Multi-objective optimization scheduling method and system for operation of power distribution network

Abstract

The invention discloses a power distribution network operation multi-target optimization scheduling method and system, and solves the problem that the existing method cannot decompose and predict high-frequency and low-frequency sequence trends of target control parameters such as wind and light power generation from the perspective of high-frequency fluctuation and low-frequency trends of the control parameters. The method comprises the following steps: acquiring target point load and wind and light power generation output parameters in the power distribution network; summing the intra-day low-frequency prediction sequence and the intra-day high-frequency prediction sequence of the target point, generating an intra-day demand response curve of the target point by the bidirectional long-short-term memory network model, and outputting a multi-target optimization scheduling scheme based on the demand response curve and the power distribution cost prediction sequence; according to the method, the power distribution cost sequence parameters are decomposed by adopting the variational mode decomposition method, so that the high-frequency cost sequence and the low-frequency cost sequence of each target point are quickly and efficiently obtained, the complexity of the load sequence is remarkably reduced, and accumulated errors during prediction are reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of distribution network optimization, and in particular relates to a multi-objective optimization scheduling method and system for distribution network operation. Background Art

[0002] The distribution network is an important part of the power system. It is mainly responsible for receiving electric energy from the transmission network or regional power plants, and distributing it locally or step by step to various users according to voltage levels through a series of distribution facilities. The distribution network optimization and dispatching refers to the reasonable adjustment and optimization of the distribution network's operating mode, equipment configuration and load distribution on the premise of ensuring the safe operation of the power grid and meeting power supply needs, so as to improve the distribution network's operating efficiency, reduce operating costs, reduce energy losses, and improve power supply reliability and power quality.

[0003] Chinese patent CN111884268B discloses an optimal power flow control method for an AC / DC hybrid distribution network taking reliability indicators into consideration, including rapid power flow calculation, optimization modeling, and optimal power flow control. Only commonly used control parameters are needed to solve the optimal power flow problem. However, in the existing method, when performing multi-objective optimization scheduling of a distribution network, the optimal power flow operation is calculated through control parameter calculation, and the high-frequency and low-frequency sequence trends of target control parameters such as wind and solar power generation cannot be decomposed and predicted from the perspective of high-frequency fluctuations and low-frequency trends of control parameters, resulting in uncertainty in the target output of wind and solar power generation during the optimization scheduling of the distribution network, and thus it is impossible to provide a variety of effective solutions with economic costs and network losses as conflicting targets. To address the above problems, we propose a multi-objective optimization scheduling method and system for distribution network operation. Summary of the invention

[0004] The purpose of the present invention is to address the deficiencies in the prior art and provide a multi-objective optimization scheduling method and system for distribution network operation, which solves the problem that the prior methods are unable to decompose and predict the high-frequency and low-frequency sequence trends of target control parameters such as wind and solar power generation from the perspective of high-frequency fluctuations and low-frequency trends of control parameters, resulting in uncertainty in the target output of wind and solar power generation during distribution network optimization scheduling.

[0005] The present invention is implemented in this way: a multi-objective optimization scheduling method for distribution network operation, the multi-objective optimization scheduling method for distribution network operation comprises:

[0006] Obtain the 24-hour load and wind and solar power output parameters of the target point in the distribution network seven days before the previous day. The target point is a controllable generator set, wind turbine, wind power, photovoltaic, energy storage, and load node. Pick up the distribution cost sequence parameters of the target point load and wind and solar power output parameters;

[0007] The variational mode decomposition method is used to decompose the distribution cost sequence parameters to obtain the high-frequency cost sequence and low-frequency cost sequence of each target point;

[0008] Obtain the low-frequency cost sequence of the target point, and perform cost prediction on the low-frequency cost sequence based on the particle swarm algorithm combined with the bidirectional long short-term memory network model to obtain the intraday low-frequency prediction sequence of the target point;

[0009] Combining big data high-dimensional analysis and convolutional neural network models to predict the high-frequency cost sequence, we can obtain the high-frequency prediction sequence of the target point;

[0010] The intraday low-frequency prediction sequence and the intraday high-frequency prediction sequence of the target point are summed up to obtain the distribution cost prediction sequence of multiple target points;

[0011] Pick up the load node demand response parameters, and predict the target point prediction sequence based on the particle swarm algorithm combined with the bidirectional long short-term memory network model. Take the target point power consumption prediction sequence as input, and the bidirectional long short-term memory network model generates the target point intraday demand response curve;

[0012] Load the demand response curve and distribution cost prediction sequence, define the minimum demand response cost and distribution network loss as the goals, build a multi-objective optimization scheduling model, and output a multi-objective optimization scheduling plan based on the demand response curve and distribution cost prediction sequence.

[0013] Preferably, the method of decomposing the distribution cost sequence parameters using the variational mode decomposition method specifically includes:

[0014] Load the distribution cost sequence parameters, which are expressed as x B =x Peak +x Valley , x Peak is the high-frequency cost parameter, x Valley is the low-frequency cost parameter;

[0015] The distribution cost sequence parameters are subjected to discrete wavelet transformation. The high-pass filter and low-pass filter of the discrete wavelet transformation decompose the distribution cost sequence parameters into subsequences representing the details of high-frequency costs and subsequences representing low-frequency costs. The discrete wavelet transformation is defined as:

[0016]

[0017] Among them, j is the discrete wavelet transform decomposition level, ψ(x) is the defined wavelet function, y(x) is the position index corresponding to the distribution cost sequence parameter, and t is the dispatch time corresponding to the distribution cost sequence parameter;

[0018]

[0019] Among them, x τ Peak is a subsequence representing the details of the high-frequency cost, x τValley is the subsequence representing the low-frequency cost, γ Peak is the wavelet transform coefficient of the high-frequency cost, γ Valley is the wavelet transform coefficient of the low-frequency cost;

[0020] The variational mode decomposition method introduces Lagrangian multipliers and quadratic penalty terms to solve the high-frequency cost sequence and low-frequency cost sequence of the target point;

[0021]

[0022] Among them, a is the quadratic penalty term, δ(t) is the Dirac trigonometric function, t is the dispatch time corresponding to the distribution cost sequence parameter, j is the discrete wavelet transform decomposition level, and λ(x) is the Lagrangian multiplier;

[0023]

[0024] Among them, λ1,λ2 are Lagrangian regularization parameters.

