Wind and solar power generation prediction method based on meteorological feature screening and transfer learning

Abstract

The present invention discloses a method for predicting wind and solar power generation based on meteorological feature screening and transfer learning, and belongs to the field of new energy power prediction. The present invention first establishes a combined trust index for screening sensitive meteorological characteristic factors, and obtains sensitive meteorological characteristic factors; then constructs a source domain prediction model, and pre-trains the source domain prediction model using sensitive meteorological characteristic factors; then obtains the historical power generation data of the photovoltaic power station to be predicted for wind and solar power generation, and constructs a target domain data set, performs transfer learning and parameter fine-tuning on the pre-trained source domain prediction model based on the target domain data set, and obtains a target domain prediction model for predicting wind and solar power generation; and uses the target domain prediction model to predict wind and solar power generation. The present invention not only effectively solves the problem of insufficient model generalization ability, but also has stronger adaptability and robustness, and significantly improves the prediction accuracy and stability.

Description

Technical Field

[0001] The present invention relates to the field of new energy prediction, and in particular to a method for predicting wind and solar power generation based on meteorological feature screening and transfer learning. Background Art

[0002] With the increasing depletion of fossil energy and the intensification of environmental pollution problems, the development of renewable energy has become a global task that needs to be urgently addressed. As green and pollution-free energy, wind energy and solar energy have steadily increased their penetration in the power grid. In the study of wind power forecasting, how to effectively use meteorological data is a key link in improving forecasting accuracy. Since meteorological characteristics are complex and diverse, not all characteristics have a positive impact on the forecast results. Therefore, screening and extracting sensitive characteristic factors that are beneficial to the forecast becomes one of the key steps. However, for newly built energy sites with insufficient data, the accuracy of the forecast is often affected. Summary of the invention

[0003] In order to address the deficiencies of the prior art, the present invention proposes a wind and solar power generation power prediction method based on meteorological feature screening and transfer learning.

[0004] The technical solution adopted by the present invention is as follows:

[0005] In a first aspect, the present invention discloses a method for predicting wind and solar power generation based on meteorological feature screening and transfer learning, comprising the following steps:

[0006] 1) Establish a combined trust index for screening sensitive meteorological characteristic factors, and screen out sensitive meteorological characteristic factors;

[0007] 2) Construct a source domain prediction model, which includes a temporal convolutional network based on a soft multi-head attention mechanism, a long short-term memory neural network based on a skip connection mechanism, and a fully connected neural network;

[0008] 3) Pre-training the source domain prediction model using the sensitive meteorological characteristic factors to obtain a pre-trained source domain prediction model;

[0009] 4) Obtain the historical power generation data of the photovoltaic power station for wind and solar power generation prediction, and construct a target domain data set. Based on the target domain data set, transfer learning and parameter fine-tuning are performed on the pre-trained source domain prediction model to obtain a target domain prediction model for predicting wind and solar power generation;

[0010] 5) Use the target domain prediction model to predict wind and solar power generation.

[0011] In a second aspect, the present invention discloses a wind and solar power generation power prediction system for implementing the method, comprising:

[0012] A screening module, which is used to establish a combined trust index for screening sensitive meteorological characteristic factors, and screen out sensitive meteorological characteristic factors;

[0013] The source domain model building module is used to build a source domain prediction model. The source domain prediction model is based on a temporal convolutional network with a soft multi-head attention mechanism, a long short-term memory neural network based on a skip connection mechanism, and a fully connected neural network.

[0014] A source domain model pre-training module, which is used to pre-train the source domain prediction model using the sensitive meteorological characteristic factors to obtain a pre-trained source domain prediction model;

[0015] The target domain prediction model acquisition module is used to obtain the historical power generation data of the photovoltaic power station for wind and solar power generation prediction, and to construct a target domain data set. Based on the target domain data set, the pre-trained source domain prediction model is transferred and parameterized to obtain a target domain prediction model for predicting wind and solar power generation.

[0016] The prediction module is used to predict the wind and solar power generation power using the target domain prediction model.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] By combining trust indicators, the most representative meteorological factors can be effectively screened out and the differences in the feature selection process can be reduced; the combined prediction model of time series convolutional network, soft multi-head attention mechanism and long short-term memory network can realize the effective extraction of power generation characteristics and efficiently model complex time series data; through the transfer learning method, the power generation characteristic knowledge of the source domain is transferred to the modeling task of the target domain, which improves the adaptability and prediction accuracy of the prediction model in a data-scarce environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the prediction method of the present invention;

[0020] Figure 2 is a schematic diagram of a residual unit in a source domain prediction model of the present invention;

[0021] Figure 3 The wind power prediction results under transfer learning and non-transfer learning methods are shown in the figure;

[0022] Figure 4 The results of photovoltaic power prediction under transfer learning and non-transfer learning methods are shown in the figure;

[0023] Figure 5 It is a structural diagram of the wind and solar power generation power prediction system of the present invention. DETAILED DESCRIPTION

[0024] The present invention is further described and illustrated below in conjunction with specific embodiments. The embodiments are merely exemplary of the present disclosure and do not define the scope of limitation. The technical features of each embodiment of the present invention may be combined accordingly without conflicting with each other.

