A multi-region wind power output scene generation method and device

Abstract

The application provides a multi-region wind power output scene generation method and device, comprising: acquiring actual power and wind power prediction data of a wind power region, calculating the prediction error data through the actual power and wind power prediction data of the wind power region; inputting the actual power data and the prediction error data into a pre-trained generator to obtain a wind power output scene; wherein the generator is based on a multi-layer convolutional network to learn and train the correlation characteristics of multi-region wind power output, establish a wind power output scene generation network overall model, and acquire the wind power output scene; the wind power output scene is acquired through the trained generator, the high-dimensional mathematical modeling of the wind power output correlation can be avoided, and thus the cost of wind power output scene generation is reduced.

Description

Technical Field

[0001] This invention belongs to the field of new energy power generation technology, specifically relating to a method and apparatus for generating multi-regional wind power output scenarios. Background Technology

[0002] As wind power accounts for an increasingly larger share of the energy mix, its inherent uncertainty poses significant challenges to the safe and stable operation of power systems, power balance analysis, and optimal dispatch. Accurately characterizing the uncertainty of wind power output based on short-term power forecasts is a crucial method for fully considering the volatility and stochasticity of wind power in power system optimal dispatch decisions. Wind power output scenario generation can provide an effective cluster of wind power output curves, forming the basis for stochastic optimal dispatch and possessing strong engineering practicality.

[0003] Power system operation is often constrained by the stability limits of transmission lines, leading to power flow congestion. Stochastic optimization scheduling based on the entire grid structure is computationally time-consuming and difficult to meet engineering application requirements. The current trend is to partition the power system's sources, grid, and loads to conduct multi-regional stochastic optimization scheduling. Due to the continuity of geographical climate, wind power output in different regions exhibits certain correlations in both time and space. Simply assuming that wind power regions are independent and ignoring their spatial correlation results in a wind power output scenario set that differs significantly from actual operation. If this correlation is considered, its time-varying nonlinearity makes mathematical modeling very difficult, especially when there are many wind power regions. Summary of the Invention

[0004] To overcome the shortcomings of the existing technology, this invention proposes a method for generating multi-regional wind power output scenarios, comprising:

[0005] The actual power and wind power prediction data of the wind power area are obtained, and the prediction error data is calculated using the actual power and wind power prediction data of the wind power area.

[0006] The actual power data and prediction error data are input into a pre-trained generator to obtain the wind power output scenario;

[0007] The generator is based on a multi-layer convolutional network to learn and train the correlation features of wind power output in multiple regions, establish an overall model of the wind power output scene generation network, and obtain wind power output scenes.

[0008] Preferably, the training process of the generator includes:

[0009] Step S1: Use the actual power of the wind power area as the condition value of the multilayer convolutional network, and use the prediction error data as the noise value of the multilayer convolutional network. Obtain the first three-dimensional matrix tensor through the condition value and the noise value to generate the input set.

[0010] Step S2: Input the input set into the generator, calculate the input set through the multi-layer convolutional network in the generator, fix the generator parameters, and obtain wind power output scenario samples;

[0011] Step S3: Obtain the second three-dimensional matrix tensor through the wind power output scenario sample and the condition value, and obtain the third three-dimensional matrix tensor through the condition value and the actual wind power output scenario sample;

[0012] Step S4: Input the third three-dimensional matrix tensor and the second three-dimensional matrix tensor into the discriminator, update the network parameters of the discriminator, and determine the Wasserstein distance through the discriminator;

[0013] Step S5: Determine whether the Wasserstein distance meets the preset convergence requirement. If it does not meet the requirement, update the generator parameters according to the wind power output scenario sample and return to step S3 to continue training. Otherwise, end the training and obtain the trained generator.

[0014] The generator and discriminator are network models trained on the input set.

[0015] Preferably, step S1 includes:

[0016] The actual power of the wind power area is used as the condition value of the multilayer convolutional network, and the prediction error data is used as the noise value of the multilayer convolutional network.

[0017] The noise values ​​are sorted according to the wind power region to generate a noise value sequence, and the condition values ​​are sorted according to the wind power region to generate a condition value sequence.

[0018] For each wind power region, the condition value sequences of each wind power region and different times in the condition value sequence and the noise value sequences of the same wind power region and different times in the noise value sequence are spliced ​​together according to the corresponding times to generate a matrix corresponding to each wind power region.

[0019] The matrices corresponding to each wind power area are horizontally spliced ​​together according to the condition values ​​and noise values ​​at the same time to obtain the first three-dimensional matrix tensor.

[0020] The first three-dimensional matrix tensor is then padded with edges on all four sides to generate the input set.