[0025] Preferably, the bidirectional long short-term memory network model is composed of a model encoder and a model decoder, wherein the model encoder includes a low-frequency cost input layer, a speed and position update mechanism, and a residual connection network, and the model decoder includes an embedding layer, a Bi-LSTM layer, a normalization processing layer, and an output layer;

[0026] The convolutional neural network model includes a sequence dimension reducer, a feature extractor, a feature analysis layer, and a temporal convolutional network;

[0027] The feature extractor uses the pre-trained YOLOV8 model as the basic architecture and freezes the Neck layer of the YOLOV8 model as the feature extractor;

[0028] The temporal convolutional network is a temporal convolutional network architecture that incorporates a multi-head attention mechanism. The temporal convolutional network architecture combines a softened threshold algorithm with a high-frequency cost sequence to iteratively train the convolutional neural network model.

[0029] Preferably, the cost prediction of the low-frequency cost sequence based on the particle swarm algorithm combined with the bidirectional long short-term memory network model specifically includes:

[0030] Obtain a low-frequency cost sequence. The low-frequency cost input layer performs position encoding and labeling on the low-frequency cost sequence to obtain a low-frequency encoding set.

[0031] Load the low-frequency code set. The speed and position update mechanism updates the low-frequency code set based on the particle swarm optimization algorithm to obtain the low-frequency update set. The low-frequency update set is calculated by the following formula:

[0032] GValley (t+1)=ω Valley *G Valley (t)+c1*r1(pbest Valley -B Valley (t))+c2*r2(gbest Valley -B Valley (t)) (5)

[0034] Among them, ω Valley is the inertia weight value, c1, c2 are the first learning factor and the second learning factor respectively, r1, r2 are random constants, pbest Valley is the target point extreme value of the low-frequency coding set, gbest Valley is the global extreme value of the low-frequency coding set, B Valley (t) is the low-frequency code set;

[0035] The residual connection network performs residual connection on the low-frequency update set based on the residual block principle to obtain a residual connection set;

[0036] The Bi-LSTM layer extracts features from the residual connection set in multiple directions to obtain a feature extraction set, calculates the probability of particle swarm clustering on the feature extraction set, and obtains the prediction probability value corresponding to the feature extraction set;

[0037] The normalization processing layer calculates the intraday low-frequency prediction value corresponding to the prediction probability value based on the softmax classification function, integrates multiple groups of intraday low-frequency prediction values, and obtains the intraday low-frequency prediction sequence;

[0038] Among them, the calculation function of the intraday low-frequency forecast value is:

[0039]

[0040] Among them, g Valley is the predicted probability value, C Valley is the feature extraction value corresponding to the feature extraction set, α is the classification coupling coefficient of the softmax classification function, and i is the number of features in the feature extraction set.

[0041] Preferably, the cost prediction of high-frequency cost series by combining big data high-dimensional analysis and convolutional neural network model specifically includes:

[0042] Obtain a high-frequency cost sequence, and use a sequence dimension reducer to reduce the dimension of the high-frequency cost sequence to obtain a reduced-dimensional sequence set;

[0043] The feature extractor performs feature enhancement on the reduced-dimensionality sequence set based on the attention enhancement mechanism to obtain an attention enhancement set;

[0044] The attention enhancement set is expressed as:

[0045]

[0046] in, is the attention enhancement set, C Peak is the input dimensionality reduction sequence set, β is the enhancement coefficient of the sigmoid function, Z Peak (t) is the gain energy value of the reduced-dimensionality sequence set, and ε is the reduced-dimensionality of the high-frequency cost sequence;

[0047]

[0048] Among them, M is the number of attention enhancement channels of the attention enhancement mechanism SimAM attention mechanism, m is the scaling factor of the input dimensionality reduction sequence set, is the variance of the reduced-dimensionality sequence set.

[0049] Preferably, the cost prediction of high-frequency cost series by combining big data high-dimensional analysis and convolutional neural network model specifically includes:

[0050] The attention enhancement set is input into the feature analysis layer. The feature analysis layer performs feature analysis on the attention enhancement set to obtain a feature analysis set. The temporal convolutional network combines the softened threshold algorithm to predict the feature analysis set to obtain a high-frequency prediction sequence of the target point.

[0051] The intraday low-frequency prediction sequence and the intraday high-frequency prediction sequence of the target point are summed up to obtain the distribution cost prediction sequence of multiple target points;

[0052] Among them, when the intraday low-frequency prediction sequence and the intraday high-frequency prediction sequence of the target point are summed up, the distribution cost prediction value is calculated based on the multi-objective solution method, and the distribution cost prediction value is integrated to obtain the distribution cost prediction sequence;

[0053]

[0054] Among them, f B is the predicted value of power distribution cost, d is the number of target points, f Peak is the intraday high-frequency forecast value, f Valley is the intraday low-frequency prediction value, κ is the hyperparameter of the multi-objective solution method, and t is the dispatch time corresponding to the distribution cost prediction value.

[0055] Preferably, the method for constructing the multi-objective optimization scheduling model specifically includes:

[0056] Obtain demand response curves and distribution cost prediction sequences, use the demand response curves and distribution cost prediction sequences as training samples and verification samples, set the loss function of the multi-objective optimization scheduling model to the cross entropy damage function, and fix the hyperparameters of the multi-objective optimization scheduling model;

[0057] The particle swarm algorithm model is used as the initial model of the multi-objective optimization scheduling model. The particle swarm algorithm model is iteratively trained, and the hyperparameters are updated using the Adam optimization algorithm until the loss function value converges. The training is stopped to obtain a trained multi-objective optimization scheduling model.

[0058] Use the test set to test the trained multi-objective optimization scheduling model to determine whether the optimization scheduling result of the test set meets the preset scheduling threshold;

[0059] If the preset scheduling threshold is met, the trained multi-objective optimization scheduling model is output;

[0060] If the preset scheduling threshold is not met, the Adam optimization algorithm is used to initialize the hyperparameters and learning rate until the loss curve converges.

[0061] On the other hand, the present invention also provides a distribution network operation multi-objective optimization scheduling system, the distribution network operation multi-objective optimization scheduling system specifically includes:

[0062] The parameter acquisition module is used to obtain the 24-hour load and wind and solar power generation output parameters of the target point in the distribution network seven days before the previous day;

[0063] The parameter decomposition module uses the variational mode decomposition method to decompose the distribution cost sequence parameters to obtain the high-frequency cost sequence and low-frequency cost sequence of each target point;

[0064] The cost prediction module is used to obtain the low-frequency cost sequence of the target point, and perform cost prediction on the low-frequency cost sequence based on the particle swarm algorithm combined with the bidirectional long short-term memory network model to obtain the intraday low-frequency prediction sequence of the target point, and perform cost prediction on the high-frequency cost sequence in combination with the big data high-dimensional analysis and the convolutional neural network model to obtain the high-frequency prediction sequence of the target point. The intraday low-frequency prediction sequence and the intraday high-frequency prediction sequence of the target point are summed and calculated to obtain the distribution cost prediction sequence of multiple target points;

[0065] The response curve generation module is used to pick up the load node demand response parameters, and predict the target point prediction sequence based on the particle swarm algorithm combined with the bidirectional long short-term memory network model. The target point power consumption prediction sequence is used as input, and the bidirectional long short-term memory network model generates the target point intraday demand response curve;

[0066] The scheduling scheme output module is used to load the demand response curve and the distribution cost prediction sequence, define the minimum demand response cost and the distribution network loss as the goal, build a multi-objective optimization scheduling model, and output the multi-objective optimization scheduling scheme based on the demand response curve and the distribution cost prediction sequence.

