Wind-solar-load coupling scene generation method and system based on LSTM-GAN

Abstract

The present invention discloses a method and system for generating a wind-solar-load coupling scene based on LSTM-GAN, the method comprising: obtaining provincial wind-solar-load data, the provincial wind-solar-load data including wind power output data, photovoltaic output data and load data; according to the provincial wind-solar-load data, using the generator in the improved generative adversarial network model to obtain a generated scene; the improved generative adversarial network model is formed by introducing a fully connected layer and a long short-term memory network LSTM into the generator of the generative adversarial network model; inputting Gaussian random noise and a real scene into the discriminator of the improved generative adversarial network model, cross-training the generator and the discriminator through adversarial training; extracting the trained generator from the improved generative adversarial network model, inputting Gaussian random noise into the trained generator, and generating a random daily wind-solar-load scene. The present invention can well characterize the timing and coupling characteristics of wind-solar-load, and the model accuracy is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of new energy output and load scenario generation, and in particular to a method and system for generating wind-solar-load coupling scenarios based on LSTM-GAN. Background Art

[0002] As the penetration rate of renewable energy generation in the existing power grid increases, how to describe the uncertainty of renewable energy output is the key to the stable operation of the power system. The common practice is to guide all parties in the power system to reasonably plan their respective production and scheduling plans by establishing a set of renewable energy output scenarios in the medium and long term. Through the scenario analysis method, the operation risk and operation cost of the renewable energy power system can be effectively reduced.

[0003] There are three common approaches to mid- to long-term scenario generation:

[0004] Method (1) Direct selection or clustering method. Select representative monthly output curves from the annual historical data, and splice the 12-month curves to obtain the final typical annual curve. Because they are all historical data, the output scenario curves obtained by this method are difficult to represent the uncertainty of renewable energy output and load power.

[0005] Method (2) Random sampling scenario generation method. A probability model that obeys the law of renewable energy generation and load distribution is established through a large amount of historical data, and a random scenario that satisfies the distribution is generated through random sampling. This method relies on a pre-assumed probability distribution, while the renewable energy output curve is complex and changeable, and often cannot be accurately described by a mathematical model, which greatly limits the adaptability of this method.

[0006] Method (3) is based on data-driven generative adversarial network learning. This method has been proven to be superior to methods (1) and (2) in terms of accuracy and applicability. Through adversarial learning between the generator and the learner, the generator can generate random scenes that are close to the real situation. Compared with method (1), this method increases the randomness of the scene, and compared with method (2), it does not need to pre-assume the probability distribution of the data.

[0007] Regarding method (3), the current common practices are:

[0008] 1) Use the GAN model to generate scenes for a single station. That is, generate random scenes for a certain wind farm or photovoltaic farm. This method is a common practice for GAN to generate scenes, but because it only targets a single station, its actual application scope is very limited.

[0009] 2) The data of different stations are artificially arranged into a two-dimensional matrix, and the convolution method is introduced to generate random scenes of joint stations. Since the stations are topologically distributed in space, it is difficult for the two-dimensional matrix to express the real geographical relationship between the stations. In addition, due to the differences in time and space of the data, it is difficult for the convolution kernel to effectively capture the inherent laws of the data.

[0010] In view of this, this application is hereby filed. Summary of the invention

[0011] The technical problem to be solved by the present invention is that the existing medium- and long-term scene generation methods are only targeted at a single station, a single element or a small area, and usually cannot achieve scene generation under the common mode of source and load. Its practical application scope is very limited, and the generation model cannot well characterize the timing and coupling characteristics of wind, light and load, and the model accuracy is not high. The purpose of the present invention is to provide a wind, light and load coupling scene generation method and system based on LSTM-GAN. The present invention takes provincial wind, light and load data as the target object, and introduces an LSTM model in the generator of the GAN model, which can effectively extract the timing feature information under data coupling, and input this information into the generator of the model to generate random scenes under wind, light and load coupling. The accuracy of the present invention is greatly improved compared with the traditional GAN.

