Multi-target load prediction adjustment method, system, device, medium and program product

Abstract

The invention discloses a multi-target load prediction adjustment method, system and device, a medium and a program product, and the method comprises the steps: carrying out the modal decomposition of power load data through employing a VMD algorithm, and obtaining the decomposed load data; a GLTN-BiGRU network model is constructed, and the GLTN-BiGRU network model is optimized and trained; inputting the decomposed load data into the optimized and trained GLTN-BiGRU network model to carry out future short-term power load prediction, and obtaining a load prediction result; and making a load configuration scheme by using the load prediction result to realize multi-target load adjustment. According to the invention, accurate prediction of future short-term power load and multi-target load adjustment are realized.

Description

Technical Field

[0001] The present invention relates to the field of intelligent electricity consumption technology, and in particular to a multi-objective load forecasting and regulation method, system, equipment, medium and program product. Background Art

[0002] With the rapid development of economy and power industry, electricity load is becoming increasingly complex and diversified. Accurate load forecasting can guide the power system to achieve supply and demand balance and ensure the efficient and stable operation of the power system, thus having a positive impact on the high-quality development of the power industry.

[0003] The load itself has strong randomness and volatility, and is easily affected by external factors, so achieving high-precision load forecasting is still a very challenging task. With the diversification of electricity consumption, multi-objective load regulation is also a challenge.

[0004] Therefore, there is an urgent need for a multi-objective load forecasting and regulation method to achieve accurate prediction of future short-term power loads and multi-objective load regulation. Summary of the invention

[0005] The present invention aims to provide a multi-objective load forecasting and regulation method, system, device, medium and program product to achieve accurate forecasting of future short-term power loads and multi-objective load regulation.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] In a first aspect, an embodiment of the present invention provides a multi-objective load forecasting and adjustment method, comprising:

[0008] The VMD algorithm is used to perform modal decomposition on the power load data to obtain the decomposed load data;

[0009] Build a GLTN-BiGRU network model, and optimize and train the GLTN-BiGRU network model;

[0010] The decomposed load data is input into the optimized and trained GLTN-BiGRU network model to perform future short-term power load forecasting and obtain the load forecasting results;

[0011] The load forecasting results are used to formulate a load configuration plan to achieve multi-objective load regulation.

[0012] In a second aspect, an embodiment of the present invention provides a multi-objective load forecasting and regulation system, the system comprising:

[0013] A decomposition unit, used for performing modal decomposition on the power load data by using a VMD algorithm to obtain decomposed load data;

[0014] A construction unit, used to construct a GLTN-BiGRU network model, and optimize and train the GLTN-BiGRU network model;

[0015] A prediction unit is used to input the decomposed load data into the optimized and trained GLTN-BiGRU network model to perform future short-term power load prediction and obtain load prediction results;

[0016] The regulating unit is used to formulate a load configuration plan using the load forecasting result to achieve multi-objective load regulation.

[0017] In a third aspect, an embodiment of the present invention further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program or instructions to implement the aforementioned multi-objective load forecasting and regulation method.

[0018] In a fourth aspect, an embodiment of the present invention further provides a computer storage medium, wherein a computer program or instruction is stored in the computer storage medium, and when the computer program or instruction is executed by a processor, the aforementioned multi-objective load forecasting and regulation method is implemented.

[0019] In a fifth aspect, an embodiment of the present invention further provides a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned multi-objective load forecasting and regulation method.

[0020] The technical effects and advantages of the present invention are as follows: the present invention adopts variational mode decomposition (VMD algorithm) to process the original load data to ensure the stability and robustness of the data, and the processed data is input into the GLTN-BiGRU network model optimized by the ADMM algorithm, thereby realizing accurate prediction of future power loads; finally, resources are reasonably adjusted according to the load prediction results, and a load configuration plan is formulated to realize multi-objective load regulation; the present invention can make more accurate predictions of regional power loads and more accurate load regulation, thereby improving the safe and stable operation of the power grid.

[0021] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0023] Figure 1 A flowchart of a multi-objective load forecasting and adjustment method according to an embodiment of the present invention;

[0024] Figure 2 It is a schematic diagram of the structure of the GLTN-BiGRU network model in an embodiment of the present invention;

[0025] Figure 3 Schematic diagram of the decomposition result of the VMD algorithm in an embodiment of the present invention;

[0026] Figure 4 Schematic diagram of comparison of optimization algorithms for objective functions in embodiments of the present invention;

[0027] Figure 5 It is a schematic diagram for comparing the load forecasting results output by various models;

[0028] Figure 6 This is a schematic diagram of multi-objective load regulation results in an embodiment of the present invention;

[0029] Figure 7 A schematic diagram of the structure of a multi-objective load forecasting and regulation system according to an embodiment of the present invention;

[0030] Figure 8 The figure is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0032] To address the deficiencies of the prior art, the present invention discloses a multi-objective load forecasting and adjustment method. Figure 1 As shown, the following steps are included:

[0033] Step S1: using VMD algorithm to perform modal decomposition on power load data to obtain decomposed load data;

[0034] Step S2: construct a GLTN-BiGRU network model, and optimize and train the GLTN-BiGRU network model;

[0035] Step S3: input the decomposed load data into the optimized and trained GLTN-BiGRU network model to perform future short-term power load forecasting and obtain load forecasting results;

[0036] Step S4: Formulate a load configuration plan using the load forecasting results to achieve multi-objective load regulation.

[0037] In some specific embodiments, before step S1, the method of the present invention further comprises:

[0038] Step S0: acquiring power load data by collecting original power load data and preprocessing it, specifically including:

[0039] The original power load data is collected by using a data monitoring and acquisition system, and the original power load data is cleaned, filled, abnormal data is eliminated and normalized to obtain the power load data;

[0040] The data filling includes: for the missing data in the power load related data, the average of the data of two similar days before and after the data is taken to fill the data;

[0041] Abnormal data elimination includes: for the data that is suddenly abnormally large or suddenly zero in the original power load data, reasonable replacement is performed based on the data of the time period before and after the data and the data of similar days;

[0042] The original power load data is normalized by the following formula:

[0043]

[0044] In the formula, x represents the power load related data, x′ represents the normalized power load related data, and x min Indicates the minimum value of power load related data, x max Indicates the maximum value of power load related data.

