Explanatable net load power probability prediction method and system

Abstract

The invention discloses an interpretable net load power probability prediction method and system. The method comprises the following steps: acquiring first net load data and first meteorological data of a distributed photovoltaic power generation system within a first preset time range; constructing a power probability prediction model based on a Kolmogorov-Arnod network, and inputting the first net load data and the first meteorological data into the power probability prediction model for parameter estimation to obtain a net load power mean value and a net load power variance of the distributed photovoltaic power generation system at each time point within a second preset time range; calculating a net load power probability density function of the distributed photovoltaic power generation system at each time point in a second preset time range based on the net load power mean value and the net load power variance; and explaining a net load power probability prediction result of the distributed photovoltaic power generation system at each time point in a second preset time range based on the net load power probability density function. According to the method, the interpretability of probability prediction model parameters is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation, and in particular to an interpretable net load power probability prediction method and system. Background Art

[0002] Photovoltaic power generation, as the main energy source for achieving energy transformation, has received widespread attention. In recent years, distributed photovoltaics have developed rapidly. However, these distributed photovoltaics are usually installed behind the user-side electricity meter. Under the combined influence of intermittent photovoltaic power generation and unstable user load demand, the net load power presents a "duck curve" shape, which injects strong uncertainty into the operation of the power system and poses a great challenge to the balance of supply and demand of the power grid. Probabilistic prediction, as a method of quantifying prediction uncertainty, can obtain the probability distribution of net load power and provide decision makers with more comprehensive prediction information. It can be applied to scenarios such as power system optimization and scheduling, and is gradually becoming an important part of modern power grid operation and management.

[0003] The complexity and depth of neural networks have enabled them to achieve significant development in the field of probability prediction, such as deep learning models based on MLP (Multi Layer Perceptron). However, deep learning models based on MLP face problems such as large demand for training data sets and poor interpretability, and their credibility in power system optimization and scheduling problems is usually low. In order to solve the problem of poor interpretability of deep learning models, existing technologies mainly include attention mechanisms, physical information neural networks, and sensitivity analysis. However, since the model parameters are usually large and the internal calculation process is too complicated, the above methods cannot change the essence of the poor interpretability of deep learning models. Summary of the invention

[0004] In order to overcome the problem of poor interpretability of the prior art in the field of net load power probability prediction, the present invention provides an interpretable net load power probability prediction method and system.

[0005] In a first aspect, an embodiment of the present invention provides an interpretable net load power probability prediction method, comprising: Acquire first net load data and first meteorological data of the distributed photovoltaic power generation system within a first preset time range, wherein the first meteorological data includes a number of meteorological characteristic data, and the meteorological characteristic data includes solar irradiance data, cloud cover data and temperature data; Building a power probability prediction model based on the Kolmogorov-Arnold network, and inputting the first net load data and the first meteorological data into the power probability prediction model for parameter estimation, so as to obtain the net load power mean and net load power variance of the distributed photovoltaic power generation system at each time point within the second preset time range; Calculating a net load power probability density function of the distributed photovoltaic power generation system at each time point within the second preset time range based on the net load power mean and the net load power variance; The net load power probability prediction result of the distributed photovoltaic power generation system at each time point within the second preset time range is explained based on the net load power probability density function.

[0006] Preferably, the constructing of a power probability prediction model based on a Kolmogorov-Arnold network comprises: Based on the initialization of the Kolmogorov-Arnold network, a first KAN model and a second KAN model are respectively constructed, and the first KAN model and the second KAN model are superimposed to obtain an initial power probability prediction model; Acquire second net load data and second meteorological data of the distributed photovoltaic power generation system within a third preset time range; A training set is constructed based on the second net load data and the second meteorological data, and the initial power probability prediction model is trained according to the training set to obtain a power probability prediction model, wherein the power probability prediction model is configured to perform parameter estimation through the first KAN model based on the first net load data and the first meteorological data to obtain the net load power mean of the distributed photovoltaic power generation system at each time point within the second preset time range, and is configured to perform parameter estimation through the second KAN model based on the first net load data and the first meteorological data to obtain the net load power variance of the distributed photovoltaic power generation system at each time point within the second preset time range.

