A weather-driven new energy power generation power prediction method

Abstract

The application discloses a weather-driven new energy power generation power prediction method, which comprises a new energy power generation power-weather factor correlation analysis method, a new energy power generation power core weather factor optimization identification method and a weather-driven new energy power generation power prediction method. The power generation power prediction model of the application has high prediction accuracy, the method considers the power generation characteristic differences caused by geographical and weather factors, and has more advantages compared with traditional prediction methods; the prediction efficiency is high, the dimension reduction measures based on weather correlation analysis can optimize the selection of key weather factors as input variables, reduce the complexity of the prediction model and improve the prediction efficiency; the method can be applied to few-data new energy stations, in view of the problem that some new energy stations lack historical power generation data, the spatial correlation and weather correlation analysis are comprehensively considered to supplement the data source of the prediction model, and the prediction accuracy of the power generation power of the few-data new energy station is improved. Therefore, the method is suitable for popularization and application.

Description

Technical Field

[0001] This invention belongs to the field of new energy power generation prediction technology, specifically, it relates to a weather-driven method for predicting new energy power generation. Background Technology

[0002] Under the strategic backdrop of the vigorous implementation of the "dual-carbon policy," traditional fossil fuels, primarily coal, natural gas, and oil, can no longer meet the requirements of low-carbon energy development. Developing clean and low-carbon energy has become an inevitable trend, necessitating a green and low-carbon energy transition and the development and utilization of clean energy to achieve more stringent energy efficiency targets. Unlike traditional thermal power units, the power generation of new energy units is related to multi-dimensional meteorological factors such as temperature, wind speed, and humidity, exhibiting fluctuations and intermittency. The large-scale integration of high-proportion clean energy sources will compress the operating space of conventional units, reducing system operational flexibility and inertia, inevitably posing significant challenges to the stability, safe operation, and economic efficiency of the power system. Against this backdrop, it is necessary to predict the power generation capacity of new energy sources. High-precision power prediction not only provides a reliable basis for grid dispatch decisions but also serves as the foundation for multi-energy coordinated development. However, new energy power generation prediction models are characterized by multi-dimensionality, multi-scale, cross-coupling, and nonlinearity. Traditional prediction methods do not consider these characteristics, resulting in disadvantages such as low accuracy and low efficiency.

[0003] Existing methods for predicting renewable energy power generation considering meteorological factors directly input the original measured multivariate meteorological factors into the prediction model. This results in a complex model structure, weak generalization ability, low training efficiency, and an inability to extract key features. Furthermore, the lack of screening for meteorological influencing factors leads to redundant information input into the model, affecting the model's learning of important correlation characteristics and reducing prediction accuracy. The impact and degree of influence of meteorological factors on renewable energy power generation vary. Due to their coupling relationships and complex and diverse influence paths, obtaining quantitative results on the degree of influence of meteorological factors on photovoltaic output, and then accurately identifying and optimizing the selection of core meteorological influencing factors, is a fundamental analysis necessary for photovoltaic power generation prediction. However, current mainstream correlation analysis algorithms are only suitable for handling linear relationships between two variables, and are difficult to handle the correlation characteristics between high-dimensional nonlinear data. They are also unsuitable for identifying high-dimensional and complex correlations between renewable energy power generation and meteorological factors such as wind speed, irradiance, and temperature.

[0004] Current methods for predicting renewable energy power generation rely solely on historical meteorological data and established power generation scenarios. However, this approach depends on the accuracy of these scenarios and fails to consider the differences in renewable energy power generation characteristics caused by geographical and climatic factors. In recent years, my country's renewable energy power generation has developed rapidly, with numerous newly established renewable energy power plants whose power generation characteristics cannot be incorporated into traditional scenarios, leading to some bias in the prediction results. As the penetration rate of renewable energy increases, this bias may result in inaccurate predictions of future power generation, consequently causing mismatches between grid planning schemes and grid connection and transmission planning schemes. Summary of the Invention