[0045] In some specific embodiments, step S1: using VMD algorithm (Variational Mode Decomposition) to perform modal decomposition on power load data to obtain decomposed load data, specifically includes:

[0046] (1) Assuming that the power load data sequence is f(t), the power load data is decomposed into k modal components;

[0047] Among them, the sum of the estimated bandwidths of each modal component is the smallest, and the constraint condition is that the sum of all modal components is equal to the original signal (i.e., the power load data sequence), then the corresponding constrained variational problem is expressed as:

[0048]

[0049] In the formula, {u k} represents the set of modal components, {ω k} represents the center frequency set, k represents the number of modal components, (k=1,2,...,K), K represents the total number of modal components; t represents the sampling time, represents partial derivative calculation, δ(t) represents unit impulse function, represents imaginary unit, * represents convolution operation, u k (t) represents the kth modal component sampled at time t.

[0050] The penalty factor α and the Lagrange multiplication operator λ(t) are introduced into the expression of the constrained variational problem to convert the constrained variational problem into an unconstrained variational problem. The expression is as follows:

[0051]

[0052] In the formula, L({u k},{ω k},λ) represents the unconstrained variational problem, ω k represents the center frequency corresponding to the kth modal component.

[0053] (2) Using the alternating direction multiplier iteration algorithm (i.e. A DMM The algorithm combines Parseval / Plancherel and Fourier isometric transform to optimize the modal components and center frequencies, solve the saddle point of the unconstrained variational problem (to achieve effective separation of signal frequencies), and alternately find the u after iteration. k ,ω k The expressions for and λ are as follows:

[0054]

[0055] In the formula, represents u after iteration k , that is, the kth modal component after the n+1th iteration Fourier transform; n represents the number of iterations, represents the Fourier transform of the power load data sequence f(t), represents the i'th modal component u i' The Fourier transform of (t), represents the Fourier transform of the Lagrange multiplication operator λ(t), α represents the penalty factor, and ω represents the center frequency; represents the iterated ω k , that is, the center frequency of the kth modal component at the n+1th iteration; represents the kth modal component u k The Fourier transform of (t), Represents the iterated That is, the Fourier transform of the Lagrange multiplication operator λ(t) after the n+1th iteration; λ(t) represents the Lagrange multiplication operator in the VMD algorithm, and T represents the noise tolerance;

[0056] In the iterative process of alternating search, determine whether the iteration termination condition is met:

[0057] If satisfied, the iteration is terminated, and the decomposed k modal components are obtained, that is, the decomposed load data (time series data) is obtained;

[0058] If it is not satisfied, repeat the above decomposition and iteration process until the iteration termination condition is met, and the decomposed k modal components are obtained, that is, the decomposed load data;

[0059] Among them, whether the iteration termination condition is met is judged by the following formula:

[0060]

[0061] In the formula, represents the kth modal component at the n+1th iteration, and ε represents the set convergence threshold.

[0062] Since the power load data is nonlinear and non-stationary, the present invention uses the VMD algorithm (i.e., variational mode decomposition method) to perform signal processing on the power load data to obtain multiple modal components;

[0063] The VMD algorithm determines the center frequency and bandwidth of each modal component by iteratively searching for the optimal solution and processes the original signal based on a non-recursive method. The obtained modal components have good mutual differences, thus effectively avoiding mutual interference between the modal components.

[0064] For example, the decomposed load data obtained by using the VMD algorithm in the embodiment of the present invention is shown in Table 1 below. In this embodiment, the values ​​of k are 3, 4, 5, 6, and 7 respectively; Figure 2 As shown, when k=3-6, no modal aliasing occurs; however, when k=7, mode 5 and mode 6 are relatively close to each other, and modal aliasing occurs; therefore, k is set to 6 in the present invention.

[0065] Table 1 Decomposition results of VMD algorithm

[0066]

[0067]

[0068] It can be seen that the VMD algorithm is used in the present invention to decompose the power load data into six modes; the trend mode, periodic mode and random mode are mainly used to perform load forecasting;

[0069] Among them, the use of trend mode can help the model better identify the trend of long-term load changes, the use of periodic mode is used to identify daily or weekly load fluctuations, and finally the use of random mode to remove some short-term fluctuations or noise in the signal that cannot be accurately predicted, reduce the complexity of the data, and make the model more focused on more predictive components.

[0070] In some specific embodiments, step S2: constructing a GLTN-BiGRU network model, and optimizing and training the GLTN-BiGRU network model using a data set (including training samples and test samples), comprises the following steps:

[0071] Step S21: constructing a GLTN-BiGRU network model based on the temporal convolutional neural network GLTN and the bidirectional gated recurrent unit BiGRU;

[0072] Step S22: using the ADMM algorithm to iteratively optimize the hyperparameters in the GLTN-BiGRU network model, find the optimal parameter values ​​of the model, and obtain the optimized GLTN-BiGRU network model; specifically including:

[0073] (1) The different modal components decomposed by the VMD algorithm are regarded as sub-problems, and a convex optimization problem of the GLTN-BiGRU network model is constructed; wherein the expression of the convex optimization problem is:

[0074]

[0075] stAx+Bz=c,

[0076] Where x represents the decision variable, z represents the auxiliary variable, f(x) represents the function to be optimized about x, g(z) represents the function to be optimized about z, A represents the relationship between the features extracted from the power load data and the load forecasting results; B represents the weight matrix of the load forecasting results, and c represents the constant vector.

[0077] (2) Based on the convex optimization problem, the Lagrangian function is proposed as:

[0078]

[0079] In the formula, represents the Lagrangian function; represents the Lagrange multiplication operator in the ADMM algorithm, Rp represents a real number; T represents noise tolerance, ρ represents a penalty parameter used to control the cost of constraint violation, ρ>0.

