Node electricity price scene generation method and device based on spatio-temporal combination and residual diffusion

Abstract

The invention provides a node electricity price scene generation method and device based on space-time combination and residual diffusion, and relates to the technical field of electricity price scene generation. A historical node electricity price sequence data set is formed, then a space-time joint neural network model is input, space-time features of node electricity price data are extracted, and after a predicted node electricity price sample fusing time sequence dynamic and space topology information and a condition information matrix are generated, a residual error denoising diffusion model is input; the power price probability distribution is subjected to refined modeling through forward noise adding and reverse denoising processes, a node power price scene covering multi-dimensional uncertainty is generated, the space-time relevance and uncertainty of the node power price can be captured, the generated node power price scene is more real and closer to reality, and the node power price is more accurate. And a reliable decision basis is provided for power market participants to optimize a scheduling strategy.

Description

Technical Field

[0001] The present invention relates to the technical field of electricity price scenario generation, and in particular to a method and device for generating a node electricity price scenario based on spatiotemporal union and residual diffusion. Background Art

[0002] Node electricity prices reflect the electricity prices at specific points in the power system at specific times. They provide critical guidance for power market participants in formulating trading strategies and assessing operating costs and benefits. Currently, neural network models are commonly used for electricity price forecasting. However, these models are often inadequate for predicting node electricity prices. This is because, in the power market, node electricity prices are influenced by a variety of complex factors, such as energy supply and demand, grid congestion, and transmission losses, exhibiting high spatiotemporal volatility and uncertainty. Unordered fluctuations in node electricity prices can disrupt the rational allocation of power resources, complicate grid scheduling, and even threaten the safe and stable operation of the system. Therefore, applying existing electricity price forecasting models to real-time prediction of power system node electricity prices fails to capture the spatiotemporal correlations and uncertainties of node electricity prices, leading to forecasts that deviate from actual scenarios. To effectively address this challenge, generating node electricity price scenarios that reflect these spatiotemporal correlations and uncertainties is crucial. This can provide a basis for subsequent decision-making by power market participants in optimizing scheduling strategies. Summary of the Invention

[0003] In order to solve the above-mentioned problems existing in the prior art, the purpose of the present invention is to provide a node electricity price scenario generation method and device based on spatiotemporal union and residual diffusion, which collaboratively generates high-quality node electricity price scenarios through a spatiotemporal union neural network model and a residual denoising diffusion model, captures the spatiotemporal correlation and uncertainty of node electricity prices, and generates node electricity price scenarios that are more realistic and close to reality, which can provide a more reliable decision-making basis for power market participants to optimize scheduling strategies.

[0004] To achieve the above object, the present invention provides the following technical solutions: In a first aspect, the present invention provides a method for generating node electricity price scenarios based on spatiotemporal union and residual diffusion, including: S1. preprocessing the acquired multi-source time series data including node electricity price data of power system nodes, wind and solar output data, load data and date data and considering spatiotemporal correlation to form a historical node electricity price series data set; S2. inputting the historical node electricity price series data set into a spatiotemporal union neural network model, extracting the spatiotemporal characteristics of the node electricity price data, and generating predicted node electricity price samples and conditional information matrices that integrate temporal dynamics and spatial topology information; S3. inputting the predicted node electricity price samples and conditional information matrix into a residual denoising diffusion model, and performing refined modeling of the electricity price probability distribution through forward denoising and reverse denoising processes to generate a node electricity price scenario covering multi-dimensional uncertainty.

[0005] In a first aspect, the present invention provides a preferred solution, S1 comprising: S11. filling missing values ​​of a short time scale not exceeding one day in multi-source time series data; S12. filling missing values ​​of a long time scale exceeding one day in multi-source time series data.

[0006] More preferably, linear interpolation is used to fill missing values ​​in multi-source time series data with a short time scale of no more than one day.

[0007] More preferably, similar day retrieval and weighted interpolation are used to fill missing values ​​of long time scales exceeding one day in multi-source time series data.

[0008] The present invention provides a preferred solution in the first aspect, S1 also includes: S13. Standardizing and max-min normalizing the node electricity price data in the multi-source time series data after filling the missing values ​​to obtain the node electricity price continuous time series feature vector, and performing max-min normalization on the wind and solar power output data and load data after filling the missing values ​​to obtain the output and load continuous time series feature vector; S14. One-hot encoding the time data after filling the missing values ​​to obtain discrete date feature variables; S15. Aligning the continuous time series feature variables with the discrete date feature variables at the time granularity to form the historical node electricity price series data set.

[0009] The present invention provides a preferred solution in the first aspect, S2 includes: S21. Arranging the pre-processed daily node electricity price data into a three-dimensional tensor according to a preset fixed node order to obtain a node electricity price tensor, and calculating the historical node electricity price information matrix of the base day based on the node electricity price tensors of similar days and neighboring days of the base day; S22. Calculating multi-dimensional statistical features of the constructed historical node electricity price information matrix; S23. Based on the node electricity price tensors of neighboring days of the base day, calculating the Pearson correlation coefficient between nodes and screening to generate a sparse adjacency matrix, and performing symmetric normalization on the sparse adjacency matrix; S24. Using a network structure with a collaborative 3D convolution layer, a graph convolution layer and a Conv LSTM layer as the spatiotemporal joint neural network model, deeply integrating the spatiotemporal features and statistical features of the node electricity price data, generating control parameters through control parameter modulation, modulating and gated the features of the input of each layer, fusing multi-scale features, and finally obtaining the predicted node electricity price sample and conditional information matrix; the spatiotemporal features include the historical node electricity price information matrix of the base day and the sparse adjacency matrix after symmetric normalization.