[0012] The present invention is achieved through the following technical solutions:

[0013] In a first aspect, the present invention provides a method for generating a wind-solar-load coupling scenario based on LSTM-GAN, the method comprising:

[0014] Obtain provincial wind and solar load data, which includes wind power output data, photovoltaic output data and load data;

[0015] According to the provincial wind and solar load data, the generator in the improved generative adversarial network model is used to obtain the generated scene; the improved generative adversarial network model is formed by introducing a fully connected layer and a long short-term memory network LSTM into the generator of the generative adversarial network model;

[0016] Gaussian random noise and real scenes are input into the discriminator of the improved generative adversarial network model, and the generator and discriminator are cross-trained through adversarial training.

[0017] The trained generator is extracted from the improved generative adversarial network model, and Gaussian random noise is input into the trained generator to generate random day-wind-solar-load scenes.

[0018] Furthermore, the method further comprises:

[0019] Different sample data are normalized based on the feature direction so that the different features of each sample are classified into the same numerical range; the sample data is formed by provincial wind and solar load data.

[0020] Furthermore, based on the provincial wind and solar load data, the generator in the improved generative adversarial network model is used to obtain the generated scenarios, including:

[0021] Based on the long short-term memory network LSTM, the temporal feature information of provincial wind-solar-load data coupling is extracted;

[0022] The temporal feature information is input into the generator of the improved generative adversarial network model, and the generated scene is output.

[0023] Furthermore, based on the long short-term memory network LSTM, the temporal feature information of provincial wind-solar-load data coupling is extracted, including:

[0024] Taking wind power output data, photovoltaic output data and load data as the three features of a sample, the time series feature information of the provincial wind, photovoltaic and load data coupling is extracted based on the long short-term memory network LSTM.

[0025] Furthermore, the long short-term memory network (LSTM) is used to expand the weight matrix in the control gate to a two-dimensional matrix, perform global learning on the input of each time step and the information of the previous time node, and output the time series feature information after fusing multiple features.

[0026] Furthermore, Wasserstein distance is used as the loss function in the discriminator, and the expression of the loss function is:

[0027]

[0028] Where W represents p r (x) and p r (x′); sup represents the least upper bound, K represents the constant of the Lipschitz function; ||D(x)|| L Indicates that the discriminator function satisfies the Lipschitz continuity condition with a constant K; Represents the true data distribution p r (x), D(x|c) represents the score of the discriminator for the real sample, and E x′～p(x′) Denotes the generated data distribution p r (x′), and D(x′|c) represents the score of the discriminator for the generated samples.

[0029] Furthermore, the network objective function expression of the improved generative adversarial network model is:

[0030]

[0031] In the formula, is the gradient penalty term, is the gradient of the discriminator with respect to the input data, Represents the L2 norm squared of the gradient.

[0032] In the second aspect, the present invention further provides a wind-solar-load coupling scene generation system based on LSTM-GAN, the system comprising:

[0033] An acquisition unit is used to acquire provincial wind and solar load data, which includes wind power output data, photovoltaic output data and load data;

[0034] The improved model unit is used to obtain a generation scenario based on provincial wind and solar load data by using a generator in an improved generative adversarial network model; the improved generative adversarial network model is formed by introducing a fully connected layer and a long short-term memory network LSTM into the generator of the generative adversarial network model;

[0035] A training unit, used for inputting Gaussian random noise and real scenes into the discriminator of the improved generative adversarial network model, and cross-training the generator and the discriminator through adversarial training;

[0036] The generation unit is used to extract the trained generator from the improved generative adversarial network model, input Gaussian random noise into the trained generator, and generate random day-wind-solar-load scenes.

[0037] Furthermore, the execution process of the improved model unit is as follows:

[0038] Based on the long short-term memory network LSTM, the temporal feature information of provincial wind-solar-load data coupling is extracted;

[0039] Input the temporal feature information into the generator of the improved generative adversarial network model and output the generated scene;

[0040] Based on the long short-term memory network LSTM, the temporal feature information of provincial wind-solar-load data coupling is extracted, including:

[0041] Taking wind power output data, photovoltaic output data and load data as the three features of a sample, the time series feature information of the provincial wind, photovoltaic and load data coupling is extracted based on the long short-term memory network LSTM.