[0005] The purpose of this invention is to provide a weather-driven method for predicting new energy power generation, thereby enabling the discovery of the correlation characteristics between new energy power generation and meteorological factors, and the construction of a precise new energy power generation prediction model that takes meteorological factors into account, thus improving the prediction accuracy of new energy power generation.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A weather-driven method for predicting the power generation of new energy sources includes the following steps:

[0008] S1, analyze the correlation characteristics between new energy power generation and meteorological factors;

[0009] S11, integrate the input new energy power generation data and meteorological data to construct a new energy power generation-meteorological factor dataset;

[0010] S12, the grey relational analysis method was used to obtain the nonlinear correlation between various meteorological elements and new energy power generation under different periods;

[0011] S13, calculate the consistency between the new energy power generation curve and the meteorological curve during the dynamic change process, and obtain the correlation characteristics between meteorological factors and photovoltaic power generation.

[0012] S2, optimizes the identification of core meteorological factors for new energy power generation;

[0013] S21, using the correlation characteristics between new energy power generation and meteorological factors as the basis for identifying and optimizing the input variables of the prediction model, and identifying the core meteorological influencing factors of power generation of each new energy power station;

[0014] S22, determine the dimensions of input variables based on the time-period characteristics of new energy power generation, and construct the set of input variables for the meteorological-driven new energy power generation prediction model based on the dimensions of input variables;

[0015] S3 enables weather-driven prediction of new energy power generation.

[0016] S31, Construct a CNN-LSTM hybrid neural network;

[0017] S32, Based on the optimized identification of core meteorological factors, a meteorological-driven new energy power generation prediction model based on the CNN-LSTM algorithm is constructed.

[0018] S33, taking into account spatial correlation and meteorological correlation characteristics, forms a reference power station combination as a supplementary data source, and uses a meteorological-driven new energy power generation prediction model to predict the power generation of new energy power stations with limited data.

[0019] Furthermore, in step S13, the correlation characteristic analysis step of the degree of correlation between meteorological factors and photovoltaic power generation includes:

[0020] S13.1 defines the new energy power generation power as the reference sequence, the multiple meteorological factors as the comparison sequence, and the difference between the new energy power generation power sequence and the meteorological factor sequence as Δ. i (k); where,

[0021] Δ i (k)=|x0(k)-x i (k)|(i=0,1…,m,k=1,2,…n)

[0022] In the formula, n is the number of reference sequences, m is the number of comparison sequences, and Δ i (k)=(Δ i (1),Δ i (2),…,Δ i (k) is essentially a sequence of differences;

[0023] S13.2, define the minimum value among all difference sequences as the minimum range min. i min k Δ i (k), where the maximum value in each difference sequence is defined as the maximum range max. i max k Δ i (k), the correlation coefficient γ between the i-th meteorological factor and photovoltaic power generation. 0i (k) is calculated as follows:

[0024]

[0025] In the formula, ξ is the resolution coefficient, ξ∈[0,1], and here it is taken as 0.5;

[0026] S13.3, the correlation coefficient γ is obtained. 0i (k) followed by the formula:

[0027]

[0028] The grey relational degree of each meteorological factor was obtained. Where n is the number of reference sequences;

[0029] S13.4, normalize the grey relational degree and calculate the factor weight coefficient γ of the i-th meteorological factor. i :

[0030]

[0031] Furthermore, in step S13, for meteorological elements with strong correlation, the kernel density estimation method is used to calculate the probability distribution between the changes in core elements and the changes in new energy power generation. If the density function of the random variable X is f(x) = F(x), a simple estimate of f(x) is obtained according to the kernel density estimation method. The calculation formula is:

[0032]

[0033] In the formula, h is the window width, which is a non-negative constant, and F(x) is the empirical distribution function of the random variable X;

[0034] Suppose there are N sample values ​​generated by the same unknown probability density: x1, x2, ... x n When choosing a nonnegative constant h such that n→+∞, h→0 and nh→+∞, the kernel density estimation function is obtained:

[0035]