[0067] Preferably, the parameter decomposition module comprises:

[0068] A wavelet conversion unit is used to load the distribution cost sequence parameters, perform discrete wavelet conversion on the distribution cost sequence parameters, and the high-pass filter and low-pass filter of the discrete wavelet conversion decompose the distribution cost sequence parameters into a subsequence representing the details of the high-frequency cost and a subsequence representing the low-frequency cost;

[0069] The cost sequence generation unit introduces Lagrangian multipliers and quadratic penalty terms based on the variational mode decomposition method to solve the high-frequency cost sequence and low-frequency cost sequence of the target point.

[0070] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0071] The embodiment of the present invention adopts the variational mode decomposition method to decompose the distribution cost sequence parameters. The decomposition process can effectively remove noise, improve data quality, and thus improve prediction accuracy, so as to quickly and efficiently obtain the high-frequency cost sequence and low-frequency cost sequence of each target point, significantly reduce the complexity of the load sequence, and reduce the cumulative error during prediction.

[0072] In an embodiment of the present invention, the distribution cost series parameters are converted and decomposed by discrete wavelet transform combined with variational mode decomposition method, so as to improve the accuracy of parameter processing, effectively suppress the phenomenon of parameter mode aliasing, and simultaneously enhance the robustness of the bidirectional long short-term memory network model and the convolutional neural network model in extracting low-frequency and high-frequency features. Lagrange multipliers and quadratic penalty terms are introduced to transform the constrained variational problem into an unconstrained variational problem, thereby reducing the computational complexity of the bidirectional long short-term memory network model and the convolutional neural network model.

[0073] In an embodiment of the present invention, a particle swarm algorithm combined with a bidirectional long short-term memory network model is provided. The bidirectional long short-term memory network model is composed of a model encoder and a model decoder. The residual connection network in the model encoder performs residual connection on the low-frequency update set through the residual block principle, which can alleviate the problem of vanishing low-frequency feature gradients and enhance low-frequency feature propagation, thereby reducing information loss and improving the expression ability of the residual connection network, so that the residual connection network can more effectively express and process low-frequency features.

[0074] In an embodiment of the present invention, a convolutional neural network model combined with high-dimensional analysis of big data is provided. The convolutional neural network model is composed of a sequence dimension reducer, a feature extractor, a feature analysis layer, and a temporal convolutional network. The convolutional neural network model can predict the feature analysis set in combination with a softened threshold algorithm to obtain a high-frequency prediction sequence of the target point, and the convolutional neural network model is suitable for feature extraction and cost prediction of high-frequency cost sequences. The convolutional neural network model and the bidirectional long short-term memory network model cooperate to process and predict high-frequency and low-frequency cost sequences respectively, strengthen probabilistic prediction, so as to reflect the uncertainty of load changes, and give full play to their respective advantages to improve prediction accuracy and speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 It is a schematic diagram of the implementation flow of the multi-objective optimization scheduling method for distribution network operation provided by the present invention.

[0076] Figure 2 The figure shows a flow chart of a method for decomposing distribution cost sequence parameters using variational mode decomposition method.

[0077] Figure 3 The figure shows a schematic diagram of the implementation process of cost prediction for low-frequency cost sequences based on the particle swarm algorithm combined with the bidirectional long short-term memory network model.

[0078] Figure 4 The figure shows a flowchart of the cost prediction method for high-frequency cost sequences by combining big data high-dimensional analysis and convolutional neural network models.

[0079] Figure 5 A schematic diagram of the implementation process of the construction method of the multi-objective optimization scheduling model is shown.

[0080] Figure 6 It is a schematic diagram of the architecture of the multi-objective optimization dispatching system for distribution network operation provided by the present invention. DETAILED DESCRIPTION

[0081] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of this application; the terms used in the specification of the application herein are only for the purpose of describing specific embodiments and are not intended to limit this application; the terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0082] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0083] Existing methods are unable to decompose and predict the high-frequency and low-frequency sequence trends of target control parameters such as wind and solar power generation from the perspective of high-frequency fluctuations and low-frequency trends of control parameters, resulting in uncertainty in target outputs such as wind and solar power generation during distribution network optimization scheduling. To address the above issues, we propose a multi-objective optimization scheduling method and system for distribution network operation. In short, the method includes obtaining the 24-hour load and wind and solar power generation output parameters of the target point in the distribution network seven days before, and then using the variational mode decomposition method to decompose the distribution cost sequence parameters to obtain the high-frequency cost sequence and low-frequency cost sequence of each target point, and using the variational mode decomposition method to optimize the distribution. The parameters of the electricity cost sequence are decomposed, and the low-frequency cost sequence is predicted based on the particle swarm algorithm combined with the bidirectional long short-term memory network model. The high-frequency cost sequence is predicted by combining the big data high-dimensional analysis and the convolutional neural network model. The intraday low-frequency prediction sequence and the intraday high-frequency prediction sequence of the target point are summed up to obtain the distribution cost prediction sequence of multiple target points. Then, the demand response curve and the distribution cost prediction sequence are loaded, and the minimum demand response cost and the distribution network loss are defined as the goals. A multi-objective optimization scheduling model is constructed, and a multi-objective optimization scheduling plan is output based on the demand response curve and the distribution cost prediction sequence. The embodiment of the present invention adopts the variational mode decomposition method to decompose the distribution cost sequence parameters. The decomposition process can effectively remove noise, improve data quality, and thus improve prediction accuracy, so as to quickly and efficiently obtain high-frequency cost sequences and low-frequency cost sequences of each target point, significantly reduce the complexity of the load sequence, and reduce the cumulative error during prediction. It overcomes the problem that in the existing method of multi-objective optimization scheduling of the distribution network, the optimal power flow operation is calculated through control parameters, and the high-frequency and low-frequency sequence trends of target control parameters such as wind and solar power generation cannot be decomposed and predicted from the perspective of high-frequency fluctuations and low-frequency trends of control parameters, resulting in uncertainty in the target output of wind and solar power generation during the optimization scheduling of the distribution network, and thus it is impossible to provide a variety of effective solutions with economic costs and network losses as conflicting goals.

