A scene generation method and system based on deep learning source load medium and long term prediction

Abstract

The application provides a scene generation method and system based on deep learning source load medium and long term prediction, which comprises the following steps: predicting future load electricity consumption through a load prediction model based on a DenseNet-Autoformer network; expanding wind and light combined scene samples according to a wind and light combined scene expansion model based on WGAN-GP; and obtaining typical wind and light combined power scenes by reducing the generated wind and light combined scene samples based on Gaussian mixture clustering. The DenseNet-Autoformer network provided by the application performs deep decomposition on a load time sequence, obtains high-dimensional time sequence features of periodic and trend components of the load time sequence, establishes a load medium and long term prediction model which comprehensively considers meteorological and economic factors, and establishes a gradual time sequence decomposition architecture. In the process of predicting future load electricity consumption, the model alternately performs prediction result optimization and sequence decomposition, combines a self-correlation mechanism, realizes sequence level linking of the load time sequence, and reduces the time complexity of the model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of source-load prediction and scenario generation, in particular to a scenario generation method and system based on deep learning source-load medium and long-term prediction. BACKGROUND

[0002] With a large number of distributed power sources connected to rural power distribution systems, and the transformation of the national economy, the economic structure of rural areas has changed greatly, and the traditional single source / load scenario modeling has been difficult to meet the needs of power system operation and dispatching, planning and design for handling uncertainty. How to accurately extract source-load uncertainty and reflect the long-term dynamic growth characteristics and short-term power fluctuation characteristics of source-load is the core challenge of source-load scenario modeling.

[0003] In the aspect of wind-solar scenario modeling, although the traditional scenario generation method can well describe the uncertainty of new energy output, it is difficult to capture the nonlinear characteristics of high-dimensional data, and this method usually relies on statistical assumptions, which has certain limitations.

[0004] As for the load prediction method, the existing artificial intelligence prediction model independently mines feature correlation information from massive data to establish a complex nonlinear model for prediction. However, the traditional recurrent neural network and its variants have low computational efficiency when processing long sequence data and are difficult to learn information related to long-term dependence. SUMMARY

[0005] Therefore, one object of the present application is to provide a scenario generation method and system based on deep learning source-load medium and long-term prediction to solve the problems mentioned in the background art and overcome the deficiencies in the prior art.

[0006] To achieve the above object, the present application adopts the following technical solutions:

[0007] In a first aspect, the present application provides a scenario generation method based on deep learning source-load medium and long-term prediction, comprising:

[0008] predicting future load consumption through a load prediction model based on a DenseNet-Autoformer network;

[0009] expanding wind-solar joint scenario samples, reducing the generated wind-solar joint scenario samples to obtain typical wind-solar joint power scenarios; integrating the predicted future load consumption and the typical wind-solar joint power scenarios to obtain source-load scenarios.

[0010] Preferably, the future load electricity consumption is predicted by the load prediction model based on the DenseNet-Autoformer network, including: inputting historical monthly meteorological, economic sequences and historical load monthly electricity consumption sequences, predicting future load electricity consumption by the load prediction model based on the DenseNet-Autoformer network, and outputting future monthly load electricity consumption prediction results.

[0011] The load prediction model based on the DenseNet-Autoformer network comprises N encoder layers and M decoder layers; each encoder layer comprises a self-correlation module, two time series decomposition modules and a deep feature extraction module; and each decoder layer comprises a self-correlation module, an encoder-decoder correlation module, three time series decomposition modules and a deep feature extraction module.

[0012] The deep feature extraction module adds a DenseNet layer after a feedforward layer, the DenseNet network structure comprises four dense connection blocks and three transfer layers, the transfer layer is a Conv(1×1)+global average pooling(2×2) layer, the feature mapping sizes transmitted between layers in the dense connection block are consistent, the non-linear transformation function in the dense connection block adopts a BN+ReLU+Conv(3×3) operation structure, and the layers in the dense connection block all obtain k feature mapping channel numbers after convolution operation, that is, the growth rate of the DenseNet network is k, and the input channel number of the lth layer is k0+k(l-1), wherein k0 is the channel number of the input layer feature mapping.

[0013] Preferably, the wind and light combined scene samples are expanded, including: expanding the wind and light combined scene samples according to the wind and light combined scene expansion model based on the WGAN-GP.

[0014] Preferably, the time series decomposition module is embedded in the encoder-decoder, and the model is alternately used for prediction result optimization and sequence decomposition to gradually separate the trend item and the periodic item from the hidden variable, so that gradual decomposition is realized, and the moving average method is used to smooth the violent fluctuations, and for an input sequence X with a length of L, the decomposition process is:

[0015] X t =AvgPool(Padding(X))

[0016] X s =X-X t

[0017] X t , X s respectively represent the trend component and the periodic component, AvgPool represents the average filtering convolution operation, and Padding represents the padding operation.

[0018] As preferred, the autocorrelation module defines the autocorrelation coefficient R t (τ) for a discrete time series {X xx (τ),

[0019]

[0020] where the autocorrelation coefficient R xx (τ) represents the correlation between the series {X t} and its τ-delayed version {X t-τ};

[0021] The first k time-delayed series with the strongest correlation are selected The time-delayed sub-series and the current time series are aggregated to establish a periodic dependence relationship, and the autocorrelation coefficient R xx (τ) is calculated by fast Fourier transform, and the calculation process is as follows:

[0022]

[0023]

[0024] where i represents an imaginary unit, t represents time in the time domain, f represents frequency in the frequency domain, τ represents time delay, F and F -1 represent fast Fourier transform and its inverse transform respectively, represents the conjugate part of F in frequency domain calculation.

[0025] As preferred, the scene reduction of the generated wind-solar combined scene samples to obtain a typical wind-solar combined power scene, and the integration of the predicted future load power consumption and the typical wind-solar combined power scene to obtain a source-load scene include: performing scene reduction on the generated wind-solar combined scene samples and load power samples based on Gaussian mixture clustering to obtain a typical source-load power scene.

[0026] As preferred, the wind-solar combined scene sample expansion based on the WGAN-GP-based wind-solar combined scene expansion model includes: taking historical wind power daily power scene and historical photovoltaic daily power scene as input, expanding wind-solar combined scene samples based on the WGAN-GP-based wind-solar combined scene expansion model, and outputting a monthly wind-solar combined power expansion scene set.

[0027] As preferred, the wind-solar combined scene sample expansion based on the WGAN-GP-based wind-solar combined scene expansion model further includes: dividing wind-solar output into monthly subsets, combining wind power and photovoltaic power according to time stamp to form wind-solar combined power for training, and expanding wind-solar output samples using GAN to generate an objective function for the min-max game in the adversarial network training process:

[0028]

[0029] wherein G(·) represents a generator, D(·) represents a discriminator, E represents expectation, x represents historical wind and light output data, p(x) represents a probability distribution of x, z is random noise, and p(z) represents a probability distribution of z;

[0030] The Wasserstein distance is used to replace the JS divergence as an optimization objective, and the definition of the Wasserstein distance is as follows:

[0031]

[0032] wherein P r represents a real data distribution, P g represents a generated data distribution, Π(P r ,P g ) is a set of all possible joint distributions combined by P r and P g , γ represents one of the joint distributions, (x, y) ~ γ represents that a sample x and y are sampled from the joint distribution γ, ‖x-y‖ is the distance between the pair of samples, and E (x,y)～γ (‖x-y‖) represents an expected value of the sample distance.

[0033] The loss function of the WGAN-GP model is as follows:

[0034]

[0035] wherein ‖·‖ p represents a p-norm; represents a gradient operator, λ is a penalty term coefficient, ζ is a uniform distribution in the range of [0, 1], represents uniform sampling of a straight line between the sampling points from the real data distribution P r (·) and the generated data distribution P g (·).

[0036] As preferred, the generated wind and light joint scene samples and load power samples are subjected to scene reduction by the Gaussian mixture clustering to obtain typical source and load power scenes, including: taking the predicted future monthly load power as input, outputting the typical load power scene and scene probability of the corresponding month in the future through GMM clustering, and taking the monthly wind and light joint power expansion scene set as input, outputting the typical wind and light joint power scene and scene probability of the corresponding month through GMM clustering.