[0010] In a second aspect, the present invention provides a node electricity price scenario generation device based on spatiotemporal union and residual diffusion, which is used to execute the above method, including: a data set establishment module, which is used to preprocess the acquired multi-source time series data including node electricity price data, wind and solar output data, load data and date data of power system nodes and considering spatiotemporal correlation to form a historical node electricity price series data set; a node electricity price sample generation module, which is used to input the historical node electricity price series data set into a spatiotemporal union neural network model, extract the spatiotemporal characteristics of the node electricity price data, and generate predicted node electricity price samples and conditional information matrices that integrate temporal dynamics and spatial topology information; a node electricity price scenario generation module, which is used to input the predicted node electricity price samples and conditional information matrix into a residual denoising diffusion model, and perform fine modeling of the electricity price probability distribution through forward denoising and reverse denoising processes to generate a node electricity price scenario covering multi-dimensional uncertainty.

[0011] Compared with the prior art, the present invention has the following advantages: The present invention provides a node electricity price scenario generation method and device based on spatiotemporal union and residual diffusion. High-quality node electricity price scenarios are generated collaboratively through a spatiotemporal union neural network model and a residual denoising diffusion model, capturing the spatiotemporal correlation and uncertainty of node electricity prices. The generated node electricity price scenarios are more realistic and closer to reality. Not only can they provide a reliable decision-making basis for optimizing scheduling strategies for power market participants, but they can also provide automated intelligent tools for subsequent effective risk assessment of power systems, thereby improving risk management capabilities and ensuring the safe, stable and efficient operation of power systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0013] Figure 1 A flowchart of a method for generating node electricity price scenarios based on spatiotemporal integration and residual diffusion, provided in accordance with a specific embodiment of the present invention; Figure 2 A module diagram of a node electricity price scenario generation device based on spatiotemporal combination and residual diffusion provided by a specific embodiment of the present invention; Figure 3 A node electricity price tensor (matrix) structure diagram is obtained in step S211 of a node electricity price scenario generation method based on spatiotemporal union and residual diffusion provided in a specific embodiment of the present invention; Figure 4A schematic diagram of the structure and processing process of a spatiotemporal joint neural network model in a node electricity price scenario generation method based on spatiotemporal joint and residual diffusion provided in a specific embodiment of the present invention; Figure 5 A schematic diagram of the main process of a residual denoising diffusion model in a node electricity price scenario generation method based on spatiotemporal combination and residual diffusion provided by a specific embodiment of the present invention; Figure 6 A structural diagram of a residual denoising diffusion model backbone using a Unet network in a node electricity price scenario generation method based on spatiotemporal union and residual diffusion provided by a specific embodiment of the present invention. DETAILED DESCRIPTION

[0014] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0015] Please refer to Figure 1 In an optional embodiment, a node electricity price scenario generation method based on spatiotemporal union and residual diffusion is provided, which is mainly implemented by the following steps: S1. Preprocess the acquired multi-source time series data, including node electricity price data, wind and solar output data, load data, and date data of power system nodes, taking into account spatiotemporal correlations, to form a historical node electricity price series dataset; S2. Input the historical node electricity price series dataset into the spatiotemporal joint neural network model, extract the spatiotemporal features of the node electricity price data, and generate predicted node electricity price samples and conditional information matrices that integrate temporal dynamics and spatial topology information. S3. Input the predicted node electricity price samples and conditional information matrix into the residual denoising diffusion model, and refine the electricity price probability distribution model through the forward denoising and backward denoising process to generate node electricity price scenarios covering multi-dimensional uncertainties.

[0016] Please refer to Figure 2 Correspondingly, in an optional embodiment, a node electricity price scenario generation device based on spatiotemporal union and residual diffusion is provided, which is used to execute the method steps of the above embodiment and mainly consists of the following modules: Dataset creation module 1 is used to pre-process the acquired multi-source time series data including node electricity price data of power system nodes, wind and solar output data, load data and date data, taking into account spatiotemporal correlation, to form a historical node electricity price series data set; Node electricity price sample generation module 2 is used to input the historical node electricity price series data set into the spatiotemporal joint neural network model, extract the spatiotemporal characteristics of the node electricity price data, and generate predicted node electricity price samples and conditional information matrices that integrate temporal dynamics and spatial topology information; The node electricity price scenario generation module 3 is used to input the predicted node electricity price samples and conditional information matrix into the residual denoising diffusion model, and to refine the modeling of the electricity price probability distribution through the forward denoising and reverse denoising process to generate a node electricity price scenario covering multi-dimensional uncertainties.

[0017] In a preferred embodiment, the following more specific and preferred implementation scheme is given corresponding to the steps of the above embodiment.