[0042] In a third aspect, the present invention further provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the above-mentioned LSTM-GAN-based wind-solar-load coupling scenario generation method is implemented.

[0043] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0044] 1. The present invention is based on the LSTM-GAN wind, light, and load coupling scene generation method and system. The present invention takes provincial wind, light, and load data as the target object, and introduces the LSTM model in the generator of the GAN model, which can effectively extract the time series feature information under data coupling, and input this information into the generator of the model to generate random scenes under wind, light, and load coupling. The present invention has a significant improvement in accuracy compared with traditional GAN;

[0045] 2. The wind-solar-load coupling scenario generation method and system based on LSTM-GAN of the present invention provides a new data normalization method, that is, normalizing the sample data according to the direction of the feature, so that the different features of each sample are returned to the same numerical range. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation of the embodiments of the present invention. In the drawings:

[0047] Figure 1 This is a flow chart of the method for generating wind-solar-load coupling scenarios based on LSTM-GAN of the present invention;

[0048] Figure 2 Schematic diagram of the internal structure of the LSTM unit of the present invention;

[0049] Figure 3 This is a structural diagram of the generator of the present invention;

[0050] Figure 4 This is a structural diagram of the discriminator of the present invention;

[0051] Figure 5 The random samples generated by the generator under different training times of the present invention;

[0052] Figure 6 This is a structural block diagram of the wind-solar-load coupling scenario generation system based on LSTM-GAN in the present invention. DETAILED DESCRIPTION

[0053] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0054] (1) In view of the problem that the existing generative adversarial network model GAN ​​only targets a single station, a single element or a small area, and is usually unable to realize scene generation under the source-load common mode, and its practical application scope is very limited, the present invention proposes a GAN scene generation method with provincial wind, light, and load data as the target object. That is, the wind, light, and load of a certain province are treated as three different stations for scene generation. Since the dispatching and planning of the power grid are usually carried out in larger administrative units, this approach has greater practical significance than the scene generation of a single station.

[0055] Different from the data processing method of the traditional GAN ​​algorithm, since the provincial wind and solar load data have large differences in magnitude, unified normalization will cause the features of different magnitudes to influence each other, and the final result generated has large volatility, which is far from the real scene. Therefore, the present invention provides a new data normalization method, that is, normalizing the sample data according to the direction of the feature, so that the different features of each sample are returned to the same numerical range.

[0056] (2) In view of the existing method of introducing convolution to generate random scenes of joint stations, the convolution kernel is difficult to effectively capture the inherent laws of data due to the differences in time and space. The present invention introduces the LSTM model in the generator of the GAN model and proposes a new LSTM-GAN model (i.e., an improved generative adversarial network model) specifically for scene generation under provincial wind-solar-load coupling. The three types of wind, light, and load data are taken as the three features of a sample. Through the LSTM model, the time series feature information under data coupling can be effectively extracted. This information is input into the generator of the model to generate random scenes under wind-solar-load coupling.

[0057] Example 1

[0058] like Figure 1 As shown, the present invention is based on LSTM-GAN wind-solar-load coupling scene generation method, the method comprising:

[0059] Step 1, obtaining provincial wind and solar load data, which includes wind power output data, photovoltaic output data and load data;

[0060] In this embodiment, the wind power output data is the sum of the wind power output data of the whole province, with a time resolution of 1 hour; the photovoltaic output data is the sum of the photovoltaic output data of the whole province, with a time resolution of 1 hour; the load data is the sum of the load output data of the whole province, with a time resolution of 1 hour.

[0061] As a further implementation, the method further comprises:

[0062] Different sample data are normalized based on the feature direction so that the different features of each sample are classified into the same numerical range; the sample data is formed by provincial wind and solar load data.