[0037] In a second aspect, the present application also provides a scene generation system based on deep learning source-load medium and long term prediction, comprising a load prediction module, a scene expansion module and a hybrid clustering module,

[0038] The load prediction module is configured to predict future load electricity consumption through a load prediction model based on a DenseNet-Autoformer network.

[0039] The scene expansion module is configured to expand wind-solar joint scene samples.

[0040] The hybrid clustering module is configured to reduce the generated wind-solar joint scene samples according to the predicted future load electricity consumption to obtain typical wind-solar joint power scenes, integrate the predicted future load electricity consumption and the typical wind-solar joint power scenes, and obtain source-load scenes.

[0041] As a preferred, the prediction of future load electricity consumption through the load prediction model based on the DenseNet-Autoformer network comprises: taking historical monthly meteorological, economic sequences and historical load monthly electricity consumption sequences as inputs, predicting future load electricity consumption through the load prediction model based on the DenseNet-Autoformer network, and outputting future monthly load electricity consumption prediction results.

[0042] The load prediction model based on the DenseNet-Autoformer network comprises N encoder layers and M decoder layers; each encoder layer comprises a self-correlation module, two time series decomposition modules and a deep feature extraction module; each decoder layer comprises a self-correlation module, an encoder-decoder correlation module, three time series decomposition modules and a deep feature extraction module.

[0043] The deep feature extraction module adds a DenseNet layer after a feedforward layer, the DenseNet network structure comprises four dense connection blocks and three transfer layers, the transfer layer is a Conv(1x1)+global average pooling(2x2) layer, the feature mapping sizes transmitted between layers in the dense connection block are consistent, the non-linear transformation function in the dense connection block adopts a operation structure of BN+ReLU+Conv(3x3), and after convolution operations are performed on layers in the dense connection block, k feature mapping channels are obtained, that is, the growth rate of the DenseNet network is k, and the input channel number of the lth layer is k0+k(l-1), wherein k0 is the channel number of the input layer feature mapping.

[0044] As a preferred, the expansion of wind-solar joint scene samples comprises: expanding wind-solar joint scene samples according to a wind-solar joint scene expansion model based on WGAN-GP.

[0045] As preferred, the timing decomposition module is embedded inside the encoder-decoder, and the prediction result optimization and sequence decomposition are alternately performed by the model to gradually separate the trend component and the periodic component from the latent variable, to realize progressive decomposition, and the moving average method is used to smooth the violent fluctuations, and for an input sequence X of length L, the decomposition process is:

[0046] X t = AvgPool(Padding(X))

[0047] X s = X-X t

[0048] Wherein, X t , X s represent the trend component and the periodic component respectively, AvgPool represents the average filtering convolution operation, and Padding represents the padding operation.

[0049] As preferred, the autocorrelation module defines the autocorrelation coefficient R t (τ) for the discrete time sequence {X xx}.

[0050]

[0051] Wherein, the autocorrelation coefficient R xx (τ) represents the correlation between the sequence {X t} and its τ delay {X t-τ}.

[0052] The first k time delay sequences with the strongest correlation are selected The time delay information aggregation method is used to establish the periodic dependence relationship between the time delay subsequence and the current time sequence, and the autocorrelation coefficient R xx (τ) is calculated by fast Fourier transform, and the calculation process is as follows:

[0053]

[0054] Wherein, i represents the imaginary unit, t represents the time in the time domain, f represents the frequency in the frequency domain, τ represents the time delay, F and F -1 represent the fast Fourier transform and its inverse transform respectively, represents the conjugate part of F in the frequency domain calculation.

[0055] As preferred, the generated wind and light combined scene samples are scene reduced to obtain typical wind and light combined power scenes, and the predicted future load power and the typical wind and light combined power scenes are integrated to obtain the source and load scene, including: based on Gaussian mixed clustering, the generated wind and light combined scene samples and the load power samples are scene reduced to obtain typical source and load power scenes.

[0056] As preferred, the wind and light combined scene sample is expanded according to the wind and light combined scene expansion model based on WGAN-GP, including: taking the historical wind power daily scene and the historical photovoltaic daily power scene as input, expanding the wind and light combined scene sample according to the wind and light combined scene expansion model based on WGAN-GP, and outputting a monthly wind and light combined power expansion scene set.

[0057] As preferred, the wind and light combined scene sample is expanded according to the wind and light combined scene expansion model based on WGAN-GP, including: taking the historical wind power daily scene and the historical photovoltaic daily power scene as input, expanding the wind and light combined scene sample according to the wind and light combined scene expansion model based on WGAN-GP, and outputting a monthly wind and light combined power expansion scene set.

[0058]

[0059] Wherein, G(·) represents the generator, D(·) represents the discriminator, E represents the expectation, x represents the historical wind and light output data, p(x) represents the probability distribution of x, z is random noise, and p(z) represents the probability distribution of z.

[0060] The Wasserstein distance is used instead of the JS divergence as the optimization target, and the definition of the Wasserstein distance is as follows:

[0061]

[0062] Wherein, P r represents the real data distribution, P g represents the generated data distribution, Π(P r ,P g ) is the set of all possible joint distributions combined by P r and P g , γ represents one of the joint distributions, (x,y) ~ γ represents that a sample x and y are obtained by sampling from the joint distribution γ, ‖x-y‖ is the distance between the pair of samples, and E (x,y)～γ (‖x-y‖) represents the expected value of the sample distance.

[0063] The loss function of the WGAN-GP model is as follows:

[0064]

[0065] Wherein, ‖·‖ p represents the p-norm; represents the gradient operator, λ is a penalty term coefficient, τ is a uniform distribution in the range of [0, 1], represents that a sample is obtained from the real data distribution Pr (·) and generate data distribution P g (·) straight line uniform sampling between sampling points.

[0066] As preferred, the Gaussian mixture clustering-based scene reduction on the generated wind-solar joint scene samples and load power samples obtains typical source-load power scenes, including: taking the predicted future monthly load power as input, outputting the typical load power scene and scene probability of the corresponding month in the future through GMM clustering, and taking the monthly wind-solar joint power expansion scene set as input, outputting the typical wind-solar joint power scene and scene probability of the corresponding month through GMM clustering.

[0067] In a third aspect, the present application also provides an electronic device, including a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the scene generation method based on deep learning source-load medium and long-term prediction as described above when executing the program.

[0068] In a fourth aspect, the present application also provides a readable storage medium having a computer program stored thereon, and the computer program implements the steps of the scene generation method based on deep learning source-load medium and long-term prediction as described above when executed by a processor.

[0069] Therefore, the present application has the following beneficial effects:

[0070] The present application performs deep decomposition on the load time series through the progressive time series decomposition architecture and the embedded time series decomposition module, the autocorrelation module and the deep feature extraction module, mines the deep periodicity, trend characteristics of the load monthly electricity consumption sequence and the correlation between external influencing factors and the load monthly electricity consumption, establishes a load medium and long-term prediction model considering meteorological and economic factors, expands the wind-solar joint scene samples in combination with the wind-solar joint scene expansion model, and obtains typical wind-solar joint power scenes through scene reduction on the generated wind-solar joint scene samples.