[0036] In the formula, Let h represent the probability density function of the population, h be the window width, N be the total number of samples, and K be the number of samples. h For kernel functions;

[0037] Choosing the Gaussian kernel function as the kernel function, the calculation formula is as follows:

[0038]

[0039] Therefore, the overall Gaussian kernel density estimation function is as follows:

[0040]

[0041] Furthermore, in step S13, the method also includes using a neural network algorithm to achieve full-state space fitting of historical new energy data, and extracting the changes in new energy power generation caused by fluctuations in core meteorological conditions at different times, thereby achieving quantitative analysis of the correlation between "new energy power generation and meteorological factors" under the influence of multiple variables. Based on the full-state space fitting results, sensitivity analysis is performed on the core meteorological factors. By observing the trend and magnitude of the changes in new energy power generation and the influence of other meteorological factors when the core meteorological factors change, the quantitative correlation between multiple meteorological factors and new energy power generation is obtained.

[0042] Furthermore, the specific steps to derive the quantitative correlation between multiple meteorological factors and new energy power generation are as follows:

[0043] a: Before using neural networks to fit new energy power, it is necessary to normalize historical data with different dimensions;

[0044] b: Select normalized wind speed or irradiance data as the horizontal axis and core meteorological factors as the vertical axis to grid the meteorological data;

[0045] c: Obtain the optimal combination of neural network model parameters through a search and traversal method;

[0046] d: Denormalize the output values ​​of the fitted results;

[0047] e: Apply sensitivity analysis to study the quantitative relationship between the coupling relationship between meteorological variables and the power generation of new energy sources.

[0048] Furthermore, in step S21, the criteria for optimizing the identification of core meteorological factors include:

[0049] 1) Reduce redundancy in input information;

[0050] 2) Appropriately select the dimensions of input variables for the prediction model;

[0051] 3) Select based on the time-period characteristics of new energy power generation.

[0052] Furthermore, the new energy power generation prediction model in step S32 is a prediction model that takes time series data containing multiple feature variables as input and outputs prediction results for a single variable. The information and related functions of each layer of the model are as follows:

[0053] S32.1, Input and Convolutional Layer, used to distribute the input of the convolutional unit to the next layer;

[0054] S32.2, the activation function layer, is used to introduce nonlinear factors into the prediction model, improve the ability to express data features, and enable neural networks to better solve nonlinear problems;

[0055] S32.3, pooling layer, is used to downsample the output vector of the convolutional layer, which is equivalent to secondary feature extraction;

[0056] S32.4, Random Deactivation Layer and Flatten Layer: The random deactivation layer is used to solve the overfitting problem in deep neural networks, and the Flatten layer is used to process the data into the format required by the LSTM layer;

[0057] S32.5, Long Short-Term Memory layer, as a variant of recurrent neural networks, enables LSTM units to operate at time 10 ...

[0058] S32.6, the fully connected layer, is used to receive the output of each LSTM unit from the previous layer. The inputs are expanded and connected to the output layer.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] The power generation prediction model of this invention has high prediction accuracy. This method takes into account the differences in power generation characteristics caused by geographical and meteorological factors, which is more advantageous than traditional prediction methods. It also has high prediction efficiency. The dimensionality reduction measures based on meteorological correlation analysis can optimize the selection of key meteorological factors as input variables, reduce the complexity of the prediction model and improve prediction efficiency. It can be applied to renewable energy power plants with limited data. In response to the problem that some renewable energy power plants lack historical power generation data, it comprehensively considers spatial correlation and meteorological correlation analysis to supplement the data sources of the prediction model, thereby improving the prediction accuracy of power generation for renewable energy power plants with limited data. Attached Figure Description

[0061] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0062] Figure 2 This is a schematic diagram of Gaussian kernel density estimation fitting in an embodiment of the present invention.

[0063] Figure 3 This is a schematic diagram of the CNN-LSTM prediction model structure in an embodiment of the present invention. Detailed Implementation

[0064] The present invention will be further described below with reference to the accompanying drawings and embodiments. The embodiments of the present invention include, but are not limited to, the following embodiments.