[0018] S1. Preprocess the acquired multi-source time series data, including node electricity price data, wind and solar output data, load data, and date data of power system nodes, taking into account spatiotemporal correlations, to form a historical node electricity price series dataset. Wind and solar output data refers to wind power output data and photovoltaic output data, and load data refers to load and net load data. Date data mainly includes: season, weekday, weekend, holiday, and non-holiday. S1 is specifically implemented through the following sub-steps:

[0019] S11. Fill missing values ​​on a short time scale of no more than one day (<24 hours) in the multi-source time series data. In a preferred embodiment, linear interpolation is used to fill missing values ​​on a short time scale of no more than one day in the multi-source time series data. The calculation formula of the linear interpolation method is as follows: ; in, As the reference time, The reference date is the observation date. The number of samples selected during interpolation. If the number of samples is 2, the value is 2. For example, if the data of point 8 is missing, the data of points 7 and 9 are used to interpolate and calculate as the data of point 8; Expressed as day Continuous time series values ​​sampled at the moment, Expressed as day Continuous time series values ​​sampled at the moment, Expressed as day Continuous time series values ​​sampled at time instants.

[0020] S12. Fill in missing values ​​of long time scales exceeding one day (≥24 hours) in multi-source time series data. In a preferred embodiment, similar day retrieval and weighted interpolation are used to fill in missing values ​​of long time scales exceeding one day in multi-source time series data, specifically as follows: for node electricity price data, weighted interpolation is performed in combination with the corresponding data of the neighboring days of the base date; for wind and solar power output data and load data, historical similarity is calculated based on the node electricity price feature vector of the base date and the node electricity price feature vector of the historical date, and the corresponding data of several similar days that are most similar to the corresponding data of the base date are selected for linear interpolation. Specifically, for node electricity price data, weighted interpolation is performed in combination with the corresponding data of the neighboring days (the previous K1 day and the same day of the week in the previous several weeks) of the base date (that is, the observation day). The following example illustrates the specific process of weighted interpolation:

[0021] Suppose the sequence of observation days (i.e. today, base day) to be filled is The previous K1 day is generally selected from the previous 3 days, and the corresponding data sequence for the previous 3 days is 、 、 , the first few weeks are generally the first two weeks, that is, the same day of the week last week and the week before last (for example, if the observation day is Tuesday, it refers to the Tuesday of last week and the Tuesday before last). The corresponding data series is , ; The calculation formula is: ;in, 、 、 、 、 、 、 are all weight parameters and satisfy: , , .

[0022] For wind and solar power output data and load data, the node electricity price feature vector based on the base date and the node electricity price feature vector of the historical day , calculate historical similarity , , select similar days, that is, the most similar k1 days (such as 3 days, or other days, in actual implementation, the selection is based on the quality and characteristics of the data set, the number of similar days, etc.) and perform linear interpolation. Among them, represents the node electricity price feature vector on the base day, The node electricity price feature vector representing the historical day; Represents the historical similarity of the node electricity price feature vectors between the base date and the historical date; is the standard deviation of the distribution of historical node electricity price data. More specifically, the first three days are the time span selected for calculation. Considering that the correlation between a day's data and historical data decreases with increasing time intervals, the data from the first three days is selected to have a higher correlation with the observation day (i.e., today, the date for which data is to be supplemented). Therefore, the first three days can be considered part of the historical day. The first two weeks, a longer timeframe than three days, have a lower correlation between daily data and the observation day (today). Furthermore, considering that daily life patterns may vary from Monday to Sunday, selecting data on the same day of the week offers higher similarity. Therefore, only data corresponding to the same day of the week from the first two weeks is selected (for example, if today's date for supplementary data is Tuesday, data from the previous Tuesday and the Tuesday before last are selected). Over timescales longer than two weeks, due to significant variations in seasons, temperature, and other factors, the data can be considered to no longer be strongly correlated.

[0023] In a more preferred embodiment, the preprocessing in S1 further performs standardization and normalization processing, and based on S11 and S12, the following steps are specifically added:

[0024] S13. Standardize and perform max-min normalization on the node electricity price data in the multi-source time series data after filling missing values ​​to obtain a continuous time series feature vector for the node electricity price. Perform max-min normalization on the wind and solar power output data and load data after filling missing values ​​to obtain continuous time series feature vectors for output and load. Preferably, standardization and max-min normalization are performed channel by channel, corresponding to the three channel components of the node electricity price: energy, congestion, and loss.

[0025] The standardized calculation formula is as follows: ; Among them, for a time series of length n , represents the value of the i-th sampling point, and ; and Represent the global mean and standard deviation of the time series respectively; Indicates the first The value of the sampling point. The sampling point here refers to any data point in the time series. Standardization is similar to normalization, which is also a way to limit the overall data to a certain range. For a time series with a length of n, if one year of data is selected as the data set, n here is 365. 24. When performing standardization, the mean and variance of a sequence consisting of n data points are calculated, and then each data point is calculated according to the above formula. Therefore, i ranges from 1 to n, which means that each data point is standardized.

[0026] The calculation formula for max-min normalization is as follows: ; in, Indicates the first The value of the sampling points, and Represent the maximum and minimum values ​​in the time series respectively; is a value normalized to the interval [0,1] by max-min. The sampling point here refers to any data point in the time series.

[0027] S14. Perform one-hot encoding on the time data after filling missing values ​​to obtain discrete date feature variables.

[0028] S15. Align the continuous time series characteristic variables with the discrete date characteristic variables at the time granularity to form the historical node electricity price series dataset.