[0071] During the prediction of future load electricity consumption by the load prediction model based on the DenseNet-Autoformer network, the model alternately performs prediction result optimization and sequence decomposition, realizes sequence-level linking of the load time series in combination with the autocorrelation mechanism, reduces the time complexity of the model, and improves the calculation efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0072] The above and / or additional aspects and advantages of the present application will become apparent and more readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings, in which:

[0073] Figure 1 is the overall flowchart of the method of the present application;

[0074] Figure 2 Fig. 3 is a schematic diagram of a source-load medium and long-term prediction and scenario generation method framework according to an embodiment of the present application;

[0075] Figure 3 Fig. 4 is a schematic diagram of a load medium and long-term prediction framework based on a DenseNet-Autoformer network according to the present application;

[0076] Figure 4 Fig. 5 is a schematic diagram of an autocorrelation mechanism structure according to an embodiment of the present application;

[0077] Figure 5 Fig. 6 is a schematic diagram of DenseNet forward propagation according to an embodiment of the present application;

[0078] Figure 6 Fig. 7 is a schematic diagram of a dense connection block structure according to an embodiment of the present application;

[0079] Figure 7 Fig. 8 is a schematic diagram of prediction results of each load prediction model according to an embodiment of the present application;

[0080] Figure 8 Fig. 9 is a wind power output probability distribution result of each wind and light scenario generation model according to an embodiment of the present application;

[0081] Figure 9 Fig. 10 is a photovoltaic output probability distribution result of each wind and light scenario generation model according to an embodiment of the present application;

[0082] Figure 10 Fig. 11 is a wind power autocorrelation coefficient result of each wind and light scenario generation model according to an embodiment of the present application;

[0083] Figure 11 Fig. 12 is a photovoltaic autocorrelation coefficient result of each wind and light scenario generation model according to an embodiment of the present application;

[0084] Figure 12 Fig. 13 is a schematic diagram of a system structure connection relationship according to the present application;

[0085] Figure 13 Fig. 14 is a schematic diagram of a computer device structure according to an embodiment of the present application. DETAILED DESCRIPTION

[0086] Embodiments of the present application are described in detail below with reference to the accompanying drawings, examples of which are shown in the drawings, wherein the same or similar notations represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to explain the present application, and cannot be understood as limiting the present application.

[0087] With the low-carbon transformation of energy structure, a large number of distributed power is connected to the rural power distribution system. The uncertainty and volatility of renewable energy output bring challenges to the safe and reliable operation of the system. At the same time, with the transformation of the national economy, the economic structure of the countryside has changed a lot. Relying on historical data of power load for load power prediction will cause a large error. Therefore, the traditional single source / load scenario modeling has been difficult to meet the needs of handling uncertainty in power system operation and dispatching, planning and design. Accurate and reliable source / load scenario description is of great significance to ensure the safe and stable operation of the power grid. How to accurately extract the uncertainty of source and load and reflect the long-term dynamic growth characteristics and short-term power fluctuation characteristics of source and load is the core challenge of source and load scenario modeling.

[0088] At present, the existing prediction model is mainly divided into parameter model and artificial intelligence model. The parameter model obtains the prediction result by establishing a mathematical parameter model between the load and its related factors, but the adaptability and accuracy of such model are insufficient for medium and long-term prediction affected by complex factors. The artificial intelligence model independently mines feature correlation information from massive data to establish a complex nonlinear model for prediction. However, the traditional recurrent neural network and its variants have low computational efficiency and are difficult to learn information related to long-term dependence when processing long sequence data. In the aspect of wind and light scene modeling, although the traditional scene generation method can well describe the uncertainty of new energy output, it is difficult to capture the nonlinear characteristics of high-dimensional data, and such method usually depends on statistical assumptions and has certain limitations. Based on this, the present application provides a scene generation method and system based on deep learning source and load medium and long-term prediction.

[0089] As shown in Figure 1 A scene generation method based on deep learning source and load medium and long-term prediction, comprising:

[0090] predicting future load power consumption through a load prediction model based on a DenseNet-Autoformer network;

[0091] expanding wind and light joint scene samples;

[0092] obtaining typical wind and light joint power scenarios by reducing the generated wind and light joint scene samples, integrating the predicted future load power consumption and the typical wind and light joint power scenarios, and obtaining source and load scenarios.

[0093] More specifically, expanding wind and light joint scene samples includes expanding wind and light joint scene samples based on a wind and light joint scene expansion model based on WGAN-GP; obtaining typical wind and light joint power scenarios by reducing the generated wind and light joint scene samples includes obtaining typical wind and light joint power scenarios by reducing the generated wind and light joint scene samples based on Gaussian mixture clustering.

[0094] The present application fully extracts source-load uncertainty characteristics, and realizes accurate and reliable characterization of source-load scenarios. First, a load medium and long-term prediction model based on a DenseNet-Autoformer network is constructed to extract periodicity and trend characteristics of the load, and the load medium and long-term prediction result is obtained by comprehensively considering meteorological and economic factors. On this basis, the typical load scenario under the corresponding month in the future is obtained by combining the load per-unit curve of the corresponding month in the history. Second, an improved adversarial generation network is used to expand wind and light scene samples. Finally, Gaussian mixture clustering is used to reduce the generated wind and light combined output samples to obtain typical wind and light scenarios.

[0095] As shown in Figure 2 , the present application constructs a source-load medium and long-term prediction and scenario generation method based on deep learning, and the overall architecture is as shown in Figure 2 , which mainly includes three parts: a load medium and long-term prediction model based on a DenseNet-Autoformer network, a wind and light combined scene expansion model based on WGAN-GP (Wasserstein GAN with gradient penalty), and a scenario reduction method based on Gaussian Mixture Model (GMM).

[0096] As a preferred embodiment, the future load electricity consumption is predicted by the load prediction model based on the DenseNet-Autoformer network, which includes: taking the historical monthly meteorological, economic sequences and historical load monthly electricity consumption sequences as inputs, predicting the future load electricity consumption by the load prediction model based on the DenseNet-Autoformer network, and outputting the prediction result of the future monthly load electricity consumption.

[0097] More specifically, the load sequence changes present trend and periodic characteristics due to the influence of economic and meteorological factors. The present application constructs a load prediction model based on the DenseNet-Autoformer network, which mines the deep periodicity and trend characteristics of the load monthly electricity consumption sequence and the correlation between external influencing factors and the load monthly electricity consumption through an embedded time series decomposition module, an autocorrelation module and a deep feature extraction module. The input is the historical monthly meteorological, economic sequences and historical load monthly electricity consumption sequences, and the output is the prediction of future monthly load electricity consumption.

[0098] The load prediction model based on the DenseNet-Autoformer network includes N encoder layers and M decoder layers; each encoder layer includes an autocorrelation module, two time series decomposition modules and a deep feature extraction module; each decoder layer includes an autocorrelation module, an encoder-decoder correlation module, three time series decomposition modules and a deep feature extraction module.

[0099] The deep feature extraction module adds a DenseNet layer after the feedforward layer. The DenseNet network structure includes 4 dense connection blocks and 3 transfer layers. The transfer layer is a Conv(1x1)+global average pooling(2x2) layer. The feature map size transmitted between layers in the dense connection block is consistent. The non-linear transformation function in the dense connection block adopts the operation structure of BN+ReLU+Conv(3x3). After the convolution operation of each layer in the dense connection block, k feature map channels are obtained, that is, the growth rate of the DenseNet network is k. The input channel number of the lth layer is k0+k(l-1), wherein k0 is the channel number of the input layer feature map.

[0100] As preferred, the time series decomposition module is embedded in the encoder-decoder, and the prediction result optimization and sequence decomposition are alternately performed by the model to gradually separate the trend item and the periodic item from the hidden variable to realize progressive decomposition. The moving average method is used to smooth the violent fluctuations. For an input sequence X of length L, the decomposition process is:

[0101] X t =AvgPool(Padding(X))

[0102] X s =X-X t

[0103] Wherein, X t , X s represent the trend component and the periodic component respectively, AvgPool represents the average filtering convolution operation, and Padding represents the padding operation.

[0104] As preferred, the autocorrelation module defines the autocorrelation coefficient R t (τ) of the discrete time sequence {X xx} as follows:

[0105]

[0106] Wherein, the autocorrelation coefficient R xx (τ) represents the correlation between the sequence {X t} and its τ delay {X t-τ}.

[0107] The first k time delay sequences with the strongest correlation are selected The time delay information aggregation method is used to establish the periodic dependence relationship between the time delay subsequence and the current time sequence. The autocorrelation coefficient R xx (τ) is calculated by fast Fourier transform, and the calculation process is as follows:

[0108]

[0109] where i denotes the imaginary unit, t denotes time in the time domain, f denotes frequency in the frequency domain, τ denotes time delay, F and F -1 denote the fast Fourier transform and its inverse, respectively, denotes the conjugate part of F in the frequency domain calculation.

[0110] As preferred, the wind-solar joint scene sample is expanded according to the wind-solar joint scene expansion model based on WGAN-GP, including: taking the historical wind power daily scene and the historical photovoltaic daily power scene as input, expanding the wind-solar joint scene sample according to the wind-solar joint scene expansion model based on WGAN-GP, and outputting a monthly wind-solar joint power expansion scene set.