[0080] (3) Using the ADMM algorithm (Alternating Direction Method of Multipliers) to iteratively optimize the parameters in the Lagrangian function, the Lagrangian function is minimized and augmented, including:

[0081] In the case of fixed auxiliary variables and Lagrange multiplication operators Under the condition of , the decision variables are updated to minimize the augmented Lagrangian function; the updated decision variables are:

[0082]

[0083] In the formula, x k′+1 represents the decision variable updated for the k′+1th time, x represents the decision variable, f(x) represents the function to be optimized about x, ρ represents the penalty parameter, A represents the relationship between the features extracted from the power load data and the load forecasting results; B represents the weight matrix of the load forecasting results, c represents the constant vector; z k′ represents the current auxiliary variable, that is, the auxiliary variable updated for the k′th time; represents the current Lagrangian multiplication operator, i.e., the Lagrangian multiplication operator updated for the k′th time; represents the Lagrange multiplication operator, R p represents a real number;

[0084] Under the condition of fixed decision variables and Lagrangian multiplication operator, the auxiliary variables are updated to minimize the augmented Lagrangian function; the updated auxiliary variables are:

[0085]

[0086] In the formula, z k′+1 represents the auxiliary variable updated for the k′+1th time, and g(z) represents the function to be optimized with respect to z;

[0087] According to the current decision variables and auxiliary variables, the Lagrangian multiplication operator is updated; wherein the updated Lagrangian multiplication operator is:

[0088]

[0089] In the formula, represents the Lagrange multiplication operator for the k′+1th update;

[0090] Finally, the convergence of the ADMM algorithm is verified. The ADMM algorithm stops iterating when the following conditions are met and obtains the optimal parameters of the model:

[0091]

[0092] In the formula, ε pri represents the tolerance threshold of the original residual, ε dual represents the tolerance threshold of the dual residual;

[0093] If the ADMM algorithm does not meet the above conditions, it returns to continue iterative updating.

[0094] (4) The weight of each modal component is independently optimized by using an alternating direction multiplier iteration algorithm, and the convex optimization problem is solved by minimizing the augmented Lagrangian function to optimize the GLTN-BiGRU network model and obtain the optimized GLTN-BiGRU network model.

[0095] In this step, the present invention uses the ADMM algorithm to optimize the hyperparameters in the GLTN-BiGRU network model, which can maximize the performance of the model and improve the prediction accuracy of the model; Figure 3 As shown in the figure, the ADMM algorithm has a faster optimization speed than the commonly used parameter optimization algorithm, and the model optimization can be completed with fewer iterations.

[0096] Step S23: Use the training samples to perform model training on the optimized GLTN-BiGRU network model to obtain the optimized trained GLTN-BiGRU network model.

[0097] In some specific embodiments, step S3: inputting the decomposed load data into the optimized and trained GLTN-BiGRU network model to perform future short-term power load forecasting to obtain load forecasting results, including the following steps:

[0098] like Figure 4 As shown, the optimized trained GLTN-BiGRU network model includes an optimized trained GLTN model, an optimized trained BiGRU model and a fully connected layer.

[0099] Step S31: input the decomposed load data (i.e., the load data to be predicted) as input data into the optimized and trained GLTN-BiGRU network model. The optimized and trained GLTN model uses the self-attention mechanism to extract the global time features and local time features of the input data, and performs feature fusion to obtain the final fusion features; specifically, it includes:

[0100] The GLTN model includes the attention mechanism layer, LocalTemporal module, GlobalTrend module and feature fusion module;

[0101] (1) The attention mechanism layer embeds the features of each time point in the input data, maps the input data to a high dimension to capture the feature information in the input data, obtains the embedded feature sequence and sends it to the attention mechanism layer and the LocalTemporal module respectively; specifically includes:

[0102] The input data is x t ={x1,x2,...,x t ,..,x T}, where xt represents the feature of the input data sequence at the tth moment;

[0103] The features of each time point in the input data are embedded, and the features after embedding are obtained for each time step (each time step represents the features of each time point):

[0104] e t =W e ·x t +b e ,

[0105] In the formula, e t represents the embedded features of the t-th time step, W e represents the embedding matrix, b e represents a constant bias;

[0106] Then, based on the embedded features of each time step, we obtain the embedded feature sequence E = {e1, e2, ..., e t ,..,e T}.

[0107] (2) The attention mechanism layer uses the self-attention mechanism to re-encode the input data to obtain the encoded input data, and calculates the output features of the attention mechanism in combination with the embedded feature sequence; specifically, it includes:

[0108] The attention mechanism layer uses the self-attention mechanism to re-encode the input data to obtain the encoded input data

[0109] The attention mechanism layer performs a linear transformation on the embedded feature sequence to obtain the weight matrices Query, Key, and Value, and combines the encoded input data to obtain the vectors Query, Key, and Value as follows:

[0110]

[0111] Where Q represents the vector Query, W q represents the weight matrix Query, K represents the vector Key, W k represents the weight matrix Key, V represents the vector Value, W v Represents the weight matrix Value.

[0112] Based on the vector Query and the vector Key, the attention score matrix A(Q, K) is calculated, and then the attention mechanism output feature h is calculated based on the attention score matrix A(Q, K) and the vector Value a ;

[0113] Among them, the attention score matrix and the attention mechanism output features are:

[0114]

[0115] h a =A(Q,K)V,

[0116] In the formula, A(Q,K) represents the attention score matrix, Q represents the vector Query, which represents the feature vector of the current input sequence (i.e., the encoded input data); K represents the vector Key, which is used to match Q to calculate the attention score matrix; Represents the scaling factor to prevent the dot product value from being too large; h a represents the output feature of the attention mechanism, V represents the vector Value, and represents the actual information of the current input sequence (i.e., the encoded input data).

[0117] (3) The attention mechanism layer transmits the encoded input data to the LocalTemporal module and the GlobalTrend module.

[0118] (4) The LocalTemporal module uses a convolutional neural network to extract detailed features of the encoded input data from a shorter time window, obtains the local temporal features of the input data and sends them to the feature fusion module.

[0119] The encoded input data is in represents the encoded input data at the Tth moment;

[0120] The local time characteristics of the input data obtained are:

[0121]

[0122] In the formula, h l represents the local temporal features of the input data, f() represents the function of extracting features, k” represents the convolution kernel size of the convolutional neural network, and Wl represents the weight parameter of the convolution kernel, represents the encoded input data, b l Represents the weight bias of the convolution in the convolutional neural network.

[0123] (5) The GlobalTrend module uses a residual network to extract global time features from the encoded input data, obtains the global time features of the input data and sends them to the feature fusion module;

[0124] Among them, the global time characteristics are obtained by the following formula:

[0125]

[0126] In the formula, h g represents the global temporal features of the input data, ResNet represents the residual network, and W g represents the weight parameter of the residual network convolution kernel, represents the encoded input data, b g Represents the weight bias of the convolution in the residual network.