[0029] S2. Input the historical node electricity price series data set into the spatiotemporal joint neural network model, extract the spatiotemporal features of the node electricity price data, and generate predicted node electricity price samples and conditional information matrices that integrate temporal dynamics and spatial topology information. S2 is specifically implemented through the following sub-steps:

[0030] S21. Arrange the preprocessed daily node electricity price data into a three-dimensional tensor according to a preset fixed node order to obtain a node electricity price tensor, and calculate the historical node electricity price information matrix of the base day based on the node electricity price tensors of similar days and adjacent days of the base day.

[0031] In a preferred embodiment, S21 is implemented by the following sub-steps:

[0032] S211. Node electricity price tensor formation: Arrange the pre-processed daily node electricity price data into a three-dimensional tensor according to a preset fixed node order; the preset fixed node order is determined based on the node geographical distance or electrical connection relationship. Specifically, arrange the pre-processed daily node electricity price data into a three-dimensional tensor according to a preset fixed node order. ,in express The node electricity price tensor of the day, Day represents every day in the dataset. The data for each day needs to be made into a tensor with the shape of Figure 3 As shown; , represents the channel component, corresponding to the three channel components of energy, congestion, and loss representing the node electricity price; , corresponding to 24 physical node numbers; the preset fixed node order is determined based on the node geographical distance or electrical connection relationship; , representing a 24-hour time series within a day. By integrating the temporal, spatial, and channel component information, three channels, 24 hours, and 24 nodes form a three-dimensional tensor sample of [3, 24, 24].

[0033] S212. Formation of historical node electricity price information matrix: Search the historical data for the node electricity price tensors of several similar days that are most similar to the node electricity price tensor of the reference day according to the feature similarity, and concatenate them with the node electricity price tensors of the adjacent days of the reference day to obtain the historical node electricity price information matrix of the reference day. Specifically, for the reference day (observation day) , use the node electricity price feature vector of the base day obtained in S1 to search the node electricity price tensors of the k2 days (such as the 5th) closest to the observation day in the historical data according to feature similarity, and splice them together with the node electricity price tensors of K2 days before the observation day (such as the previous 3 days) to obtain the base day (observation day) Historical node electricity price information matrix The feature similarity retrieval in this step of this embodiment refers to comparing the feature vectors element by element, which can be understood as calculating the Euclidean distance for the matrix. The closer the distance, the more similar they are. Specifically, let the matrix shape be , then calculate the distance between the two matrices (subtract element by element, then calculate the sum of squares). The smaller the distance, the more similar it is. Suppose the two matrices are both of shape , the calculation formula is as follows: distance ; 、 are the elements at each position in the two matrices respectively.

[0034] S22. Calculate multi-dimensional statistical features for the constructed historical node electricity price information matrix. Preferably, the multi-dimensional statistical features are calculated by channel and node.

[0035] In a preferred embodiment, the statistical features are mainly basic fluctuation features, extreme fluctuation features and mutation features. S22 is specifically implemented by the following steps:

[0036] S221. Obtaining basic fluctuation characteristics: Quantify the basic fluctuation characteristics of the historical node electricity price information matrix through mean and variance. The calculation formula of mean is as follows: ; in, , represents the length of the time series of 24 hours per day, Indicates the Day, Node exist The moment The electricity price data of each channel component.

[0037] At the mean Based on this, the variance is calculated as follows: .

[0038] S222. Obtaining extreme fluctuation characteristics: Use the maximum upper deviation and maximum lower deviation to characterize the extreme cases where the electricity price deviates from the mean.

[0039] The maximum upper deviation reflects the maximum extent to which the electricity price exceeds the mean, while the maximum lower deviation indicates the maximum extent to which the electricity price is lower than the mean. The combination of the two can intuitively present the extreme range of electricity price fluctuations. The calculation formulas for the maximum upper and lower deviations are as follows: ; .

[0040] S223. Mutation feature acquisition: Capture the sudden change features of electricity prices through the peak position and peak duration. The calculation formula for the peak position is as follows: ; in, For the Day, Node In the The peak position of each channel component indicates the time when the electricity price reaches its peak in a day (that is, the data point where the electricity price reaches the maximum in the 24-hour series); Indicates the time when the return function reaches its maximum value in the 24-hour time dimension , which is used to locate the specific moment when the electricity price of each channel component of each node reaches its peak within a day.

[0041] The calculation formula for peak duration is as follows: ; in, For the Day, Node In the The peak duration of each channel component indicates the length of time the node electricity price is greater than the set value; is the threshold coefficient, Indicates the length of time that the node electricity price is greater than the set value, Indicates the Day, Node exist The moment Electricity price data of each channel component; 、 Respectively represent the mean and standard deviation calculated previously in S221.

[0042] S23. Based on the node electricity price tensor of the adjacent days of the base date (such as within the previous two weeks), calculate the Pearson correlation coefficient between nodes and screen to generate a sparse adjacency matrix, and perform symmetric normalization on the sparse adjacency matrix.