[0021] The preferred expression for wind power output scenarios is as follows:

[0022]

[0023] In the formula, Let c be the wind power output value of the wind power area w2 at time t2; ij Let z be the conditional value of region i at time j; ij The noise value of region i at time j, where i = 1, 2, 3, j = 1, 2, 3.

[0024] Preferably, the generator loss function is as follows:

[0025] Loss G =-E x′～p(x′) [D(x′|c)]

[0026] In the formula, Loss G denoted by ; x′ represents the wind power output scenario sample generated by generator G; E represents the expected value of the wind power output scenario sample and the corresponding wind power output scenario sample data distribution; D represents the discriminator function; p(x′) represents the wind power output scenario sample data distribution; and c represents the condition value.

[0027] Preferably, the discriminator loss function is as follows:

[0028] Loss D =E x′～p(x′) [D(x′|c)]-E x～p(x) [D(x|c)])

[0029] In the formula, Loss D denoted by the discriminator loss function; x represents the actual wind power output scenario sample; x′ represents the wind power output scenario sample generated by generator G; E represents the expected value of the wind power output scenario sample and the corresponding wind power output scenario sample data distribution; D represents the discriminator function; p(x′) represents the wind power output scenario sample data distribution; p(x) represents the actual wind power output scenario sample data distribution; c represents the condition value.

[0030] Preferably, the Wasserstein distance calculation formula is as follows:

[0031]

[0032] In the formula, W represents the Wasserstein distance; x represents the actual wind power output scenario sample; x′ represents the wind power output scenario sample generated by generator G; γ represents the joint distribution of the optimal path of the wind power output scenario sample and the actual wind power output scenario sample; p(x) represents the distribution of the actual wind power output scenario sample data; p(x′) represents the distribution of the wind power output scenario sample data.

[0033] W(p(x),p(x′)) represents the distance between the actual wind power output scenario sample data distribution p(x) and the wind power output scenario sample data distribution p(x′); ∏(p(x),p(x′)) represents the joint distribution of the actual wind power output scenario sample data distribution p(x) and the wind power output scenario sample data distribution p(x′); ||xx′|| represents the distance between the actual wind power output scenario sample and the wind power output scenario sample; E represents the expected value of the corresponding distribution of the actual wind power output scenario sample and the wind power output scenario sample; inf represents the lower bound of the expected value E.

[0034] Based on the same inventive concept, this application also provides a multi-regional wind power output scenario generation device, comprising:

[0035] The data calculation module is used to acquire the actual power and wind power prediction data of the wind power area, and to calculate the prediction error data using the actual power and wind power prediction data of the wind power area.

[0036] The scene generation module is used to input the actual power data and prediction error data into a pre-trained generator to obtain a wind power output scene;

[0037] The generator is based on a multi-layer convolutional network to learn and train the correlation features of wind power output in multiple regions, establish an overall model of the wind power output scene generation network, and obtain wind power output scenes.

[0038] Preferably, the generator training process in the scene generation module includes:

[0039] Step W1: Use the actual power of the wind power area as the condition value of the multilayer convolutional network, and use the prediction error data as the noise value of the multilayer convolutional network. Obtain the first three-dimensional matrix tensor through the condition value and the noise value to generate the input set.

[0040] Step W2: Input the input set into the generator, perform calculations on the input set through the multi-layer convolutional network in the generator, fix the generator parameters, and obtain wind power output scenario samples;

[0041] Step W3: Obtain the second three-dimensional matrix tensor using the wind power output scenario sample and the condition value, and obtain the third three-dimensional matrix tensor using the condition value and the actual wind power output scenario sample;

[0042] Step W4: Input the third three-dimensional matrix tensor and the second three-dimensional matrix tensor into the discriminator, update the network parameters of the discriminator, and determine the Wasserstein distance through the discriminator;

[0043] Step W5: Determine whether the Wasserstein distance meets the preset convergence requirements. If it does not meet the requirements, update the generator parameters based on the wind power output scenario samples and return to step S3 to continue training. Otherwise, end the training and obtain the trained generator.

[0044] The generator and discriminator are network models trained on the input set.

[0045] Preferably, step W1 includes:

[0046] The actual power of the wind power area is used as the condition value of the multilayer convolutional network, and the prediction error data is used as the noise value of the multilayer convolutional network.

[0047] The noise values ​​are sorted according to the wind power region to generate a noise value sequence, and the condition values ​​are sorted according to the wind power region to generate a condition value sequence.

[0048] For each wind power region, the condition value sequences of each wind power region and different times in the condition value sequence and the noise value sequences of the same wind power region and different times in the noise value sequence are spliced ​​together according to the corresponding times to generate a matrix corresponding to each wind power region.

[0049] The matrices corresponding to each wind power area are horizontally spliced ​​together according to the condition values ​​and noise values ​​at the same time to obtain the first three-dimensional matrix tensor.