[0127] (6) The feature fusion module uses a weighted average fusion method to fuse the global time feature and the local time feature to obtain a fused feature, and concatenates the fused feature with the output feature of the attention mechanism to obtain a final fused feature.

[0128] Among them, the fusion feature is obtained by the following formula:

[0129] h f =α·h g +(1-α)·h l ,

[0130] In the formula, h f represents the fusion feature, α represents the weight parameter of feature fusion, α∈[0,1], which is used to control the relative importance of global and local features; h g represents the global temporal characteristics of the input data, h l Represents the local temporal characteristics of the input data.

[0131] The final fusion feature is obtained by the following formula:

[0132] H f =Concat(h a ,h f ),

[0133] In the formula, H f represents the final fusion feature, Concat represents the concatenation operation, and h a Represents the output features of the attention mechanism.

[0134] Step S32: The optimized trained GLTN model transfers the final fusion features to the optimized trained BiGRU model.

[0135] Step S33: The optimized trained BiGRU model captures the bidirectional information flow of the final fusion feature through the bidirectional GRU layer, and obtains the final temporal feature by splicing, which specifically includes:

[0136] (1) The optimized trained BiGRU model captures the forward and backward information regularities and data dependencies in the final fusion features through the bidirectional GRU layer, so as to fully mine the temporal features in the final fusion features and obtain forward hidden information and reverse hidden information.

[0137] The bidirectional GRU layer includes a forward GRU layer and a reverse GRU layer. The GRU layer is a simplified model of LSTM. Each GRU layer only includes two gating mechanisms: a reset gate and an update gate. This mechanism enables the GRU layer to effectively extract timing information.

[0138] The present invention optimizes the final fusion feature H obtained by the trained GLTN model f As the input information of the optimized trained BiGRU model, Represents the input of the load characteristics at each moment, and passes it to the BiGRU model as the input information at each moment.

[0139] Therefore, the optimized trained BiGRU model captures the forward information regularity in the final fusion feature through the forward GRU layer, and obtains the forward hidden information as follows:

[0140]

[0141] In the formula, represents the state of the update gate in the forward GRU layer, σ represents the Sigmoid function, and Both represent the weights of the update gate of the forward GRU layer, H f Represents the final fusion features output by the GLTN model after optimization training; Represents the final fusion feature H f The features at time t, that is, the input information of the optimized trained BiGRU model at time t (i.e., the current time); represents the forward hidden information at the t-1th moment (i.e., the previous moment), r t fwd represents the state of the reset gate in the forward GRU layer, and Both represent the weights of the reset gate of the forward GRU layer, Represents the state of the candidate state information in the forward GRU layer, U fwd and W fwd Both represent the weights of the candidate state information of the forward GRU layer, represents the Hadmard product, Represents the forward hidden information at time t;

[0142] The optimized trained BiGRU model captures the reverse information regularity in the final fusion feature through the reverse GRU layer, and obtains the backward hidden information as follows:

[0143]

[0144]

[0145] In the formula, represents the state of the update gate in the reverse GRU layer, σ represents the Sigmoid function, and Both represent the weights of the update gate of the reverse GRU layer, represents the reverse hidden information at the t+1th moment (i.e. the previous moment), r t bwd represents the state of the reset gate in the reverse GRU layer, and Both represent the weights of the reset gate of the reverse GRU layer, The state representing the candidate state information of the reverse GRU layer, Indicates reverse hidden information.

[0146] The calculation process of the reverse GRU layer is symmetrical with that of the forward GRU layer, but the reverse GRU layer is calculated based on the reverse time series.

[0147] (2) The optimized trained BiGRU model concatenates the forward hidden information and the reverse hidden information to obtain the final time series feature:

[0148]

[0149] In the formula, h t Represents the final timing characteristics, represents the forward hidden information at time t, Indicates reverse hidden information.

[0150] Step S33: Based on the final time series characteristics, load prediction is performed through the fully connected layer to obtain the load prediction result (i.e. the power load value at a certain moment) to achieve prediction of future time steps.

[0151] The load forecast results are:

[0152] yt =W out ·h t +b out ,

[0153] In the formula, y t represents the output of the optimized trained GLTN-BiGRU network model at time t, that is, the load prediction of the optimized trained GLTN-BiGRU network model for the current time step; W out represents the weight of the fully connected layer, b out represents the bias of the fully connected layer, h t Represents the final timing characteristics.

[0154] In this step, the GLTN-BiGRU network model mainly includes the GLTN model and the BiGRU model. Although the BiGRU model can extract sequence information from beginning to end, the feature extraction ability of the BiGRU model will be affected by the length of the sequence, especially the particularly long sequence; although the BiGRU can theoretically see the global context, it may be limited in capturing long-distance dependencies and extracting deep global features. To address this defect, the present invention introduces the GLTN model.

[0155] The present invention mainly extracts the local features of the input data through the LocalTemporal module in the GLTN model, and then extracts the global features of the input data through the GlobalTrend module in the GLTN model, and finally fuses the local features and the time features to obtain the fusion features, and then splices them with the output features of the attention mechanism to obtain the final fusion features and input them into the BiGRU model; the BiGRU model captures the bidirectional information flow of the final fusion features (i.e., input information) through the bidirectional GRU layer, deeply explores the dynamic change rules of the input information, obtains richer time series features, and improves the accuracy of the prediction. Finally, the model output is dimensionalized through the fully connected layer to achieve accurate prediction.

[0156] In some specific embodiments, step S4: formulating a load configuration plan using the load forecasting result to achieve multi-objective load regulation includes the following steps:

[0157] Step S41: Based on the load forecast result, the historical power consumption data of the user's demand-side response period is analyzed to obtain the comprehensive demand-side response potential of each user, specifically including:

[0158] (1) Based on the load forecasting result, analyzing the load forecasting result curve (such as the peak of the curve, the trough of the curve and the rate of change of the curve, etc.)

[0159] (2) Based on the load forecast result curve, statistically analyze the historical electricity consumption data of users during the demand-side response period on typical days, and use the standard deviation formula to measure the load dispersion of each user;

[0160] The load dispersion of each user includes: the power consumption volatility of each user α, the minimum power load x of each user’s demand side response period hmin , and the average power load x of each user's demand-side response period av ;

[0161] Among them, the power consumption volatility of each user is calculated by the following formula:

[0162]

[0163] In the formula, α v Indicates the user volatility of a certain user in a certain period of time, n v represents the number of time periods during which demand-side response is carried out, xi represents the power load of the user in the i-th demand-side response period, x av Indicates the average power load of the user during the demand-side response period.