[0084] The embodiment of the present invention provides a multi-objective optimization scheduling method for distribution network operation. Figure 1 The schematic diagram of the implementation process of the multi-objective optimization scheduling method for distribution network operation is shown, and the multi-objective optimization scheduling method for distribution network operation specifically includes:

[0085] Step S10, obtaining the 24-hour load and wind and solar power generation output parameters of the target point in the distribution network seven days before, and picking up the distribution cost sequence parameters of the load and wind and solar power generation output parameters of the target point;

[0086] It should be noted that the target points include but are not limited to controllable generator sets, wind turbines, wind power, photovoltaics, energy storage, and load nodes, and the wind and solar power generation output parameters include but are not limited to the three-phase active output parameters, reactive output parameters, start and stop cost parameters, abandoned wind and power abandonment cost data, shiftable load scheduling data, and distribution network interaction parameters of controllable generator sets, wind turbines, wind power, photovoltaics, energy storage, and load nodes, and the target point load parameters include but are not limited to the target point's power generation, load demand, frequency, voltage, current, temperature, switch start and stop, power trading data, gas phase data related to power generation and distribution, calendar data, etc.

[0087] Step S20, using variational mode decomposition method to decompose the distribution cost sequence parameters to obtain high-frequency cost sequence and low-frequency cost sequence of each target point;

[0088] Step S30, obtaining a low-frequency cost sequence of the target point, and performing cost prediction on the low-frequency cost sequence based on a particle swarm algorithm combined with a bidirectional long short-term memory network model to obtain an intraday low-frequency prediction sequence of the target point;

[0089] Step S40, combining big data high-dimensional analysis and a convolutional neural network model to perform cost prediction on the high-frequency cost sequence to obtain a high-frequency prediction sequence of the target point;

[0090] Step S50, summing up the intraday low-frequency prediction sequence and the intraday high-frequency prediction sequence of the target point to obtain the distribution cost prediction sequence of multiple target points;

[0091] Step S60, picking up the load node demand response parameters, predicting the target point prediction sequence for the load node demand response parameters based on the particle swarm algorithm combined with the bidirectional long short-term memory network model, taking the target point power consumption prediction sequence as input, and using the bidirectional long short-term memory network model to generate the target point intraday demand response curve;

[0092] It should be noted that when generating the intraday demand response curve of the target point, the Bi-LSTM layer and the normalization processing layer of the bidirectional long short-term memory network model calculate the intraday demand response value, and then the intraday demand response value and the corresponding time are input into the output layer of the bidirectional long short-term memory network model to generate the intraday demand response curve of the corresponding target point, thereby providing a basis for the multi-objective optimization scheduling model to formulate the corresponding multi-objective optimization scheduling plan according to the prediction results.

[0093] Step S70, load the demand response curve and the distribution cost prediction sequence, define the minimum demand response cost and the distribution network loss as the goals, build a multi-objective optimization scheduling model, and output a multi-objective optimization scheduling plan based on the demand response curve and the distribution cost prediction sequence.

[0094] It should be noted that the multi-objective optimization scheduling model defines the minimum demand response cost and distribution network loss as the objectives, and the constraints can be energy storage operation condition constraints, system flow distribution constraints, and grid node voltage condition constraints.

[0095] The embodiment of the present invention adopts the variational mode decomposition method to decompose the distribution cost sequence parameters. The decomposition process can effectively remove noise, improve data quality, and thus improve prediction accuracy, so as to quickly and efficiently obtain the high-frequency cost sequence and low-frequency cost sequence of each target point, significantly reduce the complexity of the load sequence, and reduce the cumulative error during prediction.

[0096] The embodiment of the present invention provides a method for decomposing distribution cost sequence parameters using variational mode decomposition method. Figure 2 The figure shows a flow chart of a method for decomposing distribution cost sequence parameters using variational mode decomposition method. The method for decomposing distribution cost sequence parameters using variational mode decomposition method specifically includes:

[0097] Step S101, load the power distribution cost sequence parameter, the power distribution cost sequence parameter is expressed as x B =x Peak +x Valley , x Peak is the high-frequency cost parameter, x Valley is the low-frequency cost parameter;

[0098] It should be noted that when loading the distribution cost series parameters, when the cost series has consecutive missing periods or the number of missing values ​​reaches more than 30%, the group of cost series is directly removed. At the same time, for scattered missing values, correlation analysis and spline interpolation methods can be used for interpolation.

[0099] Step S102, discrete wavelet transform is performed on the power distribution cost sequence parameters. The high-pass filter and low-pass filter of the discrete wavelet transform decompose the power distribution cost sequence parameters into a subsequence representing the details of the high-frequency cost and a subsequence representing the low-frequency cost. The discrete wavelet transform is defined as:

[0100]

[0101] Wherein, j is the discrete wavelet transform decomposition level. In this embodiment, the transform decomposition level may be 3-5 levels of decomposition, ψ(x) is the defined wavelet function, y(x) is the position index corresponding to the distribution cost sequence parameter, and t is the scheduling time corresponding to the distribution cost sequence parameter;

[0102]

[0103] Among them, x τ Peak is a subsequence representing the details of the high-frequency cost, xτ Valley is the subsequence representing the low-frequency cost, γ Peak is the wavelet transform coefficient of the high-frequency cost, γ Valley is the wavelet conversion coefficient of the low-frequency cost. In this embodiment, the wavelet conversion coefficient of the high-frequency cost and the wavelet conversion coefficient of the low-frequency cost may be 0.1-0.2;

[0104] Step S103, the variational mode decomposition method introduces Lagrangian multipliers and quadratic penalty terms to solve the high-frequency cost sequence and low-frequency cost sequence of the target point;

[0105]

[0106] Wherein, a is a quadratic penalty term. In this embodiment, the parameter of the quadratic penalty term can be set to 3-5, δ(t) is a Dirac trigonometric function, t is the dispatch time corresponding to the distribution cost sequence parameter, j is the discrete wavelet transform decomposition level, and λ(x) is a Lagrangian multiplier;

[0107]

[0108] Wherein, λ1 and λ2 are Lagrangian regularization parameters. In this embodiment, the Lagrangian regularization parameters are set to 0.01-0.03.

