Virtual power plant source load prediction model construction method, source load prediction method and device

Abstract

The invention discloses a virtual power plant source load prediction model construction method and device and a source load prediction method and device, relates to the technical field of power systems, and at least solves the problem that neural network parameters of a source load prediction model cannot be automatically adjusted. The method comprises the following steps: taking historical source load data and historical weather factor data of a plurality of scenes as sample data; an initial model representing the mapping relation between the sample data and the source load curve is obtained; the initial model comprises a first initial optimization layer and a second initial optimization layer; the first initial optimization layer is used for selecting a modal decomposition algorithm associated with the initial model to carry out first optimization adjustment on the sample data to obtain key feature data; the second initial optimization layer is used for performing second optimization adjustment on the model structure and the model parameters of the initial model based on the key feature data; and repeatedly executing the first optimization adjustment and the second optimization adjustment according to the sample data to obtain a prediction model matched with each scene and a preset mode decomposition algorithm associated with the prediction model.

Description

Technical Field

[0001] The present application relates to the technical field of power systems, and in particular to a method for constructing a virtual power plant source-load prediction model, a source-load prediction method and a device. Background Art

[0002] With the large-scale application of renewable energy such as solar energy and wind energy, their volatility and uncertainty have brought challenges to the power system. Therefore, improving the accuracy of source-load prediction has become a research hotspot. Advanced technologies such as deep learning neural networks can effectively process nonlinear and high-dimensional data. In order to improve the stability and reliability of source-load prediction and the prediction accuracy of renewable energy and load, the source-load data characteristics are usually observed first, and then the source-load prediction model is constructed based on the data characteristics. The neural network parameters of the above source-load prediction model are fixed and cannot be automatically adjusted according to changes in the prediction sample set. After changing the sample set, the neural network parameters must be manually adjusted again, and the whole process has a lot of trial and error time cost. Summary of the invention

[0003] The present invention provides a method for constructing a virtual power plant source-load prediction model, a source-load prediction method and a device, so as to at least solve the problem that the neural network parameters of the source-load prediction model in the related art are fixed and unchanged, and the parameters cannot be automatically adjusted according to the changes in the prediction sample set. After the sample set is changed, the neural network parameters must be manually adjusted again, and the whole process has a large amount of trial and error time cost. The technical solution of the present invention is as follows:

[0004] According to a first aspect of an embodiment of the present invention, a method for constructing a source-load prediction model of a virtual power plant is provided, wherein historical source-load data under multiple scenarios and corresponding historical weather factor data are used as sample data; initial models representing the mapping relationship between the sample data and the source-load curves of each scenario in the multiple scenarios are obtained in a one-to-one correspondence; the initial model includes a first initial optimization layer and a second initial optimization layer; the first initial optimization layer is used to select a modal decomposition algorithm associated with the initial model from multiple modal decomposition algorithms, and perform a first optimization adjustment on the sample data based on the selected associated modal decomposition algorithm, so as to extract key feature data adapted to the scenario corresponding to the initial model from the sample data; the second initial optimization layer is used to perform a second optimization adjustment on the model structure and model parameters of the initial model based on the key feature data; for any initial model of each initial model corresponding to each scenario, the first optimization adjustment and the second optimization adjustment are repeatedly performed according to the sample data until the corresponding first optimization condition and the second optimization condition are met, and a prediction model matching the scenario corresponding to any initial model and a preset modal decomposition algorithm associated with the prediction model are obtained.

[0005] In one implementation, the first optimization condition is that the first adjustment times of executing the first optimization adjustment reaches the adjustment threshold and the preset modal decomposition algorithm is the modal decomposition algorithm with the shortest adjustment time for the first optimization adjustment; the second optimization condition is that the second adjustment times of executing the second optimization adjustment reaches the adjustment threshold and the prediction model is that the source load curve output by any initial model after the first optimization adjustment and the second optimization adjustment differs from the actual source load curve within a preset error range; the actual source load curve is the real source load curve of any initial model under the corresponding scenario and historical weather factors.

[0006] In another implementation, historical source load data in multiple scenarios and corresponding historical weather factor data are used as sample data, including: fitting the historical source load data to obtain a fitting curve, and using the fitting curve and the historical weather factor data as sample data.

[0007] In another implementation, performing the first optimization adjustment includes: using the selected associated modal decomposition algorithm to decompose the fitting curve of the historical source load data into multiple subsequences; based on the weights of the multiple subsequences, removing the noise sequences in the multiple subsequences to obtain multiple useful sequences; and, based on the weights of the weather factor characteristics of the scene corresponding to the initial model, selecting historical weather characteristic factors associated with the scene corresponding to the initial model from the historical weather factor data; the weather factor characteristics include temperature, humidity, wind speed, air pressure, and precipitation; using the historical weather characteristic factors and the multiple useful sequences as key feature data, or normalizing the historical weather characteristic factors and the multiple useful sequences to obtain the key feature data.

[0008] In another implementation, based on the weights of multiple subsequences, noise sequences are removed from multiple subsequences to obtain multiple useful sequences; and based on the weights of the weather factor characteristics of the scene corresponding to the initial model, before selecting historical weather characteristic factors associated with the scene corresponding to the initial model from the historical weather factor data, the method also includes: calling the improved whale optimization algorithm to respectively determine the weights of the weather factor characteristics of the scene corresponding to the initial model and the weights of multiple subsequences; the improved whale optimization algorithm is optimized using a chaos operator, a Levy flight strategy, and an "observer" strategy.

[0009] In another implementation, the initial model is a neural network model, and the model structure of the initial model includes one or more of a convolutional layer, a pooling layer, a fully connected layer, an attention mechanism, and a basic long short-term memory neural network or a bidirectional long short-term memory neural network; the model parameters of the initial model include the size of the convolution kernel, the step size, the zero padding, the number of filters, the pooling step size, the filter size, the pooling type, the weight between the input layer and the hidden layer, and the number of neurons in each layer.

[0010] In another implementation, performing the second optimization adjustment specifically includes: calling the improved whale optimization algorithm to respectively determine the weight of the model structure of the initial model and the weight of the model parameters of the initial model; based on the weight of the model structure of the initial model, determining the model structure of the initial model in the second optimization adjustment process; and, based on the weight of the model parameters of the initial model, determining the model parameters of the initial model in the second optimization adjustment process.