[0025] The wind power and photovoltaic data of the present invention are derived from the annual large-scale new energy base planning scheme in a certain area in western Inner Mongolia, China in 2030. The source domain dataset used is the actual power and meteorological element data collected from a photovoltaic power station and a wind farm in the southern part of the Kubuqi Desert within 3 years. The target domain dataset contains the actual power and measured meteorological elements of the newly built new energy units in the Kubeihui (photovoltaic) and Hangjinhui (wind power) areas from March 1, 2024 to March 30, 2024, with a sampling interval of 1 hour. The dataset is divided into a training set and a test set in a ratio of 7:3.

[0026] like Figure 1 As shown, step 1: 1.1) Establish a combined trust index for screening sensitive meteorological characteristic factors; the combined trust index includes the Pearson correlation coefficient , mutual information entropy index and Spearman correlation coefficient .

[0027] Pearson correlation index : The Pearson correlation index reflects the waveform similarity of two data sets and can be used to measure the conflict between meteorological data and power.

[0028] ;

[0029] in, Pearson correlation index. The closer its value is to 1, the stronger the positive correlation between the two data is. The closer it is to -1, the stronger the negative correlation between the two data is. is a meteorological characteristic sequence; is the wind and solar power sequence; is the dimension of the corresponding sequence; is the mean of the meteorological characteristic sequence; is the mean value of the wind and solar power series; is the meteorological feature of the ith dimension in the meteorological feature sequence; is the i-th dimension wind-solar power in the wind-solar power sequence.

[0030] Mutual information entropy index: Mutual information provides a measurement index for analyzing the correlation between random variables, but information itself cannot be measured. Therefore, the present invention again uses the concept of entropy and uses mutual information entropy as an index to quantify the uncertainty between variables.

[0031] ;

[0032] in, for and The mutual information entropy index of for Information entropy of for Information entropy of for and The mutual information entropy of is the time series period. The mutual information entropy index reflects the correlation between two variables. The smaller the correlation, the smaller the mutual information entropy index. When the two variables are independent of each other, the mutual information entropy index is the smallest (equal to zero).

[0033] Spearman correlation index: Spearman correlation index It is a non-parametric rank statistical indicator that determines the degree of association between two parameters. Its calculation method is similar to the Pearson correlation coefficient, but it uses the ranking of the variables instead of the original value.

[0034] ;

[0035] in, is the value level of the meteorological feature of the i-th dimension in the meteorological feature sequence; is the value level of the i-th dimension wind-solar power in the wind-solar power sequence; is the mean value of the meteorological characteristic sequence; is the average value of the wind and solar power sequence.

[0036] The combined trust index can evaluate the relationship between variables from multiple perspectives by combining multiple indicators such as the Spearman correlation coefficient, the Pearson correlation coefficient, and the mutual information entropy index. The Spearman correlation coefficient focuses on analyzing nonlinear and monotonic relationships, the Pearson correlation coefficient focuses on the strength of linear relationships, and the mutual information entropy index provides an additional perspective on data uncertainty and complexity. By integrating these different correlation indicators, the potential relationship between data can be revealed more comprehensively and deeply, and the accuracy, robustness, and applicability of correlation analysis can be improved, especially in complex systems.

[0037] 1.2) Using the virtual cloud model and based on the Pearson correlation coefficient , mutual information entropy index and Spearman correlation coefficient Calculate and get the estimated expectation , estimated entropy and the estimated excess entropy ; and perform weighted summation on the three results to obtain the meteorological characteristics The corresponding combined trust.

[0038] The meteorological characteristics The corresponding combined trusts are sorted from high to low, and the first k meteorological characteristics whose cumulative combined trust values ​​reach the preset threshold are obtained to obtain sensitive meteorological characteristic factors, thereby completing the screening of sensitive meteorological characteristic factors.

[0039] The combined trust index proposed in this section is based on the construction of the virtual cloud model. The virtual cloud model is a mathematical model designed to represent fuzzy concepts and can achieve effective conversion between qualitative and quantitative information. The model describes the characteristics of fuzzy concepts and quantifies the degree of uncertainty through three numerical features: expectation, entropy, and hyperentropy. The "cloud" in the model is composed of a series of data points distributed around the expected value, but with a certain entropy and hyperentropy, which reflects the fuzziness and uncertainty in real concepts. The cloud model performs well in processing fuzzy and uncertain information and is widely used in fuzzy systems, intelligent decision-making, and complex system analysis.