[0007] Preferably, the initialization of the Kolmogorov-Arnold network is based on respectively constructing a first KAN model and a second KAN model, and superimposing the first KAN model and the second KAN model to obtain an initial power probability prediction model, including: configuring the first Kolmogorov-Arnold network to obtain an initial first KAN model, wherein the configuration includes a structure configuration, an activation function configuration, and a trainable parameter configuration; Regularizing the initial first KAN model to obtain a first KAN model; configuring the second Kolmogorov-Arnold network to obtain an initial second KAN model, wherein the configuration includes a structure configuration, an activation function configuration, and a trainable parameter configuration; Regularizing the initial second KAN model to obtain a second KAN model; The first KAN model and the second KAN model are superimposed in parallel to obtain an initial power probability prediction model.

[0008] Preferably, the regularization includes L1 regularization and entropy regularization.

[0009] Preferably, constructing a training set based on the second net load data and the second meteorological data, and training the initial power probability prediction model according to the training set to obtain a power probability prediction model includes: The second meteorological data is filtered based on the second net load data, and the filtered second meteorological data and the second net load data are reorganized by dimensionality reduction to obtain first one-dimensional vector data; Acquire actual net load power data corresponding to the first one-dimensional vector data, so that the first one-dimensional vector data and the actual net load power data constitute a training set; The initial power probability prediction model is supervisedly trained based on the training set to obtain a power probability prediction model, wherein the initial power probability prediction model constructs a loss function based on negative log-likelihood.

[0010] Preferably, the filtering the second meteorological data based on the second net load data includes: The maximum information coefficient between each meteorological characteristic data in the second meteorological data and the second net load data is calculated, and the meteorological characteristic data whose maximum information coefficient is higher than a preset threshold is selected.

[0011] Preferably, the performing supervised training on the initial power probability prediction model based on the training set to obtain the power probability prediction model includes: Based on the steps of N iterative training, a power probability prediction model is obtained, wherein N is an integer ≥ 1; The steps of each iterative training include: Inputting the training set into the initial power probability prediction model for parameter estimation to obtain net load power prediction data; Calculating the loss function based on the net load power prediction data and the net load power actual data; The trainable parameters of the initial power probability prediction model are adjusted by maximum likelihood estimation based on the loss function.

[0012] Preferably, before inputting the first net load data and the first meteorological data into the power probability prediction model for parameter estimation, the method further includes: The first meteorological data is filtered based on the first net load data, and the filtered first meteorological data and the first net load data are reorganized by dimensionality reduction to obtain second one-dimensional vector data.

[0013] Preferably, explaining the net load power probability prediction result of the distributed photovoltaic power generation system at each time point within the second preset time range based on the net load power probability density function includes: Based on the net load power probability density function, the net load power values ​​obtained by random sampling are used to explain the net load power probability prediction result of the distributed photovoltaic power generation system at each time point within the second preset time range.

[0014] In a second aspect, an embodiment of the present invention provides an interpretable net load power probability prediction system, comprising: A historical data acquisition module, used to acquire first net load data and first meteorological data of a distributed photovoltaic power generation system within a first preset time range, wherein the first meteorological data includes a number of meteorological characteristic data, and the meteorological characteristic data includes solar irradiance data, cloud cover data and temperature data; a parameter estimation module, configured to construct a power probability prediction model based on a Kolmogorov-Arnold network, and input the first net load data and the first meteorological data into the power probability prediction model for parameter estimation, so as to obtain a net load power mean and a net load power variance of the distributed photovoltaic power generation system at each time point within a second preset time range; A probability density function determination module, configured to calculate a net load power probability density function of the distributed photovoltaic power generation system at each time point within the second preset time range based on the net load power mean and the net load power variance; A prediction result interpretation module is used to interpret the net load power probability prediction result of the distributed photovoltaic power generation system at each time point within the second preset time range based on the net load power probability density function.

