Power grid dispatching method and device based on flexible resource adjustment capability

Abstract

The invention discloses a power grid dispatching method and device based on flexibility resource adjustment capability, and belongs to the field of power systems, and the method comprises the steps: obtaining historical adjustment capability data and covariable data of flexibility resources, and constructing an input data set; inputting the input data set into an adjustment capability prediction model constructed based on an Encoder-Decoder architecture, a Bayesian neural network and a Transform algorithm to obtain an adjustment capability prediction result set of each future time period; verifying by adopting a probability prediction theory and a Bayesian gating cycle unit comparison algorithm, and taking the result as a final regulation capability probability distribution result after passing verification; therefore, by implementing the method and the device, the problems of difficulty in quantitative estimation of the adjustment capability of the flexible resources in the power grid and insufficient scheduling accuracy in the prior art can be solved.

Description

Technical Field

[0001] The present invention relates to the field of power systems, and in particular to a power grid dispatching method and device based on flexibility resource regulation capability. Background Art

[0002] In modern power systems, the high volatility caused by the integration of high levels of renewable energy and the widening load peak-to-valley gap present a severe challenge: an imbalance between the supply and demand of flexibility resources. As a key regulatory tool for grid stability, the precise scheduling of flexibility resources relies on accurate predictions of their adjustable capacity.

[0003] In the existing technology, for example, patent CN119324449A can realize the quantitative evaluation of flexibility resources, but it focuses on static capacity analysis rather than dynamic prediction; although patent CN119647731A introduces the attention mechanism to improve the prediction model, it still adopts a deterministic prediction framework and cannot quantify the uncertainty caused by the randomness of user charging behavior. Therefore, there is a problem that the prediction results lack the ability to measure risk. Specifically, it is manifested as follows: ① The deterministic prediction outputs a single value, which cannot reflect the actual fluctuation range; ② The dispatcher cannot know the reliability of the predicted value. Therefore, there is an urgent need for a technical solution that can solve the problem in the existing technology that it is difficult to quantitatively estimate the regulation capacity of flexibility resources in the power grid, while improving the accuracy of the scheduling of flexibility resources in the power grid. Summary of the Invention

[0004] The present invention provides a power grid dispatching method and device based on the flexibility resource regulation capability, which can solve the problem in the prior art that it is difficult to quantitatively estimate the regulation capability of flexibility resources in the power grid while improving the accuracy of flexibility resource dispatching in the power grid.

[0005] In a first aspect, an embodiment of the present invention provides a power grid scheduling method based on flexibility resource adjustment capability, comprising: Acquiring historical adjustment capability data and covariate data of the flexibility resource, and constructing an input data set based on the historical adjustment capability data and the covariate data; Inputting the input data set into a regulation capacity prediction model, outputting regulation capacity prediction results of the flexibility resource in each future time period, and combining all regulation capacity prediction results into a regulation capacity prediction result set; wherein the regulation capacity prediction model is constructed based on an encoder-decoder architecture, a Bayesian neural network, and a transformer algorithm; According to the adjustment capacity prediction result set, each adjustment capacity prediction result in the adjustment capacity prediction result set is verified by using a probability prediction theory and a Bayesian gate recurrent unit comparison algorithm, and if the verification is passed, the adjustment capacity prediction result set is taken as a final adjustment capacity probability distribution result; The flexible resources in the power grid are dispatched according to the final adjustment capacity probability distribution result.

[0006] The embodiments of the present application construct an input data set by obtaining historical adjustment capacity data and related covariates of the flexible resources, to provide sufficient data basis for prediction; input the input data set into an adjustment capacity prediction model of a fusion Encoder-Decoder architecture, a Bayesian neural network and a Transformer algorithm, which can capture the time sequence dependence of the adjustment capacity by means of the self-attention mechanism of the Transformer, and define the model parameters as a probability distribution through the BNN, to realize the quantification of uncertainty while ensuring the prediction accuracy, and output multi-period prediction results; then verify by using a probability prediction theory and a Bayesian gate recurrent unit comparison algorithm, and ensure the result reliability by combining the deterministic and probabilistic indicators, to finally obtain the adjustment capacity probability distribution result, and dispatch the flexible resources in the power grid according to the final adjustment capacity probability distribution result. Therefore, the present application can solve the problem that it is difficult to quantitatively estimate the adjustment capacity of the flexible resources in the power grid in the prior art, and improve the accuracy of the flexible resource scheduling in the power grid.

[0007] As a preferred example of the first aspect, the historical adjustment capacity data and the covariate data of the flexible resources are obtained, and the input data set is constructed according to the historical adjustment capacity data and the covariate data, specifically: The historical adjustment capacity data and the covariate data of the flexible resources in a preset historical period are obtained; wherein the historical adjustment capacity data includes flexible resource quantity data and flexible resource adjustment capacity data, and the covariate data includes time feature data, weather feature data and electricity behavior feature data, and the covariate data includes flexible resource quantity; The historical adjustment capacity data and the covariate data are aligned according to time sequence to obtain the input data set.

[0008] In the preferred example, the historical adjustment capacity data and the covariate data in the preset historical period are obtained, which can comprehensively cover the core factors and related features that affect the adjustment capacity of the flexible resources, to ensure the integrity and relevance of the input data; the two types of data are aligned according to time sequence, which can ensure the consistency of the data in the time dimension, and avoid information deviation caused by time sequence misalignment.

[0009] As a preferred example of the first aspect, the regulation capacity prediction model is constructed based on an Encoder-Decoder architecture, a Bayesian neural network, and a Transformer algorithm, specifically: A basic model is established based on the Encoder-Decoder architecture as a basic framework; The multi-head self-attention mechanism of the Transformer algorithm and the Bayesian neural network are integrated into the basic model to obtain the regulation capacity prediction model.

[0010] In this preferred example, the basic model is established based on the Encoder-Decoder architecture as a basic framework, which can effectively realize the mapping of historical data sequences to future prediction sequences and provide a structured input-output processing mechanism for regulation capacity prediction. After integrating the multi-head self-attention mechanism of the Transformer algorithm, the model can capture the dependency relationship of flexible resource regulation capacity at different time points, focus on key timing information to improve learning effectiveness, and combine the Bayesian neural network to define model parameters as probability distributions, which can quantify the uncertainty in regulation capacity prediction, so that the final regulation capacity prediction model can accurately capture the timing variation of regulation capacity and reliably represent the prediction uncertainty caused by the randomness of flexible resource electricity consumption behavior.