[0040] In specific application scenarios, the digital features of each basic cloud are first calculated through the virtual cloud model, and these features are further integrated to generate new digital features. This paper uses the Pearson correlation coefficient, mutual information entropy index, and Spearman correlation coefficient, which are interrelated. They can be regarded as "cloud drops" in the cloud model, and they are combined and operated by the reverse cloud generator to finally obtain the combined trust.

[0041] First, the evaluation index matrix is ​​constructed based on the Pearson correlation coefficient, mutual information entropy index and Spearman correlation coefficient. , the formula is:

[0042] ;

[0043] in, represents the Pearson correlation coefficient; represents the mutual information entropy index; represents the Spearman correlation coefficient;

[0044] The virtual cloud model includes two processes: the forward generator and the reverse cloud emitter. For the forward generator, firstly, based on the evaluation index matrix Calculating Expectation ,entropy and super entropy (expect ,entropy and super entropy is a digital feature), the calculation formulas are:

[0045] ;

[0046] ;

[0047] ;

[0048] In the formula, , are the maximum and minimum values ​​of the evaluation index matrix respectively; is a constant coefficient matrix.

[0049] Next, the above expectations and super entropy Convert to For expectation, The absolute value of V cloud droplet samples as the standard deviation ( , ), thus forming a cloud model. Among them:

[0050] ;

[0051] In the formula, represents the ordinate of the cloud droplet sample; For For expectations, is a random number with standard deviation; represents the horizontal coordinate of the cloud droplet sample, So For expectation, The absolute value of is the random number of the standard deviation; V is the total number of cloud droplet samples.

[0052] For the reverse cloud emitter, it is necessary to calculate V cloud droplet samples ( , ) , estimated entropy and the estimated excess entropy First, we need to calculate V cloud droplet samples ( , ) , the second-order sample center distance and the sample variance , the calculation formulas are:

[0053] ;

[0054] ;

[0055] ;

[0056] In the formula, is the horizontal coordinate of the vth cloud droplet sample.

[0057] Next, the estimated expectation is calculated using the mean, second-order sample center distance, and sample variance , estimated entropy and the estimated excess entropy :

[0058] ;

[0059] ;

[0060] ;

[0061] In summary, through the feedback of the virtual cloud model, the corrected digital features can be obtained , that is, to obtain the estimated expectation of the estimated , estimated entropy and the estimated excess entropy In addition, in the combined trust, whether the advantages and disadvantages of each indicator can be effectively integrated, the reasonable setting of the weight coefficient is crucial. In order to solve the hierarchical needs of complex evaluation problems, this paper adopts the fuzzy hierarchical analysis method and combines the expert knowledge base to prioritize the sample individuals. Under the guidance of expert experience, firstly, based on Sort by size; if The same, then according to To distinguish, the smaller the entropy value, the higher the homogeneity of the sample individuals, the more stable the sample, and the higher the ranking; if they are still the same, further consideration , The smaller it is, the lower the randomness of the sample individuals is, and the higher the ranking is.

[0062] Based on this logic, a fuzzy complementary judgment matrix between the target layer and the criterion layer is constructed. It varies between 0 and 1, and the sum of the elements symmetric about the diagonal in the fuzzy complementary judgment matrix is ​​0, and the elements on the diagonal are 0.5.

[0063] Calculation of fuzzy consistency matrix based on fuzzy complementary judgment matrix , the calculation formula is:

[0064] ;

[0065] in, is the element of row i and column g in the fuzzy consistent matrix; is the element of row j and column g in the fuzzy consistent matrix; .

[0066] Next, the fuzzy consistent matrix Calculate the weight coefficient, the expression is:

[0067] ;

[0068] in, They are , and ; is the estimated super entropy The weight coefficient of is the estimated entropy The weight coefficient of The expected The weight coefficient of is a constant.

[0069] According to the three weight coefficients obtained, the expected , estimated entropy and the estimated excess entropy , and obtain the combined trust , combined trust The calculation formula is:

[0070] ;

[0071] The virtual cloud model is used to optimize the combination of the three evaluation indicators and calculate the final combined trust. By introducing the concept of cumulative contribution rate (i.e. sorting the combined trust index of meteorological characteristics from high to low), meteorological characteristics with a cumulative contribution rate of more than 85% are selected, and the screening of sensitive meteorological characteristic factors is successfully achieved.