[0011] According to a second aspect of an embodiment of the present invention, a method for predicting source and load of a virtual power plant is provided, the method comprising: obtaining current source and load data and current weather factor data; and, based on the association between the scene and the prediction model, selecting a target model that matches the current scene from multiple prediction models; wherein the prediction model is used to generate a source and load curve of a source and load relationship based on the source and load data and the weather factor data; the current source and load data and the current weather factor data are input into the target model, and target key feature data associated with the current scene in the current source and load data and the current weather factor data are extracted using a target modal decomposition algorithm associated with the target model, and a target source and load curve characterizing the target source and load relationship is obtained based on the target key feature data; the prediction model is constructed using the method for constructing a virtual power plant source and load prediction model as in any one of the implementations of the first aspect above.

[0012] According to a third aspect of an embodiment of the present invention, a device for constructing a source-load prediction model of a virtual power plant is provided, the device comprising: a first acquisition unit, used to use historical source-load data and corresponding historical weather factor data under multiple scenarios as sample data; the first acquisition unit, also used to obtain each initial model that represents the mapping relationship between the sample data and the source-load curve of each scenario in the multiple scenarios in a one-to-one correspondence; the initial model comprises a first initial optimization layer and a second initial optimization layer; the first initial optimization layer is used to select a modal decomposition algorithm associated with the initial model from multiple modal decomposition algorithms, and perform a first optimization adjustment on the sample data based on the selected associated modal decomposition algorithm, so as to extract key feature data adapted to the scenario corresponding to the initial model from the sample data; the second initial optimization layer is used to perform a second optimization adjustment on the model structure and model parameters of the initial model based on the key feature data; a construction unit, used to repeatedly perform the first optimization adjustment and the second optimization adjustment on any initial model of each initial model corresponding to each scenario according to the sample data, until the corresponding first optimization condition and the second optimization condition are met, and the corresponding prediction model matching the scenario corresponding to any initial model and the preset modal decomposition algorithm associated with the prediction model are obtained.

[0013] According to a fourth aspect of an embodiment of the present invention, a virtual power plant source-load prediction device is provided, the device comprising: a second acquisition unit, for acquiring current source-load data and current weather factor data; and, based on the association between the scene and the prediction model, selecting a target model matching the current scene from multiple prediction models; wherein the prediction model is used to generate a source-load curve of the source-load relationship based on the source-load data and the weather factor data; a prediction unit, for inputting the current source-load data and the current weather factor data into the target model, so as to extract target key feature data associated with the current scene from the current source-load data and the current weather factor data using a target modal decomposition algorithm associated with the target model, and obtain a target source-load curve characterizing the target source-load relationship based on the target key feature data; the prediction model is constructed using the virtual power plant source-load prediction model construction method in any one of the implementations of the first aspect above.

[0014] According to a fifth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which instructions are stored. When the instructions in the computer-readable storage medium are executed by a processor of a computer device, the computer device is enabled to execute a virtual power plant source-load prediction model construction method or a virtual power plant source-load curve prediction method as described in the first aspect and any possible implementation thereof.

[0015] According to the sixth aspect of an embodiment of the present application, a computer program product is provided, which includes computer instructions. When the computer instructions are executed on a computer device, the computer device executes the method for constructing a virtual power plant source and load prediction model of the above-mentioned first aspect and any possible implementation method thereof.

[0016] The technical solution provided by the embodiment of the present invention brings at least the following beneficial effects: using a combination of source load data and weather data of different scenarios as sample data, training the initial model of any scenario under different scenarios, and in the training process, based on the first initial optimization layer, being able to adaptively select the modal decomposition algorithm associated with the initial model, and using the selected modal decomposition algorithm, extracting key feature data for training the initial model under the scenario from the sample data, and based on the second initial optimization layer, using the key feature data to train the neural network model and its model parameters of the initial model, so that for the prediction sample sets of different scenarios, the optimal modal decomposition algorithm, neural network model and its model parameters corresponding to each scenario can be automatically selected, so that the prediction models of different scenarios are obtained by training the initial models of different network model structures and model parameters selected based on different key feature data screened by different decomposition algorithms, thereby improving the adaptability of the constructed source load prediction model (i.e., the initial model) to different source load scenario prediction sample sets, optimizing the source load prediction accuracy and time, and reducing the trial and error cost of manually building a neural network prediction model.

[0017] Through the above-mentioned automatic construction process of the prediction model, based on the first initial optimization layer and the second initial optimization layer of the initial model, the key feature data matching the scene requirements can be automatically selected as the model input data during the initial model training process, and the adaptive model structure and model parameters can be automatically selected during the training process, that is, the neural network parameters can be adaptively modified. Therefore, during the model training process, the input data, model structure and model parameters can be adaptively adjusted dynamically, rather than being fixed, and there is no need to manually adjust the neural network parameters for sample data of different scenes. Therefore, when the scene of the source load data changes, only the changed scene and the source load data under the changed scene are added to the sample data, and the initial model is trained to obtain a prediction model matching the changed scene and a preset modal decomposition algorithm for optimizing the input parameters input to the corresponding prediction model, so as to achieve adaptive matching of any source load data while also improving the universality, accuracy and operation efficiency of the initial model with source load prediction function.

[0018] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings herein are incorporated into the specification and constitute a part of the specification, illustrate embodiments consistent with the present application, and together with the specification are used to explain the principles of the present application, and do not constitute an improper limitation on the present application.

[0020] Figure 1 is a flow chart of a method for constructing a virtual power plant source-load prediction model according to an exemplary embodiment;

[0021] Figure 2 is a flow chart of solving a weight parameter set using an improved whale optimization algorithm according to an exemplary embodiment;

[0022] Figure 3 is a flow chart of a method for predicting source and load of a virtual power plant according to an exemplary embodiment;

[0023] Figure 4 is a block diagram of a device for constructing a virtual power plant source-load prediction model according to an exemplary embodiment;

[0024] Figure 5 is a block diagram of a virtual power plant source-load prediction device according to an exemplary embodiment;

[0025] Figure 6 is a schematic diagram of a prediction device according to an exemplary embodiment. DETAILED DESCRIPTION

[0026] In order to enable ordinary persons in the art to better understand the technical solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings.