[0015] Compared with the prior art, the method and system for interpretable net load power probability prediction of the embodiment of the present invention have the following beneficial effects: an initial power probability prediction model based on the Kolmogorov-Arnold network is constructed, and the initial power probability prediction model is trained to obtain a power probability prediction model, which can effectively improve the interpretability of model parameters and improve the accuracy of net load power probability prediction; the initial power probability prediction model is trained with maximum likelihood estimation as the objective function, which can achieve accurate fitting of the net load power probability distribution, perform feature screening based on the maximum information coefficient, and enhance the mapping ability of the power probability prediction model to various meteorological characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart of a method for predicting net load power probability that can be explained according to an embodiment of the present invention; Figure 2 It is a schematic diagram of a process of constructing a power probability prediction model according to an embodiment of the present invention; Figure 3 is a schematic diagram of the structure of a two-layer KAN model according to an embodiment of the present invention; Figure 4 It is a schematic diagram of the structure of the power probability prediction model according to an embodiment of the present invention; Figure 5 It is a structural schematic diagram of an interpretable net load power probability prediction system according to an embodiment of the present invention; Reference numerals: , the set of functions in the first KAN layer; , the set of functions in the second KAN layer. DETAILED DESCRIPTION

[0017] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0018] In the description of the present invention, it should be understood that the terms "first" and "second" etc. are used in the present invention to distinguish different objects rather than to describe a specific order.

[0019] In the description of the present invention, it should be noted that, unless otherwise defined, all technical and scientific terms used in the present invention have the same meaning as those commonly understood by those skilled in the art. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood by specific circumstances.

[0020] like Figure 1 As shown, an embodiment of the present invention provides an interpretable net load power probability prediction method, comprising the steps of: S1. Acquire first net load data and first meteorological data of a distributed photovoltaic power generation system within a first preset time range; The first preset time range is a historical period. That is to say, the first net load data and the first meteorological data are both historical data. The first net load data reflects the difference between the user load demand and the distributed photovoltaic power generation supply at different times in the past, and the first meteorological data will affect the distributed photovoltaic power generation, thereby affecting the net load. Specifically, the first meteorological data includes a number of meteorological characteristic data, and the meteorological characteristic data includes solar irradiance data, cloud cover data and temperature data.

[0021] It is understandable that photovoltaic power generation is very susceptible to weather conditions. Short-term fluctuations in solar irradiance can cause changes in ambient temperature, which directly affects photovoltaic conversion efficiency. Long-term changes in solar irradiance can also be affected by changes in cloud cover.

[0022] S2, constructing a power probability prediction model based on the Kolmogorov-Arnold network, and inputting the first net load data and the first meteorological data into the power probability prediction model for parameter estimation, so as to obtain the net load power mean and net load power variance of the distributed photovoltaic power generation system at each time point within the second preset time range; Specifically, Figure 2 As shown, a power probability prediction model is constructed based on the Kolmogorov-Arnold network, including the following steps: S201, constructing a first KAN model and a second KAN model respectively based on the initialization of the Kolmogorov-Arnold network, and superimposing the first KAN model and the second KAN model to obtain an initial power probability prediction model; Kolmogorov-Arnold Networks (KAN) are a neural network structure based on the Kolmogorov-Arnold (KA) theorem, which aims to solve the problem of decomposition and representation of multidimensional functions. The KA theorem shows that any multidimensional continuous function can be represented as a combination of multiple one-dimensional functions, which provides a theoretical basis for neural networks to learn multidimensional functions.

[0023] For ease of understanding, the following is a detailed description of KAN: The KA theorem states that any smooth multivariable continuous function on a bounded domain can be exactly represented by a finite polynomial combination of a series of univariate continuous functions. For any smooth function defined on a bounded domain , there exists a set of continuous univariate functions and , so that: in, , for The mapping of for mapping, and they are all continuous functions.

[0024] Therefore, the learning process of any high-dimensional function can be simplified to learning a polynomial combination of a single variable function. Considering that there are only two nonlinear layers in the original equation of the KA representation theorem, and the hidden variables are limited to 2n+1 items, it is not conducive to using deep learning techniques (such as back propagation) to train the network. Therefore, KAN generalizes the original KA representation theorem to arbitrary width and depth. Among them, KAN has a learnable single variable activation function , and all activation functions are set at the edge of the network. Therefore, KAN can Interpretation as basis functions Adding spline function In the form of: in, and is the weight coefficient. Spline function Can be further expressed as a B-spline curve and the learnable coefficient A linear combination of: where G is the number of grid intervals and k is the piecewise polynomial order of the spline function. Therefore, Input and The KAN network layer with 1 output can be defined as a set of univariate activation functions: Among them, the function With learnable parameters.