[0072] Table 1 shows the screening results of source domain meteorological characteristics. Source domain meteorological characteristics are divided into wind power meteorological characteristics and photovoltaic meteorological characteristics. The wind power meteorological characteristics include cloud coverage, radiation flux, wind speed and wind direction at different heights, etc. Photovoltaic meteorological characteristics include sunshine intensity, photovoltaic panel temperature detection, global horizontal irradiance and ultraviolet index, etc. Among the wind power meteorological characteristics, wind speed contributes the most and large-scale precipitation contributes the least; among the photovoltaic characteristics, sunshine intensity contributes the most and air pressure contributes the least, indicating that the screening strategy results of the present invention are reasonable. This is because wind speed and wind turbine power generation are in a cubic relationship, while precipitation has little effect on equipment and only has an indirect effect on equipment such as turbines or sensors. Photovoltaic power generation is determined by sunshine intensity. The stronger the sunshine, the higher the power generation. However, changes in air pressure are usually accompanied by the movement of weather systems. Although areas with low air pressure may be accompanied by more clouds or rainfall, the impact of air pressure itself on photovoltaic power generation is indirect, so the contribution is relatively low. In addition, sensitive meteorological characteristic factors with a cumulative contribution of more than 85% in the wind power and photovoltaic tables are selected as inputs for subsequent source domain prediction models. The analysis shows that by screening out key sensitive meteorological characteristic factors, this not only reduces the complexity of the input features but also ensures the preservation of key information, thereby reducing the time required for source domain prediction model training to a certain extent.

[0073]

[0074] Step 2: Establish a source domain prediction model, which includes a temporal convolutional network based on a soft multi-head attention mechanism, a long short-term memory neural network based on a skip connection mechanism, and a fully connected neural network.

[0075] a) Temporal Convolutional Network Based on Soft Multi-Head Attention Mechanism

[0076] The temporal convolutional network is an algorithm that can be used to solve time series prediction. It combines the dilated causal convolution layer and the residual connection structure, and has the following characteristics: ① It can still effectively capture the inherent influence relationship of non-continuous data and historical data with a long time interval. ② It will not omit historical information and use future information. ③ Information is not easily lost when processing complex time series.

[0077] The output power of wind power and photovoltaic clusters is a typical complex time series structure with a long time span, and the output power prediction at each moment should only consider the power information relative to the previous moment. The time series convolutional network has a high adaptability to the extraction of these features. Therefore, when designing the source domain prediction model, the time series convolutional network is selected to first process the source domain sensitive meteorological characteristic factors to obtain the weight coefficient matrix; specifically, the sensitive meteorological characteristic factors constitute a sensitive meteorological characteristic sequence, and the time series convolutional network takes the sensitive meteorological characteristic sequence as input and outputs the weight coefficient matrix.

[0078] The calculation formula of the dilated convolution of the temporal convolutional network is:

[0079] ;

[0080] In the formula, The data output of the expanded convolutional network; It is a sensitive meteorological characteristic sequence; is the dilated convolution position; is the convolution kernel; is the convolution operation; is the convolution kernel size; is the expansion coefficient; is the data input at time W-di, is the moment W; is the ith filter.

[0081] The multi-head attention mechanism uses multiple attention heads to independently calculate and splice to capture different features, allowing the model to focus on input information from multiple perspectives. On this basis, this paper proposes a soft multi-head attention mechanism that allows different heads to dynamically adjust weights, thereby flexibly identifying and focusing on more effective information in different tasks to improve prediction performance. This mechanism is applied to the time domain convolutional network (TCN) to enhance the robustness of the model and accelerate convergence. The TCN based on the soft multi-head attention mechanism constitutes the residual unit of the source domain prediction model (see Appendix Figure 2 ), first convolve the input sensitive meteorological features, then obtain the weight change matrix, then obtain the average weight change matrix, and finally obtain the weight coefficient matrix, that is, obtain the final output.

[0082] In this architecture, the soft multi-head attention mechanism not only improves the model's ability to identify key moments, but also ensures the stable propagation of gradients through residual units, avoiding the gradient vanishing problem, thereby significantly accelerating the convergence of training. In addition, the soft multi-head attention mechanism calculates multiple weight change matrices in parallel and uses Softmax to generate probability distribution, so that the model can smoothly distribute attention at each time step, avoid single focus, and better adapt to dynamic time series data. The specific manifestations are as follows:

[0083] Sensitive meteorological characteristic factors constitute multiple sensitive meteorological characteristic sequences and are input into the temporal convolutional network based on the soft multi-head attention mechanism. The temporal convolutional network convolves the input multiple sensitive meteorological characteristic sequences to obtain the convolution results of each sensitive meteorological characteristic sequence, and then uses the soft multi-head attention mechanism to perform multiple linear changes on all the convolution results to obtain the weight change matrix :