[0164] (2) Analyze the user's power consumption volatility and set a power consumption volatility threshold. When the user's power consumption volatility is higher than the power consumption volatility threshold, it indicates that the user's load fluctuation is large during this period, that is, the user's load adjustability rate is high during this period;

[0165] Therefore, the load forecast results are used to analyze the user's maximum response capacity, and combined with the load dispersion of each user, a user's single-hour demand-side response capacity determination model is set to calculate the demand-side response potential of each user in a single response period.

[0166] The user's demand-side response capability determination model for a single hour is:

[0167]

[0168] In the formula, β j represents the demand side response potential of user j in the i-th demand side response period, γi represents the maximum response capacity of user j in the i-th demand side response period; x hmin represents the minimum power load of user j in the i-th demand-side response period, x av represents the average power load of user j during the demand-side response period.

[0169] (3) Based on the demand-side response potential of each user in a single response period, the comprehensive demand-side response potential of each user is calculated as:

[0170]

[0171] In the formula, δ j represents the comprehensive demand-side response potential of user j, β j represents the demand-side response potential of user j in the i-th demand-side response period, and represents the total number of users.

[0172] Step S42: According to the comprehensive demand-side response potential of each user and the given total load energy-saving index, formulate the energy-saving index of each user to achieve multi-objective load regulation; specifically including:

[0173] Taking into account the comprehensive demand-side response potential of each user, the willingness level of each user, the response coordination degree of each user, the load importance, and the economic level of each user, the energy-saving index of each user is set based on the given total load energy-saving index as follows:

[0174] W j =α'×δ j ×(x av -x hmin ),

[0175] Where W j represents the energy saving index of user j, α' represents the comprehensive weight of user j's actual willingness level, response coordination, load importance and economic level in the i-th response period, δ j represents the comprehensive demand-side response potential of user j, x av represents the average power load of user j during the demand-side response period, x hmin represents the minimum power load of user j in the i-th demand-side response period.

[0176] In some specific embodiments, in the multi-objective regulation process, constructing a load optimization objective function and load constraint conditions includes:

[0177] In the multi-objective regulation process, the goal is to minimize load volatility (that is, to make the difference between the maximum load after peak shaving and the minimum load after valley filling as small as possible), and the load optimization objective function is constructed as follows:

[0178]

[0179] Where φ(t) represents the load optimization objective function, w1 and w2 represent weight factors, P t represents the load in the tth period, t=1,2,...,T; U t represents the user satisfaction with load allocation in period t, g(U t ) represents user satisfaction U t The reward function of

[0180] The load constraints include: upper and lower limit constraints, and load reduction and filling constraints;

[0181] The upper and lower limit constraints are: after load adjustment, the load shall not exceed the maximum load, and the minimum load shall not be lower than the minimum load; wherein, the upper and lower limit constraints are expressed as:

[0182] P min ≤P t +ΔP t ≤P max ≤P t ×η,

[0183] In the formula, p min represents the minimum capacity, Δp represents the load change, that is, the value after load adjustment; P max represents the maximum capacity, i.e. the maximum load that the user can use; η represents the maximum acceptable adjustment ratio of the user;

[0184] The load reduction and filling constraints are: for each peak period t1, if the load Exceeding the maximum capacity P max , then the reduction part is:

[0185]

[0186] In the formula, Indicates that the power is cut during the peak period t1. represents the load during the peak period t1, P max Indicates the maximum capacity;

[0187] This reduced load will be accumulated to be filled during the off-peak period:

[0188]

[0189] In the formula, Indicates the load at the trough moment t2, ΔP h Indicates the load change at peak load, N lw Indicates the number of load troughs.

[0190] Step S43: Iterative adjustment: If the load distribution after the initial multi-objective load adjustment does not achieve the desired balanced effect (i.e., the load volatility is still large), the load value after the multi-objective load adjustment is continuously iterated and optimized until the optimization condition is met and the iteration is stopped;

[0191] Wherein, the optimization condition is:

[0192]

[0193] In the formula, represents the load at the trough moment t2, and ò represents the acceptable threshold of load fluctuation.

[0194] The embodiment of the present invention mainly trains and tests the VMD-ADMM-GLTN-BiGRU model constructed by the present invention by collecting power load data with a sampling time interval of 15 minutes and a collection time from January 1, 2023 to December 31, 2023 as a sample data set, and obtains the GLTN-BiGRU network model after optimized training; at the same time, in order to illustrate the reliability and feasibility of the GLTN-BiGRU network model after optimized training, relevant personnel compare the prediction effects of the model of the present invention with the BiLSTM model, BiGRU model, GLTN model, and conventional GLTN-BiGRU model;

[0195] like Figure 5 As shown in the figure, we can see the prediction results obtained by various models for load forecasting. Figure 5 The experimental results show that compared with other methods, the proposed method is closer to the actual value at most sample points and has a smaller prediction error, especially in the area where the power rises sharply near the sample point 400, where the fitting effect is particularly outstanding; this shows that the proposed method has higher accuracy and superiority in the active power prediction task, and also shows that the prediction model of the proposed method is better than other comparison models; and in the verification process, the multi-objective load regulation results of 100 users as an example are as follows Figure 6 As shown, comprehensive Figure 5 and Figure 6 From the above, the method of the present invention can make more precise predictions of regional power loads and more accurate load regulation, thereby improving the safe and stable operation of the power grid.

[0196] The present invention uses variational mode decomposition (VMD algorithm) to process the original load data to ensure the stability and robustness of the data. The processed data is input into the GLTN-BiGRU network model optimized by the ADMM algorithm, thereby realizing accurate prediction of future power loads. Finally, resources are rationally adjusted according to the load prediction results, and a load configuration plan is formulated to achieve multi-objective load regulation. The present invention can make more accurate predictions of regional power loads and more accurate load regulation, thereby improving the safe and stable operation of the power grid.