[0111] As preferred, expanding the wind-solar joint scene sample according to the wind-solar joint scene expansion model based on WGAN-GP further includes: dividing the wind-solar output into subsets by month, and training the wind power and the photovoltaic power according to the timestamp to form the wind-solar joint power, using GAN to expand the wind-solar output sample, and generating the objective function of the adversarial network training process min-max game as:

[0112]

[0113] where G(·) denotes the generator, D(·) denotes the discriminator, E denotes the expectation, x denotes the historical wind-solar output data, p(x) denotes the probability distribution of x, z is random noise, and p(z) denotes the probability distribution of z;

[0114] The Wasserstein distance is used instead of the JS divergence as the optimization objective, and the definition of the Wasserstein distance is:

[0115]

[0116] where P r denotes the real data distribution, P g denotes the generated data distribution, Π(P r ,P g ) is a set of all possible joint distributions combined by P r and P g , γ represents one of the joint distributions, (x,y) ~ γ represents sampling from the joint distribution γ to obtain a sample x and y, and ‖x-y‖ is the distance between the pair of samples, E (x,y)～γ (‖x-y‖) represents the expected value of the sample distance.

[0117] The loss function of the WGAN-GP model is:

[0118]

[0119] where ‖·‖ p denotes the p-norm; denotes the gradient operator, and λ is a penalty term coefficient, ζ is uniformly distributed in the range of [0, 1], denotes the straight-line uniform sampling between the sampling points of the real data distribution P r (·) and the generated data distribution P g (·).

[0120] As preferred, the generated wind-solar combined scene samples and load scene samples are subjected to scene reduction based on Gaussian mixture clustering to obtain typical source-load power scenes, including: taking the predicted future monthly load power consumption as input, outputting the typical load power scene and scene probability of the corresponding month in the future through GMM clustering, and taking the monthly wind-solar combined power expansion scene set as input, outputting the typical wind-solar combined power scene and scene probability of the corresponding month through GMM clustering.

[0121] The DenseNet-Autoformer network provided by the application performs deep decomposition on the load time series through the embedded time series decomposition module, the autocorrelation module and the deep feature extraction module, mines the deep periodicity, trend characteristics of the load monthly power consumption sequence and the correlation between external influencing factors and the load monthly power consumption, and establishes a load medium and long-term prediction model that comprehensively considers meteorological and economic factors.

[0122] The application establishes a progressive time series decomposition architecture, in the process of predicting future load power consumption through the load prediction model based on the DenseNet-Autoformer network, the model alternately performs prediction result optimization and sequence decomposition, combines the autocorrelation mechanism, realizes sequence-level linking of the load time series, reduces the time complexity of the model, and improves the calculation efficiency.

[0123] The application adopts the Wasserstein distance generated adversarial network with gradient penalty term to expand the wind-solar combined output data, does not need to specify the model or fit the probability distribution, and strengthens the stability and generation effect of model training.

[0124] The application adopts the Gaussian mixture clustering model to perform scene reduction on the generated wind-solar combined output samples, uses multiple Gaussian distributions to fit the wind-solar data distribution, and can be applicable to various complex data distributions.

[0125] In this embodiment, a scene generation method based on deep learning source-load medium and long-term prediction is described in detail in combination with the drawings.

[0126] The application constructs a source-load medium and long-term prediction and scene generation method based on deep learning, and the overall architecture is as shown in Figure 2As shown, mainly includes 3 parts, respectively based on DenseNet-Autoformer network's load medium and long term prediction model, based on WGAN-GP(Wasserstein GAN with gradient penalty) wind and light joint scene expansion model and based on Gaussian mixture clustering (Gaussian Mixture Model, GMM) scene reduction.

[0127] The load medium and long term prediction model based on DenseNet-Autoformer network: influenced by economic and meteorological factors, the load sequence change presents trend and periodic characteristics. The application constructs a load prediction model based on DenseNet-Autoformer network, mines the deep periodicity, trend characteristics of the load monthly electricity consumption sequence and the correlation between external influencing factors and the load monthly electricity consumption through the embedded time series decomposition module, autocorrelation module and deep feature extraction module, inputs the historical monthly meteorological, economic sequence and historical load monthly electricity consumption sequence, and outputs the prediction of future monthly load electricity consumption.

[0128] The wind and light joint scene expansion model based on WGAN-GP (Wasserstein GAN with gradient penalty): due to the randomness and uncertainty of wind and light output, the wind power and photovoltaic historical power scene are combined into a wind and light joint power scene according to the timestamp alignment according to the monthly label division, and a wind and light sample expansion model based on WGAN-GP is constructed, the distribution deviation is quantified by the Wasserstein distance, and a gradient penalty term is added to improve the training efficiency and avoid gradient disappearance, the input is the historical wind power daily power scene and the historical photovoltaic daily power scene, and the output is a monthly wind and light joint power expansion scene set.

[0129] The scene reduction method based on Gaussian mixture clustering (Gaussian Mixture Model, GMM): according to the predicted monthly electricity consumption, the historical load unit power curve of the corresponding month is combined, the typical load power scene and scene probability of the corresponding month in the future are obtained through GMM clustering, and the typical wind and light joint power scene and scene probability of the corresponding month are obtained through GMM clustering according to the monthly wind and light joint power expansion scene set.

[0130] 1. The load medium and long term prediction model based on DenseNet-Autoformer network:

[0131] The load medium and long term prediction model based on the DenseNet-Autoformer network constructed by the application is an encoder-decoder architecture model, and the time series decomposition module, the autocorrelation mechanism and the deep feature extraction module are embedded into the encoder and the decoder as independent modules, and the overall architecture of the model is as shown in Figure 3 .

[0132] (1) The embedded time series decomposition module:

[0133] The application embeds the time series decomposition unit into the encoder-decoder, optimizes the prediction result and sequence decomposition through the model alternately, separates the trend item and the periodic item from the hidden variable step by step, so that the gradual decomposition is realized. Further, the moving average method is used to smooth the violent fluctuation, so as to reduce the influence on the stability of the model training. For an input sequence X with a length of L, the decomposition process can be represented as:

[0134] X t =AvgPool(Padding(X))

[0135] X s =X-X t

[0136] In the formula, X t , X s represent the trend component and the periodic component respectively, AvgPool represents the average filtering convolution operation, and Padding represents the padding operation to keep the sequence length unchanged. The whole decomposition process can be represented as:

[0137] X t ,X s =SeriesDevomp(X).

[0138] (2) Autocorrelation mechanism / module:

[0139] The application uses the autocorrelation mechanism instead of the traditional self-attention mechanism to enhance the feature extraction ability of similar weather processes and adjacent months. The change trend of the load monthly electricity consumption to be predicted is highly correlated with the historical monthly electricity consumption sequence, and the current time pattern can be frequently observed in the historical subsequence. For a discrete time sequence {X t}, in order to model the correlation between the time sequences, the autocorrelation coefficient R xx (τ) is defined, which can be calculated as follows:

[0140]

[0141] In the formula, the autocorrelation coefficient R xx (τ) represents the sequence {X t} and its τ delay {X t-τsimilarity between them.

[0142] Selecting the top k time delay sequences with the strongest correlation The time delay information aggregation method is used to establish the period dependence relationship between the time delay subsequence and the current time sequence. Similar to the self-attention mechanism, the self-correlation mechanism also adopts a multi-head mechanism. For a head, the length L time sequence is calculated after high-dimensional projection, the query vector Q, the key vector K and the value vector V, and then the self-correlation coefficient R xx (τ) is calculated by Fast Fourier Transform (FFT), as shown in the following formula: Figure 4 The calculation process is as follows:

[0143]

[0144] Where i represents the imaginary unit, t represents the time in the time domain, f represents the frequency in the frequency domain, τ represents the time delay, F and F -1 -1 represent Fast Fourier Transform and its inverse transform respectively, F represents the conjugate part of F in the frequency domain calculation.