[0084] ;

[0085] in, is the weight change matrix obtained by the sth linear change; is the characteristic value of the mth sensitive meteorological characteristic sequence after the sth linear change;

[0086] Calculate the weight change matrix The average value of the average weight change matrix :

[0087] ;

[0088] in, is the total number of linear changes; To carry out The average characteristic value of the mth sensitive meteorological characteristic sequence after sublinear change;

[0089] Then activate the average weight change matrix The weight coefficient is restricted to a non-negative interval, and then the weight coefficient is normalized by the activation function Sigmoid; the normalized weight coefficient is then restricted to a non-negative interval and normalized, and the final weight coefficient matrix is ​​obtained. ; The specific calculation formula is:

[0090] ;

[0091] Among them, Relu is a linear rectification function; is the weight coefficient of the mth sensitive meteorological feature sequence.

[0092] b) Long short-term memory neural network based on skip connection mechanism

[0093] Long short-term memory neural network (LSTM) can effectively process long-period time series and retain historical information that conforms to the characteristics of the sequence. In the present invention, the research modeling of wind and solar power belongs to a time series problem with correlations on adjacent time scales. LSTM can well extract the changing characteristics in these time associations. Therefore, LSTM is added to the pre-trained network to model and predict the long-term and short-term dependencies in the sequence. In addition, on this basis, a LSTM based on a jump connection mechanism is designed to process the weight coefficient matrix output by the time series convolutional network based on the soft multi-head attention mechanism. After the feature is processed by nonlinear transformation, it is superimposed with the original input and output. The jump connection mechanism can not only maximize information utilization and diversified information transmission, but also effectively solve the difficulties in model training, especially the gradient explosion problem.

[0094] Take the first process of model training as an example:

[0095] The input of the long short-term memory neural network is the weight coefficient matrix output by the temporal convolutional network, and the output of the long short-term memory neural network is , is the nonlinear mapping inside the long short-term memory neural network based on the skip connection mechanism. The final fitting target of the network is , H is the nonlinear mapping inside the ideal long short-term memory neural network. With the introduction of the jump mechanism in this paper, the target is modified to ,get The introduction of such a jump mechanism can enhance the flexibility of the model, and a small disturbance can be reflected in the output.

[0096] c) Fully connected neural network

[0097] The fully connected neural network takes the output of the long short-term memory neural network based on the skip connection mechanism as input, and maps the output of the long short-term memory neural network based on the skip connection mechanism to obtain the predicted wind and solar power generation power.

[0098] Step 3, pre-training the source domain prediction model using the sensitive meteorological characteristic factors to obtain a pre-trained source domain prediction model;

[0099] In view of the randomness of the prediction results of the source domain prediction model, the training error standard deviation evaluation index is introduced as the loss function of the source domain prediction model, that is, as the loss function of the long short-term memory neural network; the calculation formula of the loss function is:

[0100] ;

[0101] in, It is the training error standard deviation evaluation indicator; is the total number of sensitive meteorological characteristic factors in the sensitive meteorological characteristic sequence; is the actual wind and solar power generation power corresponding to the i-th dimension sensitive meteorological characteristic factor in the sensitive meteorological characteristic sequence; The wind and solar power generation power is predicted by the source domain prediction model for the i-th dimension sensitive meteorological characteristic factor in the sensitive meteorological characteristic sequence; is the mean training error.

[0102] Taking minimizing the loss function as the goal, the updated long short-term memory neural network is obtained, that is, the pre-trained source domain prediction model is obtained.

[0103] Step 4: Obtain historical power generation data of the photovoltaic power station for wind and solar power generation prediction, and construct a target domain data set. Based on the target domain data set, transfer learning and parameter fine-tuning are performed on the pre-trained source domain prediction model to obtain a target domain prediction model for predicting wind and solar power generation;

[0104] The present invention uses the pre-trained source domain prediction model and the rich meteorological data in the source domain wind and solar database to realize the migration of public knowledge of power generation mode. In order to improve the efficiency of migration model training, the structure of the target domain prediction model and the parameters in the model are consistent with the structure of the migrated source domain model and the parameters in the model; after the model migration is completed, the parameters of the target domain model are fine-tuned based on the target domain data set to obtain the target domain prediction model; wherein, when fine-tuning the parameters of the target domain model, the network parameter optimization objective function of the target domain is used as the loss function to fine-tune the parameters of the network.