[0011] As a preferred example of the first aspect, the input data set is input into the regulation capacity prediction model, and the regulation capacity prediction results of the flexible resource in each future time period are output, and all the regulation capacity prediction results are combined to form a regulation capacity prediction result set, specifically: The neural network weight and bias parameters in the regulation capacity prediction model are set as a prior distribution, and the posterior distribution is inferred using the Bayesian formula according to the input data set; The posterior distribution is approximated to a real posterior distribution using a Gaussian distribution; Sampling is performed from the real posterior distribution to obtain a plurality of sampling parameters, and the sampling parameters are forward propagated to output the regulation capacity prediction results of the flexible resource in each future time period, and all the regulation capacity prediction results are combined to form a regulation capacity prediction result set.

[0012] In this preferred example, the posterior distribution is approximated to a Gaussian distribution, which can simplify the complexity of solving the posterior distribution while retaining its core features, providing a feasible path for quantifying the uncertainty of model parameters. Sampling from the approximated real posterior distribution and forward propagating the sampling parameters can generate multiple sets of regulation capacity prediction results, which can comprehensively reflect the probability distribution characteristics of the regulation capacity of flexible resources in each future time period and effectively capture the uncertainty caused by the randomness of electricity consumption behavior.

[0013] As a preferred example of the first aspect, the adjustment capacity prediction result set is verified by using a probability prediction theory and a Bayesian gated recurrent unit comparison algorithm, and if the verification is passed, the adjustment capacity prediction result set is taken as the final adjustment capacity probability distribution result, specifically: According to the adjustment capacity prediction result set, the probability prediction theory and the Bayesian gated recurrent unit comparison algorithm are used for processing respectively to obtain a probabilistic index and a deterministic index; If the probabilistic index and the deterministic index both satisfy a preset condition, the adjustment capacity prediction result set is taken as the final adjustment capacity probability distribution result.

[0014] In the preferred example, the probability prediction theory and the Bayesian gated recurrent unit comparison algorithm are used for processing respectively according to the adjustment capacity prediction result set to obtain a probabilistic index and a deterministic index, which can comprehensively test the reliability and accuracy of the prediction result from different dimensions, paying attention to the deviation between the predicted value and the actual value and also valuing the quantification effect of uncertainty; when both types of indexes satisfy the preset condition, it indicates that the prediction result reaches the expected requirement in terms of deterministic accuracy and probability distribution representation, and at this time, it is taken as the final adjustment capacity probability distribution result, which can ensure that the result can reliably reflect the future change rule and uncertainty characteristics of the flexibility resource adjustment capacity.

[0015] In the second aspect, the application provides a power grid dispatching device based on flexibility resource adjustment capacity, which comprises a data acquisition module, a data processing module, a data verification module and a dispatching module. The data acquisition module is used for acquiring historical adjustment capacity data and covariant data of the flexibility resource, and constructing an input data set according to the historical adjustment capacity data and the covariant data; The data processing module is used for inputting the input data set into an adjustment capacity prediction model, outputting adjustment capacity prediction results of the flexibility resource in each future time period, and composing an adjustment capacity prediction result set by using all the adjustment capacity prediction results; wherein the adjustment capacity prediction model is constructed based on an Encoder-Decoder architecture, a Bayesian neural network and a Transformer algorithm; The data verification module is configured to verify each regulation capacity prediction result in the regulation capacity prediction result set by using a probability prediction theory and a Bayesian gate recurrent unit comparison algorithm, and if the verification is passed, the regulation capacity prediction result set is taken as a final regulation capacity probability distribution result, and the flexible resource in the power grid is scheduled according to the final regulation capacity probability distribution result.

[0016] The scheduling module is configured to schedule the flexible resource in the power grid according to the final regulation capacity probability distribution result.

[0017] As a preferred example of the second aspect, the data acquisition module comprises a first acquisition unit and a second acquisition unit. The first acquisition unit is configured to acquire historical regulation capacity data and covariate data of the flexible resource in a preset historical period, wherein the historical regulation capacity data comprises flexible resource quantity data and flexible resource regulation capacity data, and the covariate data comprises time characteristic data, weather characteristic data and electricity consumption behavior characteristic data, and the covariate data comprises flexible resource quantity. The second acquisition unit is configured to align the historical regulation capacity data and the covariate data according to a time sequence to obtain the input data set.

[0018] As a preferred example of the second aspect, the data processing module comprises a first processing unit and a second processing unit. The first processing unit is configured to establish a basic model by taking the Encoder-Decoder architecture as a basic framework. The second processing unit is configured to integrate a multi-head self-attention mechanism of the Transformer algorithm and a Bayesian neural network into the basic model to obtain the regulation capacity prediction model.

[0019] As a preferred example of the second aspect, the data processing module further comprises a third processing unit, a fourth processing unit and a fifth processing unit. The third processing unit is configured to set each layer neural network weight and bias parameter in the regulation capacity prediction model as a prior distribution, and infer a posterior distribution according to the input data set by using a Bayesian formula. The fourth processing unit is configured to perform real posterior distribution approximation processing on the posterior distribution by using a Gaussian distribution to obtain a real posterior distribution. The fifth processing unit is configured to sample from the real posterior distribution to obtain a plurality of sampling parameters, perform forward propagation on each sampling parameter, output an adjustment capability prediction result of the flexible resource in each future time period, and group all the adjustment capability prediction results into an adjustment capability prediction result set.

[0020] As a preferred example of the second aspect, the data verification module comprises a first verification unit and a second verification unit. The first verification unit is configured to process the adjustment capability prediction result set by a probability prediction theory and a Bayesian gate recurrent unit comparison algorithm respectively to obtain a probabilistic index and a deterministic index according to the adjustment capability prediction result set. The second verification unit is configured to take the adjustment capability prediction result set as a final adjustment capability probability distribution result if the probabilistic index and the deterministic index both satisfy a preset condition.

[0021] In summary, the embodiments of the present application construct an input data set by obtaining historical adjustment capability data of flexible resources and related covariates, providing sufficient data basis for prediction; input the input data set into an adjustment capability prediction model of a fusion Encoder-Decoder architecture, a Bayesian neural network and a Transformer algorithm, which can capture the time sequence dependence of adjustment capability by means of the self-attention mechanism of the Transformer, and at the same time, define the model parameters as a probability distribution through the BNN to realize the quantification of uncertainty while ensuring the prediction accuracy, and output multi-period prediction results; then verify by a probability prediction theory and a Bayesian gate recurrent unit comparison algorithm, combine the deterministic and probabilistic indexes to ensure the reliability of the results, finally obtain the adjustment capability probability distribution result, and schedule the flexible resources in the power grid according to the final adjustment capability probability distribution result. Therefore, the present application can solve the problem that it is difficult to quantitatively estimate the adjustment capability of flexible resources in the power grid in the prior art, and improve the accuracy of scheduling flexible resources in the power grid.