[0065] Example

[0066] like Figure 1 As shown, the present invention discloses a weather-driven method for predicting new energy power generation, comprising three parts: a method for analyzing the correlation between new energy power generation and meteorological factors, a method for optimizing and identifying core meteorological factors of new energy power generation, and a weather-driven method for predicting new energy power generation.

[0067] The method for analyzing the correlation between new energy power generation and meteorological factors is as follows: Grey relational analysis is used to obtain the nonlinear correlation between various meteorological elements and new energy power generation at different times; by calculating the consistency between the new energy power generation curve and the meteorological curve during dynamic changes, the correlation characteristics between meteorological factors and photovoltaic power generation are obtained. Specifically, the following steps are included:

[0068] 1) Define the new energy power generation power as the reference sequence, the multiple meteorological factors as the comparison sequence, and the difference between the new energy power generation power sequence and the meteorological factor sequence as Δ. i (k); where,

[0069] Δ i (k)=|x0(k)-x i (k)|(i=0,1…,m,k=1,2,…n)

[0070] In the formula, n is the number of reference sequences, m is the number of comparison sequences, and Δ i (k)=(Δ i (1),Δ i (2),…,Δ i (k) is essentially a sequence of differences.

[0071] 2) Define the minimum value in each difference sequence as the minimum range (min). i min k Δ i (k), where the maximum value in each difference sequence is defined as the maximum range max. i max k Δ i (k), the correlation coefficient γ between the i-th meteorological factor and photovoltaic power generation. 0i (k) is calculated as follows:

[0072]

[0073] In the formula, ξ is the resolution coefficient, ξ∈[0,1], and here it is taken as 0.5.

[0074] 3) Obtain the correlation coefficient γ 0i (k) followed by the formula:

[0075]

[0076] The grey relational degree of each meteorological factor was obtained. Where n is the number of reference sequences.

[0077] 4) Normalize the grey relational degree and calculate the factor weight coefficient γ of the i-th meteorological factor. i :

[0078]

[0079] For meteorological elements with strong correlations, the kernel density estimation method is used to calculate the probability distribution between the changes in core elements and the changes in new energy power generation. If the density function of the random variable X is f(x) = F(x), a simple estimate of f(x) can be obtained according to the kernel density estimation method. The calculation formula is:

[0080]

[0081] In the formula, h is the window width, which is a non-negative constant, and F(x) is the empirical distribution function of the random variable X;

[0082] like Figure 2 As shown, suppose there are N sample values ​​generated by the same unknown probability density: x1, x2, ... x n When choosing a nonnegative constant h such that n→+∞, h→0 and nh→+∞, the kernel density estimation function is obtained:

[0083]

[0084] In the formula, Let h represent the probability density function of the population, h be the window width, N be the total number of samples, and K be the number of samples. h For kernel functions;

[0085] Choosing the Gaussian kernel function as the kernel function, the calculation formula is as follows:

[0086]

[0087] Therefore, the overall Gaussian kernel density estimation function is as follows:

[0088]

[0089] When different confidence intervals are selected, such as 50% and 90% confidence intervals, nonparametric estimation models such as Gaussian kernel density estimation can be used to more intuitively observe the specific range of fluctuations in the change of renewable energy power generation, thereby analyzing the correlation characteristics between renewable energy power generation and other meteorological factors. Therefore, using nonparametric estimation models such as Gaussian kernel density estimation can achieve a more effective fit for daily varying meteorological data, thus avoiding data gaps caused by fitting with probability distribution models.