[0197] The present invention performs corresponding load regulation according to the obtained peak shaving index and the predicted volatility of each user, adjusts the load interval with large volatility, and moves the load to the trough interval to reduce overload electricity consumption; and actively introduces user feedback in the iterative process, collects users' views and feelings on the load regulation strategy through the user interface or direct communication, and this information can be used to further adjust the objective function and the regulation strategy; for example, if users report that certain regulation measures have affected their normal electricity consumption, it can be considered to reduce the influence of these measures in the next round of adjustments.

[0198] The embodiment of the present invention discloses a multi-objective load forecasting and regulation system. Figure 7 As shown, the system comprises:

[0199] A decomposition unit, used for performing modal decomposition on the power load data by using a VMD algorithm to obtain decomposed load data;

[0200] A construction unit, used to construct a GLTN-BiGRU network model, and optimize and train the GLTN-BiGRU network model;

[0201] A prediction unit is used to input the decomposed load data into the optimized and trained GLTN-BiGRU network model to perform future short-term power load prediction and obtain load prediction results;

[0202] The regulating unit is used to formulate a load configuration plan using the load forecasting result to achieve multi-objective load regulation.

[0203] Regarding the system in the above embodiment, the specific manner in which each unit module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0204] Based on the same inventive concept, an embodiment of the present invention further provides an electronic device, whose structure is as follows: Figure 8 As shown, it includes a memory, a processor and a computer program stored in the memory, and the processor executes the computer program or instructions to implement the aforementioned multi-objective load forecasting and regulation method.

[0205] Based on the same inventive concept, an embodiment of the present invention further provides a computer storage medium, wherein a computer program or instruction is stored in the computer storage medium, and when the computer program or instruction is executed by a processor, the aforementioned multi-objective load forecasting and adjustment method is implemented.

[0206] Based on the same inventive concept, an embodiment of the present invention further provides a computer program product, including a computer program or instructions, which implement the aforementioned multi-objective load forecasting and regulation method when executed by a processor.

[0207] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A multi-objective load forecasting and regulation method, characterized in that: include: The VMD algorithm is used to perform modal decomposition on the power load data to obtain the decomposed load data; Build the GLTN-BiGRU network model, and optimize and train the GLTN-BiGRU network model; The decomposed load data is input into the optimized and trained GLTN-BiGRU network model to perform future short-term power load forecasting and obtain the load forecasting results; The load forecasting results are used to formulate a load configuration plan to achieve multi-objective load regulation.

2. A multi-objective load forecasting and adjustment method according to claim 1, characterized in that: The power load data is obtained by collecting original power load data and preprocessing it, including: The original power load data is collected by using a data monitoring and acquisition system, and the original load data is cleaned, filled, abnormal data is eliminated and normalized to obtain the power load data.

3. A multi-objective load forecasting and adjustment method according to claim 1, characterized in that: The VMD algorithm is used to perform modal decomposition on the power load data to obtain the decomposed load data, including: Assuming that the power load data sequence is f(t), the power load data is decomposed into k modal components, then the corresponding constrained variation problem expression is: In the formula, {u k } represents the set of modal components, {ω k } represents the center frequency set, k represents the number of modal components, (k=1,2,...,K), K represents the total number of modal components; t represents the sampling time, represents partial derivative calculation, δ(t) represents unit impulse function, represents imaginary unit, * represents convolution operation, u k (t) represents the kth modal component sampled at time t; The penalty factor α and the Lagrange multiplication operator λ(t) are introduced into the expression of the constrained variational problem to convert the constrained variational problem into an unconstrained variational problem. The expression is as follows: In the formula, L({u k },{ω k },λ) represents the unconstrained variational problem, ω k represents the center frequency corresponding to the kth modal component; Solve the saddle point of the unconstrained variational problem and alternately find u after iteration k ,ω k The expressions for and λ are as follows: In the formula, represents u after iteration k , that is, the kth modal component after the n+1th iteration Fourier transform of ; n represents the number of iterations, represents the Fourier transform of the power load data sequence f(t), represents the i'th modal component u i' The Fourier transform of (t), represents the Fourier transform of the Lagrange multiplication operator λ(t), α represents the penalty factor, and ω represents the center frequency; represents the iterated ω k , that is, the center frequency of the kth modal component at the n+1th iteration; represents the kth modal component u k The Fourier transform of (t), Represents the iterated That is, the Fourier transform of the Lagrange multiplication operator λ(t) after the n+1th iteration; λ(t) represents the Lagrange multiplication operator in the VMD algorithm, and T represents the noise tolerance; During the iteration process, determine whether the iteration termination condition is met: If it is satisfied, the iteration is terminated, and the decomposed k modal components are obtained, that is, the decomposed load data is obtained; If it is not satisfied, repeat the above decomposition and iteration process until the iteration termination condition is met, and the decomposed k modal components are obtained, that is, the decomposed load data; Among them, whether the iteration termination condition is met is judged by the following formula: In the formula, represents the kth modal component at the n+1th iteration, and ε represents the set convergence threshold.

4. A multi-objective load forecasting and adjustment method according to claim 1 or 3, characterized in that: Build the GLTN-BiGRU network model, and optimize and train the GLTN-BiGRU network model, including: Based on the temporal convolutional neural network GLTN and the bidirectional gated recurrent unit BiGRU, a GLTN-BiGRU network model is constructed; The ADMM algorithm is used to iteratively optimize the hyperparameters in the GLTN-BiGRU network model, find the optimal parameter values ​​of the model, and obtain the optimized GLTN-BiGRU network model; The optimized GLTN-BiGRU network model is trained to obtain the optimized trained GLTN-BiGRU network model.

5. A multi-objective load forecasting and adjustment method according to claim 4, characterized in that: The ADMM algorithm is used to iteratively optimize the hyperparameters in the GLTN-BiGRU network model to find the optimal value of the model and obtain the optimized GLTN-BiGRU network model, including: The decomposed different modal components are regarded as sub-problems to construct the convex optimization problem of the GLTN-BiGRU network model; Based on the convex optimization problem, a Lagrangian function is proposed, and the parameters in the Lagrangian function are iteratively optimized using the ADMM algorithm to achieve minimization and augmentation of the Lagrangian function; The ADMM algorithm is used to independently optimize the weight of each modal component, and the convex optimization problem is solved by minimizing the augmented Lagrangian function to optimize the GLTN-BiGRU network model.