[0022] Another embodiment of the present application further provides a terminal device, comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, and when the computer program is executed by the processor, the steps of the power grid scheduling method based on the adjustment capability of flexible resources are realized.

[0023] Another embodiment of the present application further provides a computer readable storage medium item, comprising a stored computer program, and when the computer program is running, the device where the computer readable storage medium is located is controlled to execute the steps of the power grid scheduling method based on the adjustment capability of flexible resources. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. Obviously, the drawings described below are only some of the embodiments of the present application, and other drawings can also be obtained by those skilled in the art without any creative effort.

[0025] Figure 1 A flowchart of an embodiment of a power grid scheduling method based on flexibility resource regulation capability provided by the present application; Figure 2 An Encoder-Decoder architecture diagram of an embodiment of a power grid scheduling method based on flexibility resource regulation capability provided by the present application; Figure 3 A Bayesian neural network diagram of an embodiment of a power grid scheduling method based on flexibility resource regulation capability provided by the present application; Figure 4 A mask principle diagram of an embodiment of a power grid scheduling method based on flexibility resource regulation capability provided by the present application; Figure 5 A module structure diagram of an embodiment of a power grid scheduling device based on flexibility resource regulation capability provided by the present application. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solutions and advantages of the present application clearer, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without any creative effort are within the scope of protection of the present application.

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs; the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit the present application; the terms "include" and "have" and any variations thereof in the specification and claims of the present application and the above description of drawings are intended to cover non-exclusive inclusion.

[0028] In the description of the embodiments of the present application, the technical terms "first", "second", etc. are only used to distinguish different objects, and cannot be understood as indicating or implying relative importance or implicitly indicating the number, specific order or primary and secondary relationship of the indicated technical features. In the description of the embodiments of the present application, the meaning of "a plurality of" is two or more, unless otherwise explicitly specified.

[0029] Reference to an "embodiment" herein means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the application. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are they necessarily mutually exclusive of one another. As will be apparent to one of ordinary skill in the art, embodiments described herein can be combined with other embodiments.

[0030] In the description of the embodiments of the application, the term "and / or" is merely an association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / " herein generally represents an "or" relationship between the associated objects before and after it.

[0031] In the description of the embodiments of the application, the term "a plurality of" refers to two or more (including two), and similarly, "a plurality of groups" refers to two or more groups (including two groups), and "a plurality of pieces" refers to two or more pieces (including two pieces).

[0032] In the description of the embodiments of the application, unless otherwise explicitly specified and limited, the technical terms "mounting", "connection", "connection", "fixing" and the like should be understood in a broad sense, for example, it can be fixedly connected, or it can be detachably connected, or it can be integrated; it can be mechanical connection, or it can be electrical connection; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be the internal communication of two elements or the interaction relationship between two elements. For those skilled in the art, the specific meaning of the above terms in the embodiments of the application can be understood according to the specific circumstances.

[0033] Embodiment one Reference Figure 1 To solve the problem that it is difficult to quantitatively estimate the regulation capacity of flexible resources in the power grid in the prior art, and to improve the accuracy of the scheduling of flexible resources in the power grid, an embodiment of the application provides a power grid scheduling method based on the regulation capacity of flexible resources, which comprises the following steps: S1, obtaining historical regulation capacity data and covariant data of the flexible resources, and constructing an input data set according to the historical regulation capacity data and the covariant data.

[0034] In some embodiments of the application, the obtaining of the historical regulation capacity data and the covariant data of the flexible resources, and the construction of the input data set according to the historical regulation capacity data and the covariant data are specifically as follows: obtain historical adjustment capability data and covariate data of the flexibility resource in a preset historical period; wherein, the historical adjustment capability data comprises flexibility resource quantity data and flexibility resource adjustment capability data, the covariate data comprises time characteristic data, weather characteristic data and electricity consumption behavior characteristic data, and the covariate data comprises flexibility resource quantity; align the historical adjustment capability data and the covariate data according to time sequence to obtain the input data set.

[0035] It should be noted that the prediction problem of the flexibility resource adjustment capability can be modeled as a time series prediction problem, and the specific formula is as follows: wherein, represents a to-be-predicted quantity, represents adjustment capability at L previous time points, represents a value of the to-be-predicted quantity at H future time points, represents covariate values at the L previous time points, and F represents a time series prediction model, represents an actual value, represents a predicted value, is the number of data points, and are mean absolute error and root mean square error respectively, used for verifying the prediction result.

[0036] S2, input the input data set into an adjustment capability prediction model, output adjustment capability prediction results of the flexibility resource in each future period, and group all the adjustment capability prediction results to obtain an adjustment capability prediction result set; wherein, the adjustment capability prediction model is constructed based on an Encoder-Decoder architecture, a Bayesian neural network and a Transformer algorithm.

[0037] It should be noted that, in order to realize flexibility resource adjustment capability prediction in a future period, the input of the prediction model is adjustment capability and related variables in a previous period, and the output is adjustment capability in the prediction period. This form of inputting a sequence and outputting a sequence is defined as a Seq2Seq model.

[0038] The Encoder-Decoder architecture is a structure widely used in Seq2Seq. With the development of deep learning, this structure has evolved into various model variants, and has outstanding performance in many fields such as machine translation, text summarization and speech recognition. The core of this architecture is to use an encoder and a decoder to process input and output data, and the specific formula is as follows: Figure 2As shown in the figure, the encoder can encode the input sequence into a fixed-length vector and extract core features from the input data through a series of neural network layers, including convolutional layers, recurrent layers, and attention layers, etc. After processing by the encoder, the original data retains the main information while removing redundant information and noise, enabling subsequent operations to focus more on important information. In the implementation process, the encoder updates its internal state according to the real-time input data at each time step, generating a fixed-length vector that captures the semantic and structural information of the input sequence. The decoder receives the fixed-length vector passed from the encoder and processes the information through multiple neural network layers to generate the final output sequence. The structure and training method of the decoder differ from that of the encoder, as it generates an element of the output sequence at each time step while updating its internal parameters. Since the output sequence of the encoder has a fixed length, the process of generating the output sequence usually starts from a start symbol and ends with an end symbol.