[0027] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the attached claims.

[0028] Before introducing in detail the method for constructing a virtual power plant source-load prediction model provided in an embodiment of the present application, a brief introduction to the application scenarios involved in the embodiment of the present application is first given.

[0029] With the large-scale application of renewable energy such as solar energy and wind energy, their volatility and uncertainty have brought challenges to the power system. Therefore, improving the accuracy of source-load prediction has become a research hotspot. At present, advanced technical means such as deep learning neural networks can effectively process nonlinear and high-dimensional data, improve the stability and reliability of predictions, and improve the prediction accuracy of renewable energy and loads. However, the original source-load data has poor readability, weak time series regularity, and many influencing factors, which can easily affect the accuracy and efficiency of the source-load prediction model.

[0030] According to research, the process of constructing a source load prediction model for different scenarios is usually as follows: first observe the source load data characteristics, use the modal decomposition method to deal with data noise, and then manually adjust the neural network parameters to finally form a source load prediction model. However, the resulting source load prediction model cannot automatically adjust parameters according to changes in the prediction sample set. After changing the sample set, the neural network parameters must be manually adjusted again, and the whole process involves a lot of trial and error time costs.

[0031] In response to the above problems, the present application provides a method for constructing a source-load prediction model for a virtual power plant, which uses a combination of source-load data and weather data of different scenarios as sample data, trains the initial model of any scenario under different scenarios, and during the training process, based on the first initial optimization layer, can adaptively select the modal decomposition algorithm associated with the initial model, and use the selected modal decomposition algorithm to extract key feature data for training the initial model under the scenario from the sample data, and based on the second initial optimization layer, use the key feature data to train the neural network model and its model parameters of the initial model. In this way, for prediction sample sets of different scenarios, the optimal modal decomposition algorithm, neural network model and its model parameters corresponding to each scenario can be automatically selected, so that the prediction models of different scenarios are based on different key feature data screened by different decomposition algorithms, and the initial models of selected different network model structures and model parameters are trained, thereby improving the adaptability of the constructed source-load prediction model (i.e., the initial model) to different source-load scenario prediction sample sets, optimizing the source-load prediction accuracy and time, and reducing the trial-and-error cost of manually building a neural network prediction model.

[0032] Through the above-mentioned automatic construction process of the prediction model, based on the first initial optimization layer and the second initial optimization layer of the initial model, the key feature data matching the scene requirements can be automatically selected as the model input data during the initial model training process, and the adaptive model structure and model parameters can be automatically selected during the training process, that is, the neural network parameters can be adaptively modified. Therefore, during the model training process, the input data, model structure and model parameters can be adaptively adjusted dynamically, rather than being fixed, and there is no need to manually adjust the neural network parameters for sample data of different scenes. Therefore, when the scene of the source load data changes, only the changed scene and the source load data under the changed scene are added to the sample data, and the initial model is trained to obtain a prediction model matching the changed scene and a preset modal decomposition algorithm for optimizing the input parameters input to the corresponding prediction model, so as to achieve adaptive matching of any source load data while also improving the universality, accuracy and operation efficiency of the initial model with source load prediction function.

[0033] The method for constructing a virtual power plant source-load prediction model provided in the embodiment of the present application can be applied to a virtual power plant with a source-load prediction function or a prediction device (such as a server) used for source-load prediction in a virtual power plant.

[0034] To facilitate understanding, the virtual power plant source-load prediction model construction method provided in this application is specifically introduced below with reference to the accompanying drawings.

[0035] Figure 1 is a flowchart of a method for constructing a virtual power plant source-load prediction model according to an exemplary embodiment. Figure 1As shown, the method for constructing the virtual power plant source-load prediction model includes the following steps.

[0036] S11, taking historical source load data under multiple scenarios and corresponding historical weather factor data as sample data.

[0037] The above scenario is also called source-load scenario. Source-load data includes wind and solar power generation data and user electricity, cooling, heating and gas load data.

[0038] The multiple scenarios may include summer, winter, rainy season, non-rainy season, working days, non-working days and other source load scenarios.

[0039] S12, obtaining each initial model representing a one-to-one correspondence between the mapping relationship between the sample data and the source-load curve of each scene in the plurality of scenes.

[0040] The initial model is a neural network model.

[0041] In some embodiments, the source-load curve can also be understood as a source-load power curve (Source-Load PowerCurve). It usually refers to a curve of the power relationship between the source (generator) and the load (electrical equipment) in a power system. It is used to show the output power changes of the generator or power supply under different load conditions. This curve is of great significance to the design, optimization and operation of the power system.

[0042] The initial model includes a first initial optimization layer and a second initial optimization layer.

[0043] The first initial optimization layer is used to select a modal decomposition algorithm associated with the initial model from multiple modal decomposition algorithms, and perform a first optimization adjustment on the sample data based on the selected associated modal decomposition algorithm to extract key feature data adapted to the scene corresponding to the initial model from the sample data.

[0044] It is understandable that the first initial optimization layer is used to perform a first optimization adjustment on the modal decomposition algorithm associated with the initial model selected from multiple modal decomposition algorithms, wherein the modal decomposition algorithm is used to extract key feature data adapted to the scene corresponding to the initial model from the sample data.

[0045] The second initial optimization layer is used to perform a second optimization adjustment on the model structure and model parameters of the initial model based on the key feature data.

[0046] It can be understood that the constructed initial model is suitable for training the source-load prediction model of each scenario. That is, each initial model with one-to-one mapping relationship between the source-load curves of each scenario is the same model at the beginning, and the neural network parameters of the initial model are changed during the training process based on the first initial optimization layer and the second initial optimization layer.

[0047] S13, for any initial model among the initial models corresponding to each scene, repeatedly perform the first optimization adjustment and the second optimization adjustment according to the sample data until the corresponding first optimization condition and the second optimization condition are met, and the corresponding prediction model matching the scene corresponding to any initial model and the preset modal decomposition algorithm associated with the prediction model are obtained.