[0043] In a preferred embodiment, S23 is specifically implemented by the following steps:

[0044] S231. Pearson correlation coefficient calculation: Calculate the Pearson correlation coefficient between nodes based on the standardized covariance formula to obtain the Pearson correlation coefficient matrix between nodes , used to characterize the spatial correlation between nodes. Specifically, the calculation node With node In the The correlation on the channel components is calculated as follows: ; Among them, the base date The calculation is based on the data from the previous 14 days. Any day within 14 days, Representation node With node In the The Pearson correlation coefficient on the channel components, represents the Pearson correlation coefficient matrix between nodes on the kth channel component, is a matrix The elements in Indicates the Day, Node exist The moment The electricity price data of each channel component, Indicates the Day, Node exist The moment The electricity price data of each channel component, Representation node In the past two weeks The time series mean of the electricity price data of the channel components, Representation node In the past two weeks The time series mean of the electricity price data of the channel components.

[0045] S232. Threshold screening and sparseness: According to the preset threshold, the Pearson correlation coefficient matrix is ​​screened to generate a sparse adjacency matrix for representing spatial topological information. Specifically, according to the preset threshold Screen the correlation matrix to generate a sparse adjacency matrix , the calculation formula is as follows: .

[0046] S233. Symmetric normalization: Perform symmetric normalization on the sparse adjacency matrix to obtain a normalized sparse adjacency matrix. The calculation formula is as follows: ; in, represents the degree matrix, whose diagonal elements . Represents the identity matrix, and the node's own characteristics are retained by introducing self-loops.

[0047] S24. A network structure that uses a 3D convolutional layer (3D Conv), a graph convolutional layer, and a Conv LSTM layer to form the spatiotemporal joint neural network model (e.g. Figure 4 As shown in FIG2 ), the spatiotemporal features of the node electricity price data (the historical node electricity price information matrix obtained in S21 and the sparse adjacency matrix obtained in S23) are deeply integrated with the statistical features, and the control parameters are generated by control parameter modulation. The features of each layer input are modulated and gated, and multi-scale features are integrated to finally obtain the predicted node electricity price samples and conditional information matrix; the spatiotemporal features include the historical node electricity price information matrix of the base date and the sparse adjacency matrix after symmetric normalization. In a preferred embodiment, S24 specifically includes the following sub-steps:

[0048] S241. Control parameter modulation: The statistical features (statistical feature vectors extracted in step S22) and the normalized sparse adjacency matrix are input into the multi-branch fully connected network to generate control parameters, including the modulated adjacency matrix. This step is also called statistical feature preprocessing. The multi-branch fully connected network can fit and output multiple control parameters. In a preferred embodiment, the control parameters mainly include: conditional batch normalization control parameters, channel gating control weights, modulated adjacency matrix, scale factor, offset factor and spike modulation factor; wherein, the conditional batch normalization control parameters and channel gating control weights are used in step S242; the scale factor, offset factor and spike modulation factor are used in step S245. A more specific process is as follows: Generate the following four sets of control parameters, and the calculation formula is as follows: ; ; ; ; in, , is the control parameter of conditional batch normalization, is the control weight of the channel gating, is the modulated adjacency matrix, enhancing the node association in the spike period, , , respectively are scale factor, offset factor and spike modulation factor, forming explicit mapping with mean, variance and spike feature in statistical feature, is the statistical feature vector, namely the basic fluctuation feature, fluctuation extreme feature and mutation feature in S22; 、 、 、 represent four fully connected network functions, used to fit the nonlinear function relationship.

[0049] S242.3D convolution: based on 3D convolution layer, the historical node price information matrix is subjected to 3D convolution operation, and statistical feature gating processing is carried out by using control parameters, and 3D convolution feature map is output, and the calculation formula is as follows: ; wherein, and respectively represent the weight matrix and bias of the 3D convolution layer, represent the activation function of the 3D convolution layer.

[0050] by means of the generated by step S241, the obtained after convolution is subjected to conditional batch normalization, and the calculation formula is as follows: ; wherein, and are the mean and variance of the current batch data, is a very small constant to prevent the denominator from being zero.

[0051] The statistical feature gating is performed on the batch normalization obtained by the gating weight , and 3D convolution feature map is output, and the calculation formula is as follows: ; wherein, represent element-wise multiplication, is generated by step S241, and by selecting the active channel, the noise feature irrelevant to the statistical law is suppressed.

[0052] S243. Graph convolution: based on the graph convolution layer, the modulated adjacency matrix is combined with the 3D convolution feature map to perform graph convolution operation, and graph convolution spatial topology fusion feature is output The calculation formula is as follows: ; in, represents the LeakyReLU activation function, represents the modulated adjacency matrix, is the weight matrix, Represents channel-level weight scaling, which adjusts the contribution of each channel to graph convolution through statistical features.

[0053] S244. Temporal Dynamic Capture: Fusion of Graph Convolutional Spatial Topology Features Based on Conv LSTM Layer and 3D convolutional feature maps After convolution and capturing the temporal dynamic features through the gating mechanism, the hidden state of the Conv LSTM layer is obtained. . Specifically include the following contents: Input Gate: ; Forget Gate: ; Output gate: ; Candidate memory element: ; Update the memory meta-state: ; Update the hidden layer state: ; in, , that is, and splicing; Convolution operation is used to extract features from the daily node electricity price matrix ; 、 、 、 、 、 、 、 is the weight parameter, 、 、 、 is the bias parameter, is the hidden layer state, represents the sigmoid activation function, represents the tanh activation function; is the memory element of the matrix.