[0109] In this embodiment, the limitations of the quadratic penalty term can be improved by introducing the Lagrangian multiplier, thereby solving the problems of numerical stability and computational efficiency that the quadratic penalty term may encounter when dealing with complex constraints or high-dimensional problems. In this embodiment, the combination of the Lagrangian multiplier and the quadratic penalty term can ensure that the quadratic penalty term has a larger penalty parameter, thereby flexibly handling different forms of constraints and not relying on the characteristics of the specific problem.

[0110] In an embodiment of the present invention, the distribution cost series parameters are converted and decomposed by discrete wavelet transform combined with variational mode decomposition method, so as to improve the accuracy of parameter processing, effectively suppress the phenomenon of parameter mode aliasing, and simultaneously enhance the robustness of the bidirectional long short-term memory network model and the convolutional neural network model in extracting low-frequency and high-frequency features. Lagrange multipliers and quadratic penalty terms are introduced to transform the constrained variational problem into an unconstrained variational problem, thereby reducing the computational complexity of the bidirectional long short-term memory network model and the convolutional neural network model.

[0111] The embodiment of the present invention provides a cost prediction method for low-frequency cost sequences based on a particle swarm algorithm combined with a bidirectional long short-term memory network model. Figure 3The figure shows a schematic diagram of the implementation process of cost prediction for low-frequency cost sequences based on a particle swarm algorithm combined with a bidirectional long short-term memory network model. The method of cost prediction for low-frequency cost sequences based on a particle swarm algorithm combined with a bidirectional long short-term memory network model specifically includes:

[0112] Step S201, obtaining a low-frequency cost sequence, and the low-frequency cost input layer performs position encoding and labeling on the low-frequency cost sequence to obtain a low-frequency code set;

[0113] It should be noted that when the low-frequency cost input layer performs position encoding and labeling on the low-frequency cost sequence, it can adopt absolute position encoding, relative position encoding or triangular position encoding.

[0114] Step S202, loading the low-frequency code set, the speed and position update mechanism updates the low-frequency code set based on the particle swarm optimization algorithm to obtain a low-frequency update set, and the low-frequency update set is calculated by the following formula:

[0115] G Valley (t+1)=ω Valley *G Valley (t)+c1*r1(pbest Valley -B Valley (t))+c2*r2(gbest Valley -B Valley (t)) (5)

[0117] Among them, ω Valley is the inertia weight value, c1 and c2 are the first learning factor and the second learning factor respectively. In this embodiment, the first learning factor and the second learning factor can take values ​​of 0.02-0.05, r1 and r2 are random constants, and pbest Valley is the target point extreme value of the low-frequency encoding set, gbest Valley is the global extreme value of the low-frequency coding set, B Valley (t) is the low-frequency code set;

[0118] Step S203, the residual connection network performs residual connection on the low-frequency update set based on the residual block principle to obtain a residual connection set;

[0119] In an embodiment of the present invention, residual connections are performed on low-frequency update sets through the residual block principle, which can alleviate the problem of vanishing low-frequency feature gradients and enhance low-frequency feature propagation, thereby reducing information loss and improving the expression ability of the residual connection network, so that the residual connection network can more effectively express and process low-frequency features.

[0120] Step S204, the Bi-LSTM layer performs feature extraction in multiple directions on the residual connection set to obtain a feature extraction set, calculates the particle swarm clustering probability of the feature extraction set, and obtains a prediction probability value corresponding to the feature extraction set;

[0121] Step S205, the normalization processing layer calculates the intraday low-frequency prediction value corresponding to the prediction probability value based on the softmax classification function, integrates multiple groups of intraday low-frequency prediction values, and obtains the intraday low-frequency prediction sequence;

[0122] Among them, the calculation function of the intraday low-frequency forecast value is:

[0123]

[0124] Among them, g Valley is the predicted probability value, C Valley is the feature extraction value corresponding to the feature extraction set, α is the classification coupling coefficient of the softmax classification function. In this embodiment, the classification coupling coefficient can be set to 0.6-0.75, and i is the number of features in the feature extraction set.

[0125] In this embodiment, the bidirectional long short-term memory network model is composed of a model encoder and a model decoder, wherein the model encoder includes a low-frequency cost input layer, a speed and position update mechanism, and a residual connection network, and the model decoder includes an embedding layer, a Bi-LSTM layer, a normalization processing layer and an output layer.

[0126] It should be noted that the low-frequency cost output layer is connected to the speed and position update mechanism, the speed and position update mechanism is connected to the residual connection network, the residual connection network is connected to the embedding layer, the embedding layer is connected to the Bi-LSTM layer, and the Bi-LSTM layer is connected to the normalization processing layer. The training and optimization methods of the bidirectional long short-term memory network model are existing technologies and are not limited here.

[0127] In an embodiment of the present invention, a particle swarm algorithm combined with a bidirectional long short-term memory network model is provided. The bidirectional long short-term memory network model is composed of a model encoder and a model decoder. The residual connection network in the model encoder performs residual connection on the low-frequency update set through the residual block principle, which can alleviate the problem of vanishing low-frequency feature gradients and enhance low-frequency feature propagation, thereby reducing information loss and improving the expression ability of the residual connection network, so that the residual connection network can more effectively express and process low-frequency features.

[0128] The embodiment of the present invention provides a cost prediction method for high-frequency cost sequences by combining big data high-dimensional analysis and convolutional neural network models. Figure 4The schematic diagram of the implementation process of the method for cost prediction of high-frequency cost series by combining big data high-dimensional analysis and convolutional neural network model is shown. The method for cost prediction of high-frequency cost series by combining big data high-dimensional analysis and convolutional neural network model specifically includes:

[0129] Step S301, obtaining a high-frequency cost sequence, and a sequence dimension reducer performs dimension reduction processing on the high-frequency cost sequence to obtain a dimension reduction sequence set;

[0130] In this embodiment, when the high-frequency cost sequence is subjected to dimensionality reduction processing, principal component analysis or local linear embedding method can be used to reduce the dimensionality of the data, thereby simplifying high-dimensional data into low-dimensional data, which helps to improve data processing efficiency and model performance.

[0131] Step S302, the feature extractor performs feature enhancement on the dimensionality reduction sequence set based on the attention enhancement mechanism to obtain an attention enhancement set;

[0132] The attention enhancement set is expressed as:

[0133]

[0134] in, is the attention enhancement set, C Peak is the input dimensionality reduction sequence set, β is the enhancement coefficient of the sigmoid function, the enhancement coefficient can be 0.1 or 0.2, Z Peak (t) is the gain energy value of the reduced-dimensionality sequence set, ε is the reduced-dimensionality of the high-frequency cost sequence, and in this embodiment, the reduced-dimensionality may be 2-4;

[0135]

[0136] Wherein, M is the number of attention enhancement channels of the attention enhancement mechanism SimAM attention mechanism, the number of attention enhancement channels can be 3-9, m is the scaling factor of the input dimensionality reduction sequence set, in this embodiment, the scaling factor can be 0.3-0.5, is the variance of the reduced-dimensionality sequence set.