[0025] KA Theorem In order to approximate any function, the network can be designed as a simple combination of two KAN layers. The input of the first layer is , the output is , and the input of the second layer is , the output is . However, in practice, KAN is not restricted by this constraint, and a KAN network with N layers stacked can be represented by an integer array: in, The KAN model The number of nodes in the layer. The output of the layer can be represented by matrix operations: in, Is the connection The i-th neuron and the The activation function of the jth neuron in the layer. Indicates The set of functions in the KAN layer, then the KAN model with L layers can be expressed as: in, represents the Hadamard product.

[0026] Figure 3 The schematic diagram of the structure of the KAN model with two layers is shown. It can be understood that represents the set of functions in the first KAN layer, Represents the function set in the second KAN layer, each black dot represents a neuron, and each box represents an activation function.

[0027] Specifically, step S201 includes: 1) Configure the first Kolmogorov-Arnold network to obtain the initial first KAN model; Similar to traditional neural networks, the configuration includes structure configuration, activation function configuration and trainable parameter configuration. In order to ensure the function approximation ability, the first Kolmogorov-Arnold network includes at least two layers.

[0028] 2) Regularize the initial first KAN model to obtain the first KAN model; In order to ensure the function approximation ability, L1 regularization and entropy regularization are introduced into the loss function of the initial first KAN model to achieve regularization and obtain the first KAN model.

[0029] For the activation function , the L1 norm is defined as the average value of the function output over all N training samples: For a Input and KAN network layer with output , the L1 regularization term can be obtained by The L1 norm of all activation functions in is summed up to get: Sparsification penalizes network structures that result in larger outputs by minimizing the L1 regularization term in the loss function, thereby making the network more sparse during training: By minimizing entropy, the training process encourages The activation functions in are more evenly distributed to reduce the number of activation functions activated at the same time, which is equivalent to introducing a Dropout mechanism between the activation functions to further sparse the KAN network. Therefore, the total loss function of the first KAN model can be defined as the sum of the prediction loss plus the two regularization terms: in, represents the net load power mean predicted loss, , and is the weight factor. Usually .

[0030] 3) configuring the second Kolmogorov-Arnold network to obtain the initial second KAN model; Similar to traditional neural networks, the configuration includes structure configuration, activation function configuration and trainable parameter configuration. In order to ensure the function approximation ability, the second Kolmogorov-Arnold network includes at least two layers.

[0031] 4) Regularizing the initial second KAN model to obtain the second KAN model; In order to ensure the function approximation ability, L1 regularization and entropy regularization are introduced into the loss function of the initial second KAN model to achieve regularization and obtain the second KAN model.

[0032] For the relevant content about L1 regularization and entropy regularization, please refer to the description in 2), which will not be repeated here. Therefore, the total loss function of the second KAN model can be defined as the sum of the prediction loss plus the two regularization terms: in, represents the net load power variance prediction loss, , and is the weight factor. Usually .

[0033] 5) The first KAN model and the second KAN model are superimposed in parallel to obtain an initial power probability prediction model.

[0034] Through parallel superposition, the first KAN model and the second KAN model can process the input data at the same time and obtain different data processing results respectively.

[0035] S202, obtaining second net load data and second meteorological data of the distributed photovoltaic power generation system within a third preset time range; The third preset time range is a historical period. That is to say, the second net load data and the second meteorological data are both historical data. The second net load data reflects the difference between the user load demand and the distributed photovoltaic power generation supply at different times in the past, and the second meteorological data will affect the distributed photovoltaic power generation, thereby affecting the net load. Specifically, the second meteorological data includes a number of meteorological characteristic data, and the meteorological characteristic data includes solar irradiance data, cloud cover data and temperature data. It can be understood that the third preset time range and the first preset time range are different historical periods.

[0036] S203: construct a training set based on the second net load data and the second meteorological data, and train an initial power probability prediction model according to the training set to obtain a power probability prediction model.