[0145] (3) Deep feature extraction module:

[0146] The present application adds a DenseNet layer after the feedforward layer to deeply extract the hidden layer periodicity and trend information of the long-time electricity consumption sequence, as shown in the dashed box in the following formula: Figure 3 The DenseNet adopts a dense connection mechanism, each layer is connected with each channel dimension of all previous layers, and is used as the input of the next layer, that is,

[0147] x l = H i ([x0,x1,...,x l-1 ])

[0148] In the formula, x l is the output of the lth layer; [x0,x1,...,x l-1 ] is the feature mapping connection of 0, 1, …, l-1 layers; is a nonlinear transformation function, which includes batch normalization (BN), rectified linear unit (ReLU), pooling layer and convolution layer (convolution) operations.

[0149] The forward propagation process of the DenseNet network is as follows: Figure 5As shown, the input of h3 is x2 of h2 and the connection of x1, x0 of the previous two layers in the channel dimension, that is, each layer can receive feature information from all previous layers.

[0150] The main part of the DenseNet network structure includes 4 dense blocks (DB) and 3 transition layers (TL). The TL is a Conv (1x1) + global average pooling (GAP) (2x2) layer. The feature maps transmitted between the layers in the DB are of the same size, which can realize the connection in the channel dimension; the non-linear transformation function H i () in the DB adopts the operation structure of BN + ReLU + Conv (3x3), as shown in Figure 6 In addition, the convolution operation of each layer in the DB obtains k channel numbers of feature maps, that is, the growth rate of the DenseNet network is k, and the input channel number of the lth layer is k0+k(l-1), where k0 is the channel number of the input layer feature map.

[0151] (4) Encoder:

[0152] As shown in Figure 3 , the DenseNet-Autoformer network has N encoder layers, each of which includes an autocorrelation module, two time series decomposition modules and a deep feature extraction module. The encoder is mainly used for periodic component modeling, and its input is a load monthly electricity consumption sequence X en of length Lseq, and the output contains historical period information of the load monthly electricity consumption sequence and will be used as interaction information to help the decoder refine the prediction results. The total process of the lth encoder layer can be expressed as as follows:

[0153]

[0154] In the formula, "_" represents the eliminated trend component. l∈{1,2,...,N} represents the output of the lth encoder layer and X en is encoded as respectively represent the periodic components output by the ith time series decomposition unit in the lth layer.

[0155] (5) Decoder:

[0156] As shown in Figure 3As shown in Figure 1, DenseNet-Autoformer has M decoder layers, each of which contains an autocorrelation module, an encoder-decoder correlation module that can use historical period information to refine the prediction, three time series decomposition modules, and a deep feature extraction module. The structure of the decoder layer can be divided into two parts, namely, an accumulation mechanism structure for accumulating trend components and a multi-layer autocorrelation mechanism block for period components. The input of the decoder is a length The periodic component X des and the trend component X det :

[0157]

[0158] X des =Concat(X ens , X0)

[0159] X det =Concat(X ent , X mean )

[0160] Where, X ens , X ent Represents the monthly load power consumption sequence X en The second half is the period and trend items obtained after time series decomposition. mean Represents a placeholder of length L0 filled with 0 and mean respectively. The overall output of the encoder is Using it as mutual information, the overall process of the l-th layer decoder can be expressed as Details are as follows:

[0161]

[0162] Where, l∈{1,2,...,M} represents the output of the l-th layer decoder; X des Encoded as And X det Encoded as i∈{1,2,3} respectively represent the periodic component and trend component obtained after the i-th time series decomposition block processing in the l-th layer decoder. l,i ,i∈{1,2,3} means extracting deep trend information in the i-th time series decomposition block Finally, the encoder outputs the periodic features and trend features extracted from the two refined structures respectively.

[0163] 2. Expanded model for wind-solar combined power scenarios based on WGAN-GP:

[0164] Since the wind and light output has obvious seasonal characteristics and correlation, the wind and light output is divided into subsets according to months in the wind and light scene generation, and the wind power and the photovoltaic power are combined into wind and light combined power according to the time stamp for training.

[0165] The GAN is used to expand the wind and light output sample, and the objective function of the adversarial network training process min-max game is:

[0166]

[0167] Wherein G(·) represents the generator, D(·) represents the discriminator, E represents the expectation, x represents the historical wind and light output data, p(x) represents the probability distribution of x, z is random noise, and p(z) represents the probability distribution of z.

[0168] However, the traditional GAN is prone to training instability, gradient disappearance and mode collapse problems. In view of the problems in the training of the GAN model, the Wasserstein distance is used instead of the JS divergence as the optimization target, and the definition of the Wasserstein distance is:

[0169]

[0170] Wherein P r represents the real data distribution, P g represents the generated data distribution, Π(P r ,P g ) is the set of all possible joint distributions combined by P r and P g , γ represents one of the joint distributions, (x,y) ~ γ represents that a sample x and y are obtained by sampling from the joint distribution γ, ‖x-y‖ is the distance between the pair of samples, and E (x,y)～γ (‖x-y‖) represents the expected value of the sample distance.

[0171] The above formula represents that the lower bound of the expected value that can be taken in all possible joint distributions is the Wasserstein distance.

[0172] The WGAN adopts weight clipping to meet the Lipschitz continuity condition, so that the weight and bias coefficients of the WGAN tend to be extreme, which limits the fitting ability of the neural network. Therefore, a gradient penalty term is added to the original network loss function, and the loss function of the established WGAN-GP model is:

[0173]

[0174] Wherein ‖·‖ p represents the p-norm; denotes the gradient operator, and λ is the penalty term coefficient, ζ is uniformly distributed in the range of [0, 1], denotes the real data distribution P r (·) and the generated data distribution P g (·) uniformly samples the straight line between the sampling points.

[0175] 3. A scene reduction method based on GMM clustering:

[0176] According to the load unit curve of the corresponding month in the history and the predicted monthly electricity consumption, the application obtains the typical load power scene in the corresponding month in the future through GMM clustering, and obtains the typical wind and light power scene in the corresponding month according to the monthly wind and light combined power expansion scene set through GMM clustering. The clustering process is described as follows:

[0177] (1) The parameters μ j and ε j of the k multivariate Gaussian distributions are initialized.

[0178] (2) All sample points are traversed, and the probability γ i of the sample point x i,j (i = 1, 2, …, m) belonging to the jth Gaussian distribution is calculated.

[0179]

[0180] (3) The update values μ' j and ε' j of the Gaussian distribution parameters μ j and ε j are calculated by the following formula.

[0181]

[0182] (4) Repeat (2) and (3) until the Gaussian parameters converge.

[0183] (5) Using the calculated Gaussian parameters, traverse all samples, and classify the samples into the class with the maximum probability γ i,j .

[0184] In this embodiment, source-load data, regional economic development and meteorological data from January 1, 2012 to December 31, 2022 in a certain rural area in northern China are used for research, the meteorological data includes temperature and humidity, the economic development data includes GDP, per capita GDP, regional population size and industry added value. The time resolution of wind, light and load power is 1 hour, and the above data is divided into training set and test set in the ratio of 4:1, and the parameter setting is shown in Table 1.

[0185] Table 1 parameter settings

[0186] Parameter Parameter value Encoder layer number 2 Decoder layer number 1 Batch size 16 Iterations 2 Epoch 10 Learning rate 0.01 Number of wind and light moon samples 1000 Number of typical scenes of wind and light moon 5

[0187] (1) Comparative model settings:

[0188] To verify the superiority of the load medium and long-term prediction method based on the DenseNet-Autoformer model proposed in the present application, three prediction models are set up for comparative analysis. Model 1: DenseNet-Autoformer model considering meteorological and economic factors, Model 2: DenseNet-Autoformer model without considering meteorological and economic factors, Model 3: LSTM model considering meteorological and economic factors, which is used to predict the load of rural areas. The test set is the monthly electricity consumption from January 2021 to December 2022.

[0189] To verify the effectiveness of the wind and light combined scene generation method based on WGAN-GP proposed in the present application, two wind and light sample generation models are set up for comparative analysis. Model 1: wind and light combined scene generation model based on WGAN-GP, Model 2: wind and light combined scene generation model based on Frank-copula.

[0190] (2) Model evaluation:

[0191] For the deterministic load prediction results, root mean square error (RMSE) and mean absolute percentage error (MAPE) are used as load prediction model precision evaluation indexes:

[0192]

[0193] (3) Comparative experimental results:

[0194] Figure 7 The load prediction results of the three models are shown. Figure 7 It can be seen that the prediction trends of load medium and long-term prediction models 1, 2 and 3 are consistent with the measured value trends, proving the sensitivity and effectiveness of the models. However, the prediction curve of model 1 has the highest fitting degree with the measured curve, and has higher prediction accuracy.