[0105] The objective function of network parameter optimization in the target domain includes minimizing the power generation prediction error of the target wind power and photovoltaic clusters and minimizing the representation divergence between the source domain and the target domain migration model. The mathematical expression is:

[0106] ;

[0107] In the formula, is the target domain dataset; is the data in the target domain dataset; is the predicted value of wind and solar power generation power of the target domain data set; is the actual value of wind and solar power generation power in the target domain dataset; are the parameters of the network in the target domain model during the lth training; is the convergence condition of the objective function; is the parameter to be optimized; is the L2 norm.

[0108] The network parameter optimization objective function of the target domain is used as an additional objective function of network training in the target domain model, with the goal of keeping the public knowledge part unchanged as much as possible. The specificity of wind and solar power output in the target domain is captured to the maximum extent through retraining of the fully connected layer, and the optimal frozen pre-trained network parameters are obtained; by setting a smaller learning rate for the feature extraction layer, the parameter changes of the temporal convolutional network layer based on the soft multi-head attention mechanism and the long short-term memory neural network layer based on the jump connection mechanism of the extracted features in the network structure are minimized, so as to achieve the effect of fine-tuning the pre-trained network parameters.

[0109] After the source domain data is preprocessed, sensitive meteorological characteristic factors are obtained, which are input into the source domain prediction model for pre-training of the source domain prediction model. The pre-trained source domain prediction model is adapted to one month of data in the target domain dataset for parameter fine-tuning, and finally a target domain prediction model adapted to the target domain is formed.

[0110] Figure 3 The prediction results of wind power and photovoltaic power under transfer learning and non-transfer learning methods are shown. Figure 4 The prediction results of photovoltaic power under transfer learning and non-transfer learning methods are shown. The difference between the target domain wind and solar power curve generated by the model and the actual power curve is calculated through indicators such as mean absolute error (MAE), root mean square error (RMSE) and fitting coefficient (R-square, R²). The relevant calculation formulas are shown in the following three formulas to measure the accuracy of the model migration effect.

[0111] ;

[0112] ;

[0113] ;

[0114] in, is the number of prediction points, is the actual value of wind and solar power generation at the i-th prediction point, is the predicted value of wind and solar power generation at the i-th prediction point; It is the average value of wind and solar power generation at all prediction points.

[0115] MAE, RMSE, and R of the wind power test set 2 The errors are 0.0811, 0.1114, and 0.8895, respectively. The MAE, RMSE, and R 2 The errors are 0.0184, 0.0379, and 0.9771, respectively. The errors mainly occur when the power generation curve value rises or falls sharply. The analysis shows that the prediction effect of the transfer learning model is good. It not only accurately predicts the overall trend, but also has a relatively high prediction accuracy at most precise time nodes. The overall verification result is good.

[0116] like Figure 5 As shown, the present invention also discloses a wind and solar power generation power prediction system based on meteorological feature screening and transfer learning, comprising:

[0117] A screening module, which is used to establish a combined trust index for screening sensitive meteorological characteristic factors, and screen out sensitive meteorological characteristic factors;

[0118] The source domain model building module is used to build a source domain prediction model. The source domain prediction model is based on a temporal convolutional network with a soft multi-head attention mechanism, a long short-term memory neural network based on a skip connection mechanism, and a fully connected neural network.

[0119] A source domain model pre-training module, which is used to pre-train the source domain prediction model using the sensitive meteorological characteristic factors to obtain a pre-trained source domain prediction model;

[0120] The target domain prediction model acquisition module is used to obtain the historical power generation data of the photovoltaic power station for wind and solar power generation prediction, and to construct a target domain data set. Based on the target domain data set, the pre-trained source domain prediction model is transferred and parameterized to obtain a target domain prediction model for predicting wind and solar power generation.

[0121] The prediction module is used to predict the wind and solar power generation power using the target domain prediction model.

[0122] For the system embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment, and the implementation methods of the remaining modules will not be repeated here. The system embodiment described above is only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of the present invention. Ordinary technicians in this field can understand and implement it without paying creative work.

[0123] The embodiments of the system of the present invention can be applied to any device with data processing capabilities, and the device with data processing capabilities can be a device or apparatus such as a computer. The system embodiments can be implemented by software, or by hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, the corresponding computer program instructions in the non-volatile memory are read into the memory by the processor of any device with data processing capabilities and run.

[0124] The above description of the embodiments is to facilitate those skilled in the art to understand and apply the present invention. The present invention is not limited to the above embodiments, and improvements and modifications made by those skilled in the art based on the disclosure of the present invention should be within the protection scope of the present invention.