[0039] In general, the Encoder-Decoder architecture reads the input sequence step by step through the encoder in the encoding phase and generates a fixed-length vector c. In the decoding phase, the decoder receives the context vector and generates the target sequence step by step The traditional Encoder-Decoder architecture has a problem of losing key information when processing long sequence inputs into fixed-length vectors. To solve this problem, the Attention mechanism dynamically focuses on different parts of the input sequence during the decoding process to improve the model's performance on long time sequence tasks. Taking an encoder composed of RNN as an example, after the encoder processes the input sequence, the correlation between the final time vector and any time vector is calculated, and the Softmax function is used to normalize the attention weight distribution. Then, based on the obtained attention score, the vectors at each time step of the encoder are weighted and summed, so that the model will pay special attention to the data at the time step with a large attention score, thus obtaining a vector with attention weighting. Through the processing of the Attention mechanism, the decoder can automatically focus on the important parts of the input sequence in the decoding phase, and finally generate a complete output sequence through iteration. The Attention mechanism enhances the model's ability to process important parts of long time sequences to improve the accuracy of the prediction results.

[0040] The Transformer is a model based on an Encoder-Decoder architecture, which uses a multi-head self-attention mechanism instead of neural network layers in the original model to enhance the model's attention to important parts in long time series. The encoder of the Transformer is composed of multiple attention layers and feedforward layers, and residual connection layers and normalization operations are added after each attention layer and feedforward layer to solve the problem of gradient disappearance in the training process, making the model training more stable and faster.

[0041] The attention mechanism in the Transformer model includes three important matrices, namely Query matrix, Key matrix and Value matrix, which represent the information demand of the current position, the information index and the true value of each position respectively. Among them, the Query matrix and the Key matrix can be matched to find the key position, and the value in the Value matrix is corresponding to the Key matrix. When the Query matches a certain Key, the corresponding Value will be used to calculate the output. In the calculation process, dot product operation is needed for weights and matrices. The dot product value in high-dimensional space will increase exponentially with the increase of dimension, so a scaling factor , i.e. the square root of the input vector dimension, is introduced to adjust the similarity score to a reasonable interval, avoiding the gradient disappearance caused by the input value being too large in the subsequent Softmax function. Then the scaled dot product result is input into the Softmax function for normalization to get the probability distribution of attention weights, ensuring that the sum of attention weights of all positions is 1. The normalized attention weight matrix and the Value matrix are weighted and summed to get the final output. The generated attention output matrix integrates global information and local features, enabling the model to adaptively focus on high-value information areas in the sequence.

[0042] In some embodiments of the present application, the adjustment ability prediction model is constructed based on the Encoder-Decoder architecture, Bayesian neural network and Transformer algorithm, specifically: The Encoder-Decoder architecture is used as the basic framework to establish a basic model. The multi-head self-attention mechanism of the Transformer algorithm and the Bayesian neural network are integrated into the basic model to obtain the adjustment ability prediction model.

[0043] It should be noted that the network parameters and the prediction results in the deterministic prediction model are a certain numerical value, however, it is difficult to accurately predict a certain value for a subject with high randomness such as flexibility resource. Therefore, the traditional deterministic prediction model is not suitable for the prediction of flexibility resource adjustment capacity, BNN defines the parameters of each layer of neural network as a specific distribution, usually a Gaussian distribution characterized by mean μ and standard deviation σ, the mean and standard deviation of the distribution are updated during the training process, and the distribution information of the predicted value is obtained by sampling the neural network parameters multiple times after the training is completed, this method can significantly improve the model's processing of uncertainty, and the BNN network can be seen from the following figure Figure 3 In this embodiment, BNN is introduced into the Transformer prediction model with multi-head attention mechanism, the parameters in the Transformer model are set as probability distribution, so that the model retains the powerful ability of Transformer in processing complex sequence data while enhancing the model's quantitative ability for uncertainty.

[0044] In some optional embodiments of the present application, the Encoder-Decoder architecture can be used as the input and output sequence of the adjustment capacity prediction model, and the specific process is as follows: The Encoder-Decoder architecture compresses the input sequence into a fixed-length context vector through the encoder in the encoding stage, and the decoder generates the target sequence step by step based on the vector; The input sequence is the adjustment capacity value and the covariate at the historical L time points, and the output sequence is the prediction value of the adjustment capacity at the future H time points.

[0045] In some optional embodiments of the present application, the features of the adjustment capacity prediction model are as follows: The similarity of Query matrix, Key matrix and Value matrix is calculated by parallel h attention heads to generate scaled dot-product attention weights: wherein, , and represent the weights of Query matrix, Key matrix and Value matrix respectively, represents a scaling factor, and the Softmax function is a function for converting a real number vector into a probability distribution with a sum of 1, given an input vector , the definition of the Softmax function is as follows: wherein, represents the i-th element of the input vector, Exponential function is represented, and the Query matrix, Key matrix and Value matrix are independently calculated by using h attention heads, and the corresponding attention weights are obtained, and then the h output results are spliced and input into a linear layer to integrate multiple outputs to generate a multi-head fused output: The full connection layer is composed of two MLPs, wherein, and are parameters of the first MLP, and are parameters of the second MLP.

[0046] The decoder of the Transformer is also composed of multiple decoder layers, and each decoding layer is composed of a masked self-attention layer, a self-attention layer and a feedforward neural network. The masked self-attention layer is based on the self-attention layer described above and adds a mask operation to ensure that only the current position and the information before it can be accessed during decoding, and the future information is shielded to ensure the rationality of the model and the accuracy of the prediction. There are two types of masks in the Transformer, namely Padding Mask (padding mask) and Sequence Mask (sequence mask). If the lengths of the input data are inconsistent, the shorter sequences will be padded to the length of the longest sequence, and the Padding Mask is used to mark these padding positions so that the model ignores these invalid information when calculating. In this embodiment, Sequence Mask is used, and the model can only rely on the current data and the information before it when predicting the next sequence, and cannot see the future data. Sequence Mask realizes this by constructing a lower triangular matrix, and the part below the diagonal line is 1, and the part above the diagonal line is covered with an inhibition symbol, so that the future information is shielded. As shown in Figure 4 , the implementation of the mask is to cover the mask area with an inhibition symbol along the diagonal line of the matrix, so that the model cannot see the future information.