[0050] The first three-dimensional matrix tensor is then padded with edges on all four sides to generate the input set.

[0051] The preferred expression for wind power output scenarios is as follows:

[0052]

[0053] In the formula, Let c be the wind power output value of the wind power area w2 at time t2; ij Let z be the conditional value of region i at time j; ij The noise value of region i at time j, where i = 1, 2, 3, j = 1, 2, 3.

[0054] Preferably, the generator loss function is as follows:

[0055] Loss G =-E x′～p(x′) [D(x′|c)]

[0056] In the formula, Loss Gdenoted by ; x′ represents the wind power output scenario sample generated by generator G; E represents the expected value of the wind power output scenario sample and the corresponding wind power output scenario sample data distribution; D represents the discriminator function; p(x′) represents the wind power output scenario sample data distribution; and c represents the condition value.

[0057] Preferably, the discriminator loss function is as follows:

[0058] Loss D =E x′～p(x′) [D(x′|c)]-E x～p(x) [D(x|c)])

[0059] In the formula, Loss D denoted by the discriminator loss function; x represents the actual wind power output scenario sample; x′ represents the wind power output scenario sample generated by generator G; E represents the expected value of the wind power output scenario sample and the corresponding wind power output scenario sample data distribution; D represents the discriminator function; p(x′) represents the wind power output scenario sample data distribution; p(x) represents the actual wind power output scenario sample data distribution; c represents the condition value.

[0060] Preferably, the Wasserstein distance calculation formula is as follows:

[0061]

[0062] In the formula, W represents the Wasserstein distance; x represents the actual wind power output scenario sample; x′ represents the wind power output scenario sample generated by generator G; γ represents the joint distribution of the optimal path of the wind power output scenario sample and the actual wind power output scenario sample; p(x) represents the distribution of the actual wind power output scenario sample data; p(x′) represents the distribution of the wind power output scenario sample data.

[0063] W(p(x),p(x′)) represents the distance between the actual wind power output scenario sample data distribution p(x) and the wind power output scenario sample data distribution p(x′); ∏(p(x),p(x′)) represents the joint distribution of the actual wind power output scenario sample data distribution p(x) and the wind power output scenario sample data distribution p(x′); ||xx′|| represents the distance between the actual wind power output scenario sample and the wind power output scenario sample; E represents the expected value of the actual wind power output scenario sample and the corresponding distribution of the wind power output scenario sample; inf represents the lower bound of the expected value E.

[0064] Compared with the closest existing technology, the present invention has the following beneficial effects:

[0065] 1. This invention provides a method and apparatus for generating wind power output scenarios in multiple regions, comprising: acquiring actual power and predicted wind power data of wind power regions; calculating prediction error data using the actual power and predicted wind power data of the wind power regions; inputting the actual power data and prediction error data into a pre-trained generator to obtain wind power output scenarios; wherein the generator is based on a multi-layer convolutional network to learn and train the correlation features of wind power output in multiple regions, establishing an overall model of the wind power output scenario generation network, and obtaining wind power output scenarios; this invention obtains wind power output scenarios by training a generator, which can avoid high-dimensional mathematical modeling of wind power output correlation, thereby reducing the cost of generating wind power output scenarios.

[0066] 2. This invention designs a structure based on wind power output characteristics and network structure, and constructs a three-dimensional matrix tensor by splicing the conditional value and noise value sequences through a convolutional network. The resulting three-dimensional matrix tensor is then used to train the generator and discriminator, which can fully learn the complex correlation of wind power output and improve the reliability of wind power output scenarios. Attached Figure Description

[0067] Figure 1 This is a schematic diagram of a method for generating multi-regional wind power output scenarios provided by the present invention;

[0068] Figure 2 This is a schematic diagram of three-dimensional data stitching and reconstruction in a multi-region wind power output scene generation method provided by the present invention;

[0069] Figure 3 This is a schematic diagram of three-dimensional convolution calculation in a multi-region wind power output scene generation method provided by the present invention;

[0070] Figure 4 This is a schematic diagram of the overall structure of the scene generation network in a multi-region wind power output scene generation method provided by the present invention;

[0071] Figure 5-(a) is a schematic diagram of the generator G network structure in a multi-region wind power output scenario generation method provided by the present invention;

[0072] Figure 5-(b) is a schematic diagram of the discriminator D network structure in a multi-region wind power output scenario generation method provided by the present invention;

[0073] Figure 6 This is a schematic diagram of the generator training strategy in a multi-region wind power output scenario generation method provided by the present invention;

[0074] Figure 7 This is a schematic diagram of the connection of a multi-region wind power output scene generation device provided by the present invention. Detailed Implementation

[0075] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0076] Example 1:

[0077] like Figure 1 As shown, the present invention provides a method for generating multi-regional wind power output scenarios, comprising:

[0078] Step 1: Obtain the actual power and wind power prediction data of the wind power area, and calculate the prediction error data using the actual power and wind power prediction data of the wind power area;

[0079] Step 2: Input the actual power data and prediction error data into the pre-trained generator to obtain the wind power output scenario;

[0080] The generator is based on a multi-layer convolutional network to learn and train the correlation features of wind power output in multiple regions, establish an overall model of the wind power output scene generation network, and obtain wind power output scenes.