[0137] Step S303, inputting the attention enhancement set into the feature analysis layer, the feature analysis layer performs feature analysis on the attention enhancement set to obtain a feature analysis set, and the temporal convolutional network combines with the softened threshold algorithm to predict the feature analysis set to obtain a high-frequency prediction sequence of the target point;

[0138] Step S304, summing up the intraday low-frequency prediction sequence and the intraday high-frequency prediction sequence of the target point to obtain the distribution cost prediction sequence of multiple target points;

[0139] Among them, when the intraday low-frequency prediction sequence and the intraday high-frequency prediction sequence of the target point are summed up, the distribution cost prediction value is calculated based on the multi-objective solution method, and the distribution cost prediction value is integrated to obtain the distribution cost prediction sequence;

[0140]

[0141] Among them, f B is the predicted value of power distribution cost, d is the number of target points, f Peak is the intraday high-frequency forecast value, f Valley is the intraday low-frequency prediction value, κ is the hyperparameter of the multi-objective solution method. In this embodiment, the hyperparameter of the multi-objective solution method is set to 0.002, and t is the scheduling time corresponding to the distribution cost prediction value.

[0142] In this embodiment, the convolutional neural network model includes a sequence dimension reducer, a feature extractor, a feature analysis layer, and a temporal convolutional network;

[0143] The feature extractor uses the pre-trained YOLOV8 model as the basic architecture and freezes the Neck layer of the YOLOV8 model as the feature extractor;

[0144] The temporal convolutional network is a temporal convolutional network architecture that incorporates a multi-head attention mechanism. The temporal convolutional network architecture combines a softened threshold algorithm with a high-frequency cost sequence to iteratively train the convolutional neural network model.

[0145] In an embodiment of the present invention, a convolutional neural network model combined with high-dimensional analysis of big data is provided. The convolutional neural network model is composed of a sequence dimension reducer, a feature extractor, a feature analysis layer, and a temporal convolutional network. The convolutional neural network model can predict the feature analysis set in combination with a softened threshold algorithm to obtain a high-frequency prediction sequence of the target point, and the convolutional neural network model is suitable for feature extraction and cost prediction of high-frequency cost sequences. The convolutional neural network model and the bidirectional long short-term memory network model cooperate to process and predict high-frequency and low-frequency cost sequences respectively, strengthen probabilistic prediction, so as to reflect the uncertainty of load changes, and give full play to their respective advantages to improve prediction accuracy and speed.

[0146] The embodiment of the present invention provides a method for constructing a multi-objective optimization scheduling model. Figure 5 The following is a schematic diagram of a method for constructing a multi-objective optimization scheduling model. The method for constructing a multi-objective optimization scheduling model specifically includes:

[0147] Step S401, obtaining a demand response curve and a distribution cost prediction sequence, using the demand response curve and the distribution cost prediction sequence as training samples and verification samples, setting the loss function of the multi-objective optimization scheduling model to a cross entropy damage function, and fixing the hyperparameters of the multi-objective optimization scheduling model;

[0148] In this embodiment, the allocation ratio of training samples to validation samples may be 4:1 or 7:2.

[0149] Step S402, using the particle swarm algorithm model as the initial model of the multi-objective optimization scheduling model, iteratively training the particle swarm algorithm model, using the Adam optimization algorithm to update the hyperparameters until the loss function value converges, stopping the training, and obtaining a trained multi-objective optimization scheduling model.

[0150] Step S403, using the test set to test the trained multi-objective optimization scheduling model, to determine whether the optimization scheduling result of the test set meets the preset scheduling threshold; it should be noted that the preset scheduling threshold can be set to 0.65-0.8.

[0151] Step S404: if the preset scheduling threshold is met, output the trained multi-objective optimization scheduling model;

[0152] If the preset scheduling threshold is not met, the Adam optimization algorithm is used to initialize the hyperparameters and learning rate, and the process returns to step S402 until the loss curve converges.

[0153] The embodiment of the present invention provides a multi-objective optimization dispatching system for distribution network operation. Figure 6 The schematic diagram of the architecture of the multi-objective optimization dispatching system for distribution network operation is shown, and the multi-objective optimization dispatching system for distribution network operation specifically includes:

[0154] The parameter acquisition module 100 is used to obtain the 24-hour load and wind and solar power generation output parameters of the target point in the distribution network seven days before;

[0155] The parameter decomposition module 200 decomposes the distribution cost sequence parameters using the variational mode decomposition method to obtain the high-frequency cost sequence and the low-frequency cost sequence of each target point;

[0156] The cost prediction module 300 is used to obtain the low-frequency cost sequence of the target point, perform cost prediction on the low-frequency cost sequence based on the particle swarm algorithm combined with the bidirectional long short-term memory network model, obtain the intraday low-frequency prediction sequence of the target point, and perform cost prediction on the high-frequency cost sequence in combination with the big data high-dimensional analysis and the convolutional neural network model to obtain the high-frequency prediction sequence of the target point, and sum the intraday low-frequency prediction sequence and the intraday high-frequency prediction sequence of the target point to obtain the distribution cost prediction sequence of multiple target points;

[0157] The response curve generation module 400 is used to pick up the load node demand response parameters, and predict the target point prediction sequence for the load node demand response parameters based on the particle swarm algorithm combined with the bidirectional long short-term memory network model. The target point power consumption prediction sequence is used as input, and the bidirectional long short-term memory network model generates the target point intraday demand response curve;

[0158] The scheduling scheme output module 500 is used to load the demand response curve and the distribution cost prediction sequence, define the minimum demand response cost and the distribution network loss as the target, build a multi-objective optimization scheduling model, and output the multi-objective optimization scheduling scheme based on the demand response curve and the distribution cost prediction sequence.

[0159] In this embodiment, the parameter decomposition module 200 includes:

[0160] The wavelet conversion unit 210 is used to load the power distribution cost sequence parameters, perform discrete wavelet conversion on the power distribution cost sequence parameters, and the high-pass filter and low-pass filter of the discrete wavelet conversion decompose the power distribution cost sequence parameters into subsequences representing the details of high-frequency high-frequency costs and subsequences representing low-frequency low-frequency costs;

[0161] The cost sequence generating unit 220 introduces Lagrangian multipliers and quadratic penalty terms based on the variational mode decomposition method to solve the high-frequency cost sequence and the low-frequency cost sequence of the target point.

