status, and working status data of the industrial park and performs cleaning, screening, data conversion, and multi-type data aggregation, and sends the processed data to the generative adversarial network module and the enriched raw data module; The generative adversarial network module receives the data sent by the data acquisition and processing module, generates an adversarial network by constructing generator and discriminator models, designing consistency regularization adding strategies and regularization terms, designing loss functions of the generator and the discriminator, and sends the data to the enriched original data module; The enriched raw data module receives data sent by the generated adversarial network module and the data acquisition and processing module, generates simulation data through the adversarial network, enriches the raw data after removing abnormalities, and sends the data to the module for building a multi-type load prediction model; The multi-type load forecasting model module is constructed by constructing a multi-factor neural network, defining a loss function, initializing model parameters, designing a gradient moment estimation function, designing a matrix offset correction term, designing a step-based parameter update method, designing an optimal initial model parameter search method, iteratively updating and constructing an optimal parameter model, and sending data to a specific operation module; The specific operation module receives data sent by the module for constructing a multi-type load prediction model, performs load prediction based on the historical load information of the industrial park, and optimizes and schedules the system based on the prediction results to achieve coordinated control of multi-type loads.

Images