[0081] Specifically, step 1 includes: acquiring actual power and wind power prediction data for wind power areas, and calculating prediction error data using the actual power and wind power prediction data for the wind power areas. This invention collected actual power and wind power prediction data for five wind power areas in Northwest my country. The dataset covers the period from 00:00 on January 1, 2019 to 23:45 on December 31, 2020, with a data acquisition time resolution of 15 minutes, and values ​​are expressed in per-unit values. Based on the large amount of collected actual wind power data, the spatiotemporal correlation characteristics between wind power output in multiple regions are analyzed. Acquiring actual power and wind power prediction data for wind power areas, and calculating prediction error data using the actual power and wind power prediction data for the wind power areas, wherein the prediction error data is the difference between the actual power data and the wind power prediction data.

[0082] The specific training process of the generator in step 2 includes: Step S1: Using the actual power of the wind power area as the condition value of the multi-layer convolutional network, and the prediction error data as the noise value of the multi-layer convolutional network, the first three-dimensional matrix tensor is obtained through the condition value and the noise value to generate the input set; Step S2: The input set is input to the generator, and the multi-layer convolutional network in the generator is used to calculate the input set, and the generator parameters are fixed to obtain wind power output scene samples; Step S3: Using the wind power output scene samples and the condition... The conditions are used to obtain the second three-dimensional matrix tensor, and the third three-dimensional matrix tensor is obtained through the condition values ​​and actual wind power output scenario samples; Step S4: Input the third three-dimensional matrix tensor and the second three-dimensional matrix tensor into the discriminator, update the network parameters of the discriminator, and determine the Wasserstein distance through the discriminator; Step S5: Determine whether the Wasserstein distance meets the preset convergence requirements. If it does not meet the requirements, update the generator parameters according to the wind power output scenario samples and return to step S3 to continue training; otherwise, end the training and obtain the trained generator; Figure 2 As shown, wind power prediction data and prediction error are used as the network's conditional value c and noise z inputs, respectively. Assume there are m wind power areas, Z... i and C i Let Z represent the random sampling sequence of the prediction error distribution and the power prediction data for the i-th wind power region, respectively. Then the input sequence set of the noise z is Z = {Z1, Z2, ..., Zn}. m-1 Z m The input sequence set for the conditional information c is C = {C1, C2, ..., C}. m-1 C m The conditional value sequence z and noise sequence c of m regions are vertically concatenated to form m planar matrices. Then, the m planar matrices are horizontally concatenated to form a three-dimensional matrix tensor from the input data.

[0083] If the correlation of wind power output between adjacent time points and adjacent regions is considered, a 3×3×3 convolution kernel is selected to perform convolution calculation on the network input. The calculation process is as follows: Figure 3 As shown. When generating wind power output scenarios, to ensure that the output and input sizes remain unchanged after each convolution calculation, this invention performs edge-padding processing on the input 3D data before the convolution calculation. If a convolution calculation is performed on wind power region 2 at time t2, the wind power output scenario value at time t2 will be... The mathematical expression for is shown in equation (1).

[0084]

[0085] In the formula, c represents the wind power output scenario value of the w2 wind power area at time t2;ij Let z be the conditional value of region i at time j; ij The noise value of region i at time j, where i = 1, 2, 3, j = 1, 2, 3.

[0086] A convolutional network is used to design the conditional generative adversarial network (GAN) structure. The GAN mainly consists of a generator G and a discriminator D, and its overall structure is as follows: Figure 4 As shown, the predicted wind power and wind power prediction error random sequences from multiple regions are integrated into a three-dimensional tensor, which serves as the input to the generator G. This tensor is then processed by a multi-layer convolutional network to learn the correlation features of wind power output across multiple regions, and finally, generated samples are output. The generated samples and real wind power data are concatenated with the predicted wind power in the same way and input into the discriminator D. The discriminator D learns to distinguish between generated and real data and outputs a discrimination result. The network continuously updates its parameters based on the output of the discriminator D, thus completing the learning process for generating multi-regional wind power output scenarios.