[0063] Step 2: Based on the provincial wind and solar load data, the generator in the improved generative adversarial network model is used to obtain the generated scene; the improved generative adversarial network model is formed by introducing a fully connected layer and a long short-term memory network LSTM into the generator of the generative adversarial network model;

[0064] Step 2 specifically includes:

[0065] Step 21, based on the long short-term memory network LSTM, extract the time series feature information under the coupling of provincial wind, photovoltaic and load data; specifically: take the wind power output data, photovoltaic output data and load data as three features of a sample, and based on the long short-term memory network LSTM, extract the time series feature information under the coupling of provincial wind, photovoltaic and load data;

[0066] Step 22, input the temporal feature information into the generator of the improved generative adversarial network model, and output the generated scene.

[0067] In this embodiment, the long short-term memory network LSTM can capture dependencies in longer time series, and the core structure includes an input gate, a forget gate, an output gate, and a cell state update. Figure 2 This is the internal structure of the LSTM unit.

[0068] Different from traditional one-dimensional time series data, the long short-term memory network LSTM at this time not only captures the time series dependency of the data, but also learns the dependency between different features at the same time. The weight matrix in the control gate is expanded to a two-dimensional matrix, and the input of each time step is combined with the information of the previous time node for global learning. The formula of the control gate in LSTM is shown below.

[0069] The input gate determines how many memory cells the current input information enters.

[0070] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0071]

[0072] In the formula, i t represents the output of the input gate, represents the cell state at the current time step, ht-1 Represents the hidden layer data of the previous time node, W i ,W C is the corresponding weight matrix, b i ,b C is the bias term, σ is the sigmoid activation function, and tanh is the tanh activation function. The weight matrix for the input value is 3*h, where h represents the size of the hidden layer and 3 represents the number of features.

[0073] The forgetfulness gate determines how much of the past memories are forgotten.

[0074] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0075] In the formula, f t represents the output of the forget gate, W f represents the weight matrix of the forget gate, h t-1 Represents the hidden layer data of the previous time node, x t Indicates the input data at that time, b f represents the bias term, σ represents the activation function. The weight matrix for the input value is 3*h, h represents the size of the hidden layer, and 3 represents the number of features.

[0076] Update memory status

[0077]

[0078] In the formula, C t-1 represents the cell state at the previous moment, f t represents the output of the forget gate, i t represents the output of the input gate, Represents a candidate cell state.

[0079] The output gate determines which memory is used for output. A 64-dimensional vector is output at a single moment as feature information before that moment, and is also used as input for the next moment. The hidden layer vector output at this time contains the temporal feature information after fusing multiple features.

[0080] o t =σ(W o ·[h t-1 , x t ]+b o )

[0081] h t =o t tanh(C t )

[0082] In the formula, o t is the activation value of the output gate, h t is the hidden layer output at time t, C t is the cell state at time t, and σ is the sigmoid activation function.

[0083] The characteristics of LSTM for multi-dimensional data are very suitable for extracting the timing and coupling feature information of wind and solar loads. The present invention introduces a fully connected layer and a long short-term memory network LSTM into the generator of the traditional GAN, and proposes an LSTM-GAN model (i.e., an improved generative adversarial network model) for wind and solar load scene generation. The generator of this model is as follows Figure 3 As shown, the discriminator is Figure 4 As shown, the detailed parameters of the model are shown in Table 1. After the random noise is input into the generator, a one-dimensional tensor of 72 dimensions is output through a fully connected layer. The data is then reshaped into a 24*3 two-dimensional tensor, with columns representing different data features, corresponding to the three types of data of wind, light, and load in the present invention, and rows representing different time points. The tensor is input into the LSTM, and a 24*64 two-dimensional tensor is output, with rows representing different time points, and columns representing the time feature information under the wind-light-load coupling at each time point. The data is then reshaped into a one-dimensional tensor, connected to two fully connected layers, and the output result is reshaped into a three-dimensional tensor of 128*3*6, and then connected to two deconvolution layers to output a generated result of 1*3*24. It is worth noting that in order to generate the authenticity of the data distribution, the last deconvolution layer is not connected to BN and RELU. The discriminator adopts the traditional GAN ​​network model, inputs sample data of shape 1*3*24, and outputs the score of the sample through two convolutional layers and two fully connected layers. The detailed parameters of the generator and discriminator are shown in Table 1. Table 1 shows the specific parameters of the network. The generator outputs 1*3*24 data, where 1 means the number of channels is 1, 3 corresponds to wind and solar load, and 24 represents the time node in a day, representing the generation of a scene under the coupling of wind and solar load for 24 hours a day. The discriminator outputs a value, which indicates the score of the sample.