[0048] It is understandable that the initial model training corresponding to each scene in the multiple scenes, the corresponding prediction model and the preset modal decomposition algorithm associated with the prediction model can be implemented by executing step S13.

[0049] The first optimization condition is that the first adjustment times of executing the first optimization adjustment reaches the adjustment threshold and the preset modal decomposition algorithm is the modal decomposition algorithm with the shortest adjustment time of the first optimization adjustment.

[0050] The second optimization condition is that the second adjustment times of performing the second optimization adjustment reaches the adjustment threshold and the prediction model is that the source-load curve output by any initial model after the first optimization adjustment and the second optimization adjustment differs from the actual source-load curve within a preset error range.

[0051] The actual source load curve is the real source load curve of any initial model under the corresponding scenario and historical weather factors. Different scenarios correspond to different real source load curves.

[0052] The initial models corresponding to the above scenarios can be trained synchronously in parallel.

[0053] Through the above implementation, a combination of source load data and weather data of different scenarios is used as sample data to train the initial model of any scenario under different scenarios. During the training process, based on the first initial optimization layer, the modal decomposition algorithm associated with the initial model can be adaptively selected, and the selected modal decomposition algorithm is used to extract key feature data for training the initial model under the scenario from the sample data, and based on the second initial optimization layer, the key feature data is used to train the neural network model and its model parameters of the initial model. In this way, for prediction sample sets of different scenarios, the optimal modal decomposition algorithm, neural network model and its model parameters corresponding to each scenario can be automatically selected, so that the prediction models of different scenarios are obtained by training the initial models of different network model structures and model parameters selected based on different key feature data screened by different decomposition algorithms, thereby improving the adaptability of the constructed source load prediction model (i.e., the initial model) to prediction sample sets of different source load scenarios, optimizing the source load prediction accuracy and time, and reducing the trial and error cost of manually building a neural network prediction model.

[0054] Through the above-mentioned automatic construction process of the prediction model, based on the first initial optimization layer and the second initial optimization layer of the initial model, the key feature data matching the scene requirements can be automatically selected as the model input data during the initial model training process, and the adaptive model structure and model parameters can be automatically selected during the training process, that is, the neural network parameters can be adaptively modified. Therefore, during the model training process, the input data, model structure and model parameters can be adaptively adjusted dynamically, rather than being fixed, and there is no need to manually adjust the neural network parameters for sample data of different scenes. Therefore, when the scene of the source load data changes, only the changed scene and the source load data under the changed scene are added to the sample data, and the initial model is trained to obtain a prediction model matching the changed scene and a preset modal decomposition algorithm for optimizing the input parameters input to the corresponding prediction model, so as to achieve adaptive matching of any source load data while also improving the universality, accuracy and operation efficiency of the initial model with source load prediction function.

[0055] As an implementation method, in order to facilitate the selected modal decomposition algorithm to decompose the historical source load data in the sample data and speed up the decomposition speed, the historical source load data is fitted to obtain a fitting curve, and the fitting curve and historical weather factor data are used as sample data.

[0056] In one implementation, the specific process of performing the first optimization adjustment is as follows.

[0057] The fitting curve of the historical source-load data is decomposed into multiple subsequences by using the selected associated modal decomposition algorithm. Based on the weights of the multiple subsequences, the noise sequences in the multiple subsequences are removed to obtain multiple useful sequences; and based on the weights of the weather factor characteristics of the scenes corresponding to the initial model, the historical weather characteristic factors associated with the scenes corresponding to the initial model are respectively selected from the historical weather factor data. The historical weather characteristic factors and the multiple useful sequences are used as key feature data, or the historical weather characteristic factors and the multiple useful sequences are normalized to obtain the key feature data.

[0058] The modal decomposition algorithm includes one or more of empirical mode decomposition, ensemble empirical mode decomposition, complementary ensemble empirical mode decomposition, and fully adaptive noise ensemble empirical mode decomposition. At the same time, a binary weight is set for the modal decomposition method. When the binary weight is 0, it means that the modal decomposition method is not used as a decomposition method for the source-load curve; when the binary weight is 1, it means that the modal decomposition method is used as a decomposition method for the source-load curve. The sum of the binary weights is 1.

[0059] Optionally, the number of iterations and the residual threshold are used as parameter variables in the mode decomposition algorithm that are optimized by the algorithm.

[0060] Optionally, singular spectrum analysis is used to embed, decompose, group, and reconstruct the decomposed subsequences, and the subsequences after modal decomposition are divided into useful sequences and noise. The specific steps of using singular spectrum analysis are as follows.

[0061] (1) Embedding. The analysis object of singular spectrum analysis is a finite-length one-dimensional time series [x1, x2, …, x N , where N is the sequence length. It is necessary to select an appropriate window length L. Usually, L < N / 2 is taken. Let K = N - L + 1, then the trajectory matrix X can be constructed, as shown in the following formula (1).

[0062]

[0063] (2) Decomposition. Perform singular value decomposition on the obtained trajectory matrix X to obtain the eigenvalues (i.e., singular values) λ i and eigenvectors U i which are respectively expressed as the following formula (2).

[0064] λ i , (i = 1, 2, …, L)

[0065] U i , (i = 1, 2, …, L) Formula (2).

[0066] (3) Grouping and reconstruction. Since grouping and reconstruction are highly computationally related, they are described together.

[0067] First, calculate the lag sequence X i 's projection on U m through the following formula (3).

[0068]

[0069] Among them, X i represents the i-th column of the trajectory matrix X, is the weight of the time evolution type reflected by X i in the x i+1 , x i+2 , …, x i+L time period of the original sequence, and is called the time principal component.

[0070] Next, reconstruction is performed through the time empirical orthogonal function and the time principal component. The specific reconstruction process is as follows formula (4).

[0071]

[0072] It can be obtained therefrom that the sum of all reconstructed sequences should be equal to the original sequence, as shown in the following formula (5).

[0073]

[0074] The useful sequence and weather factor data are normalized using the following formula (6).