[0054] S245. Multi-scale feature fusion and prediction output: 3D convolution feature map , graph convolution spatial topology fusion features and Conv LSTM layer hidden state Splicing into multi-scale fusion features , combined with the control parameters and processed by the fully connected layer to generate the predicted node electricity price sample and the conditional information matrix , as the initialization input and conditional parameter of the subsequent residual denoising diffusion model. The calculation formula is as follows: ; ; ; in, 、 、 Represent the predicted basic electricity price matrix (that is, the predicted node electricity price sample ), the log-variance matrix and the spike intensity matrix, Represents the fully connected layer function of the neural network, where represents the basic electricity price fully connected layer function, represents the log-variance fully connected layer function, represents the spike intensity fully connected layer function; 、 represents the mean and variance characteristics of the node electricity price samples extracted in step S22, 、 、 This represents the scaling factor, offset factor, and peak modulation factor generated by the control parameter modulation step. Conditional information refers to the constraints imposed on node electricity prices given weather conditions, load conditions, and other factors. The conditional information matrix (logarithmic variance matrix and peak intensity matrix) here refers to the conditional information used in the final residual diffusion step. Wind and solar power output, load, and other data serve as the conditional information for generating the predicted node electricity price samples.

[0055] S3. Input the predicted node electricity price samples and conditional information matrix into the residual denoising diffusion model, and refine the electricity price probability distribution through the forward denoising and reverse denoising process to generate node electricity price scenarios covering multi-dimensional uncertainties. Specifically, the main process of the model and the backbone Unet network structure are as follows: Figure 5 and Figure 6 As shown. By adding noise forward ( Figure 5 Forward) and reverse denoising ( Figure 5The Backward process in the conditional information matrix performs a refined modeling of the probability distribution of electricity price fluctuations, thereby generating node electricity price scenario samples covering multi-dimensional uncertainties and generating high-credibility node electricity price scenario samples. The logarithmic variance matrix in the conditional information matrix ensures that the generated samples have a certain variance, that is, the generated results are not single; the spike intensity matrix in the conditional information matrix enables the generated results to fit the peaks at each specified position. The multi-dimensional uncertainty of the node electricity price scenario is portrayed by these different variances and different peaks. In a preferred embodiment, S3 specifically includes the following sub-steps:

[0056] S31. Forward noise addition process: The predicted node electricity price samples and conditional information matrix generated in step S2 are input into the residual denoising diffusion model, and noise is gradually added to achieve the initial disturbance of the electricity price probability distribution. The calculation formula is as follows: ; in is the noise scheduling parameter, is random Gaussian noise, Indicates the The samples after adding noise.

[0057] S32. Conditionally guided reverse denoising process: using residual denoising network The noise sample is restored and the calculation formula is as follows: ; in, is the sample after adding noise The sample after denoising again, Indicates the The denoised samples, , , is random noise, represents the Unet denoising network, is the control parameter for the intensity of the added noise; is a series of control parameters The product of is the random noise added in step t, that is, a matrix of random numbers randomly generated by the system that conforms to the Gaussian distribution; Represents the conditional information matrix, that is, step S245 outputs the conditional information matrix 、 Through the complete After a denoising process of 10 steps, the node electricity price scenario generation result is obtained. T is set by the model. A larger T means more sub-steps are involved in the entire process, resulting in a more refined fit. A more complex model leads to higher generation quality, but also requires higher hardware performance and slower generation. A complete generation process consists of T steps. Figure 5 middle, It represents the starting point of diffusion, the result of preliminary fusion of the input predicted node electricity price samples and condition information; Indicates the end point of the noise addition process and the starting point of the denoising process. The above is obtained by adding noise for T steps. Mathematically, it approximately conforms to the Gaussian distribution. It can be considered that after adding noise for T steps, it has lost its original characteristics and turned into pure noise. The denoising process is to continuously remove noise from it and then restore the original sample. Represents the difference between the input sample and the target sample, that is, the residual is added in the noise addition and denoising process. In machine learning, adding residual processing can usually improve model performance; Represents the distribution function of the forward noise addition process, that is, it is known 、 In the case of The distribution that should be satisfied; Represents the distribution function of the reverse denoising process, that is, it is known In the case of The distribution that should be satisfied; here That is, the noisy sample in the t-th step of the diffusion model is an intermediate result. The noise addition process is to execute these t steps forward, from 0 to t, and the denoising process is to execute these t steps in reverse.