[0084] Table 1 Detailed parameters of the LSTM-GAN model of the present invention

[0085]

[0086] Among them, Generator: generator, Discriminator: discriminator, Input: input data, Layer: neural network layer of the model, FC: fully connected network, Conv: convolutional network, BN: normalization layer, RELU: RELU activation layer, LSTM: long short-term memory network, Conv_transpose: deconvolution network, and the numbers are the data sizes of input or output.

[0087] The optimization of traditional GAN ​​introduces Jensen-Shannon divergence and relative entropy divergence to identify the similarity between the generated data distribution and the real data distribution. However, during the training process, the probability of the generated data and the real data perfectly overlapping in the high-dimensional space is very low, and it is easy to have mode collapse and loss non-convergence. GAN introduces Wasserstein distance to characterize the similarity between two distributions. Using this as LOSS can solve the mode collapse caused by JS divergence in the traditional GAN ​​training process. The formula is as follows:

[0088]

[0089] In the formula, γ represents a joint distribution representing the optimal path; W represents p r (x) and p r The distance between (x′);

[0090] π(p r (x),p r (x′)) represents the joint distribution between the two; ||xx′|| represents the distance between the real sample and the generated sample. It is not easy to be solved directly. By limiting the absolute value range of the discriminator network parameters, the discriminator loss function satisfies Lipschitz continuity, and the above formula becomes:

[0091]

[0092] Where W represents p r (x) and p r (x′); sup represents the least upper bound, K represents the constant of the Lipschitz function; ||D(x)|| L Indicates that the discriminator function satisfies the Lipschitz continuity condition with a constant K; Represents the true data distribution p r (x), D(x|c) represents the score of the discriminator for the real sample, and E x′～p(x′) Denotes the generated data distribution p r (x′), and D(x′|c) represents the score of the discriminator for the generated samples.

[0093] The network objective function expression of the final improved generative adversarial network model is:

[0094]

[0095] In the formula, is the gradient penalty term, is the gradient of the discriminator with respect to the input data, Represents the L2 norm squared of the gradient.

[0096] Figure 3 and Figure 4 They represent the model structures of the generator and discriminator respectively, black represents input data, yellow represents convolutional layers, fully connected layers, and deconvolutional layers, green represents normalization, pink represents activation functions, gray represents reorganized shapes, blue represents intermediate output data, red represents generated samples, and blue represents scoring. The brackets represent the size of the tensor channel, row, and column, respectively, and 0 means that the dimension does not exist.

[0097] Figure 3 and Figure 4 In the present invention, the different structures in the model are described as follows:

[0098] (1) Convolutional layers (Conv2, De_Conv2)

[0099] The convolution operation is used to extract local features in neural networks, and its mathematical expression is:

[0100]

[0101] Where x[i,j] is the value of the input feature map at position (i,j); w[m,n] is the convolution kernel k h ×k w The weight value of; b is the bias; k h ,k w are the height and width of the convolution kernel respectively; y[i,j]: the value of the output feature map at position (i,j); c is the number of channels

[0102] (2) Normalization layer (BN)

[0103] BN is used to speed up training and stabilize the network by normalizing and scaling input features. Its formula is:

[0104] Compute the mean and variance:

[0105] Normalization: Here, ∈ is a small number used to prevent the denominator from being zero.

[0106] Zoom and Pan: γ is a scaling parameter (learnable); β is a translation parameter (learnable).

[0107] (3) Leaky ReLU activation function

[0108] Leaky ReLU is a variant of ReLU, which is used to alleviate the "dead neuron" problem of standard ReLU. The formula is:

[0109]

[0110] Where x is the input value; α is the negative slope (usually set to a small value, such as 0.01).

[0111] (4) Fully connected layer (FC)

[0112] The fully connected layer is one of the most basic layers in a neural network, where every input neuron is fully connected to the output neuron.