[0087] The network structures of generator G and discriminator D are shown in Figures 5-(a) and 5-(b). In the input and intermediate layers of generator G, batch normalization and activation functions are used after the convolutional layers. Because the distribution of output data changes continuously in each layer, the data distribution between the input and output layers of the generator network will differ significantly after processing through six convolutional layers. Batch normalization ensures that the output data distribution of each layer remains constant, avoiding the problem of intermediate layers changing with the input data distribution. The activation function layer uses the LeakyReLU function to prevent entering the training dead zone. In the output layer of the generator network, batch normalization is not required after convolution; it can be directly connected to the activation function layer. The activation function layer is set to ReLU, ensuring that the network output data meets the condition that the wind power output is not less than 0.

[0088] The discriminator network's output and intermediate layers are all designed using convolutional networks. Each convolutional layer undergoes batch normalization after convolution, followed by activation using the LeakyReLU function. The LeakyReLU activation function is chosen for each layer to avoid gradient vanishing during training. The discriminator network's output layer consists of two fully connected networks using full-plane weights to calculate the final discriminant value.

[0089] Training a conditional generative adversarial network mainly involves training the generator G and training the discriminator D. The training strategy is as follows: Figure 6As shown in the figure. The generator G learns from the input data and outputs the generated sample x′=G(z|c). The discriminator D distinguishes the input sample, judges the similarity between the generated sample x′ and the real sample x, and outputs the discriminant value D(·|c). The generator loss function and the discriminator loss function are shown in equation (2) and equation (3), respectively.

[0090] Loss G =-E x′～p(x′) [D(x′|c)](2)

[0091]

[0092] In the formula, Loss D Represents the discriminator loss function; Loss G Let G represent the generator loss function; x represent the actual wind power output scenario sample; x′ represent the wind power output scenario sample generated by generator G; E represent the expected value of the distribution of wind power output scenario sample and actual wind power output scenario sample data; D represent the discriminator function; p(x′) represent the distribution of wind power output scenario sample data; p(x) represent the distribution of actual wind power output scenario sample data; c represents the condition value.

[0093] The optimization objective function of the generative adversarial network is shown in equation (4).

[0094]

[0095] During network training, the discriminator D's training optimization objective is to continuously decrease the D(x′|c) of the generated sample x′ and continuously increase the D(x|c) of the real sample x, maximizing the difference between the generated sample x′ and the real sample x. The training optimization direction is max(V(D,G)). The generator G's training objective is to continuously increase the D(x′|c) value of the generated sample x′, minimizing the difference between the generated sample x′ and the real sample x. The optimization direction is min(V(D,G)).

[0096] To avoid problems such as training difficulties and pattern collapse, this invention introduces Wasserstein distance into the discriminator loss function to measure the distance between generated samples and real samples, as defined in equation (5).

[0097]

[0098] In the formula, γ represents a joint distribution representing the optimal path; W represents p r The distance between (x) and p(x′);

[0099] ∏(p(x),p(x′)) represents the joint distribution of p(x) and p(x′); ||xx′|| represents the distance between the real sample and the generated sample. The improved discriminator loss function is shown in Equation (6).

[0100]

[0101] In the formula, This indicates that random sampling is performed on the line connecting the real sample x and the generated sample x′; λ represents a constant coefficient.

[0102] This invention evaluates the rationality and effectiveness of the generated scenarios from the perspectives of volatility, temporal correlation, spatial correlation, and reliability. The evaluation metrics are as follows:

[0103] 1) Maximum fluctuation range: The maximum fluctuation range refers to the difference between the maximum and minimum values ​​within a time T (15min, 30min, 60min, etc.), as defined in equation (7).

[0104]

[0105] In the formula, P tmax =Max{P t+i},P tmin =Min{P t+i}, i=0,1,2,…,T / △t-1,t ptmax t ptmin P respectively tmax P tmin The time elapsed since the beginning of the time series is used to represent P. tmax P tmin At the moment of appearance, P N This refers to the rated total installed capacity.

[0106] 2) Autocorrelation coefficient: describes the correlation of wind power output over time, and is defined as shown in equation (8).

[0107]

[0108] In the formula, t represents time and k represents time interval; This represents the average wind power output. This represents the autocorrelation of wind power output over a time interval of k, with a value range of [-1, 1].

[0109] 3) Spearman correlation coefficient: describes the spatial correlation between the power output of two wind power areas, and is defined as shown in equation (9).

[0110]

[0111] In the formula, Px P y P represents the actual wind power output in regions x and y, respectively; x(i) and P y(i) These represent the sample values ​​with sort number i. This represents the mean of the wind power output sequence.

[0112] 4) Coverage: describes the inclusion of the generated scene set in the actual wind power output, and is defined as shown in Equation (10).

[0113]

[0114] In the formula, T (1-α) The number of times the actual wind power output falls within the confidence level 1-α range; T represents the number of times in the wind power output sequence; CR (1-α) This represents the coverage rate at a confidence level of 1-α.

[0115] 5) Average width of power interval: evaluates the random uncertainty of the generated scene distribution, and is defined as shown in Equation (11).