[0162] It can be understood that the distribution network operation multi-objective optimization scheduling system provided in the embodiment of the present invention corresponds to the above-mentioned distribution network operation multi-objective optimization scheduling method, and the explanations, examples, beneficial effects and other parts of its relevant contents can refer to the corresponding contents in the distribution network operation multi-objective optimization scheduling method, which will not be repeated here.

[0163] In summary, the present invention provides a multi-objective optimization scheduling method and system for distribution network operation. The embodiment of the present invention adopts the variational mode decomposition method to decompose the distribution cost sequence parameters. The decomposition process can effectively remove noise and improve data quality, thereby improving prediction accuracy, so as to quickly and efficiently obtain the high-frequency cost sequence and low-frequency cost sequence of each target point, significantly reduce the complexity of the load sequence, and reduce the cumulative error during prediction.

[0164] It should be noted that, for the above-mentioned embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should know that the present invention is not limited by the described order of actions, because according to the present invention, some steps may be performed in other orders or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.

[0165] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on these embodiments, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in this field can still combine, add, delete or make other adjustments to the features in the various embodiments of the present invention according to the circumstances without conflict, without making creative work, so as to obtain different other technical solutions that do not deviate from the concept of the present invention in essence, and these technical solutions also belong to the scope of protection of the present invention.

Claims

1. A multi-objective optimization scheduling method for distribution network operation, characterized in that: The multi-objective optimization scheduling method for distribution network operation includes: Obtain the 24-hour load and wind and solar power output parameters of the target point in the distribution network seven days before the previous day. The target point is a controllable generator set, wind turbine, wind power, photovoltaic, energy storage, and load node. Pick up the distribution cost sequence parameters of the target point load and wind and solar power output parameters; The variational mode decomposition method is used to decompose the distribution cost sequence parameters to obtain the high-frequency cost sequence and low-frequency cost sequence of each target point; Obtain the low-frequency cost sequence of the target point, and perform cost prediction on the low-frequency cost sequence based on the particle swarm algorithm combined with the bidirectional long short-term memory network model to obtain the intraday low-frequency prediction sequence of the target point; Combining big data high-dimensional analysis and convolutional neural network models to predict the high-frequency cost sequence, we can obtain the high-frequency prediction sequence of the target point; The intraday low-frequency prediction sequence and the intraday high-frequency prediction sequence of the target point are summed up to obtain the distribution cost prediction sequence of multiple target points; Pick up the load node demand response parameters, and predict the target point prediction sequence based on the particle swarm algorithm combined with the bidirectional long short-term memory network model. Take the target point power consumption prediction sequence as input, and the bidirectional long short-term memory network model generates the target point intraday demand response curve; Load the demand response curve and distribution cost prediction sequence, define the minimum demand response cost and distribution network loss as the goals, build a multi-objective optimization scheduling model, and output a multi-objective optimization scheduling plan based on the demand response curve and distribution cost prediction sequence.

2. The multi-objective optimization scheduling method for distribution network operation according to claim 1, characterized in that: The method of decomposing the distribution cost sequence parameters using the variational mode decomposition method specifically includes: Load the distribution cost sequence parameters, which are expressed as x B =x Peak +x Valley , x Peak is the high-frequency cost parameter, x Valley is the low-frequency cost parameter; The distribution cost sequence parameters are subjected to discrete wavelet transformation. The high-pass filter and low-pass filter of the discrete wavelet transformation decompose the distribution cost sequence parameters into subsequences representing the details of high-frequency costs and subsequences representing low-frequency costs. The discrete wavelet transformation is defined as: Among them, j is the discrete wavelet transform decomposition level, ψ(x) is the defined wavelet function, y(x) is the position index corresponding to the distribution cost sequence parameter, and t is the dispatch time corresponding to the distribution cost sequence parameter; Among them, x τ Peak is a subsequence representing the details of the high-frequency cost, x τ Valley is the subsequence representing the low-frequency cost, γ Peak is the wavelet transform coefficient of the high-frequency cost, γ Valley is the wavelet transform coefficient of the low-frequency cost; The variational mode decomposition method introduces Lagrangian multipliers and quadratic penalty terms to solve the high-frequency cost sequence and low-frequency cost sequence of the target point; Among them, a is the quadratic penalty term, δ(t) is the Dirac trigonometric function, t is the dispatch time corresponding to the distribution cost sequence parameter, j is the discrete wavelet transform decomposition level, and λ(x) is the Lagrangian multiplier; Among them, λ1,λ2 are Lagrangian regularization parameters.

3. The multi-objective optimization scheduling method for distribution network operation according to claim 1, characterized in that: The bidirectional long short-term memory network model is composed of a model encoder and a model decoder, wherein the model encoder includes a low-frequency cost input layer, a speed and position update mechanism, and a residual connection network, and the model decoder includes an embedding layer, a Bi-LSTM layer, a normalization processing layer, and an output layer; The convolutional neural network model includes a sequence dimension reducer, a feature extractor, a feature analysis layer, and a temporal convolutional network; The feature extractor uses the pre-trained YOLOV8 model as the basic architecture and freezes the Neck layer of the YOLOV8 model as the feature extractor; The temporal convolutional network is a temporal convolutional network architecture that incorporates a multi-head attention mechanism. The temporal convolutional network architecture combines a softened threshold algorithm with a high-frequency cost sequence to iteratively train the convolutional neural network model.

4. The multi-objective optimization scheduling method for distribution network operation according to claim 3, characterized in that: The method of predicting the cost of low-frequency cost series based on the particle swarm algorithm combined with the bidirectional long short-term memory network model specifically includes: Obtain a low-frequency cost sequence. The low-frequency cost input layer performs position encoding and labeling on the low-frequency cost sequence to obtain a low-frequency encoding set. Load the low-frequency code set. The speed and position update mechanism updates the low-frequency code set based on the particle swarm optimization algorithm to obtain the low-frequency update set. The low-frequency update set is calculated by the following formula: G Valley (t+1)=ω Valley *G Valley (t)+c1*r1(pbest Valley -B Valley (t))+c2*r2(gbest Valley -B Valley (t)) (5) Among them, ω Valley is the inertia weight value, c1, c2 are the first learning factor and the second learning factor respectively, r1, r2 are random constants, pbest Valley is the target point extreme value of the low-frequency coding set, gbest Valley is the global extreme value of the low-frequency coding set, B Valley (t) is the low-frequency code set; The residual connection network performs residual connection on the low-frequency update set based on the residual block principle to obtain a residual connection set; The Bi-LSTM layer extracts features from the residual connection set in multiple directions to obtain a feature extraction set, calculates the probability of particle swarm clustering on the feature extraction set, and obtains the prediction probability value corresponding to the feature extraction set; The normalization processing layer calculates the intraday low-frequency prediction value corresponding to the prediction probability value based on the softmax classification function, integrates multiple groups of intraday low-frequency prediction values, and obtains the intraday low-frequency prediction sequence; Among them, the calculation function of the intraday low-frequency forecast value is: Among them, g Valley is the predicted probability value, C Valley is the feature extraction value corresponding to the feature extraction set, α is the classification coupling coefficient of the softmax classification function, and i is the number of features in the feature extraction set.