[0113]

[0114] In the formula, x j is the value of the input neuron; w ij is the weight connecting input j to output i; b i is the bias of output neuron i; y i is the value of output neuron i; n is the number of input neurons.

[0115] (5) ReLU activation function (ReLU)

[0116] ReLU is the most commonly used activation function in neural networks. It performs a nonlinear transformation on the output value of each layer to increase the depth of the entire network. The formula is:

[0117] f(x)=max(0,x)

[0118] Where x is the input value and f(x) is the output value.

[0119] Step 3: Input Gaussian random noise and real scenes into the discriminator of the improved generative adversarial network model, and cross-train the generator and discriminator through adversarial training;

[0120] Among them, the real scene is the provincial wind and solar load data.

[0121] Step 4: extract the trained generator from the improved generative adversarial network model, input Gaussian random noise into the trained generator, and generate random day-wind-solar-load scenes.

[0122] In step 3, the random samples generated by the generator under different training times are as follows Figure 5 As shown, Figure 5The upper left in the middle is a random scene generated by the generator when training 0 times, the upper right is a random scene generated by the generator when training 100 times, the lower left is a random scene generated by the model when training 500 times, and the lower right is a random scene generated by the model when training 2000 times. It can be clearly seen that this method can generate random scenes close to the actual situation for provincial wind and solar load data.

[0123] It has been verified that for scene generation under wind-solar-load coupling, LSTM-GAN has greatly improved the accuracy compared with traditional GAN ​​and can describe real scenes more accurately.

[0124] Example 2

[0125] like Figure 6 As shown, the difference between this embodiment and embodiment 1 is that this embodiment provides a wind-solar-load coupling scene generation system based on LSTM-GAN, and the system includes:

[0126] An acquisition unit is used to acquire provincial wind and solar load data, which includes wind power output data, photovoltaic output data and load data;

[0127] The improved model unit is used to obtain a generation scenario based on provincial wind and solar load data by using a generator in an improved generative adversarial network model; the improved generative adversarial network model is formed by introducing a fully connected layer and a long short-term memory network LSTM into the generator of the generative adversarial network model;

[0128] A training unit, used for inputting Gaussian random noise and real scenes into the discriminator of the improved generative adversarial network model, and cross-training the generator and the discriminator through adversarial training;

[0129] The generation unit is used to extract the trained generator from the improved generative adversarial network model, input Gaussian random noise into the trained generator, and generate random day-wind-solar-load scenes.

[0130] As a further implementation, the execution process of the improved model unit is:

[0131] Based on the long short-term memory network LSTM, the temporal feature information of provincial wind-solar-load data coupling is extracted;

[0132] Input the temporal feature information into the generator of the improved generative adversarial network model and output the generated scene;

[0133] Based on the long short-term memory network LSTM, the temporal feature information of provincial wind-solar-load data coupling is extracted, including:

[0134] Taking wind power output data, photovoltaic output data and load data as the three features of a sample, the time series feature information of the provincial wind, photovoltaic and load data coupling is extracted based on the long short-term memory network LSTM.

[0135] Among them, the execution process of each unit can be executed according to the process steps of the wind-solar-load coupling scenario generation method based on LSTM-GAN in Example 1, and will not be repeated one by one in this embodiment.

[0136] At the same time, the present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the above-mentioned LSTM-GAN-based wind-solar-load coupling scenario generation method is implemented.

[0137] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0138] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0139] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0140] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0141] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. The wind-solar-load coupling scenario generation method based on LSTM-GAN is characterized by: The method includes: Obtain provincial wind and solar load data, wherein the provincial wind and solar load data includes wind power output data, photovoltaic output data and load data; According to the provincial wind and solar load data, an improved generative adversarial network model is constructed; the improved generative adversarial network model is formed by introducing a fully connected layer and a long short-term memory network LSTM into the generator of the generative adversarial network model; Inputting Gaussian random noise and real scenes into the discriminator of the improved generative adversarial network model, and cross-training the generator and the discriminator through adversarial training; A trained generator is extracted from the improved generative adversarial network model, and Gaussian random noise is input into the trained generator to generate a random day-wind-solar-load scene.