[0116]

[0117] In the formula, PIAW (1-α) This represents the average width of the power interval at a confidence level of 1-α; This represents the upper bound of the wind power output scenario set at time t under confidence level 1-α; This represents the lower bound of the wind power output scenario set at time t under confidence level 1-α.

[0118] 6) Pinball loss function: It comprehensively evaluates the performance of the generated wind power output scenario in terms of coverage and average width of power range, and provides a comprehensive quantitative index for the reliability of the generated scenario. It is defined as shown in Equation (12) and Equation (13).

[0119]

[0120]

[0121] In the formula, S represents the loss function value between wind power output and the generated scenario distribution; S(P t Q t ) represents the loss function value at time t; m represents the number of quantiles; α i Represents the i-th quantile; P represents the wind power output scenario value corresponding to the i-th quantile; realThis represents the actual wind power output. A three-dimensional matrix tensor is constructed by concatenating the sequence of conditional and noise values ​​using a convolutional network. This three-dimensional matrix tensor is then used to train the generator and discriminator, enabling the generator to fully learn the complex relationships between wind power output and improving the reliability of wind power output scenarios.

[0122] Example 2:

[0123] like Figure 7 As shown, the present invention provides a multi-region wind power output scenario generation device, comprising:

[0124] The data calculation module is used to acquire the actual power and wind power prediction data of the wind power area, and to calculate the prediction error data using the actual power and wind power prediction data of the wind power area.

[0125] The scene generation module is used to input the actual power data and prediction error data into a pre-trained generator to obtain a wind power output scene;

[0126] The generator is based on a multi-layer convolutional network to learn and train the correlation features of wind power output in multiple regions, establish an overall model of the wind power output scene generation network, and obtain wind power output scenes.

[0127] Specifically, the data calculation module is used to acquire actual power and wind power prediction data for wind power areas, and to calculate prediction error data using this data. This invention collected actual power and wind power prediction data for five wind power areas in Northwest my country. The dataset spans from 00:00 on January 1, 2019 to 23:45 on December 31, 2020, with a data acquisition time resolution of 15 minutes and per-unit values. Based on the large amount of collected actual wind power data, the spatiotemporal correlation characteristics between wind power output in multiple regions are analyzed. The actual power and wind power prediction data for wind power areas are acquired, and prediction error data is calculated using this data, where the prediction error data is the difference between the actual power data and the predicted wind power data.

[0128] Specifically, the training process of the generator in the scene generation module includes:

[0129] Step W1: Use the actual power of the wind power area as the condition value of the multilayer convolutional network, and use the prediction error data as the noise value of the multilayer convolutional network. Obtain the first three-dimensional matrix tensor through the condition value and the noise value to generate the input set.

[0130] Step W2: Input the input set into the generator, perform calculations on the input set through the multi-layer convolutional network in the generator, fix the generator parameters, and obtain wind power output scenario samples;

[0131] Step W3: Obtain the second three-dimensional matrix tensor using the wind power output scenario sample and the condition value, and obtain the third three-dimensional matrix tensor using the condition value and the actual wind power output scenario sample;

[0132] Step W4: Input the third three-dimensional matrix tensor and the second three-dimensional matrix tensor into the discriminator, update the network parameters of the discriminator, and determine the Wasserstein distance through the discriminator;

[0133] Step W5: Determine whether the Wasserstein distance meets the preset convergence requirements. If it does not meet the requirements, update the generator parameters based on the wind power output scenario samples and return to step S3 to continue training. Otherwise, end the training and obtain the trained generator.

[0134] The generator and discriminator are network models trained on the input set.

[0135] The following uses five wind power areas as examples to further illustrate the implementation of the present invention. Here, the illustrative embodiments and descriptions of the present invention are used to explain the present invention, but are not intended to limit the present invention. Among them, (1) the wind power predicted power data and actual power data are preprocessed to analyze and understand the spatiotemporal correlation characteristics of wind power output and extract the power prediction error distribution characteristics; (2) the wind power predicted power data and prediction error are used as the condition value c and noise z input of the generator G, respectively. Set the wind power output scenario sequence length to t, reconstruct the input data of the 5 wind power regions into a three-dimensional data tensor of size t×5×2, and input it into the generator G. The generator outputs the wind power output scenario of size t×5×1; (3) Reconstruct the generated wind power output scenario and the condition values ​​c corresponding to the 5 regions into a three-dimensional data tensor of size t×5×2, and input it into the discriminator D. The discriminator outputs the discrimination value between the actual wind power output and the wind power output scenario; (4) Calculate the loss function of the generator G and the discriminator D, and use the RMSprop optimization algorithm to update the network parameters of the generator and the discriminator respectively. The Wasserstein distance is determined. If the convergence requirement is met, the training ends; otherwise, return to (2) for the next round of training. (5) After training ends, the trained generator G is extracted. The predicted power data of the five wind power regions and the 50 noise distributions corresponding to the regions are reconstructed into 50 three-dimensional tensors of size t×5×2. These tensors are then input into the generator G, which outputs 50 regional wind power output scenarios. By training the generator and discriminator, wind power output scenarios can be obtained, which avoids high-dimensional mathematical modeling of wind power output correlation and thus reduces the cost of generating wind power output scenarios.