[0090] Given the randomness and volatility of historical renewable energy data, a neural network algorithm is used to achieve full-state space fitting of historical data, eliminating the impact of data fluctuations. Then, the changes in renewable energy power generation caused by fluctuations in core meteorological conditions under different periods are extracted, enabling a quantitative analysis of the correlation between renewable energy power generation and meteorological factors under multivariate influences. Based on the analysis results of the degree of influence of meteorological factors, wind speed, irradiance, and core meteorological influencing factors are selected as the horizontal and vertical axes, respectively, and the meteorological data is presented in a gridded form, with the coordinate axis boundaries representing the minimum and maximum values ​​of historical meteorological data. This gridded meteorological data can cover all possible meteorological conditions that may occur at the analyzed power station. The gridded meteorological data is used as the input to the model in matrix form, and the model output is the gridded fitting result of renewable energy power generation under the corresponding meteorological conditions, achieving full-state space fitting of historical data. Based on this fitting result, sensitivity analysis is performed on the core meteorological factors. By observing the trend and magnitude of changes in renewable energy power generation and the influence of other meteorological factors when core meteorological factors change, a quantitative correlation between multiple meteorological factors and renewable energy power generation is obtained. The specific implementation steps are as follows:

[0091] 1) Normalization. Different variables in the raw meteorological and processed data have different units, and the differences in magnitude are usually significant. Algorithm models are sensitive to numerical ranges; excessively large numerical ranges will negatively impact the training of neural network algorithm models. Therefore, before using neural networks to fit new energy power, it is necessary to normalize historical data with different dimensions to avoid the negative impact of different dimensions on the fitting. In the data preprocessing step, the historical data has already been scaled to the range [0,1].

[0092] 2) Meteorological data gridding. Normalized wind speed or irradiance data are selected as the horizontal axis, and core meteorological factors are selected as the vertical axis. The grid precision is set to 0.05, resulting in a 20-row * 20-column meteorological condition matrix as the input to the model.

[0093] 3) Model parameter settings. The optimal combination of neural network model parameters is obtained through a search and traversal method. The Adam algorithm is selected for model training optimization, the training batch size is 128, and the total number of training iterations is set to 200.

[0094] 4) Inverse normalization of fitting results. To obtain fitting results with normal dimensions, the output values ​​in the range [0,1] need to be inverse normalized. The formula for inverse normalization is as follows: The fitting results are displayed in the form of a heatmap. The color intensity of each grid can intuitively show the power generation of new energy sources under the given meteorological conditions.

[0095] 5) Sensitivity Analysis. Sensitivity analysis is applied to study the quantitative relationship between the coupling relationship between meteorological variables and the impact of renewable energy power generation. A specific meteorological factor is selected for analysis, and the irradiance value is kept constant. The difference in renewable energy power generation is calculated when this meteorological factor changes by a unit. The results of the sensitivity analysis are presented in the form of a heat map, with the grid depth representing the trend and magnitude of the change in renewable energy power generation under these meteorological conditions.

[0096] The optimization and identification method for core meteorological factors in new energy power generation is as follows: The correlation characteristics between new energy power generation and meteorological factors are used as the basis for identifying and optimizing the input variables of the prediction model. The core meteorological influencing factors of power generation at each new energy power station are identified. The dimensions of the input variables are determined based on the time-period characteristics of new energy power generation, and a set of input variables for a meteorology-driven new energy power generation prediction model is constructed accordingly. The basis for optimizing and identifying core meteorological factors is as follows:

[0097] 1) Reduce redundancy in input information. Selecting inappropriate meteorological factors may introduce excessive redundant information into the model, increasing the computational complexity.

[0098] 2) Choose the appropriate dimensions of the input variables for the prediction model. If there are too many meteorological factors input into the prediction model, the model structure will become complex, and the sparse distribution of limited input data in a high-dimensional space will prevent the prediction model from being trained effectively.

[0099] 3) Select input variables based on the time-specific characteristics of renewable energy power generation. The correlation between renewable energy power generation and meteorological factors varies significantly across different time periods. Therefore, the wet season and dry season need to be considered separately when selecting input variables.