[0037] The power probability prediction model is configured to perform parameter estimation through the first KAN model based on the first net load data and the first meteorological data, and obtain the net load power mean of the distributed photovoltaic power generation system at each time point within the second preset time range, and is configured to perform parameter estimation through the second KAN model based on the first net load data and the first meteorological data, and obtain the net load power variance of the distributed photovoltaic power generation system at each time point within the second preset time range. It should be noted that the second preset time range is a future period. In other words, the net load power mean obtained by parameter estimation through the first KAN model and the net load power variance obtained by parameter estimation through the second KAN model are both predicted values ​​for future moments.

[0038] Specifically, step S203 includes: 1) Filtering the second meteorological data based on the second net load data, and performing dimensionality reduction and reorganization on the filtered second meteorological data and the second net load data to obtain first one-dimensional vector data; Specifically, in order to improve the mapping capability of the power probability prediction model to the input features, feature screening is required, that is, by quantifying the impact of multiple meteorological features such as solar irradiance, cloud cover and temperature on the net load, the meteorological features that have a significant impact on the net load are selected. Specifically, the maximum information coefficient between each meteorological feature data in the second meteorological data and the second net load data is calculated, and the meteorological feature data whose maximum information coefficient is higher than a preset threshold is selected.

[0039] It should be noted that the Maximum Information Coefficient (MIC) is a feature correlation evaluation method designed based on the idea of ​​grid division and mutual information. It can effectively identify the linear or nonlinear functional relationship between two variables, is not easily affected by outliers, and has strong universality. The value range of MIC is [0,1]. The higher the value, the stronger the correlation between the two variables. The calculation method can be expressed as: Among them, a and b are the number of grid divisions of independent variables and latent variables respectively, G(a,b) is the grid division method, is the mutual information of data set D under G(a,b), is the normalized mutual information, function for , is the size of dataset D, which represents the specific dataset used to calculate the maximum information coefficient and related statistics. In the calculation process of MIC, dataset D is used to evaluate the relationship and correlation between different features or variables.

[0040] Furthermore, the filtered second meteorological data and the second net load data are expanded into a one-dimensional vector to obtain first one-dimensional vector data.

[0041] 2) obtaining actual net load power data corresponding to the first one-dimensional vector data, so that the first one-dimensional vector data and the actual net load power data constitute a training set; It can be understood that the actual net load power data is equivalent to the label data of the first one-dimensional vector data. The first one-dimensional vector data and the actual net load power data constitute a training set, which is actually a typical supervised learning problem.

[0042] 3) Perform supervised training on the initial power probability prediction model based on the training set to obtain the power probability prediction model.

[0043] Specifically, based on N iterative training steps, a power probability prediction model is obtained, where N is an integer ≥1.

[0044] The steps of each iterative training include: 1) Input the training set into the initial power probability prediction model for parameter estimation to obtain net load power prediction data; After the training set is input into the initial power probability prediction model, the first KAN model and the second KAN model respectively perform parameter estimation on the training set, wherein the net load power prediction mean is obtained by parameter estimation of the first KAN model, and the net load power prediction variance is obtained by parameter estimation of the second KAN model.

[0045] 2) Calculate the loss function based on the net load power prediction data and the net load power actual data; The initial power probability prediction model constructs a loss function based on negative log-likelihood. Specifically, the following formula is used to characterize the loss function: in, is the actual value of net load power at time i, is the net load power forecast mean value at time i output by the initial power probability forecast model, is the predicted variance of the net load power at time i output by the initial power probability prediction model, reflecting the uncertainty of the predicted distribution. It can be understood that if the probability distribution parameters (mean and variance) predicted by the initial power probability prediction model are closer to the actual situation, the negative log-likelihood value will be smaller; conversely, the negative log-likelihood value will be larger.

[0046] 3) Adjusting the trainable parameters of the initial power probability prediction model through maximum likelihood estimation based on the loss function.

[0047] During the training process, the initial power probability prediction model takes the samples in the first one-dimensional vector data as input, and hopes that the probability distribution composed of the output parameters of the initial power probability prediction model will have the maximum probability value at the actual data of the net load power. Therefore, the objective function of the initial power probability prediction model during training is , which can be obtained through the log-likelihood function express: in, is the jth sample in the first one-dimensional vector data, is the jth sample in the actual data of net load power.