[0136] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0137] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0138] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0139] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0140] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or equivalent substitutions to the specific implementation methods of the application, but these changes, modifications or equivalent substitutions are all within the scope of protection of the claims pending approval.

Claims

1. A method for generating multi-regional wind power output scenarios, characterized in that, include: The actual power and wind power prediction data of the wind power area are obtained, and the prediction error data is calculated using the actual power and wind power prediction data of the wind power area. The actual power data and prediction error data are input into a pre-trained generator to obtain the wind power output scenario; The generator is based on a multi-layer convolutional network to learn and train the correlation features of wind power output in multiple regions, establish an overall model of the wind power output scene generation network, and obtain wind power output scenes. The training process of the generator includes: Step S1: Use the actual power of the wind power area as the condition value of the multilayer convolutional network, and use the prediction error data as the noise value of the multilayer convolutional network. Obtain the first three-dimensional matrix tensor through the condition value and the noise value to generate the input set. Step S2: Input the input set into the generator, calculate the input set through the multi-layer convolutional network in the generator, fix the parameters of the generator, and obtain wind power output scenario samples; Step S3: Obtain the second three-dimensional matrix tensor through the wind power output scenario sample and the condition value, and obtain the third three-dimensional matrix tensor through the condition value and the actual wind power output scenario sample; Step S4: Input the third three-dimensional matrix tensor and the second three-dimensional matrix tensor into the discriminator, update the network parameters of the discriminator, and determine the Wasserstein distance through the discriminator; Step S5: Determine whether the Wasserstein distance meets the preset convergence requirement. If it does not meet the requirement, update the generator parameters according to the wind power output scenario sample and return to step S3 to continue training. Otherwise, end the training and obtain the trained generator. The generator and discriminator are network models trained on the input set. Step S1 includes: The actual power of the wind power area is used as the condition value of the multilayer convolutional network, and the prediction error data is used as the noise value of the multilayer convolutional network. The noise values ​​are sorted according to the wind power region to generate a noise value sequence, and the condition values ​​are sorted according to the wind power region to generate a condition value sequence. For each wind power region, the condition value sequences of each wind power region and different times in the condition value sequence and the noise value sequences of the same wind power region and different times in the noise value sequence are spliced ​​together according to the corresponding times to generate a matrix corresponding to each wind power region. The matrices corresponding to each wind power area are horizontally spliced ​​together according to the condition values ​​and noise values ​​at the same time to obtain the first three-dimensional matrix tensor. The first three-dimensional matrix tensor is then padded with edges on all four sides to generate the input set.

2. The method as described in claim 1, characterized in that, The expression for wind power output scenarios is as follows: In the formula, for t Wind power output scenario value of wind power area w2 at time 2; For the region i exist j The condition value at time; For the region i exist j The noise value at time, where i =1, 2, 3 j =1, 2, 3.

3. The method as described in claim 1, characterized in that, The generator loss function is as follows: In the formula ， Represents the generator loss function; Represents generator Generated wind power output scenario samples; This represents the expected value of the wind power output scenario sample and the corresponding wind power output scenario sample data distribution; This represents the discriminator function; This indicates the distribution of sample data for wind power output scenarios; Indicates the condition value.

4. The method as described in claim 1, characterized in that, The discriminator loss function is as follows: In the formula, This represents the discriminator loss function; x This represents a sample of actual wind power output scenarios; Represents generator Generated wind power output scenario samples; This represents the expected value of the wind power output scenario sample and the corresponding wind power output scenario sample data distribution; This represents the discriminator function; This indicates the distribution of sample data for wind power output scenarios; This represents the distribution of sample data for actual wind power output scenarios; Indicates the condition value.

5. The method as described in claim 1, characterized in that, The Wasserstein distance calculation formula is as follows: In the formula, W represents the Wasserstein distance; x This represents a sample of actual wind power output scenarios; This represents a wind power output scenario sample generated by generator G; This represents the joint distribution of the optimal paths for wind power output scenario samples and actual wind power output scenario samples; This represents the distribution of sample data for actual wind power output scenarios; This indicates the distribution of sample data for wind power output scenarios; This represents the distribution of sample data for actual wind power output scenarios. Distribution of sample data for wind power output scenarios The distance between them; This represents the distribution of sample data for actual wind power output scenarios. Distribution of sample data for wind power output scenarios The joint distribution of the two; This represents the distance between the actual wind power output scenario sample and the wind power output scenario sample. This represents the expected value of the actual wind power output scenario sample and the corresponding distribution of the wind power output scenario sample; Indicates expected value The lower bound value.