[0100] The meteorological-driven method for predicting renewable energy power generation is as follows: It combines the advantages of Convolutional Neural Network (CNN) feature extraction and Long Short-Term Memory Neural Network (LSTM) in effectively processing time series data to form a CNN-LSTM hybrid neural network. Based on the optimized identification of core meteorological factors, dimensionality reduction measures are implemented for the input variables of the prediction model, thereby constructing a meteorological-driven renewable energy power generation prediction model based on the CNN-LSTM algorithm. For renewable energy power plants with limited data, a reference power plant combination is formed by comprehensively considering spatial correlation and meteorological correlation characteristics as a supplementary data source. The meteorological-driven renewable energy power generation prediction model is then used to predict the power generation of renewable energy power plants with limited data. The CNN-LSTM prediction model constructed in this method takes time series data containing multiple feature variables as input and outputs a single-variable prediction result. The information and related functions of each layer of the model are described below:

[0101] (1) Input and Convolutional Layers. The purpose of the input layer is to distribute the input of the convolutional units to the next layer, accepting one data sample at a time. Each data sample has n features, including historical renewable energy power generation, irradiance, wind speed, temperature, etc. CNNs are suitable for processing grid-like data, including multivariate time series datasets used for feature extraction. Therefore, the dataset represents a one-dimensional grid of data samples collected at equal time steps. Since CNNs are convolutional neural networks, the layer units use convolution operations instead of traditional multiplication operations.

[0102]

[0103] In the formula, * represents convolution operation, I is the input of dimension M, and K is the convolution kernel of dimension M. In the prediction model constructed in this paper, M equals the number of input meteorological factor variables, and features k1 to k... M Each corresponds to a different meteorological factor selected and entered.

[0104] This layer contains multiple convolutional kernels of size 1, stride 1, and no padding. Kernel weight initialization uses the Glorot unified initializer to prevent over-saturation of the activation function.

[0105]

[0106] In the formula, U[-i,i] represents a uniform distribution within the interval [-i,i], and S(=M) is the size of the previous layer. The bias value is initialized to 0.

[0107] (2) Activation Function. The role of the activation function is to introduce nonlinear factors into the prediction model, improve the ability to express data features, and enable the neural network to better solve nonlinear problems. Commonly used activation functions include the sigmoid function, tanh function, and ReLU function. The ReLU function is chosen as the activation function because its calculation speed and convergence speed are faster than the sigmoid function and tanh function. The specific expression is:

[0108] a l(i,j) =f(y l(i,j) )=max{a l(i,j)}

[0109] In the formula, l is the feature matrix of the convolutional network, and y l(i,j ) represents the output value of the convolutional layer, a l(i,j) This is the activation value.

[0110] (3) Pooling Layer. The pooling layer performs a downsampling operation on the output vector of the convolutional layer, which can be equivalent to secondary feature extraction. This paper uses max pooling, taking the maximum value in the neuron as the output, to effectively reduce the network parameters. The calculation formula is shown below:

[0111]

[0112] In the formula, a l(i,j) p is the activation value. l(i,j) The result is the pooling result, and W is the width of the pooling window.

[0113] (4) Dropout layer and Flatten layer. The Dropout layer is used to solve the common overfitting problem in deep neural networks. Its principle is to temporarily discard units from the network according to a certain probability, so that the activation values ​​are discarded in a certain proportion during the network training process, thereby improving the generalization performance of the neural network. The Flatten layer is used to format the data required by the LSTM layer.

[0114] (5) Long Short-Term Memory Layer.

[0115] As a variant of recurrent neural networks, LSTM's units retain memory at time t. It also outputs the hidden state of the loop to solve the gradient vanishing problem.

[0116]

[0117] In the formula, The output of the LSTM network, obtained through dynamic control of the memory units, is calculated using the following formula:

[0118]

[0119] In the formula, σ is an sigmoid function, x t W o U o and V o These are the input vector, weight matrix, and diagonal matrix of the LSTM unit, respectively. The memory unit uses the state information at the current time step. The memory information at the current moment is calculated from the state information of the previous moment using the formula.

[0120]

[0121]

[0122] In the formula, and These are the activation vectors for the forget gate and the input gate, respectively.