[0048] In order to With the maximum value, the initial power probability prediction model can obtain the optimal network parameters by training the network through maximum likelihood estimation : During the training process, the initial power probability prediction model will continuously adjust the trainable parameters of the initial power probability prediction model through maximum likelihood estimation based on the loss value calculated by the negative log-likelihood function, so that the initial power probability prediction model can achieve the optimal degree of fit for the training set.

[0049] Furthermore, the first net load data and the first meteorological data are input into the power probability prediction model for parameter estimation to obtain the net load power mean and net load power variance of the distributed photovoltaic power generation system at each time point within the second preset time range. Figure 4 As shown, it is a schematic diagram of the structure of the power probability prediction model. It can be understood that each black dot represents a neuron and each box represents an activation function. Among them, the first KAN model is used for mean prediction and the second KAN model is used for variance prediction.

[0050] It should be noted that before the first net load data and the first meteorological data are input into the power probability prediction model for parameter estimation, the process also includes: filtering the first meteorological data based on the first net load data, and reducing the dimension of the filtered first meteorological data and the first net load data to obtain the second one-dimensional vector data. In other words, the second one-dimensional vector data is actually input into the power probability prediction model for parameter estimation. For the relevant content about obtaining the second one-dimensional vector data, please refer to the process of obtaining the first one-dimensional vector data in the previous text, which will not be repeated here.

[0051] S3, based on the net load power mean and the net load power variance, calculating the net load power probability density function of the distributed photovoltaic power generation system at each time point within the second preset time range; The conditional probability distribution of the net load power of the distributed photovoltaic power generation system at each time point within the second preset time range can usually be expressed by a probability density function. In a specific embodiment, the starting time point is t, given the input (C is the number of feature types in the second one-dimensional vector data), As a covariate, the probability distribution of net load power after T steps is predicted, which can be expressed as: in, is the predicted value of net load power in the next T steps, is the predicted value of net load power at step t, are the trainable parameters of the power probability prediction model, is the mapping function.

[0052] The uncertainty of net load power at each time point follows a Gaussian distribution Therefore, at each time point The probability density function of the Gaussian distribution can be expressed as: It should be noted that, based on the net load power mean and net load power variance of the distributed photovoltaic power generation system at each time point within the second preset time range, the specific form of the corresponding probability density function can be determined when the probability distribution type obeyed by the uncertainty of the net load power at each time point is known. The probability distribution type obeyed by the uncertainty of the net load power at each time point within the second preset time range of the distributed photovoltaic power generation system of the present invention is not limited to the Gaussian distribution in the above-mentioned embodiment.

[0053] S4. Based on the net load power probability density function, explain the net load power probability prediction result of the distributed photovoltaic power generation system at each time point within the second preset time range.

[0054] Based on the net load power probability density function, the net load power values ​​obtained by random sampling are used to explain the net load power probability prediction results of the distributed photovoltaic power generation system at each time point within the second preset time range.

[0055] Specifically, after obtaining the net load power probability density function at each time point, random sampling is used to obtain a number of net load power values. Through these net load power values, the possible future situations of net load power can be estimated and analyzed. For example, by calculating the probability that the net load power falls within a certain interval, the probability of the net load power within this interval can be predicted, which is the probability prediction result of the net load power.

[0056] In order to verify the interpretability of the power probability prediction model and enhance the transparency of the power probability prediction model, in a specific embodiment, the feature importance analysis method is used to analyze the input features, latent space features, and network parameter weights between the output mean and output variance of the power probability prediction model. The learnable coefficient The number of In order to effectively quantify the influence of each input feature on the output result, each spline function of the power probability prediction model after parameter fitting is The mean of the learnable coefficients is calculated, and a heat map of the average weight coefficient of the spline function is drawn. By analyzing the network weights of the input features to the hidden features, it is possible to identify which features are more critical to the prediction process of the power probability prediction model, thereby understanding the decision-making process of the power probability prediction model and providing a theoretical basis for improving the power probability prediction model.

[0057] The embodiment of the present invention provides an interpretable net load power probability prediction method, which constructs an initial power probability prediction model based on the Kolmogorov-Arnold network, and trains the initial power probability prediction model to obtain a power probability prediction model, which can effectively improve the interpretability of model parameters and improve the accuracy of net load power probability prediction; the initial power probability prediction model is trained with maximum likelihood estimation as the objective function, which can achieve accurate fitting of the net load power probability distribution, perform feature screening based on the maximum information coefficient, and enhance the power probability prediction model's mapping capability to various meteorological characteristics.