Claims

1. A wind and solar power generation power prediction method based on meteorological feature screening and transfer learning, characterized in that: The following steps are involved: 1) Establish a combined trust index for screening sensitive meteorological characteristic factors, and screen out sensitive meteorological characteristic factors; 2) Construct a source domain prediction model, which includes a temporal convolutional network based on a soft multi-head attention mechanism, a long short-term memory neural network based on a skip connection mechanism, and a fully connected neural network; 3) Pre-training the source domain prediction model using the sensitive meteorological characteristic factors to obtain a pre-trained source domain prediction model; In step 3), the sensitive meteorological characteristic factors constitute a sensitive meteorological characteristic sequence, and the temporal convolutional network takes multiple sensitive meteorological characteristic sequences as input and outputs a weight coefficient matrix; including: The temporal convolutional network convolves each sensitive meteorological feature sequence to obtain the convolution results of each sensitive meteorological feature sequence; then the soft multi-head attention mechanism is used to perform multiple linear changes on all the convolution results to obtain the weight change matrix , ； in, is the weight change matrix obtained by the sth linear change; is the characteristic value of the mth sensitive meteorological characteristic sequence after the sth linear change; Calculate the weight change matrix The average value of the average weight change matrix : ； in, is the total number of linear changes; To carry out The average characteristic value of the mth sensitive meteorological characteristic sequence after sublinear change; Activation mean weight change matrix The weight coefficient is restricted to a non-negative interval, and then the weight coefficient is normalized by the activation function Sigmoid; the normalized weight coefficient is restricted to a non-negative interval and normalized in turn to obtain the weight coefficient matrix ; The specific calculation formula is: ； Among them, Relu is a linear rectification function; is the weight coefficient of the mth sensitive meteorological feature sequence; 4) Obtain the historical power generation data of the photovoltaic power station for wind and solar power generation prediction, and construct a target domain data set. Based on the target domain data set, transfer learning and parameter fine-tuning are performed on the pre-trained source domain prediction model to obtain a target domain prediction model for predicting wind and solar power generation; 5) Use the target domain prediction model to predict wind and solar power generation.

2. The method for predicting wind and solar power generation based on meteorological feature screening and transfer learning according to claim 1 is characterized in that: In step 1), the combined trust index includes the Pearson correlation coefficient , mutual information entropy index and Spearman correlation coefficient ; The Pearson correlation coefficient The calculation formula is: ； in, is the meteorological characteristic sequence; is the wind and solar power sequence; is the dimension of the meteorological characteristic sequence or the wind and solar power sequence; is the meteorological feature of the ith dimension in the meteorological feature sequence; is the i-th dimension wind-solar power in the wind-solar power sequence; is the mean of the meteorological characteristic sequence; is the mean value of the wind and solar power series; The calculation formula of the mutual information entropy index is: ； in, for and The mutual information entropy index of for Information entropy of for Information entropy of for and The mutual information entropy of is the time series period; The Spearman correlation coefficient The calculation formula is: ； in, is the value level of the meteorological feature of the i-th dimension in the meteorological feature sequence; is the value level of the i-th dimension wind-solar power in the wind-solar power sequence; is the mean value of the meteorological characteristic sequence; is the average value of the wind and solar power sequence.

3. The method for predicting wind and solar power generation based on meteorological feature screening and transfer learning according to claim 2 is characterized in that: In step 1), the screening of sensitive meteorological characteristic factors includes: a) Using the virtual cloud model and based on the Pearson correlation coefficient , mutual information entropy index and Spearman correlation coefficient Calculate and get the expected value , estimated entropy and the estimated excess entropy ; Perform weighted summation on the three results to obtain meteorological characteristics The corresponding combined trust; b) Meteorological characteristics The corresponding combined trusts are sorted from high to low, and the first k meteorological characteristics whose cumulative combined trust values ​​reach the preset threshold are obtained to obtain sensitive meteorological characteristic factors, thereby completing the screening of sensitive meteorological characteristic factors.