[0123] (6) Fully connected layer. The single neuron in this layer receives the output from each LSTM unit in the previous layer. The input is expanded and connected to the output layer via an expression:

[0124] o = f(Wx + b)

[0125] In the formula, f, W, x, and b are the activation function, weights, input, and bias vector, respectively.

[0126] like Figure 3 As shown, based on the CNN-LSTM prediction model structure, the input layer time step is set to 12, meaning it predicts the data for the next hour using data from the previous 12 hours. Typically, for simple mappings between input and output variables, a single-hidden-layer neural network is sufficient. As the number of hidden layers and neurons in each hidden layer increases, the model's ability to fit nonlinear functions improves, allowing for the extraction of more complex mappings. However, this also increases the model's complexity and reduces training efficiency. This embodiment uses one one-dimensional convolutional layer with 128 convolutional units and one LSTM layer with 100 units.

[0127] Through the above design, the power generation prediction model of this invention has high prediction accuracy. This method takes into account the differences in power generation characteristics caused by geographical and meteorological factors, which is more advantageous than traditional prediction methods. It also has high prediction efficiency. The dimensionality reduction measures based on meteorological correlation analysis can optimize the selection of key meteorological factors as input variables, reduce the complexity of the prediction model and improve prediction efficiency. It can be applied to renewable energy power plants with limited data. In response to the problem that some renewable energy power plants lack historical power generation data, the model comprehensively considers spatial correlation and meteorological correlation analysis to supplement the data sources of the prediction model, thereby improving the prediction accuracy of power generation for renewable energy power plants with limited data.

[0128] The above embodiments are merely one of the preferred embodiments of the present invention and should not be used to limit the scope of protection of the present invention. Any modifications or refinements made to the main design concept and spirit of the present invention that are not of substantial significance, but solve the same technical problem as the present invention, should be included within the scope of protection of the present invention.

Claims

1. A weather-driven new energy power generation power prediction method, characterized in that, The method comprises the following steps: S1, analyzing the correlation characteristics between new energy power generation and meteorological factors; S11, integrating the input new energy power generation data and meteorological data, and constructing a new energy power generation-meteorological factor data set; S12, using a grey correlation analysis method to obtain the nonlinear correlation between each meteorological factor and new energy power generation in different periods; S13, calculating the consistency of the new energy power generation curve and the meteorological curve in the dynamic change process to obtain the correlation characteristics of the correlation degree between the meteorological factors and the new energy power generation; S2, optimizing and identifying the core meteorological factors of new energy power generation; S21, taking the correlation characteristics of new energy power generation and meteorological factors as the input variable identification optimization basis of the prediction model, and identifying the core meteorological influencing factors of each new energy station power generation; S22, determining the input variable dimension according to the time period characteristics of new energy power generation, and constructing an input variable set of the meteorological-driven new energy power generation prediction model based on the input variable dimension; S3, realizing meteorological-driven new energy power generation prediction; S31, constructing a CNN-LSTM hybrid neural network; S32, constructing a meteorological-driven new energy power generation prediction model based on the CNN-LSTM algorithm according to the optimization identification of the core meteorological factors; S33, comprehensively considering the spatial correlation and meteorological correlation characteristics to form a reference station combination as a supplementary data source, and using the meteorological-driven new energy power generation prediction model to realize the power generation prediction of the new energy station with less data.

2. The weather-driven new energy power generation power prediction method according to claim 1, characterized in that, In step S13, the correlation characteristics analysis step of the correlation degree between the meteorological factors and the new energy power generation comprises: S13.1, defining the new energy power generation as a reference sequence, a plurality of meteorological factors as a comparison sequence, and the difference between the new energy power generation sequence and the meteorological factor sequence as ; wherein, (i = 0, 1,..., m, k = 1, 2,... n) In the formula, is the number of reference sequences, is the number of comparison sequences, is the essence of the sequence is poor; S13.2 defines the minimum value among the differences in each sequence as the minimum range. The maximum value in each difference sequence is defined as the maximum range. , No. Correlation coefficients between meteorological factors and new energy power generation The calculation formula is as follows: (i = 0, 1,..., m, k = 1, 2,... n) wherein is the resolution coefficient, wherein the value is 0.