[0058] Based on the above-mentioned interpretable net load power probability prediction method, such as Figure 5 As shown, an embodiment of the present invention provides an interpretable net load power probability prediction system, including: The historical data acquisition module 1 is used to acquire first net load data and first meteorological data of the distributed photovoltaic power generation system within a first preset time range, wherein the first meteorological data includes a number of meteorological characteristic data, and the meteorological characteristic data includes solar irradiance data, cloud cover data and temperature data; Parameter estimation module 2 is used to construct a power probability prediction model based on the Kolmogorov-Arnold network, and input the first net load data and the first meteorological data into the power probability prediction model for parameter estimation, so as to obtain the net load power mean and net load power variance of the distributed photovoltaic power generation system at each time point within the second preset time range; A probability density function determination module 3 is used to calculate the net load power probability density function of the distributed photovoltaic power generation system at each time point within a second preset time range based on the net load power mean and the net load power variance; The prediction result interpretation module 4 is used to interpret the net load power probability prediction result of the distributed photovoltaic power generation system at each time point within the second preset time range based on the net load power probability density function.

[0059] It should be noted that each module in the above-mentioned interpretable net load power probability prediction system can be fully or partially implemented by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above-mentioned modules. For the specific definition of an interpretable net load power probability prediction system, please refer to the definition of an interpretable net load power probability prediction method above. The two have the same functions and effects, which will not be repeated here.

[0060] In summary, the embodiment of the present invention provides an interpretable net load power probability prediction method and system, which constructs an initial power probability prediction model based on the Kolmogorov-Arnold network, and trains the initial power probability prediction model to obtain a power probability prediction model, which can effectively improve the interpretability of model parameters and improve the accuracy of net load power probability prediction; the initial power probability prediction model is trained with maximum likelihood estimation as the objective function, which can achieve accurate fitting of the net load power probability distribution, perform feature screening based on the maximum information coefficient, and enhance the power probability prediction model's mapping ability to various meteorological characteristics.

[0061] Each embodiment in this specification is described in a progressive manner, and the same or similar parts of each embodiment can be directly referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. It should be noted that the technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, all possible combinations of the technical features in the above-mentioned embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0062] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and substitutions can be made without departing from the technical principles of the present invention. These improvements and substitutions should also be regarded as the scope of protection of the present invention.

Claims

1. A method for predicting net load power probability that can be explained, characterized in that: include: Acquire first net load data and first meteorological data of the distributed photovoltaic power generation system within a first preset time range, wherein the first meteorological data includes a number of meteorological characteristic data, and the meteorological characteristic data includes solar irradiance data, cloud cover data and temperature data; Building a power probability prediction model based on the Kolmogorov-Arnold network, and inputting the first net load data and the first meteorological data into the power probability prediction model for parameter estimation, so as to obtain the net load power mean and net load power variance of the distributed photovoltaic power generation system at each time point within the second preset time range; Calculating a net load power probability density function of the distributed photovoltaic power generation system at each time point within the second preset time range based on the net load power mean and the net load power variance; The net load power probability prediction result of the distributed photovoltaic power generation system at each time point within the second preset time range is explained based on the net load power probability density function.

2. The explainable net load power probability prediction method according to claim 1 is characterized in that: The power probability prediction model is constructed based on the Kolmogorov-Arnold network, including: Based on the initialization of the Kolmogorov-Arnold network, a first KAN model and a second KAN model are respectively constructed, and the first KAN model and the second KAN model are superimposed to obtain an initial power probability prediction model; Acquire second net load data and second meteorological data of the distributed photovoltaic power generation system within a third preset time range; A training set is constructed based on the second net load data and the second meteorological data, and the initial power probability prediction model is trained according to the training set to obtain a power probability prediction model, wherein the power probability prediction model is configured to perform parameter estimation through the first KAN model based on the first net load data and the first meteorological data to obtain the net load power mean of the distributed photovoltaic power generation system at each time point within a second preset time range, and is configured to perform parameter estimation through the second KAN model based on the first net load data and the first meteorological data to obtain the net load power variance of the distributed photovoltaic power generation system at each time point within the second preset time range.