[0075]

[0076] Among them, f represents the parameter value to be normalized, min f and max f They are the minimum and maximum values ​​of the parameters on all samples, and the normalized parameter value is f new .

[0077] The above weather factors characteristics include temperature, humidity, wind speed, air pressure and precipitation.

[0078] Exemplarily, a binary weight is set for the historical weather factor data. If the binary weight is 0, it means that the historical weather factor data is not used; if the binary weight is 1, it means that the historical weather factor data is used. The weights of the historical weather factors are all initialized to 1.

[0079] The specific process of the above-selected associated modal decomposition algorithm is as follows. A binary weight is set for each modal decomposition algorithm. When the binary weight is 0, it means that the modal decomposition algorithm is not used as a decomposition method; when the binary weight is 1, it means that the modal decomposition method is used as a decomposition method. The sum of the binary weights is 1. The number of iterations and the residual threshold are used as the weight variables optimized by the algorithm in the modal decomposition method.

[0080] Optionally, the specific process of determining the key feature data in the above-mentioned first optimization adjustment process is as follows. Call the improved whale optimization algorithm to determine the weights of the weather factor characteristics of the scene corresponding to the initial model and the weights of multiple subsequences. Among them, the improved whale optimization algorithm is optimized by using the chaos operator, Levy flight strategy, and "observer" strategy.

[0081] In one embodiment, performing the second optimization adjustment specifically includes: calling the improved whale optimization algorithm to respectively determine the weight of the model structure of the initial model and the weight of the model parameters of the initial model; based on the weight of the model structure of the initial model, determining the model structure of the initial model in the second optimization adjustment process; and, based on the weight of the model parameters of the initial model, determining the model parameters of the initial model in the second optimization adjustment process.

[0082] In one embodiment, the model structure of the initial model includes one or more optimized features of a convolutional layer, a pooling layer, a fully connected layer, an attention mechanism, and a network structure of a basic long short-term memory neural network or a bidirectional long short-term memory neural network.

[0083] The model parameters of the above initial model include hyperparameters in the optimization features. Exemplarily, the model parameters of the initial model include hyperparameters such as the size of the convolution kernel, step size, zero padding, number of filters, pooling step size, filter size, pooling type, weights between the input layer and the hidden layer, and the number of neurons in each layer.

[0084] The model structure optimization process of the initial model can be understood as a combined selection process of selecting optimization features and selecting a network structure of a long short-term memory neural network.

[0085] Specifically, a selection parameter is set for the convolution layer and the pooling layer, and the selection parameter is the number of convolution layers and the pooling layer. A binary weight is set for the fully connected layer and the attention mechanism. When the binary weight is 0, it means that the initial model does not use this optimization feature, and when the binary weight is 1, it means that the initial model uses this optimization feature.

[0086] Based on the weights involved in the above implementation methods, the whale optimization algorithm can be used to solve the weight parameter set. The specific process of using the improved whale optimization algorithm to solve the weight parameter set is as follows: Figure 2 shown.

[0087] It can be understood that useful sequences, subsequences and historical source load data related to various source load data, various historical weather factor data, model structure, model parameters, optimization features and hyperparameters related to model structure and model parameters are all called whales, and the specific weight parameter sets of their parameters are as follows.

[0088] S21, initialize algorithm parameters.

[0089] Among them, the algorithm parameters include whale location, population size, number of iterations, upper and lower limits of weights, modal decomposition algorithm parameters of the first initial optimization layer and weather factor selection variables, neural network parameters of the second initial optimization layer, etc.

[0090] The first initial optimized layer is also referred to as an outer layer, and the second initial optimized layer is also referred to as an inner layer.

[0091] S22, using chaos operator to reset the position of whale population.

[0092] Its whale population position is the position of the population in the variable space.

[0093] S23, calculate the fitness of each whale, find the position of the current optimal whale and retain it.

[0094] S24, updating the modal decomposition algorithm parameter a, parameter probability p and coefficient vectors A1 and C1 corresponding to adjacent step sizes.

[0095] S25, determine whether the probability p is less than 50%, if yes, directly transfer to S26, otherwise transfer to S29.

[0096] S26, determine whether the absolute value of the coefficient vector A1 is less than 1, if so, proceed to S27; otherwise, proceed to S28.

[0097] S27, surround the prey, use the hunting mechanism to update the position, and go to S30.

[0098] S28, global random search for prey, use Levy flight strategy to update the position, and go to S30.

[0099] S29, uses the bubble net predation mechanism for position update.

[0100] S30, calculating the population fitness value and updating the optimal whale position.

[0101] Specifically, after the position update is completed, the fitness of each whale is calculated and compared with the position of the previously retained optimal whale. If it is better than the previous position, it is replaced with the new optimal solution.

[0102] S31, using the “observer” strategy, randomly observes the positions around the whale population, re-ranks the fitness of the current population, and updates the optimal whale position.

[0103] S32, determine whether the current calculation reaches the maximum number of iterations, if yes, proceed to step S33, otherwise proceed to the next iteration and return to S24.

[0104] S33, output the current optimal whale position and build the optimal source-load combination neural network prediction model.

[0105] The chaos operator adopts the improved tent chaotic map. By adding a disturbance point r to the original tent map to jump out of the periodic rand cycle, the tent map returns to chaos. The mathematical model is shown in formula (7) and formula (8).

[0106]

[0107] X new =X+x change Formula (8).

[0108] Among them, X new represents the updated position of the whale population, X represents the position of the whale population, and x change Indicates the change in the location of the whale population.

[0109] The above Levy flight strategy can enhance the global search capability of the algorithm. The mathematical model is as follows: formula (10) to formula (12).

[0110] x i,new =x i +A1×0.01×step⊕(xi -C1×x rand ) formula (10).

[0111]

[0112] C1=2rand(0,1) formula (12).

[0113] In the above formula: x i,new For whale individual x i The updated position, n is the current iteration number, N is the maximum iteration number, ⊕ represents the dot product, x rand represents the position of a randomly selected whale individual in the current population, the coefficient vectors A1 and C1 corresponding to adjacent steps, and step represents the random step. Formula (13) generated by the Mantegna method is as follows.

