A non-crossing quantile wind power probability prediction method based on integrated deep learning

Abstract

The application discloses a non-cross quantile wind power probability prediction method based on integrated deep learning, which firstly establishes a quantile regression model based on deep learning by using Huber loss function; then proposes an exponential stack mapping method for avoiding the quantile cross problem of the deep learning quantile regression model; finally, a quantile loss index for evaluating the overall performance of the predicted quantile is constructed for two-stage adaptive integration of the deep learning quantile regression model, and a hybrid integrated model of homogeneous deep learning and heterogeneous deep learning is constructed. The method uses the nonlinear mapping ability and adaptive feature extraction ability of deep learning, uses integrated learning to improve the generalization ability of the prediction model, ensures the strict monotonicity of the predicted quantile, realizes the non-parametric probability prediction of wind power, effectively quantifies the prediction uncertainty, and provides key information support for the safe, reliable and stable operation of a new energy power system.

Description

TECHNICAL FIELD

[0001] The application relates to a non-cross quantile wind power probability prediction method based on integrated deep learning, and belongs to the field of power system power prediction. BACKGROUND

[0002] As one of the most important renewable energy sources, the proportion of wind power in modern power systems has been growing rapidly worldwide. However, the uncertainty of wind power poses a serious challenge to the safe and stable operation of power systems. Accurate and reliable wind power prediction is of great significance to the safe and economic operation of power systems with high wind power penetration.

[0003] Traditional deterministic prediction only provides a single point expected value for future wind power generation. However, due to the chaotic nature of the meteorological system, the prediction error of wind power generation is inevitable. Probability prediction can provide the complete probability distribution of the prediction object and has become a hot topic in the research of new energy power systems, providing key information support for demand response, unit commitment, reserve scheduling, and power market transactions. The classical probability prediction model artificially sets the prediction error distribution of the prediction object, which is difficult to accurately describe the prediction uncertainty. The quantile crossing problem in the non-parametric adaptive quantile regression model has not been solved. SUMMARY

[0004] In order to overcome the shortcomings of the prior art, the purpose of the present application is to provide a non-cross quantile wind power probability prediction method based on integrated deep learning, which avoids the quantile crossing problem in quantile prediction, improves the model generalization ability and overall performance by using two-stage integration, and realizes the non-parametric quantization of wind power prediction uncertainty.

[0005] In order to achieve the above purpose, the application adopts the following technical scheme:

[0006] A non-cross quantile wind power probability prediction method based on integrated deep learning, comprising the following steps.

[0007] (1) Constructing a deep learning quantile regression model based on Huber loss

[0008] The training of the deep learning model relies on the gradient descent method. Since the classical quantile regression loss is not differentiable at zero, the Huber loss function is used as the objective function of the deep learning quantile regression model. The constructed deep learning quantile regression model is represented as:

[0009]

[0010]

[0011]

[0012] where x n is the input vector of the prediction model, y n is the actual power output value, DL∈{M,L,D} represents three heterogeneous ensemble deep learning models of multi-layer perceptron, long short-term memory neural network and deep belief neural network respectively, where f DL is the prediction output quantile vector, and θ DL is the parameter of the deep learning model. is the Huber loss function under a given quantile level α i :

[0013]

[0014] where: is the prediction quantile under a given quantile level α i , and η is an arbitrarily small positive number. ρ is defined as:

[0015]

[0016] (2) Constructing a non-crossing quantile regression probability prediction model based on exponential stack mapping

[0017] The prediction model outputs the prediction quantile under different quantile levels, which satisfies the following constraints:

[0018]

[0019] where α l and α u are quantile levels, is a given set of quantile levels.

[0020] To ensure the above non-crossing characteristics of the prediction quantile, a non-crossing quantile generation strategy based on exponential stack mapping is constructed. First, the deep learning quantile regression model is used to generate quantile elements Then the prediction quantile is generated, which is represented as:

[0021]

[0022] Combining the exponential stack mapping method with the deep learning quantile regression model can realize the output of non-crossing quantiles using the deep learning model, as shown in Figure 1 .

[0023] (3) Constructing a quantile loss index for evaluating the overall performance of quantiles

[0024] A quantile loss index Q n is constructed for evaluating the overall performance of quantiles:

[0025]

[0026] wherein:

[0027]

[0028]

[0029]

[0030]

[0031]

[0032] (4) Constructing a homogeneous ensemble deep learning prediction model based on adaptive integrated weights

[0033] The homogeneous deep learning model ensemble comprises the following steps:

[0034] a. Given a set of homogeneous deep learning model indexes training sample set training sample set index set minimum iteration number N min , maximum iteration number N max , loss index threshold L th , initial training sample sampling weight Homogeneous deep learning model index m = 1;

[0035] b. Construct a new resampling training set according to the training sample sampling weight Train the homogeneous deep learning model f m (·) for N iterations min , set the training number l = N min ;

[0036] c. Use the initial training sample set to test the homogeneous deep learning model f m (·) to calculate the quantile loss index Q n , Continue to train f m (·) once, and update the training number l = l + 1;

[0037] d. Judgment condition: l ≤ N max and If satisfied, repeat step c; if not satisfied, proceed to the subsequent steps;

[0038] e. Define the quantile loss index of the homogeneous deep learning model f m (·) as update the training sample sampling weight wherein updating the homogeneous deep learning model index m = m + 1;

[0039] f. Judgment condition: m <= M, if satisfied, repeat steps b-e, if not satisfied, calculate model integration weight constructing a homogeneous integrated deep learning model G(x n ) = g(f m (x n ; theta m ); tau m ),

[0040] The homogeneous integrated deep learning process is shown in Figure 2

[0041] (5) Constructing a heterogeneous integrated deep learning prediction model based on adaptive integration weight

[0042] The same type of deep learning model is used as the base model in the homogeneous integrated deep learning model, and three types of deep learning models, multilayer perceptron, long short-term memory neural network and deep belief neural network, are selected in the heterogeneous integrated model. After completing the first stage of homogeneous integration, the second stage of heterogeneous model integration is carried out.