3. The explainable net load power probability prediction method according to claim 2 is characterized in that: The method of constructing a first KAN model and a second KAN model based on the initialization of the Kolmogorov-Arnold network, and superimposing the first KAN model and the second KAN model to obtain an initial power probability prediction model includes: configuring the first Kolmogorov-Arnold network to obtain an initial first KAN model, wherein the configuration includes a structure configuration, an activation function configuration, and a trainable parameter configuration; Regularizing the initial first KAN model to obtain a first KAN model; configuring the second Kolmogorov-Arnold network to obtain an initial second KAN model, wherein the configuration includes a structure configuration, an activation function configuration, and a trainable parameter configuration; Regularizing the initial second KAN model to obtain a second KAN model; The first KAN model and the second KAN model are superimposed in parallel to obtain an initial power probability prediction model.

4. The explainable net load power probability prediction method according to claim 3 is characterized in that: The regularization includes L1 regularization and entropy regularization.

5. The explainable net load power probability prediction method according to claim 2, characterized in that: The constructing a training set based on the second net load data and the second meteorological data, and training the initial power probability prediction model according to the training set to obtain a power probability prediction model, includes: The second meteorological data is filtered based on the second net load data, and the filtered second meteorological data and the second net load data are reorganized by dimensionality reduction to obtain first one-dimensional vector data; Acquire actual net load power data corresponding to the first one-dimensional vector data, so that the first one-dimensional vector data and the actual net load power data constitute a training set; The initial power probability prediction model is supervisedly trained based on the training set to obtain a power probability prediction model, wherein the initial power probability prediction model constructs a loss function based on negative log-likelihood.

6. The explainable net load power probability prediction method according to claim 5, characterized in that: The screening of the second meteorological data based on the second net load data includes: The maximum information coefficient between each meteorological characteristic data in the second meteorological data and the second net load data is calculated, and the meteorological characteristic data whose maximum information coefficient is higher than a preset threshold is selected.

7. The explainable net load power probability prediction method according to claim 5, characterized in that: The method of performing supervised training on the initial power probability prediction model based on the training set to obtain a power probability prediction model includes: Based on the steps of N iterative training, a power probability prediction model is obtained, wherein N is an integer ≥ 1; The steps of each iterative training include: Inputting the training set into the initial power probability prediction model for parameter estimation to obtain net load power prediction data; Calculating the loss function based on the net load power prediction data and the net load power actual data; The trainable parameters of the initial power probability prediction model are adjusted by maximum likelihood estimation based on the loss function.

8. The interpretable net load power probability prediction method according to claim 1, characterized in that: Before inputting the first net load data and the first meteorological data into the power probability prediction model for parameter estimation, the method further includes: The first meteorological data is filtered based on the first net load data, and the filtered first meteorological data and the first net load data are reorganized by dimensionality reduction to obtain second one-dimensional vector data.

9. The explainable net load power probability prediction method according to claim 1, characterized in that: Explaining the net load power probability prediction result of the distributed photovoltaic power generation system at each time point within the second preset time range based on the net load power probability density function includes: Based on the net load power probability density function, the net load power values ​​obtained by random sampling are used to explain the net load power probability prediction result of the distributed photovoltaic power generation system at each time point within the second preset time range.

10. An interpretable net load power probability prediction system, characterized in that: include: A historical data acquisition module, used to acquire first net load data and first meteorological data of a distributed photovoltaic power generation system within a first preset time range, wherein the first meteorological data includes a number of meteorological characteristic data, and the meteorological characteristic data includes solar irradiance data, cloud cover data and temperature data; a parameter estimation module, configured to construct a power probability prediction model based on a Kolmogorov-Arnold network, and input the first net load data and the first meteorological data into the power probability prediction model for parameter estimation, so as to obtain a net load power mean and a net load power variance of the distributed photovoltaic power generation system at each time point within a second preset time range; A probability density function determination module, configured to calculate a net load power probability density function of the distributed photovoltaic power generation system at each time point within the second preset time range based on the net load power mean and the net load power variance; A prediction result interpretation module is used to interpret the net load power probability prediction result of the distributed photovoltaic power generation system at each time point within the second preset time range based on the net load power probability density function.

Images