5. S13.3, to obtain the correlation index After, according to the formula: = The grey correlation degrees of various meteorological factors are obtained wherein n is the number of reference sequences; S13.4, the grey correlation degree is normalized, and a factor weight coefficient of the ith meteorological factor is calculated : 。 3. The weather-driven new energy power generation power prediction method according to claim 2, characterized in that, In step S13, for the weather elements with strong correlation, the kernel density estimation method is adopted to calculate the probability distribution between the core element change and the new energy power generation change. If the density function of the random variable X is f(x)=F(x), according to the kernel density estimation method, the simple estimation of f(x) is obtained , and the calculation formula is: In the formula, h is the window width, which is a non-negative constant, and F(x) is the empirical distribution function of the random variable X; Let there be N sample values generated by the same unknown probability density: x1, x2, … xN n When selecting the non-negative constant h, when n→+∞, h→0 and nh→+∞, the kernel density estimation function is obtained: wherein represents the overall probability density function, h is the window width, N represents the total number of samples, K h is the kernel function; The kernel function is selected as a Gaussian kernel function, and the calculation formula is as follows: Therefore, the total Gaussian kernel density estimation function is as follows: 。 4. The weather-driven new energy power generation power prediction method according to claim 1, characterized in that, In step S13, based on the neural network algorithm, the full state space fitting of the historical new energy data is realized, and the new energy power generation changes caused by the fluctuation of the core meteorological conditions in different periods are extracted, so that the quantitative analysis of the correlation between the new energy power generation and the meteorological factors under the influence of multiple variables is realized; on the basis of the full state space fitting result, the sensitivity analysis of the core meteorological factors is carried out, and the quantitative correlation between the multiple meteorological factors and the new energy power generation is obtained through the change trend, amplitude and influence of the new energy power generation when the core meteorological factors change.

5. The weather-driven new energy power generation power prediction method according to claim 4, characterized in that, The specific steps of obtaining the quantitative correlation between the multiple meteorological factors and the new energy power generation are as follows: a: the historical data with different dimensions need to be normalized before using the neural network to fit the new energy power; b: select the normalized wind speed or irradiance data as the horizontal axis and the core meteorological factor as the vertical axis to grid the meteorological data; c: the best combination of neural network model parameters is obtained by searching and traversing; d: the output value of the fitting result is denormalized; e: apply sensitivity analysis to study the quantitative relationship between the coupling relationship between meteorological variables and the influence on new energy power generation.

6. The weather-driven new energy power generation power prediction method according to claim 1, characterized in that, In step S21, the core meteorological factor optimization identification basis comprises: 1) reduce the redundancy of input information; 2) reasonably select the input variable dimension of the prediction model; 3) select according to the time period characteristics of new energy power generation.

7. The weather-driven new energy power generation power prediction method according to claim 6, characterized in that, The new energy power generation prediction model in step S32 is a prediction model that inputs time series data containing multiple characteristic variables and outputs a single variable prediction result. The information and related functions of each layer of the model are as follows: S32.1, input and convolution layer, used to distribute the input of the convolution unit to the next layer; S32.2, activation function layer, used to introduce non-linear factors into the prediction model, improve the expression ability of data characteristics, and make the neural network better solve nonlinear problems; S32.3, pooling layer, used to realize the downsampling operation of the output vector of the convolution layer, which is equivalent to secondary feature extraction; S32.4, random inactivation layer and Flatten layer, the random inactivation layer is used to solve the overfitting problem in the deep neural network, and the Flatten layer is used to process the data into the format required by the LSTM layer; S32.5, long short-term memory layer, as a variant of recurrent neural network, makes the LSTM unit at time S32.6, fully connected layer, used to receive the output of each LSTM unit from the previous layer, the input is expanded and connected to the output layer.

Images