[0114]

[0115] where u(u1,u2,…,u d ) and v(v1,v2,…,v d ) is a d-dimensional space vector, and β is a constant, which can be β=1.5. Each component of u and v obeys the following normal distribution.

[0116]

[0117] Among them, r() is the input function of different variables.

[0118] The “observer” strategy is to randomly observe the surrounding positions after each whale population completes its position update. The mathematical model of the random observation position is as follows: Formula (16).

[0119]

[0120] X new =X+x change Formula (17).

[0121] As an optional implementation, the inner neural network prediction model takes prediction accuracy as the objective function, which is specifically expressed by using the entropy weight method to calculate the mean absolute error E MAE , Root mean square error E RMSE , Square absolute percentage error E MAPE , The coefficient is weighted f in , the mathematical model is as follows.

[0122] f in =δ1E MAE +δ2E RMSE +δ3E MAPE+δ4E R2 Formula (18).

[0123]

[0124] Where: δ1, δ2, δ3 and δ4 are the weights of each target determined by the entropy weight method, y t represents the true value of the source load data at time t, represents the predicted value of the neural network prediction model at time t, n represents the number of predicted samples, Represents the average value of the true value of the source load data.

[0125] As an optional implementation, the outer neural network prediction model, that is, the first initial optimization layer, takes the prediction time as the objective function, which is specifically expressed as the following formula.

[0126] f out =min(t end -t start ) formula (23).

[0127] Where: f2 is the minimum fault diagnosis time, t end is the end time of the prediction model, t start is the start time of the prediction model, t start and t end All can be obtained through the timing function.

[0128] Through the above implementation, the modal decomposition process of source load data is added to the neural network prediction model, useful sequences and weather influencing factors in the data are selected to form a comprehensive sample set, and the algorithm is used to optimize the parameters of the modal decomposition method and the neural network model, thereby improving the prediction accuracy and reducing the prediction time. This solves the problem that the previous technology requires manual selection of modal decomposition methods and neural network models, and requires a lot of time to debug the method model parameters. According to the characteristics of the source load data, the appropriate modal decomposition method and neural network model are automatically selected to form a combined source load prediction model, which reduces labor costs and trial and error time costs, and improves the efficiency of source load data prediction.

[0129] Based on the construction of a virtual power plant source load prediction model in the above implementation mode, that is, a prediction model, the present application also proposes a virtual power plant source load prediction method, the method comprising: Figure 3 Steps shown.

[0130] S301, obtaining current source load data and current weather factor data; and, based on the association between the scene and the prediction model, selecting a target model that matches the current scene from multiple prediction models.

[0131] Among them, the prediction model is used to generate a source-load curve of the source-load relationship based on the source-load data and weather factor data.

[0132] S302, input the current source load data and the current weather factor data into the target model, and use the target modal decomposition algorithm associated with the target model to extract the target key feature data associated with the current scene in the current source load data and the current weather factor data, and based on the target key feature data, obtain the target source load curve that characterizes the target source load relationship.

[0133] In order to achieve the above functions, the virtual power plant source-load prediction model construction device includes hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art should easily realize that, in combination with the algorithm steps of each example described in the embodiments disclosed herein, the present application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in the form of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0134] The present application also provides a method Figure 4 The virtual power plant source-load prediction model construction device shown includes: a first acquisition unit 401 and a construction unit 402.

[0135] The first acquisition unit 401 is used to use historical source load data in multiple scenarios and corresponding historical weather factor data as sample data.

[0136] The first acquisition unit 401 is further used to acquire each initial model representing a one-to-one correspondence between the mapping relationship between the sample data and the source-load curve of each scene in the multiple scenes.

[0137] The initial model includes a first initial optimization layer and a second initial optimization layer; the first initial optimization layer is used to select a modal decomposition algorithm associated with the initial model from multiple modal decomposition algorithms, and perform a first optimization adjustment on the sample data based on the selected associated modal decomposition algorithm, so as to extract key feature data adapted to the scene corresponding to the initial model from the sample data; the second initial optimization layer is used to perform a second optimization adjustment on the model structure and model parameters of the initial model based on the key feature data.

[0138] The construction unit 402 is used to repeatedly perform the first optimization adjustment and the second optimization adjustment for any initial model among the initial models corresponding to each scene according to the sample data until the corresponding first optimization condition and the second optimization condition are met, thereby obtaining a prediction model matching the scene corresponding to any initial model and a preset modal decomposition algorithm associated with the prediction model.

[0139] In one embodiment, the first optimization condition is that the first adjustment times of executing the first optimization adjustment reaches the adjustment threshold and the preset modal decomposition algorithm is the modal decomposition algorithm with the shortest adjustment time for the first optimization adjustment; the second optimization condition is that the second adjustment times of executing the second optimization adjustment reaches the adjustment threshold and the prediction model is that the source load curve output by any initial model after the first optimization adjustment and the second optimization adjustment is within a preset error range from the actual source load curve; the actual source load curve is the real source load curve of any initial model under the corresponding scenario and historical weather factors.

[0140] In another implementation manner, the first acquisition unit 401 is specifically used to: fit the historical source load data to obtain a fitting curve, and use the fitting curve and the historical weather factor data as sample data.

[0141] In another implementation, the construction unit 402 is specifically used to: use the selected associated modal decomposition algorithm to decompose the fitting curve of the historical source load data into multiple subsequences; based on the weights of the multiple subsequences, remove the noise sequences in the multiple subsequences to obtain multiple useful sequences; and, based on the weights of the weather factor characteristics of the scene corresponding to the initial model, select the historical weather characteristic factors associated with the scene corresponding to the initial model from the historical weather factor data; the weather factor characteristics include temperature, humidity, wind speed, air pressure, and precipitation; use the historical weather characteristic factors and multiple useful sequences as key feature data, or normalize the historical weather characteristic factors and multiple useful sequences to obtain key feature data.

[0142] In another implementation, based on the weights of multiple subsequences, noise sequences are removed from multiple subsequences to obtain multiple useful sequences; and based on the weights of the weather factor characteristics of the scene corresponding to the initial model, before selecting historical weather characteristic factors associated with the scene corresponding to the initial model from the historical weather factor data, the method also includes: calling the improved whale optimization algorithm to respectively determine the weights of the weather factor characteristics of the scene corresponding to the initial model and the weights of multiple subsequences; the improved whale optimization algorithm is optimized using a chaos operator, a Levy flight strategy, and an "observer" strategy.