4. The method for predicting wind and solar power generation based on meteorological feature screening and transfer learning according to claim 3 is characterized in that: In step a), the estimated expectation is obtained by calculating , estimated entropy and the estimated excess entropy ;include: Construct an evaluation index matrix based on Pearson correlation coefficient, mutual information entropy index and Spearman correlation coefficient , the formula is: ； in, represents the Pearson correlation coefficient; represents the mutual information entropy index; represents the Spearman correlation coefficient; Calculating Expectation ,entropy and super entropy , the formulas are: ； ； ； in, is the maximum value of the evaluation index matrix; is the minimum value of the evaluation index matrix; is a constant coefficient matrix; The expectation and super entropy Convert to For expectation, The absolute value of V cloud droplet samples as the standard deviation ( , ); ； in, represents the ordinate of the cloud droplet sample; For For expectations, is a random number with standard deviation; represents the horizontal coordinate of the cloud droplet sample, So For expectation, The absolute value of is a random number with standard deviation; V is the total number of cloud droplet samples; Calculate the mean of cloud droplet samples , the second-order sample center distance and the sample variance , the calculation formulas are: ； ； ； in, is the horizontal coordinate of the vth cloud droplet sample; Expectations ,entropy and super entropy Make an estimate and get the expected value , estimated entropy and the estimated excess entropy , the calculation formulas are: ； ； 。 5. The method for predicting wind and solar power generation based on meteorological feature screening and transfer learning according to claim 4 is characterized in that: In step a), obtaining meteorological characteristics The corresponding combined trust level includes: Construct a fuzzy complementary judgment matrix, where the elements in the fuzzy complementary judgment matrix It varies between 0 and 1, and the sum of the elements symmetrical about the diagonal in the fuzzy complementary judgment matrix is ​​0, and the elements on the diagonal are 0.5; The fuzzy consistency matrix is ​​calculated based on the fuzzy complementary judgment matrix , the calculation formula is: ； in, is the element of row i and column g in the fuzzy consistent matrix; is the element of row j and column g in the fuzzy consistent matrix; ; Next, the weight coefficient is calculated through the fuzzy consistency matrix, and the calculation formula is: ； in, They are , and ; is the estimated super entropy The weight coefficient of is the estimated entropy The weight coefficient of The expected The weight coefficient of is a constant; According to the three weight coefficients obtained, the expected , estimated entropy and the estimated excess entropy , get the combined trust , the calculation formula is: 。 6. The method for predicting wind and solar power generation based on meteorological feature screening and transfer learning according to claim 1 is characterized in that: In step 3), the long short-term memory neural network takes the weight coefficient matrix as input and outputs , It is a nonlinear mapping inside the long short-term memory neural network based on the skip connection mechanism; the output of the long short-term memory neural network is input to the fully connected neural network for mapping to obtain the predicted wind and solar power generation power; a loss function is constructed, and the formula of the loss function is: ； in, It is the training error standard deviation evaluation indicator; is the total number of sensitive meteorological characteristic factors in each sensitive meteorological characteristic sequence; is the actual wind and solar power generation power corresponding to the i-th dimension sensitive meteorological characteristic factor in the sensitive meteorological characteristic sequence; The wind and solar power generation power is predicted by the source domain prediction model for the i-th dimension sensitive meteorological characteristic factor in the sensitive meteorological characteristic sequence; is the average training error; Taking minimizing the loss function as the goal, the updated long short-term memory neural network is obtained, and the pre-trained source domain prediction model is obtained.

7. The method for predicting wind and solar power generation based on meteorological feature screening and transfer learning according to claim 1 is characterized in that: In step 4), the transfer learning and parameter fine-tuning of the source domain prediction model based on the target domain dataset includes: Directly migrating the pre-trained source domain prediction model to obtain a target domain model, wherein the structure of the target domain model and the parameters in the model are equal to the structure of the pre-trained source domain prediction model and the parameters in the model; Then, the parameters of the target domain model are fine-tuned based on the target domain data set to obtain the target domain prediction model.

8. The method for predicting wind and solar power generation based on meteorological feature screening and transfer learning according to claim 7 is characterized in that: When fine-tuning the parameters of the target domain model, the network parameters are fine-tuned using the network parameter optimization objective function of the target domain as the loss function; the formula of the network parameter optimization objective function of the target domain is: ； in, is the target domain dataset; is the data in the target domain dataset; is the predicted value of wind and solar power generation power of the target domain data set; is the actual value of wind and solar power generation power in the target domain dataset; are the parameters of the network in the target domain model during the lth training; is the target convergence threshold; is the network parameter to be optimized; is the L2 norm.

9. A wind and solar power generation power prediction system for implementing the method described in claim 1, characterized in that: include: A screening module, which is used to establish a combined trust index for screening sensitive meteorological characteristic factors, and screen out sensitive meteorological characteristic factors; The source domain model building module is used to build a source domain prediction model. The source domain prediction model is based on a temporal convolutional network with a soft multi-head attention mechanism, a long short-term memory neural network based on a skip connection mechanism, and a fully connected neural network. A source domain model pre-training module, which is used to pre-train the source domain prediction model using the sensitive meteorological characteristic factors to obtain a pre-trained source domain prediction model; The target domain prediction model acquisition module is used to obtain the historical power generation data of the photovoltaic power station for wind and solar power generation prediction, and to construct a target domain data set. Based on the target domain data set, the pre-trained source domain prediction model is transferred and parameterized to obtain a target domain prediction model for predicting wind and solar power generation. The prediction module is used to predict the wind and solar power generation power using the target domain prediction model.

Images