[0047] In some embodiments of the present application, the input data set is input into the regulation capacity prediction model, and the regulation capacity prediction results of the flexible resource in each future period are output, and all the regulation capacity prediction results are combined into a regulation capacity prediction result set, specifically: The neural network weights and bias parameters of each layer of the regulation capacity prediction model are set as a prior distribution, and the posterior distribution is inferred according to the input data set by using the Bayesian formula; The posterior distribution is approximated to the real distribution by using Gaussian distribution to obtain the real posterior distribution; Sampling from the real posterior distribution, obtaining a plurality of sampling parameters, and forward propagating each sampling parameter to output the regulation capacity prediction result of the flexible resource in each future time period, and combining all regulation capacity prediction results to form a regulation capacity prediction result set.

[0048] In some optional embodiments of the present application, the inputting of the input data set into the regulation capacity prediction model and the outputting of the regulation capacity prediction result of the flexible resource in each future time period can be implemented by the following method: On the basis of preserving the structure of the Transformer model, a prior distribution is set for the weight and bias parameters of each layer of neural network Based on the training data set D, the model infers the posterior distribution by calculation Wherein, the prior distribution of each weight and bias adopts the form of a mixed Gaussian distribution, which is expressed as: Wherein, the mean of each distribution is 0, and the variances are and The mixing coefficients are and The posterior distribution is calculated based on the training data set data, and the formula is: The posterior distribution cannot be directly solved, and the variational inference is introduced to use the parameter to construct a Gaussian distribution to approximate the real posterior distribution, and the formula is: Wherein, the parameter is composed of the mean and the standard deviation , and the objective function is further simplified to the lower bound of the evidence: The backpropagation Bayesian method is used for solving, and the model can solve the optimal distribution after completing the training, so that the test stage of the model samples the distribution of the prediction result from the posterior distribution of the optimized network parameters, and outputs the regulation capacity prediction result of different confidence intervals in the future time period.

[0049] S3, according to the regulation capacity prediction result set, using the probability prediction theory and the Bayesian gate recurrent unit comparison algorithm, verifying each regulation capacity prediction result in the regulation capacity prediction result set, if the verification is passed, the regulation capacity prediction result set is taken as the final regulation capacity probability distribution result.

[0050] In certain embodiments of the present application, the set of regulatory ability prediction results is verified using probability prediction theory and Bayesian gated recurrent unit comparison algorithm based on the set of regulatory ability prediction results. If the verification passes, the set of regulatory ability prediction results is used as the final regulatory ability probability distribution result, specifically: According to the set of regulatory capacity prediction results, probabilistic prediction theory and Bayesian gated recurrent unit comparison algorithm are used to process them respectively to obtain a probabilistic index and a deterministic index; If both the probabilistic index and the deterministic index meet the preset conditions, the set of adjustment capability prediction results is used as the final adjustment capability probability distribution result.

[0051] It's important to note that probabilistic forecasting methods can be broadly categorized into three types: quantile forecasting, interval forecasting, and density forecasting. These methods provide prediction results at specific quantiles, specific confidence intervals, and probability distribution characteristics, respectively. Compared to traditional deterministic forecasting, probabilistic forecasting not only provides predicted values ​​but also characterizes the fluctuation range and uncertainty of the predicted results, providing a more reliable and comprehensive representation of the forecast results.

[0052] Currently, a widely used probabilistic forecasting method is quantile regression, which focuses on predicting specific quantiles of the data rather than the overall mean. Quantile regression offers the advantages of high flexibility and adaptability. It can be combined with various established point forecasting methods through simple operations such as modifying the loss function, allowing users to select the appropriate method based on their specific needs. Furthermore, this method accurately adapts to complex distributional characteristics such as data skewness and kurtosis, overcoming the limitations of traditional regression methods. Quantiles divide the probability distribution range of a random variable into several equal points. Commonly used are the median (i.e., bisectors), quartiles (the 25th, 50th, and 75th percentiles), and percentiles. Traditional time series forecasting methods typically regress the data to the mean, but quantile regression can analyze forecasts at different quantiles and estimate the relationship between different quantitative values ​​of the conditional distribution of the independent and dependent variables.

[0053] In terms of technical implementation, quantile regression based on gradient descent requires adjusting the loss function to the following form so that the learning function can better approximate the quantiles of the real samples: in, is the output, is the target value, is the quantile; The prediction result of interval prediction can output the upper and lower bounds of the prediction variable at a given confidence level to reflect the fluctuation range of the prediction value. In practical applications, interval prediction is often used in combination with a probability density function to represent the probability distribution characteristics of the variable value, thereby effectively quantifying the uncertainty of the prediction result. For a given confidence level , the prediction interval of the predicted variable at time t can be expressed as follows: The upper and lower bounds of the prediction interval and are usually represented by the quantiles of .

[0054] To complete interval prediction, first, a preliminary prediction result should be obtained based on a deterministic prediction method, then the upper and lower bounds of the prediction interval are calculated assuming that the prediction overall error is normally distributed, and finally, the deterministic prediction result and the error distribution are integrated to construct the prediction interval.

[0055] Exemplarily, the deterministic index obtained by the probability prediction theory and the Bayesian gated recurrent unit comparison algorithm can be implemented in the following preferred manner: The probability prediction theory further includes probability density prediction, which represents the prediction result with a probability density function and a cumulative distribution function, and is mainly divided into three categories: maximum likelihood estimation, maximum a priori estimation, and Bayesian inference. For the maximum likelihood estimation method, let be the actual value, be the model estimated value, then the maximum likelihood estimation is denoted as: In the parameter estimation process, first, the to-be-estimated parameter is taken as the independent variable, and by traversing its domain, the value that maximizes the probability of the known observation result is found, which is the maximum likelihood estimation value of the model parameter; In the maximum a priori estimation method, the parameter value that maximizes the prior probability is selected as the estimation result, denoted as: The Bayesian gated recurrent unit is used to estimate the posterior distribution based on the results of the next random trial based on the existing observation data. According to the Bayesian formula, the posterior distribution cannot be directly calculated and needs to be approximately calculated by combining the prior distribution and the likelihood function. By integrating prior knowledge and observation data to obtain more accurate probability inference, the final adjustment ability probability distribution result is obtained.