[0058] Compared with the prior art, the present invention has the following beneficial effects based on the above embodiments: 1. This invention uses a spatiotemporal joint neural network model and a residual denoising diffusion model to collaboratively generate high-quality node electricity price scenarios, capturing the spatiotemporal correlation and uncertainty of node electricity prices. The generated node electricity price scenarios are more realistic and close to reality. This not only provides a reliable decision-making basis for optimizing scheduling strategies for power market participants, but also provides an automated intelligent tool for subsequent effective risk assessment of the power system, improving risk management capabilities and ensuring the safe, stable and efficient operation of the power system. 2. The present invention adopts a combination of linear interpolation and similar day weighted interpolation to fill in missing values ​​at different time scales of multi-source time series data, avoiding data deviation caused by a single method. Compared with traditional single interpolation methods, it can more accurately retain the original trend of the data. The node electricity price series is standardized and normalized by channel, and normalized according to the characteristics of photovoltaic output and other data, effectively retaining the original fluctuation characteristics of the data and solving the dimensional confusion problem caused by traditional unified processing methods. After one-hot encoding, the discrete date features are aligned with the continuous variables to achieve time granularity, converting multi-source heterogeneous data into structured data sets, and comprehensively improving data availability compared to unintegrated data; 3. The present invention constructs daily node electricity price data into a three-dimensional tensor, integrating time, space and channel component information. Compared with the traditional two-dimensional representation, it can more intuitively reflect the physical relationship between the power grid topology and electricity price. Multi-dimensional statistical features are calculated by channel and node, and the electricity price characteristics are fully characterized from mean variance to peak characteristics. Compared with the conventional method of only extracting basic statistics, the information obtained is richer. Based on the Pearson correlation coefficient, a sparse adjacency matrix is ​​generated and processed, and node spatial dependencies are adaptively captured. Combined with 3D convolution, graph convolution, convolutional LSTM collaborative network, and control parameter modulation (statistical feature preprocessing) mechanism, compared with a single network structure, it achieves deep fusion and efficient extraction of spatiotemporal characteristics of node electricity prices; 4. This invention utilizes the forward noise addition and backward denoising processes of the residual diffusion model to generate multidimensional uncertainty scenario samples based on statistical characteristics. This method overcomes the limitations of traditional point prediction methods and can fully cover the long-tail distribution and extreme cases of electricity price fluctuations. The optimized noise scheduling strategy can output diverse and highly reliable forecast results, providing strong support for real-time risk assessment and decision-making in the power market.

[0059] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a computer program. The program can be stored in a computer-readable storage medium, and when executed, the program can include the processes in the above-described method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0060] The technical features of the above-described embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification. The above-described embodiments only express several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they cannot be understood as limiting the scope of the present invention. For those of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A node electricity price scenario generation method based on spatiotemporal union and residual diffusion, characterized in that: include: S1. Preprocess the acquired multi-source time series data, including node electricity price data, wind and solar output data, load data, and date data of power system nodes, taking into account spatiotemporal correlations, to form a historical node electricity price series dataset; S2. Input the historical node electricity price series dataset into the spatiotemporal joint neural network model, extract the spatiotemporal features of the node electricity price data, and generate predicted node electricity price samples and conditional information matrices that integrate temporal dynamics and spatial topology information. S3. Input the predicted node electricity price samples and conditional information matrix into the residual denoising diffusion model, and refine the electricity price probability distribution model through the forward denoising and backward denoising process to generate node electricity price scenarios covering multi-dimensional uncertainties.

2. The node electricity price scenario generation method based on spatiotemporal combination and residual diffusion according to claim 1 is characterized in that: S1 includes: S11. Fill missing values ​​of short time scales not exceeding one day in multi-source time series data; S12. Fill missing values ​​on a time scale longer than one day in multi-source time series data.

3. The node electricity price scenario generation method based on spatiotemporal combination and residual diffusion according to claim 2 is characterized in that: Linear interpolation is used to fill in missing values ​​of short time scales not exceeding one day in multi-source time series data. The calculation formula of the linear interpolation method is as follows: ； in, is the number of samples selected during interpolation; Expressed as day Continuous time series values ​​sampled at the moment, Expressed as day Continuous time series values ​​sampled at the moment, Expressed as day Continuous time series values ​​sampled at time instants.

4. The node electricity price scenario generation method based on spatiotemporal combination and residual diffusion according to claim 2 is characterized in that: Similar day retrieval and weighted interpolation are used to fill missing values ​​in multi-source time series data with a time scale longer than one day, as follows: For node electricity price data, weighted interpolation is performed in combination with the corresponding data of the adjacent days of the base date; for wind and solar power output data and load data, the historical similarity is calculated based on the node electricity price feature vector of the base date and the node electricity price feature vector of the historical day, and the corresponding data of several similar days that are most similar to the corresponding data of the base date are selected for linear interpolation.

5. The node electricity price scenario generation method based on spatiotemporal combination and residual diffusion according to claim 2 is characterized in that: The S1 also includes: S13. Standardize and max-min normalize the node electricity price data in the multi-source time series data after filling missing values ​​to obtain the node electricity price continuous time series feature vector. Perform max-min normalization on the wind and solar power output data and load data after filling missing values ​​to obtain the output and load continuous time series feature vectors. S14. Perform one-hot encoding on the time data after filling in missing values ​​to obtain discrete date feature variables; S15. Align the continuous time series characteristic variables with the discrete date characteristic variables at the time granularity to form the historical node electricity price series dataset.