[0195] Table 2 gives the root mean square error, mean absolute error evaluation index comparison of the model and two kinds of comparative model. As can be seen from Table 2, the load prediction model of the present application is better than the comparative prediction model in two evaluation indexes. From the root mean square error, the load prediction model of the present application is improved by 5.31%, 5.62% compared with model 1, model 2 respectively; from the average absolute error, the load prediction model of the present application is improved by 5.1%, 5.77% compared with model 1, model 2 respectively, which effectively proves the advantage of the load prediction method of the present application in the aspect of prediction accuracy.

[0196] Table 2 evaluation index

[0197] Evaluation index Model 1 Model 2 Model 3 RMSE 5.15％ 10.46％ 10.77％ MAPE 3.91％ 9.01％ 9.68％

[0198] From Figure 8 and Figure 9 It can be seen that the probability density function curve of the sample generated by the model of the present application is closer to the probability density function curve of the real sample, which shows that the WGAN-GP model of the present application can effectively fit the wind and light output characteristics. Further, the autocorrelation coefficient of the generated data is calculated with a time interval of 1-6h, and the results are shown in Figure 10 and Figure 11 The generated data of the model and the comparative model have certain autocorrelation within 1-6h time interval, and the autocorrelation coefficient gradually decreases with the increase of time interval, but the autocorrelation coefficient of the generated data of the model of the present application is closer to the autocorrelation coefficient of the real data, which shows that the model of the present application has the ability to simulate the time series of wind and light output.

[0199] (4) Conclusion:

[0200] The results of the present embodiment show that:

[0201] The load medium and long term prediction method based on DenseNet-Autoformer model of the present application is better than the comparative model in the two evaluation indexes of root mean square error and mean absolute error. Compared with model 1 which is better in the comparative model, the root mean square error and the mean absolute error are improved by 5.31%, 5.1% respectively.

[0202] Compared with the traditional Frank-copula wind and light joint scene generation method, the wind and light joint scene generation method based on WGAN-GP of the present application obtains the probability density curve and autocorrelation coefficient curve of wind and light output which are closer to the real sample, which effectively verifies that the wind and light joint scene generation method of the present application can effectively fit the wind and light output characteristics.

[0203] The method of the present application can be used for generating typical source and load scenes.

[0204] As shown in Figure 12 The application also provides a scene generation system based on deep learning source-load medium and long-term prediction, which comprises a load prediction module, a scene expansion module and a hybrid clustering module.

[0205] The load prediction module is configured to predict future load electricity consumption through a load prediction model based on a DenseNet-Autoformer network.

[0206] The scene expansion module is configured to expand wind-solar joint scene samples.

[0207] The hybrid clustering module is configured to reduce the generated wind-solar joint scene samples according to the predicted future load electricity consumption to obtain typical wind-solar joint power scenes, integrate the predicted future load electricity consumption and the typical wind-solar joint power scenes, and obtain source-load scenes.

[0208] Specifically, the load prediction module contains a load prediction model based on a DenseNet-Autoformer network, which is used to take historical monthly meteorological and economic sequences and historical load monthly electricity consumption sequences collected by the data acquisition module as input, predict future load electricity consumption through the load prediction model based on the DenseNet-Autoformer network, and output future monthly load electricity consumption prediction results.

[0209] Specifically, the scene expansion module contains a wind-solar joint scene expansion model based on WGAN-GP, which is used to take historical wind power daily power scenes and historical photovoltaic daily power scenes collected by the data acquisition module as input, expand wind-solar joint scene samples according to the wind-solar joint scene expansion model based on WGAN-GP, and output a monthly wind-solar joint power expansion scene set.

[0210] Specifically, the hybrid clustering module is used to take the predicted future monthly load electricity consumption as input, output typical load power scenes and scene probabilities in the corresponding month through GMM clustering, and take the monthly wind-solar joint power expansion scene set as input, output typical wind-solar joint power scenes and scene probabilities in the corresponding month through GMM clustering.

[0211] The scene generation system based on deep learning source-load medium and long-term prediction of the application is applicable to the scene generation method based on deep learning source-load medium and long-term prediction described above.

[0212] As an implementation mode, the system further comprises a data acquisition module, and specifically, the data acquisition module is configured to collect data required by the load prediction module, the scene expansion module and the hybrid clustering module, including historical monthly meteorological and economic sequence data, historical load monthly electricity consumption sequence data, historical wind power daily power scenes and historical photovoltaic daily power scenes.

[0213] Specifically, the medium- and long-term load forecasting model based on the DenseNet-Autoformer network constructed in the present invention is an encoder-decoder architecture model. The time series decomposition module, autocorrelation mechanism and deep feature extraction module are embedded in the encoder and decoder as independent modules. The overall architecture of the model is as follows: Figure 3 shown.

[0214] This paper embeds a time series decomposition unit within the encoder-decoder. By alternating between prediction optimization and sequence decomposition, the model gradually separates trend terms and periodic terms from latent variables, achieving progressive decomposition. Furthermore, a moving average method is used to smooth out sharp fluctuations to reduce their impact on model training stability. For an input sequence X of length L, the decomposition process can be expressed as:

[0215] X t =AvgPool(Padding(X))

[0216] X s =XX t

[0217] Where, X t 、X s Represent the trend component and the period component respectively, AvgPool represents the average filter convolution operation, and Padding represents the padding operation to keep the sequence length unchanged. The entire decomposition process can be expressed as:

[0218] X t ,X s =SeriesDevomp(X).

[0219] The present invention uses the autocorrelation mechanism instead of the traditional self-attention mechanism to enhance the feature extraction capability of similar weather processes and adjacent months. The monthly power consumption trend of the load to be predicted is highly correlated with the historical monthly power consumption sequence, and the current time pattern can be frequently observed in the historical subsequence. For the discrete time series {X t In order to model the correlation between its time series, the present invention defines the autocorrelation coefficient R xx (τ), can be calculated by the following formula:

[0220]

[0221] In the formula, the autocorrelation coefficient R xx (τ) represents the sequence {X t} with its τ delay {X t-τ} similarities between them.

[0222] Select the top k time delay sequences with the strongest correlation The time delay information aggregation method is used to establish the periodic dependence relationship between the time delay subsequence and the current time sequence. Similar to the self-attention mechanism, the self-correlation mechanism also adopts a multi-head mechanism. For one head, the length L time sequence is calculated after high-dimensional projection, the query vector Q, the key vector K and the value vector V are calculated, and the self-correlation coefficient R xx (τ) is calculated by Fast Fourier Transform (FFT), as shown in the following formula: Figure 4

[0223]

[0224] Where i represents an imaginary unit, t represents time in the time domain, f represents frequency in the frequency domain, τ represents time delay, F and F -1 represent Fast Fourier Transform and its inverse transform respectively, F* represents the conjugate part of F in the frequency domain calculation.

[0225] The present application adds a DenseNet layer after the feedforward layer to deeply extract the hidden layer periodicity and trend information of the long-time electricity consumption sequence, as shown in the dashed box in the following formula: Figure 3 The DenseNet adopts a dense connection mechanism, each layer is connected with each channel dimension of all previous layers, and is used as the input of the next layer, that is,

[0226] x l = H i ([x0,x1,...,x l-1 ])

[0227] Where x l is the output of the lth layer; [x0,x1,...,x k-1 ] is the feature mapping connection of 0, 1,..., l-1 layers; is a nonlinear transformation function, which includes batch normalization (BN), rectified linear unit (ReLU), pooling layer and convolution layer operations.

[0228] The forward propagation process of the DenseNet network is shown in the following formula: Figure 5 The input of h3 is the connection of x2 of h2 and x1, x0 of the previous two layers in the channel dimension, that is, each layer can receive feature information from all previous layers.