[0056] wherein the sampling number is assumed to be and the sampling weight is denoted as​ The input data x can be passed to the prediction model The multiple prediction results are obtained and the expectation value and variance thereof are calculated, and in addition, the range of flexibility resource adjustment capability within a specified confidence interval can be determined according to the multiple sampling results. Assuming that the adjustment capability of the flexibility resource obeys a Gaussian distribution, the mean value and variance obtained through multiple sampling can calculate the probability distribution of the adjustment potential. The mean value and variance of the adjustment prediction can be estimated by an unbiased estimator: Therefore, the model is a modification of the Transformer architecture using a Bayesian neural network, which quantitatively represents the uncertainty while preserving the prediction accuracy, and realizes the probability prediction of the flexibility resource adjustment potential. The embodiment can effectively represent the uncertainty caused by the uncontrollable flexibility resource power consumption behavior by fitting the distribution of neural network parameters, and the probability prediction curve provided by the sampling method provides more comprehensive and reliable prediction results.

[0057] Experimental data comparison: The present application compares the existing probability prediction algorithm, and predicts in the case of single time step and multiple time steps. The prediction results of the algorithm of the present application have obvious advantages.

[0058] (1) Single time step Table 1: Comparison of evaluation indexes of single time step deterministic prediction of different algorithms As shown in Table 1, in terms of mean absolute error MAE, the MAE of the method of the present application is the optimal value of 55.53872, which is improved by 18.92% compared with the QLSTM algorithm and 40.13% compared with the BGRU algorithm. In terms of root mean square error RMSE, the method of the present application is still the optimal value, which is improved by 22.17% compared with the QLSTM algorithm and 43.57% compared with the BGRU algorithm. The above comparison results show that the prediction method proposed in the present application effectively improves the prediction accuracy.

[0059] Table 2: Comparison of evaluation indexes of single time step probability prediction of different algorithms The three indexes of Pinball, PICP and MPIW are used to evaluate the probability prediction result. Pinball can be used to evaluate the prediction effect under different confidence intervals. It can be observed from the prediction result that the proposed method is slightly inferior to QLSTM algorithm but superior to BGRU algorithm under the confidence interval of 95%. The proposed algorithm performs best under the confidence interval of 85%, which is 11.71977, followed by QLSTM. To further analyze the probability prediction result, the other two indexes are analyzed. PICP represents the prediction interval coverage probability, which reflects the reliability of prediction. The greater the value, the more the true value falls within the prediction interval. However, when the data fluctuates greatly, a higher PICP may lead to a wider prediction interval, thereby reducing the sharpness of prediction. It can be found from the prediction result that the PICP index of QLSTM algorithm is best, showing the best reliability, followed by the proposed algorithm. MPIW represents the average prediction interval width, in which the MPIW of the proposed algorithm is 189.0293, which performs best and is far superior to 419.7942 of QLSTM algorithm and 233.4119 of BGRU algorithm. According to the comparison results of the above indexes, it can be found that under the condition of single time step, the reliability of QLSTM algorithm is the highest, but the sharpness is greatly sacrificed. The reliability and sharpness of BGRU algorithm are general. Although the reliability of the proposed algorithm is slightly inferior to that of QLSTM algorithm, the sharpness is far better than that of QLSTM algorithm. In summary, the probability prediction performance of the proposed algorithm is the best.

[0060] (2) Multi-time step The prediction of 48 time steps is performed. It can be found from Table 3 that the certainty prediction evaluation indexes MAE and RMSE of the proposed method are better than those of QLSTM algorithm, which are 125.0222 and 163.9 respectively. Then the obtained results are compared with the evaluation indexes of single time step. It can be found from the comparison results that the prediction accuracy of both the proposed method and the comparative algorithm decreases with the increase of prediction time step. Because the flexible resource electricity consumption behavior has great randomness, its adjustment ability is more affected by the characteristics of nearby time period, and the multi-time step prediction model itself also has error accumulation effect. However, it can be found from the comparison results that the proposed algorithm still performs better in multi-time step, thereby verifying the effectiveness of the proposed algorithm. In practical application, for example, the flexible resource participates in the regulation of power grid and the dispatch of power market, which can make up for the error of trial prediction by real-time prediction update, so as to realize the friendly interaction between flexible resource and power grid in multi-time scale and help to maintain the real-time power balance of power system.

[0061] Table 3: Comparison of probability prediction evaluation indexes of different algorithms under single time step; As can be seen from Table 4, in terms of probability prediction evaluation indexes, the method is slightly inferior to the QLSTM algorithm in terms of Pinball and PICP, but significantly better in terms of MPIW.

[0062] Table 4: Comparison of single time step probability prediction evaluation indexes of different algorithms S4, scheduling the flexible resource in the power grid according to the final regulation capacity probability distribution result.

[0063] In summary, the embodiment of the application constructs an input data set by obtaining historical regulation capacity data and related covariates of the flexible resource, providing sufficient data basis for prediction; inputs the input data set into a regulation capacity prediction model based on the Encoder-Decoder architecture, Bayesian neural network and Transformer algorithm, can capture the time sequence dependence of the regulation capacity with the help of the self-attention mechanism of the Transformer, and at the same time defines the model parameters as a probability distribution through the BNN, realizes the quantification of uncertainty while ensuring the prediction accuracy, and outputs the multi-period prediction result; then verifies the result through the comparative algorithm of probability prediction theory and Bayesian gating recurrent unit, combines the deterministic and probabilistic indexes to ensure the reliability of the result, finally obtains the regulation capacity probability distribution result, and schedules the flexible resource in the power grid according to the final regulation capacity probability distribution result. Therefore, through the application, the problem of being difficult to quantitatively estimate the regulation capacity of the flexible resource in the power grid in the prior art can be solved, and the accuracy of the scheduling of the flexible resource in the power grid is improved.

[0064] Embodiment Two As Figure 5 shown, on the basis of the above method item embodiment, a corresponding device item embodiment is provided; An embodiment of the application provides a power grid scheduling device based on regulation capacity of flexible resource, comprising: comprising: data acquisition module 51, data processing module 52, data verification module 53 and scheduling module 54; The data acquisition module 51 is used for acquiring historical regulation capacity data and covariate data of the flexible resource, and constructing an input data set according to the historical regulation capacity data and the covariate data; The data processing module 52 is used for inputting the input data set into a regulation capacity prediction model, outputting regulation capacity prediction results of the flexible resource at each future period, and combining all regulation capacity prediction results into a regulation capacity prediction result set; wherein the regulation capacity prediction model is constructed based on the Encoder-Decoder architecture, the Bayesian neural network and the Transformer algorithm; The data verification module 53 is configured to verify each of the regulation capacity prediction results in the regulation capacity prediction result set by using a probability prediction theory and a Bayesian gate recurrent unit comparison algorithm, and if the verification is passed, the regulation capacity prediction result set is taken as a final regulation capacity probability distribution result, and the flexible resource in the power grid is dispatched according to the final regulation capacity probability distribution result.