2. The method for generating wind-solar-load coupling scenarios based on LSTM-GAN according to claim 1 is characterized in that: The method further includes: Different sample data are normalized based on the feature direction so that different features of each sample are returned to the same numerical range; the sample data is a sample formed by provincial wind and solar load data.

3. The method for generating wind-solar-load coupling scenarios based on LSTM-GAN according to claim 1 is characterized in that: According to the provincial wind and solar load data, the generator in the improved generative adversarial network model is used to obtain the generated scenario, including: Based on the long short-term memory network LSTM, the temporal characteristic information under the coupling of the provincial wind-solar-load data is extracted; The temporal feature information is input into the generator of the improved generative adversarial network model, and a generated scene is output.

4. The method for generating wind-solar-load coupling scenarios based on LSTM-GAN according to claim 3 is characterized in that: Based on the long short-term memory network LSTM, the temporal feature information of the provincial wind-solar-load data coupling is extracted, including: The wind power output data, photovoltaic output data and load data are taken as three features of a sample, and based on the long short-term memory network LSTM, the temporal characteristic information under the coupling of the provincial wind, photovoltaic and load data is extracted.

5. The method for generating wind-solar-load coupling scenarios based on LSTM-GAN according to claim 4 is characterized in that: The long short-term memory network LSTM is used to expand the weight matrix in the control gate to a two-dimensional matrix, perform global learning on the input of each time step and the information of the previous time node, and output the time series feature information after fusing multiple features.

6. The method for generating wind-solar-load coupling scenarios based on LSTM-GAN according to claim 1 is characterized in that: The Wasserstein distance is used as the loss function in the discriminator, and the expression of the loss function is: Where W represents p r (x) and p r (x′); sup represents the least upper bound, K represents the constant of the Lipschitz function; ||D(x)||| L Indicates that the discriminator function satisfies the Lipschitz continuity condition with a constant K; Represents the true data distribution p r (x), D(x|c) represents the score of the discriminator for the real sample, Denotes the generated data distribution p r (x′), and D(x′|c) represents the score of the discriminator for the generated samples.

7. The method for generating wind-solar-load coupling scenarios based on LSTM-GAN according to claim 1 is characterized in that: The network objective function expression of the improved generative adversarial network model is: In the formula, is the gradient penalty term, is the gradient of the discriminator with respect to the input data, Represents the L2 norm squared of the gradient.

8. The wind-solar-load coupling scenario generation system based on LSTM-GAN is characterized by: The system includes: An acquisition unit, used to acquire provincial wind and solar load data, wherein the provincial wind and solar load data includes wind power output data, photovoltaic output data and load data; An improved model unit is used to obtain a generation scenario according to the provincial wind and solar load data by using a generator in an improved generative adversarial network model; the improved generative adversarial network model is formed by introducing a fully connected layer and a long short-term memory network LSTM into the generator of the generative adversarial network model; A training unit, used for inputting Gaussian random noise and real scenes into the discriminator of the improved generative adversarial network model, and cross-training the generator and the discriminator through adversarial training; A generation unit is used to extract a trained generator from the improved generative adversarial network model, input Gaussian random noise into the trained generator, and generate a random day-wind-solar-load scene.

9. The LSTM-GAN-based wind-solar-load coupling scenario generation system according to claim 8, characterized in that: The execution process of the improved model unit is: Based on the long short-term memory network LSTM, the temporal characteristic information under the coupling of the provincial wind-solar-load data is extracted; Inputting the temporal feature information into a generator of an improved generative adversarial network model, and outputting a generated scene; The long short-term memory network LSTM is used to extract the temporal feature information of the provincial wind-solar-load data coupling, including: The wind power output data, photovoltaic output data and load data are taken as three features of a sample, and based on the long short-term memory network LSTM, the temporal characteristic information under the coupling of the provincial wind, photovoltaic and load data is extracted.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the wind-solar-load coupling scenario generation method based on LSTM-GAN is implemented.

Images