[0229] ​The main part of the DenseNet network structure includes four dense connection blocks (DBs) and three transition layers (TLs). The TL is a Conv (1x1) + global average pooling (GAP) (2x2) layer. The feature map sizes transmitted between the layers in the DB are consistent, which can realize the connection in the channel dimension; the non-linear transformation function H i () in the DB adopts the operation structure of BN + ReLU + Conv (3x3), as shown in Figure 6 In addition, the channel number of k feature maps is obtained after the convolution operation of each layer in the DB, that is, the growth rate of the DenseNet network is k, and the input channel number of the lth layer is k0+k(l-1), where k0 is the channel number of the input layer feature map.

[0230] As shown in Figure 3 , the DenseNet-Autoformer network has N encoder layers, and each encoder layer includes an autocorrelation module, two time series decomposition modules, and a deep feature extraction module. The encoder is mainly used for periodic component modeling, and its input is a load monthly electricity consumption sequence X en , and the output contains historical periodic information of the load monthly electricity consumption sequence and will be used as interaction information to help the decoder refine the prediction result. The total process of the lth layer encoder can be expressed as Details are as follows:

[0231]

[0232] In the formula, "_" represents the eliminated trend component. l∈{1,2,...,N} represents the output of the lth layer encoder and X en is encoded as i∈{1,2} respectively represent the periodic components output by the ith time series decomposition unit of the lth layer.

[0233] As shown in Figure 3 , the DenseNet-Autoformer has M decoder layers, and each decoder includes an autocorrelation module, an encoder-decoder correlation module that can refine the prediction using historical periodic information, three time series decomposition modules, and a deep feature extraction module. The structure of the decoder layer can be divided into two parts, i.e., an accumulation mechanism structure for accumulating trend components, and a multi-layer autocorrelation mechanism block for periodic components. The input of the decoder is the periodic component X and the trend component X des , and the output is the final prediction result X det:

[0234]

[0235] X des =Concat(X ens ,X0)

[0236] X det =Concat(X ent ,X mean )

[0237] In the formula, X ens , X ent represent the period and trend items obtained by time series decomposition of the second half of the load monthly electricity consumption sequence X en . X0, X mean respectively represent placeholders with a length of L0 filled by 0 and mean value. The overall output of the encoder is Using it as mutual information, the total process of the lth layer decoder can be expressed as In detail as follows:

[0238]

[0239] In the formula, l∈{1,2,...,M} represents the output of the lth layer decoder; X des is encoded as and X det is encoded as i∈{1,2,3} respectively represent the period component and the trend component obtained after the i th time series decomposition block processing in the lth layer decoder. W l,i ,i∈{1,2,3} represent the high-dimensional projection when the deep trend information is extracted in the i th time series decomposition block. Finally, the encoder outputs the periodic features and the trend features extracted by the two refinement structures respectively.

[0240] Since the wind and light output has obvious seasonal characteristics and correlation, the wind and light output is divided into subsets according to months in the wind and light scene generation, and the wind power and photovoltaic power are combined into wind and light combined power according to the time stamp for training.

[0241] GAN is used to expand the wind and light output sample, and the objective function of the training process of the generative adversarial network min-max game is:

[0242]

[0243] Wherein, G(·) represents a generator, D(·) represents a discriminator, E represents expectation, x represents historical wind light output data, p(x) represents a probability distribution of x, z is random noise, and p(z) represents a probability distribution of z.

[0244] However, the conventional GAN is prone to training instability, gradient disappearance and mode collapse. In view of the problems in the training of the GAN model, the present application uses a Wasserstein distance instead of a JS divergence as an optimization target, and the definition of the Wasserstein distance is as follows:

[0245]

[0246] Wherein, P r represents a real data distribution, P g represents a generated data distribution, Π(P r ,P g ) is a set of all possible joint distributions combined by P r and P g , γ represents one of the joint distributions, (x,y) ~ γ represents sampling from the joint distribution γ to obtain a sample x and y, ‖x-y‖ is the distance between the pair of samples, E (x,y)～γ (‖x-y‖) represents the expected value of the sample distance.

[0247] The above formula indicates that the lower bound of the expected value that can be taken in all possible joint distributions is the Wasserstein distance.

[0248] The WGAN adopts a weight clipping method to meet the Lipschitz continuity condition, which causes the weight and bias coefficients of the WGAN to tend to be extreme, thereby limiting the fitting ability of the neural network. Therefore, a gradient penalty term is added to the original network loss function, and the loss function of the established WGAN-GP model is as follows:

[0249]

[0250] Wherein, ‖·‖ p represents a p-norm; represents a gradient operator, λ is a penalty term coefficient, ζ is a uniform distribution in the range of [0, 1], represents uniform sampling of a straight line between sampling points from the real data distribution P r (·) and the generated data distribution P g (·).

[0251] The application obtains the typical load power scene in the corresponding month in the future through GMM clustering according to the load unit curve of the corresponding month in history and the predicted monthly power consumption, and obtains the typical wind-solar power scene in the corresponding month according to the monthly wind-solar combined power expansion scene set through GMM clustering. The clustering process is described as follows:

[0252] Step 1: initialize the parameters μ of k multivariate Gaussian distributions j and ε j , assuming that each mixed element has a respective diagonal matrix.

[0253] Step 2: traverse all sample points, and calculate the probability γ i of the sample point x i,j (i=1, 2, …, m) belonging to the jth Gaussian distribution.

[0254]

[0255] Step 3: calculate the updated values μ' and ε' of the Gaussian distribution parameters μ and ε by the following formula. j j j j

[0256]

[0257] Step 4: repeat step 2 and step 3 until the Gaussian parameters converge.

[0258] Step 5: traverse all samples using the calculated Gaussian parameters, and classify the samples into the class with the maximum probability γ i,j .

[0259] To solve the above technical problems, the embodiment of the application further provides an electronic device, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the scene generation method based on deep learning source load medium and long term prediction as described above when executing the program.

[0260] As Figure 13 ​​​​As shown, the computer device includes a memory, a processor, and a network interface that are communicatively connected to each other through a system bus. It should be noted that the computer device shown in the figure has components of memory, processor, network interface, and operating system, but it should be understood that not all of the components shown need to be implemented, and more or fewer components can be implemented instead. Among them, those skilled in the art can understand that the computer device here is a device that can automatically perform numerical calculation and / or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to microprocessors, application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0261] The computer device can be a desktop computer, a notebook computer, a palm computer, a cloud server, and the like. The computer device can interact with the user through a keyboard, a mouse, a remote controller, a touchpad, a voice control device, and the like.

[0262] The memory can be one or more, and at least includes one type of readable storage medium, including a flash memory, a hard disk, a multimedia card, a card-type memory (e.g., an SD or DX memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, an optical disk, etc. In some embodiments, the memory can be an internal storage unit of the computer device, such as a hard disk or a memory of the computer device. In other embodiments, the memory can also be an external storage device of the computer device, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped with the computer device. Of course, the memory can also include both the internal storage unit and the external storage device of the computer device. In this embodiment, the memory is generally used to store the operating system and various application software installed on the computer device, such as program codes based on the scene generation method for long-term prediction of deep learning source load, etc. In addition, the memory can also be used to temporarily store various data that have been output or will be output.

[0263] The processor can be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip in some embodiments. The processor is generally used to control the overall operation of the computer device. In the present embodiment, the processor is used to run program codes or process data stored in the memory, for example, program codes of the scene generation method based on deep learning source load medium and long term prediction.

[0264] The network interface can include a wireless network interface and / or a wired network interface, which is generally used to establish a communication connection between the computer device and other electronic devices.

[0265] The present application also provides another implementation, that is, to provide a readable storage medium, which stores a computer program, and the computer program is executed by the processor to realize the steps of the scene generation method based on deep learning source load medium and long term prediction as described above.

[0266] The purpose of the present application is to fully extract the source load uncertainty features and realize the accurate and reliable characterization of the source load scene. First, a load medium and long term prediction model based on DenseNet-Autoformer network is constructed to extract the periodicity and trend features of the load, and the load medium and long term prediction result is obtained by comprehensively considering meteorological and economic factors. On this basis, combined with the load standard curve of the corresponding month in history, the typical load scene under the corresponding month in the future is obtained. Secondly, the improved generative adversarial network is used to expand the wind and light scene samples. Finally, the typical wind and light scene is obtained by Gaussian mixture clustering on the generated wind and light joint output samples.