[0143] In another implementation, the initial model is a neural network model, and the model structure of the initial model includes one or more of a convolutional layer, a pooling layer, a fully connected layer, an attention mechanism, and a basic long short-term memory neural network or a bidirectional long short-term memory neural network; the model parameters of the initial model include the size of the convolution kernel, the step size, the zero padding, the number of filters, the pooling step size, the filter size, the pooling type, the weight between the input layer and the hidden layer, and the number of neurons in each layer.

[0144] In another implementation, the construction unit 402 is specifically used to: call the improved whale optimization algorithm to respectively determine the weight of the model structure of the initial model and the weight of the model parameters of the initial model; determine the model structure of the initial model in the second optimization and adjustment process based on the weight of the model structure of the initial model; and determine the model parameters of the initial model in the second optimization and adjustment process based on the weight of the model parameters of the initial model.

[0145] The present application also provides a method Figure 5 The virtual power plant source-load prediction device shown includes: a second acquisition unit 501 and a prediction unit 502.

[0146] The second acquisition unit 501 is used to acquire current source load data and current weather factor data; and, based on the association between the scene and the prediction model, select a target model that matches the current scene from multiple prediction models.

[0147] Among them, the prediction model is used to generate a source-load curve of the source-load relationship based on the source-load data and weather factor data.

[0148] The prediction unit 502 is used to input the current source load data and the current weather factor data into the target model, and to extract the target key feature data associated with the current scene in the current source load data and the current weather factor data using a target modal decomposition algorithm associated with the target model, and to obtain a target source load curve characterizing the target source load relationship based on the target key feature data; the prediction model is constructed using the virtual power plant source load prediction model construction method in any one of the implementation methods of the first aspect above.

[0149] Regarding the device in the above embodiment, the specific manner in which each unit module performs the operation has been described in detail in the embodiment of the method, and will not be elaborated here.

[0150] Figure 6 is a schematic diagram of a prediction device provided by this application. Figure 6 The prediction device 60 may include at least one processor 601 and a memory 603 for storing processor executable instructions. The processor 601 is configured to execute instructions in the memory 603 to implement the virtual power plant source-load prediction model construction method in the following embodiment.

[0151] In addition, the prediction device 60 may further include a communication bus 602 , at least one communication interface 604 , an input device 606 , and an output device 605 .

[0152] The processor 601 may be a central processing unit (CPU), a microprocessing unit, an ASIC, or one or more integrated circuits for controlling the execution of the program of the present application.

[0153] Communication bus 602 may include a pathway for transmitting information between the above-mentioned components.

[0154] The communication interface 604 uses any transceiver or other device for communicating with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0155] The input device 606 is used to receive input signals and the output device 605 is used to output signals.

[0156] The memory 603 may be a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM) or other types of dynamic storage devices that can store information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical disc, laser disc, optical disc, digital versatile disc, Blu-ray disc, etc.), a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store the desired program code in the form of an instruction or data structure and can be accessed by a computer, but is not limited thereto. The memory may exist independently and be connected to the processing unit through a bus. The memory may also be integrated with the processing unit.

[0157] The memory 603 is used to store instructions for executing the solution of the present application, and the execution is controlled by the processor 601. The processor 601 is used to execute the instructions stored in the memory 603, so as to realize the functions in the method of the present application.

[0158] In a specific implementation, as an embodiment, the processor 601 may include one or more CPUs, such as Figure 6 CPU0 and CPU1 in.

[0159] In a specific implementation, as an embodiment, the prediction device 60 may include multiple processors, such as Figure 6601 and processor 607 in FIG. Each of these processors may be a single-CPU processor or a multi-CPU processor. A processor herein may refer to one or more devices, circuits, and / or processing cores for processing data (e.g., computer program instructions).

[0160] The prediction device Figure 6 The system includes: a processor 601 and a memory 603 for storing executable instructions of the processor 601; wherein the processor 601 is configured to execute the executable instructions to implement a method for constructing a virtual power plant source-load prediction model according to any of the above possible implementations. The same technical effect can be achieved, and to avoid repetition, it will not be described here.

[0161] The embodiment of the present application also provides a computer-readable storage medium, when the instructions in the computer-readable storage medium are executed by the processor of the control device or control equipment, the control device or control equipment can execute the method for constructing a virtual power plant source-load prediction model as described in any possible implementation method described above. And the same technical effect can be achieved, so it will not be repeated here to avoid repetition.

[0162] The embodiment of the present application also provides a computer program product, including a computer program or instruction, which is executed by a processor to construct a virtual power plant source-load prediction model according to any of the above possible implementations. And the same technical effect can be achieved, so it will not be repeated here to avoid repetition.

[0163] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the invention disclosed herein. The present application is intended to cover any modification, use or adaptation of the present application, which follows the general principles of the present application and includes common knowledge or customary techniques in the art that are not disclosed in the present application. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present application are indicated by the following claims.

[0164] It should be understood that the present application is not limited to the precise structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims

1. A method for constructing a virtual power plant source-load prediction model, characterized in that: The method comprises: The historical source load data under multiple scenarios and the corresponding historical weather factor data are used as sample data; Acquire each initial model that represents the mapping relationship between the sample data and the source-load curve of each scene in the multiple scenes in a one-to-one correspondence; the initial model includes a first initial optimization layer and a second initial optimization layer; the first initial optimization layer is used to select a modal decomposition algorithm associated with the initial model from multiple modal decomposition algorithms, and perform a first optimization adjustment on the sample data based on the selected associated modal decomposition algorithm, so as to extract key feature data adapted to the scene corresponding to the initial model from the sample data; the second initial optimization layer is used to perform a second optimization adjustment on the model structure and model parameters of the initial model based on the key feature data; For any of the initial models corresponding to the various scenes, the first optimization adjustment and the second optimization adjustment are repeatedly performed according to the sample data until the corresponding first optimization condition and the second optimization condition are met, and a prediction model matching the scene corresponding to any of the initial models and a preset modal decomposition algorithm associated with the prediction model are obtained.