[0065] The dispatching module 54 is configured to dispatch the flexible resource in the power grid according to the final regulation capacity probability distribution result.

[0066] In some embodiments of the present application, the data acquisition module 51 comprises a first acquisition unit and a second acquisition unit. The first acquisition unit is configured to acquire historical regulation capacity data and covariate data of the flexible resource in a preset historical period, wherein the historical regulation capacity data comprises flexible resource quantity data and flexible resource regulation capacity data, and the covariate data comprises time characteristic data, weather characteristic data and electricity consumption behavior characteristic data, and the covariate data comprises flexible resource quantity. The second acquisition unit is configured to align the historical regulation capacity data and the covariate data according to a time sequence to obtain the input data set.

[0067] In some embodiments of the present application, the data processing module 52 comprises a first processing unit and a second processing unit. The first processing unit is configured to establish a basic model by taking the Encoder-Decoder architecture as a basic framework. The second processing unit is configured to integrate a multi-head self-attention mechanism of the Transformer algorithm and a Bayesian neural network into the basic model to obtain the regulation capacity prediction model.

[0068] In some embodiments of the present application, the data processing module 52 further comprises a third processing unit, a fourth processing unit and a fifth processing unit. The third processing unit is configured to set each layer neural network weight and bias parameter in the regulation capacity prediction model as a prior distribution, and infer a posterior distribution according to the input data set by using a Bayesian formula. The fourth processing unit is configured to perform real posterior distribution approximation processing on the posterior distribution by using a Gaussian distribution to obtain a real posterior distribution. The fifth processing unit is configured to sample from the real posterior distribution to obtain a plurality of sampling parameters, forward propagate each sampling parameter, output an adjustment capability prediction result of the flexible resource in each future time period, and group all the adjustment capability prediction results into an adjustment capability prediction result set.

[0069] In some embodiments of the present application, the data verification module 53 comprises a first verification unit and a second verification unit. The first verification unit is configured to process the adjustment capability prediction result set by a probability prediction theory and a Bayesian gate recurrent unit comparison algorithm respectively to obtain a probabilistic index and a deterministic index according to the adjustment capability prediction result set. The second verification unit is configured to take the adjustment capability prediction result set as a final adjustment capability probability distribution result if the probabilistic index and the deterministic index both satisfy a preset condition.

[0070] The more detailed step flow and working principle of the present embodiment can be but not limited to referring to the related description of Embodiment One.

[0071] In summary, the embodiments of the present application construct an input data set by obtaining historical adjustment capability data of flexible resources and related covariates to provide sufficient data basis for prediction; input the input data set into an adjustment capability prediction model of a fusion Encoder-Decoder architecture, a Bayesian neural network and a Transformer algorithm, which can capture the time sequence dependence relationship of adjustment capability by means of the self-attention mechanism of the Transformer, and at the same time, define the model parameters as a probability distribution through the BNN to realize the quantification of uncertainty while ensuring the prediction accuracy, and output multi-period prediction results; then verify by a probability prediction theory and a Bayesian gate recurrent unit comparison algorithm, combine the deterministic and probabilistic indexes to ensure the reliability of the results, finally obtain the adjustment capability probability distribution result, and schedule the flexible resources in the power grid according to the final adjustment capability probability distribution result. Therefore, the present application can solve the problem of difficult quantification and estimation of the adjustment capability of flexible resources in the power grid in the prior art, and improve the accuracy of the scheduling of flexible resources in the power grid.

[0072] It can be understood that the above-mentioned device item embodiments correspond to the method item embodiments of the present application, and can realize the power grid scheduling method based on the adjustment capability of flexible resources provided by any one of the above-mentioned method item embodiments of the present application.

[0073] It should be noted that the apparatus embodiments described above are only illustrative, and part or all of the modules can be selected to achieve the purpose of the embodiment of the present embodiment. In addition, in the apparatus embodiment provided by the present application, the connection relationship between the modules indicates that there is a communication connection between them, which can be implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement it without creative labor.

[0074] Embodiment three On the basis of the above-mentioned embodiment of the power grid scheduling method based on flexibility resource adjustment capability, another embodiment of the present application provides a terminal device, which comprises a processor, a memory and a computer program stored in the memory and configured to be executed by the processor, and when the processor executes the computer program, the power grid scheduling method based on flexibility resource adjustment capability of any embodiment of the present application is realized.

[0075] For example, in this embodiment, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present application. The one or more modules can be a series of computer program instruction segments capable of completing a specific function, which are used to describe the execution process of the computer program in the terminal device.

[0076] The terminal device can be a desktop computer, a notebook computer, a palm computer and a cloud server, etc. The terminal device can include, but is not limited to, a processor and a memory.

[0077] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The processor is the control center of the terminal device, and connects all parts of the terminal device through various interfaces and lines.

[0078] Embodiment four On the basis of the above method embodiment, another embodiment of the present application provides a computer readable storage medium, comprising a stored computer program, wherein the computer readable storage medium controls the device to execute the power grid scheduling method based on flexibility resource adjustment capability according to any one of the above method embodiments of the present application when the computer program runs.

[0079] The modules / units integrated in the apparatus / terminal device, if realized in the form of software function units and sold or used as independent products, can be stored in a computer readable storage medium. Based on this understanding, all or part of the processes in the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware, and the computer program can be stored in a computer readable storage medium. When the processor executes the computer program, the steps of each method embodiment can be implemented. The computer program includes computer program code, which can be in the form of source code, object code, executable file or some intermediate form. The computer readable medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc.

[0080] The above is the preferred embodiment of the present application. It should be noted that, for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which are also considered within the scope of protection of the present application.

Claims

1. A power grid dispatching method based on flexibility resource regulation capability, characterized in that: include: Acquiring historical adjustment capability data and covariate data of the flexibility resource, and constructing an input data set based on the historical adjustment capability data and the covariate data; Inputting the input data set into a regulation capacity prediction model, outputting regulation capacity prediction results of the flexibility resource in each future time period, and combining all regulation capacity prediction results into a regulation capacity prediction result set; wherein the regulation capacity prediction model is constructed based on an encoder-decoder architecture, a Bayesian neural network, and a transformer algorithm; Based on the set of regulatory ability prediction results, each regulatory ability prediction result in the set of regulatory ability prediction results is verified using probability prediction theory and a Bayesian gated recurrent unit comparison algorithm. If the verification passes, the set of regulatory ability prediction results is used as the final regulatory ability probability distribution result; The flexibility resources in the power grid are scheduled according to the final regulation capability probability distribution result.