[0267] In the description of the present application, the description of the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present application, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0268] It is not difficult for those skilled in the art to understand that the present application includes any combination of the above description of the summary and specific embodiments and the parts shown in the drawings, which is limited in length and is brief in the description. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0269] Although the embodiments of the present application have been shown and described above, it is understood that the above-described embodiments are exemplary and are not to be construed as limiting the present application, and that modifications, substitutions, replacements and variations of the above-described embodiments can be made by those skilled in the art without departing from the principles and spirit of the present application. The scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A scene generation method based on deep learning source load medium and long term prediction, characterized in that, include: Predict future load power consumption through a load forecasting model based on the DenseNet-Autoformer network; Expand the wind-solar joint scenario samples, reduce the generated wind-solar joint scenario samples to obtain the typical wind-solar joint power scenario, integrate the predicted future load power consumption and the typical wind-solar joint power scenario to obtain the source-load scenario; The method of predicting future load power consumption by using a load forecasting model based on a DenseNet-Autoformer network includes: taking historical monthly average meteorological and economic sequences and historical monthly load power consumption sequences as input, predicting future load power consumption by using a load forecasting model based on a DenseNet-Autoformer network, and outputting load power consumption prediction results for each future month; The load forecasting model based on the DenseNet-Autoformer network includes N encoder layers and M decoder layers; each encoder layer includes an autocorrelation module, two time series decomposition modules and a deep feature extraction module; each decoder layer includes an autocorrelation module, an encoder-decoder correlation module, three time series decomposition modules and a deep feature extraction module; The deep feature extraction module adds a DenseNet layer after the feedforward layer. The DenseNet network structure includes 4 dense connection blocks and 3 transfer layers. The transfer layer is Conv1×1+global average pooling layer. The feature maps transferred between each layer in the dense connection block are of the same size. The nonlinear conversion function in the dense connection block adopts the BN+ReLU+Conv3×3 operation structure. After the convolution operation, each layer in the dense connection block obtains the number of channels of k feature maps, that is, the growth rate of the DenseNet network is k, and the number of input channels of the lth layer is k0+k(l-1), where k0 is the number of channels of the input layer feature map.

2. The scene generation method based on deep learning source load medium and long term prediction according to claim 1, characterized in that, The expanding the wind-solar joint scene samples includes: expanding the wind-solar joint scene samples according to a wind-solar joint scene expansion model based on WGAN-GP.

3. The scene generation method based on deep learning source load medium and long term prediction according to claim 1, characterized in that, The time series decomposition module is embedded in the encoder-decoder. It alternates between optimizing prediction results and decomposing sequences through the model, gradually separating trend terms and periodic terms from latent variables to achieve progressive decomposition. The moving average method is used to smooth out sharp fluctuations. For an input sequence X of length L, the decomposition process is as follows: X t = AvgPool(Padding(X)) X s = X - X t wherein X t , X s represent the trend component and the periodic component, respectively, AvgPool denotes an average filtering convolution operation, and Padding denotes a padding operation.

4. The scene generation method based on deep learning source load medium and long term prediction according to claim 1, characterized in that, The autocorrelation module defines the autocorrelation coefficients R t (τ) for a discrete time series {X xx (τ), where the autocorrelation coefficient R xx (τ) represents the correlation between the sequence {X t} and its τ delay {X t-τ} selecting the strongest correlation of the first c time delay sequence The time delay information aggregation method is used to establish the period dependence relationship between the time delay subsequence and the current time sequence, and the autocorrelation coefficient R xx (τ) is calculated as follows: where i denotes the imaginary unit, t denotes time in the time domain, f denotes frequency in the frequency domain, τ denotes a time delay, F and F -1 denote the fast Fourier transform and its inverse, respectively, denotes the conjugate part of F in the frequency domain calculation.

5. The scene generation method based on deep learning source load medium and long term prediction according to claim 1, characterized in that, The generated wind-solar joint scenario samples are reduced to obtain a typical wind-solar joint power scenario, and the predicted future load power consumption and the typical wind-solar joint power scenario are integrated to obtain the source-load scenario, including: based on Gaussian mixture clustering, the generated wind-solar joint scenario samples and load power samples are reduced to obtain a typical source-load power scenario.

6. The scene generation method based on deep learning source load medium and long term prediction according to claim 2, characterized in that, The method of expanding the wind-solar joint scenario samples according to the wind-solar joint scenario expansion model based on WGAN-GP includes: taking historical wind power daily power scenarios and historical photovoltaic daily power scenarios as input, expanding the wind-solar joint scenario samples according to the wind-solar joint scenario expansion model based on WGAN-GP, and outputting a monthly wind-solar joint power expansion scenario set.

7. The scene generation method based on deep learning source load medium and long term prediction according to claim 6, characterized in that, The wind and light combined scene sample expanded according to the wind and light combined scene expansion model based on WGAN-GP further includes: dividing wind and light output into subsets according to months, and training wind power and photovoltaic power combined into wind and light combined power according to timestamps, using GAN to expand wind and light output samples, and generating a target function of a min-max game of an adversarial network training process as: Wherein, G(·) represents a generator, D(·) represents a discriminator, E represents an expectation, x represents historical wind and light output data, p(x) represents a probability distribution of x, z is random noise, and p(z) represents a probability distribution of z; The Wasserstein distance is used to replace the JS divergence as an optimization target, and the definition of the Wasserstein distance is: where P r represents the real data distribution, P g represents the generated data distribution, Π(P r , P g ) is the set of all possible joint distributions that P r and P g are combined together, γ represents one of the joint distributions, (x, y) ~ γ represents that a sample x and y are sampled from the joint distribution γ, and ||x - y|| is the distance between the pair of samples, E (x，y)～γ (‖x - y‖) represents the expected value of the sample distance. The loss function of the WGAN-GP model is: where ||·|| denotes the p-norm; p denotes the p-norm; denotes the gradient operator, and λ is a penalty term coefficient, ζ is uniformly distributed in the range [0, 1], denotes sampling from the real data distribution P r (·) and the generated data distribution P g uniformly samples a straight line between the sampled points of (·) and (·).

8. The scene generation method based on deep learning source load medium and long term prediction according to claim 5, characterized in that, The typical source and load power scene is obtained by scene reduction on the generated wind and light combined scene sample and load power sample based on Gaussian mixture clustering, including: taking the predicted future monthly load power as input, outputting the typical load power scene and scene probability of the corresponding month in the future through GMM clustering, and taking the monthly wind and light combined power expansion scene set as input, outputting the typical wind and light combined power scene and scene probability of the corresponding month through GMM clustering. 9.A system for generating a scenario based on deep learning source load medium and long term prediction, characterized in that, It includes: a load prediction module, a scene expansion module and a mixed clustering module, The load prediction module is configured to predict future load power through a load prediction model based on a DenseNet-Autoformer network; The scene expansion module is configured to expand wind and light combined scene samples; The mixed clustering module is configured to reduce the generated wind and light combined scene samples to typical wind and light combined power scenes according to the predicted future load power, and integrate the predicted future load power and the typical wind and light combined power scenes to obtain a source and load scene; The future load power is predicted through the load prediction model based on the DenseNet-Autoformer network, including: taking historical monthly meteorological, economic sequences and historical load monthly power consumption sequences as input, predicting future load power through the load prediction model based on the DenseNet-Autoformer network, and outputting future monthly load power prediction results; The load prediction model based on the DenseNet-Autoformer network includes N encoder layers and M decoder layers; each encoder layer includes a self-correlation module, two time decomposition modules and a deep feature extraction module; each decoder layer includes a self-correlation module, an encoder-decoder correlation module, three time decomposition modules and a deep feature extraction module; The deep feature extraction module adds a DenseNet layer after a feedforward layer, the DenseNet network structure comprises four dense connection blocks and three transfer layers, the transfer layer is a Conv1x1+global average pooling layer, the feature mapping sizes transmitted between layers in the dense connection block are consistent, a BN+ReLU+Conv3x3 operation structure is adopted for a nonlinear conversion function in the dense connection block, after convolution operation is performed on layers in the dense connection block, k feature mapping channel numbers are obtained, that is, a growth rate of the DenseNet network is k, and an input channel number of the lth layer is k0+k(l-1), wherein k0 is a feature mapping channel number of an input layer.

Images