6. The node electricity price scenario generation method based on spatiotemporal combination and residual diffusion according to claim 1 is characterized in that S2 include: S21. Arrange the preprocessed daily node electricity price data into a three-dimensional tensor according to a preset fixed node order to obtain a node electricity price tensor. Calculate the historical node electricity price information matrix for the base date based on the node electricity price tensors of similar and adjacent days to the base date. S22. Calculate multi-dimensional statistical characteristics of the constructed historical node electricity price information matrix; S23. Based on the node electricity price tensor of the adjacent days of the base date, calculate the Pearson correlation coefficient between nodes and screen to generate a sparse adjacency matrix, and perform symmetric normalization on the sparse adjacency matrix; S24. A network structure composed of 3D convolutional layers, graph convolutional layers, and Conv LSTM layers is used as the spatiotemporal joint neural network model to deeply integrate the spatiotemporal features and statistical features of the node electricity price data. Control parameters are generated through control parameter modulation, and the features of the input of each layer are modulated and gated. Multi-scale features are integrated to finally obtain the predicted node electricity price samples and conditional information matrix. The spatiotemporal features include the historical node electricity price information matrix of the base date and the sparse adjacency matrix after symmetric normalization.

7. The node electricity price scenario generation method based on spatiotemporal combination and residual diffusion according to claim 6 is characterized in that: S21 includes: S211. Node electricity price tensor formation: Arrange the pre-processed daily node electricity price data into a three-dimensional tensor according to a preset fixed node order; the preset fixed node order is determined based on the node geographical distance or electrical connection relationship; S212. Formation of historical node electricity price information matrix: retrieve the corresponding node electricity price tensors of several similar days that are most similar to the corresponding node electricity price tensors of the base day according to feature similarity in the historical data, and splice them together with the node electricity price tensors of the adjacent days of the base day to obtain the historical node electricity price information matrix of the base day.

8. The node electricity price scenario generation method based on spatiotemporal combination and residual diffusion according to claim 6 is characterized in that: The statistical characteristics include basic fluctuation characteristics, fluctuation extreme value characteristics and mutation characteristics; S22 specifically includes: S221. Obtaining basic fluctuation characteristics: Quantifying the basic fluctuation characteristics of the historical node electricity price information matrix through mean and variance; S222. Obtaining extreme fluctuation characteristics: Use the maximum upper deviation and maximum lower deviation to characterize the extreme deviation of electricity prices from the mean; S223. Mutation feature acquisition: Capture the electricity price mutation features through peak location and peak duration.

9. The node electricity price scenario generation method based on spatiotemporal combination and residual diffusion according to claim 6 is characterized in that: S23 specifically includes: S231. Pearson correlation coefficient calculation: Calculate the Pearson correlation coefficient between nodes based on the standardized covariance formula to obtain the Pearson correlation coefficient matrix between nodes, which is used to represent the spatial correlation between nodes. S232. Threshold Screening and Sparsification: Filter the Pearson correlation coefficient matrix according to a preset threshold to generate a sparse adjacency matrix for representing spatial topological information. S233. Symmetric normalization: Perform symmetric normalization on the sparse adjacency matrix to obtain a normalized sparse adjacency matrix.

10. The node electricity price scenario generation method based on spatiotemporal combination and residual diffusion according to claim 9 is characterized in that: S24 specifically includes: S241. Control parameter modulation: Inputting the statistical features and the normalized sparse adjacency matrix into a multi-branch fully connected network to generate control parameters, including the modulated adjacency matrix; S242.3D Convolution: Performs a 3D convolution operation on the historical node electricity price information matrix based on a 3D convolution layer, and uses control parameters to perform statistical feature gating processing to output a 3D convolution feature map. S243. Graph Convolution: Based on the graph convolution layer, the modulated adjacency matrix is ​​combined with the 3D convolution feature map to perform a graph convolution operation, and the output is the graph convolution spatial topology fusion feature. S244. Temporal Dynamic Capture: Based on the Conv LSTM layer, the graph convolution spatial topology fusion features and the 3D convolution feature map are convolved and the temporal dynamic features are captured through a gating mechanism to obtain the hidden state of the Conv LSTM layer. S245. Multi-scale feature fusion and prediction output: The 3D convolution feature map, graph convolution spatial topology fusion features and ConvLSTM layer hidden state are spliced ​​into multi-scale fusion features, combined with the control parameters and processed by the fully connected layer to generate the predicted node electricity price samples and conditional information matrix, which serve as the initialization input and conditional parameters of the subsequent residual denoising diffusion model.

11. The node electricity price scenario generation method based on spatiotemporal combination and residual diffusion according to claim 10, characterized in that: The control parameters include control parameters of conditional batch normalization, control weights of channel gating, a modulated adjacency matrix, a scale factor, an offset factor, and a peak modulation factor; wherein the control parameters of conditional batch normalization and the control weights of channel gating are used in step S242; and the scale factor, the offset factor, and the peak modulation factor are used in step S245.

12. A node electricity price scenario generation device based on spatiotemporal union and residual diffusion, used to execute the method according to any one of claims 1 to 11, characterized in that: include: The data set establishment module is used to pre-process the acquired multi-source time series data including node electricity price data of power system nodes, wind and solar output data, load data and date data, taking into account the temporal and spatial correlation, to form a historical node electricity price series data set; The node electricity price sample generation module is used to input the historical node electricity price series data set into the spatiotemporal joint neural network model, extract the spatiotemporal characteristics of the node electricity price data, and generate predicted node electricity price samples and conditional information matrices that integrate temporal dynamics and spatial topology information; The node electricity price scenario generation module is used to input the predicted node electricity price samples and conditional information matrix into the residual denoising diffusion model, and to refine the modeling of the electricity price probability distribution through the forward denoising and reverse denoising process to generate a node electricity price scenario covering multi-dimensional uncertainties.

Images