6. A multi-objective load forecasting and adjustment method according to claim 5, characterized in that: The expression of the convex optimization problem is: s.t .A x+ B z＝c , In the formula, x represents the decision variable, z represents the auxiliary variable, f(x) represents the function to be optimized about x, g(z) represents the function to be optimized about z, A represents the relationship between the features extracted from the power load data and the load forecasting results; B represents the weight matrix of the load forecasting results, and c represents the constant vector; The Lagrangian function is: In the formula, represents the Lagrangian function; represents the Lagrange multiplication operator in the ADMM algorithm, R p represents a real number; T represents the noise tolerance, and ρ represents the penalty parameter.

7. A multi-objective load forecasting and adjustment method according to claim 5 or 6, characterized in that: The ADMM algorithm is used to iteratively optimize the parameters in the Lagrangian function, including: In the case of fixed auxiliary variables and Lagrange multiplication operators Under the condition of , the decision variables are updated to minimize the augmented Lagrangian function; the updated decision variables are: In the formula, x k′+1 represents the decision variable updated for the k′+1th time, x represents the decision variable, f(x) represents the function to be optimized about x, ρ represents the penalty parameter, A represents the relationship between the features extracted from the power load data and the load forecasting results; B represents the weight matrix of the load forecasting results, c represents the constant vector; Z k′ represents the current auxiliary variable, that is, the auxiliary variable updated for the k′th time; represents the current Lagrangian multiplication operator, i.e., the Lagrangian multiplication operator updated for the k′th time; represents the Lagrange multiplication operator in the ADMM algorithm, R p represents a real number; With fixed decision variables and Lagrange multiplication operators Under the condition of , the auxiliary variables are updated to minimize the augmented Lagrangian function; the updated auxiliary variables are: In the formula, z k′+1 represents the auxiliary variable updated for the k′+1th time, and g(z) represents the function to be optimized with respect to z; Update the Lagrange multiplication operator according to the current decision variables and auxiliary variables Among them, the updated Lagrange multiplication operator is: In the formula, represents the Lagrange multiplication operator for the k′+1th update; Verify the convergence of the ADMM algorithm. The ADMM algorithm stops iterating when the following conditions are met to obtain the optimal parameters of the model: In the formula, ε pri represents the tolerance threshold of the original residual, ε dual Represents the tolerance threshold of the dual residual.

8. A multi-objective load forecasting and adjustment method according to claim 1, characterized in that: The decomposed load data is input into the optimized and trained GLTN-BiGRU network model to calculate the future short-term power load and obtain the load forecast results, including: The optimized trained GLTN-BiGRU network model includes an optimized trained GLTN model, an optimized trained BiGRU model and a fully connected layer; The decomposed load data is input as input data into the optimized and trained GLTN-BiGRU network model. The optimized and trained GLTN model uses the self-attention mechanism to extract the global and local time features of the input data, and performs feature fusion to obtain the final fusion feature. The optimized trained GLTN model transfers the final fusion features to the optimized trained BiGRU model. The optimized trained BiGRU model captures the bidirectional information flow of the final fusion features through the bidirectional GRU layer and obtains the final time series features through splicing. Based on the final time series features, the power load forecasting is performed through the fully connected layer to obtain the load forecasting results; The load forecast result is obtained by the following formula: y t =W out ·h t +b out , In the formula, y t represents the output of the optimized trained GLTN-BiGRU network model at time t, that is, the load prediction of the optimized trained GLTN-BiGRU network model for the current time step; W out represents the weight of the fully connected layer, b out represents the bias of the fully connected layer, h t Represents the final timing characteristics.

9. A multi-objective load forecasting and adjustment method according to claim 8, characterized in that: The optimized trained GLTN model uses the self-attention mechanism to extract the global and local temporal features of the input data, and performs feature fusion to obtain the final fused features, including: The GLTN model includes the attention mechanism layer, LocalTemporal module, GlobalTrend module and feature fusion module; The attention mechanism layer maps the input data to a high dimension to capture the feature information in the input data to obtain the embedded feature sequence, and sends it to the attention mechanism layer and the LocalTemporal module respectively; The attention mechanism layer uses the self-attention mechanism to re-encode the input data to obtain the encoded input data, and calculates the output features of the attention mechanism in combination with the embedded feature sequence; The attention mechanism layer transmits the encoded input data to the LocalTemporal module and the GlobalTrend module; The LocalTemporal module uses a convolutional neural network to extract detail features from the encoded input data, obtains the local temporal features of the input data and sends them to the feature fusion module; The GlobalTrend module uses a residual network to extract global time features from the encoded input data, obtains the global time features of the input data and sends them to the feature fusion module; The feature fusion module uses a weighted average fusion method to fuse the global time features and the local time features to obtain a fused feature, and concatenates the fused feature with the output feature of the attention mechanism to obtain a final fused feature.

10. A multi-objective load forecasting and adjustment method according to claim 8 or 9, characterized in that: The optimized trained BiGRU model captures the bidirectional information flow of the final fusion features through the bidirectional GRU layer, and obtains the final temporal features through splicing, including: The optimized trained BiGRU model captures the forward and backward information regularity and data forward and backward dependencies in the final fusion features through the bidirectional GRU layer, so as to fully mine the temporal features in the final fusion features and obtain forward hidden information and reverse hidden information. The optimized trained BiGRU model concatenates the forward hidden information and the reverse hidden information to obtain the final time series features: In the formula, h t Represents the final timing characteristics, represents the forward hidden information at time t, Indicates reverse hidden information; The bidirectional GRU layer includes a forward GRU layer and a reverse GRU layer, and each GRU layer includes a reset gate and an update gate.

11. A multi-objective load forecasting and adjustment method according to claim 10, characterized in that: The optimized trained BiGRU model captures the forward information regularity in the final fusion feature through the forward GRU layer, and obtains the forward hidden information as follows: In the formula, represents the state of the update gate in the forward GRU layer, σ represents the Sigmoid function, and Both represent the weights of the update gate of the forward GRU layer, H f Represents the final fusion features output by the GLTN model after optimization training; Represents the final fusion feature H f The features at time t, that is, the input information of the optimized trained BiGRU model at time t; represents the forward hidden information at time t-1, r t fwd represents the state of the reset gate in the forward GRU layer, and Both represent the weights of the reset gate of the forward GRU layer, Represents the state of the candidate state information in the forward GRU layer, U fwd and W fwd Both represent the weights of the candidate state information of the forward GRU layer, represents the Hadmard product, Represents the forward hidden information at time t; The optimized trained BiGRU model captures the reverse information regularity in the final fusion feature through the reverse GRU layer, and obtains the backward hidden information as follows: In the formula, represents the state of the update gate in the reverse GRU layer, σ represents the Sigmoid function, and Both represent the weights of the update gate of the reverse GRU layer, represents the reverse hidden information at the t+1th moment (i.e. the previous moment), r t bwd represents the state of the reset gate in the reverse GRU layer, and Both represent the weights of the reset gate of the reverse GRU layer, The state representing the candidate state information of the reverse GRU layer, Indicates reverse hidden information.