5. The multi-objective optimization scheduling method for distribution network operation according to claim 3, characterized in that: The method of combining big data high-dimensional analysis and convolutional neural network model to predict the cost of high-frequency cost series specifically includes: Obtain a high-frequency cost sequence, and use a sequence dimension reducer to reduce the dimension of the high-frequency cost sequence to obtain a reduced-dimensional sequence set; The feature extractor performs feature enhancement on the reduced-dimensionality sequence set based on the attention enhancement mechanism to obtain an attention enhancement set; The attention enhancement set is expressed as: Among them, x Peak is the attention enhancement set, C Peak is the input dimensionality reduction sequence set, β is the enhancement coefficient of the sigmoid function, Z Peak (t) is the gain energy value of the reduced-dimensionality sequence set, and ε is the reduced-dimensionality of the high-frequency cost sequence; Among them, M is the number of attention enhancement channels of the attention enhancement mechanism SimAM, m is the scaling factor of the input dimensionality reduction sequence set, and C Peak is the variance of the reduced-dimensionality sequence set.

6. The multi-objective optimization scheduling method for distribution network operation according to claim 5, characterized in that: The method of combining big data high-dimensional analysis and convolutional neural network model to predict the cost of high-frequency cost series specifically includes: The attention enhancement set is input into the feature analysis layer. The feature analysis layer performs feature analysis on the attention enhancement set to obtain a feature analysis set. The temporal convolutional network combines the softened threshold algorithm to predict the feature analysis set to obtain a high-frequency prediction sequence of the target point. The intraday low-frequency prediction sequence and the intraday high-frequency prediction sequence of the target point are summed up to obtain the distribution cost prediction sequence of multiple target points; Among them, when the intraday low-frequency prediction sequence and the intraday high-frequency prediction sequence of the target point are summed up, the distribution cost prediction value is calculated based on the multi-objective solution method, and the distribution cost prediction value is integrated to obtain the distribution cost prediction sequence; Among them, f B is the predicted value of power distribution cost, d is the number of target points, f Peak is the intraday high-frequency forecast value, f Valley is the intraday low-frequency prediction value, κ is the hyperparameter of the multi-objective solution method, and t is the dispatch time corresponding to the distribution cost prediction value.

7. The multi-objective optimization scheduling method for distribution network operation according to claim 1, characterized in that: The method for constructing the multi-objective optimization scheduling model specifically includes: Obtain demand response curves and distribution cost prediction sequences, use the demand response curves and distribution cost prediction sequences as training samples and verification samples, set the loss function of the multi-objective optimization scheduling model to the cross entropy damage function, and fix the hyperparameters of the multi-objective optimization scheduling model; The particle swarm algorithm model is used as the initial model of the multi-objective optimization scheduling model. The particle swarm algorithm model is iteratively trained, and the hyperparameters are updated using the Adam optimization algorithm until the loss function value converges. The training is stopped to obtain a trained multi-objective optimization scheduling model.

8. The multi-objective optimization scheduling method for distribution network operation according to claim 7, characterized in that: The method for constructing the multi-objective optimization scheduling model specifically includes: Use the test set to test the trained multi-objective optimization scheduling model to determine whether the optimization scheduling result of the test set meets the preset scheduling threshold; If the preset scheduling threshold is met, the trained multi-objective optimization scheduling model is output; If the preset scheduling threshold is not met, the Adam optimization algorithm is used to initialize the hyperparameters and learning rate until the loss curve converges.

9. A distribution network operation multi-objective optimization dispatching system, used to implement the distribution network operation multi-objective optimization dispatching method according to any one of claims 1 to 8, characterized in that: The distribution network operation multi-objective optimization dispatching system specifically includes: The parameter acquisition module is used to obtain the 24-hour load and wind and solar power generation output parameters of the target point in the distribution network seven days before the previous day; The parameter decomposition module uses the variational mode decomposition method to decompose the distribution cost sequence parameters to obtain the high-frequency cost sequence and low-frequency cost sequence of each target point; The cost prediction module is used to obtain the low-frequency cost sequence of the target point, and perform cost prediction on the low-frequency cost sequence based on the particle swarm algorithm combined with the bidirectional long short-term memory network model to obtain the intraday low-frequency prediction sequence of the target point, and perform cost prediction on the high-frequency cost sequence in combination with the big data high-dimensional analysis and the convolutional neural network model to obtain the high-frequency prediction sequence of the target point. The intraday low-frequency prediction sequence and the intraday high-frequency prediction sequence of the target point are summed and calculated to obtain the distribution cost prediction sequence of multiple target points; The response curve generation module is used to pick up the load node demand response parameters, and predict the target point prediction sequence based on the particle swarm algorithm combined with the bidirectional long short-term memory network model. The target point power consumption prediction sequence is used as input, and the bidirectional long short-term memory network model generates the target point intraday demand response curve; The scheduling scheme output module is used to load the demand response curve and the distribution cost prediction sequence, define the minimum demand response cost and the distribution network loss as the goal, build a multi-objective optimization scheduling model, and output the multi-objective optimization scheduling scheme based on the demand response curve and the distribution cost prediction sequence.

10. The distribution network operation multi-objective optimization dispatching system according to claim 9, characterized in that: The parameter decomposition module comprises: A wavelet conversion unit is used to load the distribution cost sequence parameters, perform discrete wavelet conversion on the distribution cost sequence parameters, and the high-pass filter and low-pass filter of the discrete wavelet conversion decompose the distribution cost sequence parameters into a subsequence representing the details of the high-frequency cost and a subsequence representing the low-frequency cost; The cost sequence generation unit introduces Lagrangian multipliers and quadratic penalty terms based on the variational mode decomposition method to solve the high-frequency cost sequence and low-frequency cost sequence of the target point.

Images