2. A power grid dispatching method based on flexibility resource regulation capability according to claim 1, characterized in that: The acquiring of historical adjustment capability data and covariate data of the flexibility resource and constructing an input data set according to the historical adjustment capability data and the covariate data is specifically as follows: Obtaining historical regulation capacity data and covariate data of the flexibility resource within a preset historical period; wherein the historical regulation capacity data includes flexibility resource quantity data and flexibility resource regulation capacity data, the covariate data includes time characteristic data, weather characteristic data, and electricity consumption behavior characteristic data, and the covariate data includes the quantity of flexibility resources; The historical regulatory ability data and the covariate data are aligned in time series to obtain the input data set.

3. A power grid dispatching method based on flexibility resource regulation capability according to claim 1, characterized in that: The regulatory capacity prediction model is constructed based on the Encoder-Decoder architecture, Bayesian neural network and Transformer algorithm, specifically: Using the Encoder-Decoder architecture as the basic framework, a basic model is established; The multi-head self-attention mechanism and Bayesian neural network of the Transformer algorithm are integrated into the basic model to obtain the regulatory ability prediction model.

4. A power grid dispatching method based on flexibility resource regulation capability according to claim 1, characterized in that: The input data set is input into the adjustment capacity prediction model, the adjustment capacity prediction results of the flexibility resource in each future time period are output, and all the adjustment capacity prediction results are combined into an adjustment capacity prediction result set, specifically: The weights and bias parameters of each layer of the neural network in the regulatory ability prediction model are set as a prior distribution, and the posterior distribution is inferred using the Bayesian formula based on the input data set; The posterior distribution is approximated by Gaussian distribution to obtain a true posterior distribution; Sampling is performed from the true posterior distribution to obtain a number of sampling parameters, and each sampling parameter is forward propagated to output the adjustment capacity prediction results of the flexibility resource in each future time period, and all the adjustment capacity prediction results are combined into an adjustment capacity prediction result set.

5. The power grid dispatching method based on flexibility resource regulation capability according to claim 1, characterized in that: According to the set of adjustment ability prediction results, the probability prediction theory and the Bayesian gated recurrent unit comparison algorithm are used to verify each adjustment ability prediction result in the set of adjustment ability prediction results. If the verification passes, the set of adjustment ability prediction results is used as the final adjustment ability probability distribution result, specifically: According to the set of regulatory capacity prediction results, probabilistic prediction theory and Bayesian gated recurrent unit comparison algorithm are used to process them respectively to obtain a probabilistic index and a deterministic index; If both the probabilistic index and the deterministic index meet the preset conditions, the set of adjustment capability prediction results is used as the final adjustment capability probability distribution result.

6. A power grid dispatching device based on flexibility resource regulation capability, characterized in that: include: Data acquisition module, data processing module, data verification module and scheduling module; The data acquisition module is configured to acquire historical adjustment capability data and covariate data of the flexibility resource, and construct an input data set based on the historical adjustment capability data and the covariate data; The data processing module is configured to input the input data set into a regulation capacity prediction model, output regulation capacity prediction results of the flexibility resource in each future time period, and combine all regulation capacity prediction results into a regulation capacity prediction result set; wherein the regulation capacity prediction model is constructed based on an encoder-decoder architecture, a Bayesian neural network, and a Transformer algorithm; The data verification module is used to verify each regulation capacity prediction result in the regulation capacity prediction result set based on the regulation capacity prediction result set using probability prediction theory and Bayesian gated recurrent unit comparison algorithm; if the verification passes, the regulation capacity prediction result set is used as the final regulation capacity probability distribution result, and the flexibility resources in the power grid are scheduled according to the final regulation capacity probability distribution result; The scheduling module is used to schedule the flexibility resources in the power grid according to the final regulation capability probability distribution result.

7. A power grid dispatching device based on flexibility resource regulation capability according to claim 6, characterized in that: The data acquisition module includes a first acquisition unit and a second acquisition unit; The first acquisition unit is configured to acquire historical regulation capability data and covariate data of the flexibility resource within a preset historical period; wherein the historical regulation capability data includes flexibility resource quantity data and flexibility resource regulation capability data, the covariate data includes time characteristic data, weather characteristic data, and electricity consumption behavior characteristic data, and the covariate data includes the quantity of flexibility resources; The second acquisition unit is configured to align the historical regulatory capability data and the covariate data in time series to obtain the input data set.

8. The power grid dispatching device based on flexibility resource regulation capability according to claim 6, characterized in that: The data processing module includes a first processing unit and a second processing unit; The first processing unit is configured to establish a basic model using the Encoder-Decoder architecture as a basic framework; The second processing unit is used to integrate the multi-head self-attention mechanism and Bayesian neural network of the Transformer algorithm into the basic model to obtain the regulatory ability prediction model.

9. The power grid dispatching device based on flexibility resource regulation capability according to claim 6, characterized in that: The data processing module further includes a third processing unit, a fourth processing unit and a fifth processing unit; The third processing unit is used to set the weights and bias parameters of each layer of the neural network in the regulatory ability prediction model as a prior distribution, and to infer the posterior distribution using the Bayesian formula based on the input data set; The fourth processing unit is configured to perform approximate true processing on the posterior distribution using a Gaussian distribution to obtain a true posterior distribution; The fifth processing unit is used to sample from the true posterior distribution to obtain a number of sampling parameters, and forward propagate each sampling parameter to output the adjustment capacity prediction results of the flexibility resource in each future time period, and form all the adjustment capacity prediction results into an adjustment capacity prediction result set.

10. The power grid dispatching device based on flexibility resource regulation capability according to claim 6, characterized in that: The data verification module includes a first verification unit and a second verification unit; The first verification unit is configured to process the set of adjustment capability prediction results using probability prediction theory and a Bayesian gated recurrent unit comparison algorithm to obtain a probability index and a deterministic index; The second verification unit is configured to use the set of adjustment capability prediction results as a final adjustment capability probability distribution result if both the probabilistic index and the deterministic index meet preset conditions.

Images