12. A multi-objective load forecasting and adjustment method according to claim 1, characterized in that: The load forecasting results are used to formulate a load configuration plan to achieve multi-objective load regulation, including: Based on the load forecast results, historical electricity consumption data of the user's demand-side response period is analyzed to obtain the comprehensive demand-side response potential of each user; According to the comprehensive demand-side response potential of each user and the given total load energy-saving index, formulate energy-saving index for each user to achieve multi-objective load regulation; If the load distribution after the initial multi-objective load adjustment does not achieve the desired balance effect, the load value after the multi-objective load adjustment is iteratively optimized until the optimization condition is met and the iteration is stopped; Wherein, the optimization condition is: In the formula, represents the load at the trough moment t2, and ò represents the acceptable threshold of load fluctuation.

13. A multi-objective load forecasting and adjustment method according to claim 12, characterized in that: Based on the load forecast results, the historical electricity consumption data of the user's demand side response period is analyzed to obtain the comprehensive demand side response potential of each user, including: Analyze the load forecast result curve based on the load forecast result; Based on the load forecast result curve, statistically analyze the historical electricity consumption data of users during the demand-side response period on typical days, and calculate the load dispersion of each user; The maximum response capacity of users is obtained by analyzing the load forecast results. Combined with the load dispersion of each user, a user's demand-side response capacity determination model for a single hour is set to calculate the demand-side response potential of each user in a single response period. Based on the demand-side response potential of each user in a single response period, the comprehensive demand-side response potential of each user is calculated; The load dispersion of each user includes: the power consumption volatility of each user α, the minimum power load x of each user’s demand side response period hmin , and the average power load x of each user's demand-side response period av .

14. A multi-objective load forecasting and adjustment method according to claim 13, characterized in that: The power consumption volatility of each user is calculated by the following formula: In the formula, α v Indicates the user volatility of a certain user in a certain period of time, n v represents the number of demand-side response sessions conducted, xi represents the power load of the user in the i-th demand-side response period, x av Indicates the average power load of the user during the demand side response period; The user's demand-side response capability determination model for a single hour is: In the formula, β j represents the demand-side response potential of user j in the demand-side response period, γi represents the maximum response capacity of user j in the ith demand-side response period; x hmin represents the minimum power load of user j during the demand side response period; The comprehensive demand-side response potential of each user is calculated by the following formula: In the formula, δ j represents the comprehensive demand-side response potential of user j, and m represents the total number of users.

15. A multi-objective load forecasting and adjustment method according to claim 12, characterized in that: According to the comprehensive demand-side response potential of each user and the given total load energy-saving index, formulate energy-saving indicators for each user, including: Taking into account the comprehensive demand-side response potential of each user, the willingness level of each user, the response coordination degree of each user, the load importance, and the economic level of each user, the energy-saving index of each user is set based on the given total load energy-saving index as follows: W j =α'×δ j ×(x av -x h min ), Where W j represents the energy saving index of user j, α' represents the comprehensive weight of user j's actual willingness level, response coordination, load importance and economic level in the i-th response period, δ j represents the comprehensive demand-side response potential of user j, x av represents the average power load of user j during the demand-side response period, x hmin represents the minimum power load of user j in the i-th demand-side response period.

16. A multi-objective load forecasting and regulation method according to claim 1 or 12, characterized in that: In the multi-objective regulation process, the load optimization objective function and load constraints are constructed, including: In the multi-objective regulation process, the objective is to minimize load volatility and construct the load optimization objective function as follows: Where φ(t) represents the load optimization objective function, w1 and w2 represent weight factors, P t represents the load at the tth moment, t=1,2,...,T; U t represents the user satisfaction with load allocation at time t, g(U t ) represents user satisfaction U t The reward function of The load constraints include: upper and lower limit constraints, and load reduction and filling constraints; The upper and lower limit constraints are: after load adjustment, the load shall not exceed the maximum load, and the minimum load shall not be lower than the minimum load; wherein, the upper and lower limit constraints are expressed as: P min ≤P t +ΔP t ≤P max ≤P t ×η, In the formula, p min represents the minimum capacity, Δp represents the load change, P max represents the maximum capacity, and η represents the maximum acceptable adjustment ratio of the user; The load reduction and filling constraints are: for each peak period t1, if the load P t1 Exceeding the maximum capacity P max , then the reduction part is: In the formula, Indicates the power cut during the peak period t1; The reduced load will be accumulated to fill in the valley period t2: In the formula, Indicates the load at the trough moment t2, ΔP h Indicates the load change at peak load, N lw Indicates the number of load troughs.

17. A multi-objective load forecasting and regulation system, characterized in that: The system comprises: A decomposition unit, used for performing modal decomposition on the power load data by using a VMD algorithm to obtain decomposed load data; A construction unit, used to construct a GLTN-BiGRU network model, and optimize and train the GLTN-BiGRU network model; A prediction unit is used to input the decomposed load data into the optimized and trained GLTN-BiGRU network model to perform future short-term power load prediction and obtain load prediction results; The regulating unit is used to formulate a load configuration plan using the load forecasting result to achieve multi-objective load regulation.

18. An electronic device, characterized in that: It includes a memory, a processor and a computer program stored in the memory, and the processor executes the computer program or instructions to implement a multi-objective load forecasting and regulation method as described in any one of claims 1-16.

19. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program or instruction, and when the computer program or instruction is executed by a processor, a multi-objective load forecasting and adjustment method as described in any one of claims 1-16 is implemented.

20. A computer program product comprising a computer program or instructions, characterized in that When the computer program or instruction is executed by a processor, a multi-objective load forecasting and adjustment method as described in any one of claims 1-16 is implemented.

Images