6. A multi-regional wind power output scenario generation device, characterized in that, include: The data calculation module is used to acquire the actual power and wind power prediction data of the wind power area, and to calculate the prediction error data using the actual power and wind power prediction data of the wind power area. The scene generation module is used to input the actual power data and prediction error data into a pre-trained generator to obtain a wind power output scene; The generator is based on a multi-layer convolutional network to learn and train the correlation features of wind power output in multiple regions, establish an overall model of the wind power output scene generation network, and obtain wind power output scenes. The generator training process in the scene generation module includes: Step W1: Use the actual power of the wind power area as the condition value of the multilayer convolutional network, and use the prediction error data as the noise value of the multilayer convolutional network. Obtain the first three-dimensional matrix tensor through the condition value and the noise value to generate the input set. Step W2: Input the input set into the generator, calculate the input set through the multi-layer convolutional network in the generator, fix the parameters of the generator, and obtain wind power output scenario samples; Step W3: Obtain the second three-dimensional matrix tensor using the wind power output scenario sample and the condition value, and obtain the third three-dimensional matrix tensor using the condition value and the actual wind power output scenario sample; Step W4: Input the third three-dimensional matrix tensor and the second three-dimensional matrix tensor into the discriminator, update the network parameters of the discriminator, and determine the Wasserstein distance through the discriminator; Step W5: Determine whether the Wasserstein distance meets the preset convergence requirements. If it does not meet the requirements, update the generator parameters based on the wind power output scenario samples and return to step W3 to continue training. Otherwise, end the training and obtain the trained generator. The generator and discriminator are network models trained on the input set. Step W1 includes: The actual power of the wind power area is used as the condition value of the multilayer convolutional network, and the prediction error data is used as the noise value of the multilayer convolutional network. The noise values ​​are sorted according to the wind power region to generate a noise value sequence, and the condition values ​​are sorted according to the wind power region to generate a condition value sequence. For each wind power region, the condition value sequences of each wind power region and different times in the condition value sequence and the noise value sequences of the same wind power region and different times in the noise value sequence are spliced ​​together according to the corresponding times to generate a matrix corresponding to each wind power region. The matrices corresponding to each wind power region are horizontally spliced ​​according to the condition values ​​and noise values ​​at the same time to obtain the first three-dimensional matrix tensor; and the first three-dimensional matrix tensor is padded with edges on all four sides to generate the input set.

7. The apparatus as claimed in claim 6, characterized in that, The expression for wind power output scenarios is as follows: In the formula, for t Wind power output scenario value of wind power area w2 at time 2; For the region i exist j The condition value at time; For the region i exist j The noise value at time, where i =1, 2, 3 j =1, 2, 3.

8. The apparatus as claimed in claim 6, characterized in that, The generator loss function is as follows: In the formula ， Represents the generator loss function; Represents generator Generated wind power output scenario samples; This represents the expected value of the wind power output scenario sample and the corresponding wind power output scenario sample data distribution; This represents the discriminator function; This indicates the distribution of sample data for wind power output scenarios; Indicates the condition value.

9. The apparatus as claimed in claim 6, characterized in that, The discriminator loss function is as follows: In the formula, This represents the discriminator loss function; x This represents a sample of actual wind power output scenarios; Represents generator Generated wind power output scenario samples; This represents the expected value of the wind power output scenario sample and the corresponding wind power output scenario sample data distribution; This represents the discriminator function; This indicates the distribution of sample data for wind power output scenarios; This represents the distribution of sample data for actual wind power output scenarios; Indicates the condition value.

10. The apparatus as claimed in claim 6, characterized in that, The Wasserstein distance calculation formula is as follows: In the formula, W represents the Wasserstein distance; x This represents a sample of actual wind power output scenarios; This represents a wind power output scenario sample generated by generator G; This represents the joint distribution of the optimal paths for wind power output scenario samples and actual wind power output scenario samples; This represents the distribution of sample data for actual wind power output scenarios; This indicates the distribution of sample data for wind power output scenarios; This represents the distribution of sample data for actual wind power output scenarios. Distribution of sample data for wind power output scenarios The distance between them; This represents the distribution of sample data for actual wind power output scenarios. Distribution of sample data for wind power output scenarios The joint distribution of the two; This represents the distance between the actual wind power output scenario sample and the wind power output scenario sample. This represents the expected value of the actual wind power output scenario sample and the corresponding distribution of the wind power output scenario sample; Indicates expected value The lower bound value.

Images