2. The method for constructing a virtual power plant source-load prediction model according to claim 1, characterized in that: The prediction model is used to generate a source-load curve of the source-load relationship based on source-load data and weather factor data; the first optimization condition is that the first adjustment times of executing the first optimization adjustment reaches the adjustment threshold and the preset modal decomposition algorithm is the modal decomposition algorithm with the shortest adjustment time for the first optimization adjustment; the second optimization condition is that the second adjustment times of executing the second optimization adjustment reaches the adjustment threshold and the prediction model is that the source load curve output by any initial model after the first optimization adjustment and the second optimization adjustment differs from the actual source load curve within a preset error range; the actual source load curve is the real source load curve of any initial model under the corresponding scenario and historical weather factors.

3. The method for constructing a virtual power plant source-load prediction model according to claim 1, characterized in that: The historical source load data and corresponding historical weather factor data in multiple scenarios are used as sample data, including: The historical source load data is fitted to obtain a fitting curve, and the fitting curve and the historical weather factor data are used as the sample data.

4. The method for constructing a virtual power plant source-load prediction model according to claim 3, characterized in that: The performing the first optimization adjustment includes: Decomposing the fitting curve of the historical source-load data into a plurality of subsequences by using the selected associated modal decomposition algorithm; removing noise sequences from the plurality of subsequences based on weights of the plurality of subsequences to obtain a plurality of useful sequences; And, based on the weights of the weather factor characteristics of the scene corresponding to the initial model, historical weather characteristic factors associated with the scene corresponding to the initial model are respectively selected from the historical weather factor data; the weather factor characteristics include temperature, humidity, wind speed, air pressure, and precipitation; The historical weather characteristic factors and the multiple useful sequences are used as the key characteristic data, or the historical weather characteristic factors and the multiple useful sequences are normalized to obtain the key characteristic data.

5. The method for constructing a virtual power plant source-load prediction model according to claim 4, characterized in that: Before removing noise sequences from the multiple subsequences based on the weights of the multiple subsequences to obtain multiple useful sequences; and selecting historical weather characteristic factors associated with the scene corresponding to the initial model from the historical weather factor data based on the weights of the weather factor characteristics of the scene corresponding to the initial model, the method further includes: The improved whale optimization algorithm is called to determine the weights of the weather factor characteristics of the scene corresponding to the initial model and the weights of the multiple subsequences respectively; the improved whale optimization algorithm is optimized by using a chaotic operator, a Levy flight strategy, and an "observer" strategy.

6. The method for constructing a virtual power plant source-load prediction model according to any one of claims 1 to 5, characterized in that: The initial model is a neural network model, and the model structure of the initial model includes one or more of a convolutional layer, a pooling layer, a fully connected layer, an attention mechanism, and a basic long short-term memory neural network or a bidirectional long short-term memory neural network; the model parameters of the initial model include the size of the convolution kernel, the step size, the zero padding, the number of filters, the pooling step size, the filter size, the pooling type, the weight between the input layer and the hidden layer, and the number of neurons in each layer.

7. The method for constructing a virtual power plant source-load prediction model according to claim 6, characterized in that: The performing of the second optimization adjustment specifically includes: Calling the improved whale optimization algorithm to respectively determine the weight of the model structure of the initial model and the weight of the model parameters of the initial model; Based on the weight of the model structure of the initial model, the model structure of the initial model in the second optimization and adjustment process is determined; and based on the weight of the model parameters of the initial model, the model parameters of the initial model in the second optimization and adjustment process are determined.

8. A virtual power plant source load prediction method, characterized in that: The method comprises: Acquire current source-load data and current weather factor data; and, based on the association between the scene and the prediction model, select a target model that matches the current scene from a plurality of the prediction models; wherein the prediction model is used to generate a source-load curve of a source-load relationship based on the source-load data and the weather factor data; The current source load data and the current weather factor data are input into the target model, and the target modal decomposition algorithm associated with the target model is used to extract the target key feature data associated with the current scene in the current source load data and the current weather factor data, and based on the target key feature data, a target source load curve characterizing the target source-load relationship is obtained; the prediction model is constructed using the virtual power plant source load prediction model construction method according to any one of claims 1 to 7.

9. A virtual power plant source-load prediction model construction device, characterized in that: The device comprises: A first acquisition unit is used to use historical source load data and corresponding historical weather factor data in multiple scenarios as sample data; The first acquisition unit is further used to acquire each initial model that represents the mapping relationship between the sample data and the source-load curve of each scene in the multiple scenes in a one-to-one correspondence; the initial model includes a first initial optimization layer and a second initial optimization layer; the first initial optimization layer is used to select a modal decomposition algorithm associated with the initial model from multiple modal decomposition algorithms, and perform a first optimization adjustment on the sample data based on the selected associated modal decomposition algorithm, so as to extract key feature data adapted to the scene corresponding to the initial model from the sample data; the second initial optimization layer is used to perform a second optimization adjustment on the model structure and model parameters of the initial model based on the key feature data; A construction unit is used to repeatedly perform the first optimization adjustment and the second optimization adjustment for any initial model among the initial models corresponding to the various scenarios according to the sample data until the corresponding first optimization condition and the second optimization condition are met, thereby obtaining a prediction model matching the scenario corresponding to any initial model and a preset modal decomposition algorithm associated with the prediction model.

10. A virtual power plant source load prediction device, characterized in that: The device comprises: A second acquisition unit is used to acquire current source-load data and current weather factor data; and, based on the association between the scene and the prediction model, select a target model that matches the current scene from a plurality of the prediction models; wherein the prediction model is used to generate a source-load curve of a source-load relationship based on the source-load data and the weather factor data; A prediction unit, used for inputting the current source load data and the current weather factor data into the target model, so as to extract the target key feature data associated with the current scene from the current source load data and the current weather factor data using a target modal decomposition algorithm associated with the target model, and obtain a target source load curve characterizing the target source load relationship based on the target key feature data; the prediction model is constructed using the virtual power plant source load prediction model construction method according to any one of claims 1 to 7.

Images