[0043] The heterogeneous integrated deep learning model is represented as:

[0044]

[0045] In the formula: lambda DL is the integration weight of the heterogeneous integrated deep learning model, which is defined as:

[0046]

[0047] In the formula: M, L, D respectively represent multilayer perceptron, long short-term memory neural network and deep belief neural network three types of deep learning model, is the quantile loss index obtained by testing the homogeneous integrated deep learning model constructed by using a certain deep learning model on the training set.

[0048] The beneficial effects of the present application are:

[0049] ​This paper proposes a non-crossing quantile wind power probabilistic prediction method based on ensemble deep learning. It utilizes the nonlinear mapping and adaptive feature extraction capabilities of deep learning to construct a nonparametric quantile regression model, achieving nonparametric quantification of wind power prediction uncertainty. An exponential stack mapping method is proposed to ensure that the predicted quantiles monotonically increase with the quantile level, avoiding the quantile crossover problem in traditional quantile regression models. A two-stage ensemble framework is constructed to achieve adaptive ensemble of homogeneous and heterogeneous ensemble deep learning, integrating the advantages of different types of deep learning models and improving the generalization ability of the prediction model. A quantile loss index for evaluating the overall performance of the predicted quantiles is proposed, based on which the adaptive construction of ensemble weights is achieved, avoiding complex modeling of the ensemble model weight optimization problem. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the exponential stack mapping method;

[0051] Figure 2 A flowchart for homogeneous ensemble deep learning;

[0052] Figure 3 Flowchart for integrating deep learning for non-cross quantile wind power probability prediction;

[0053] Figure 4 This is a comparison chart of the probability prediction results and the actual values ​​of this invention. Detailed Implementation

[0054] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0055] See appendix Figure 3 This is the overall flowchart for integrated deep learning non-cross quantile wind power probability prediction. The specific execution process is described below.

[0056] (1) First, given the training dataset Where x n y is the input variable for the prediction model composed of historical wind power measurement data. n The actual value of the wind power to be predicted;

[0057] (2) Construct M deep learning quantile regression models based on multilayer perceptron, long short-term memory neural network, and deep belief neural network, respectively. The loss function of each model is the mean of Hubel loss across all training samples and all quantile levels. Each deep learning quantile regression model is constructed as follows:

[0058]

[0059]

[0060]

[0061] In the formula: x n y is the input vector for the prediction model. n The actual power output value is given by DL∈{M, L, D}, which represent three heterogeneous ensemble deep learning models: multilayer perceptron, long short-term memory neural network, and deep belief neural network, respectively. To predict the output quantile vector, f DL and θ DL These are the deep learning model and its parameters; Given a quantile level α i Hubell loss function:

[0062]

[0063] In the formula: Given a quantile level α i The predicted quantiles below, Given a set of r quantile levels, satisfying α i <α i+1 η is an arbitrarily small positive number. Defined as:

[0064]

[0065] The predicted quantiles are generated using an exponential stack mapping method to ensure the non-crossing property of the predicted quantiles, and are represented as follows:

[0066]

[0067] In the formula: This is the output layer result of a deep learning quantile regression model. The activation function of the output layer of the deep learning quantile regression model uses the ReLU function to ensure non-negativity. The ReLU function is defined as:

[0068]

[0069] The predicted quantiles generated by the above-described non-cross quantile generation strategy based on exponential stack mapping satisfy the monotonically increasing property, that is:

[0070]

[0071] (3)Use Figure 2 The process shown constructs homogeneous ensemble deep learning models G based on the three deep learning models described above. M (x n ), G L (x n ), G D(x n );

[0072] (4) According to the test of the three homogeneous integrated deep learning models on the initial training sample set, obtain the quantile loss index, and determine the integration weight of the heterogeneous integrated deep learning model according to the quantile loss index, to obtain the heterogeneous integrated deep learning model E(x n );

[0073] (5) According to the test data set, the prediction quantile is calculated by using E(x n ), and the reliability and overall performance of the obtained prediction quantile are evaluated by using the following index:

[0074] a. The reliability of the prediction quantile is evaluated according to the average proportion deviation (APD), which is defined as the difference between the empirical proportion and the nominal quantile level α:

[0075]

[0076] In the formula:

[0077]

[0078] The indicator function is defined as:

[0079]

[0080] The smaller the absolute value of the average proportion deviation, the closer the empirical proportion of the prediction quantile to the nominal quantile level, and the higher the prediction reliability.

[0081] b. The sharpness of the prediction quantile can be reflected by the comprehensive performance score, and the commonly used quantile skill score S Q is expressed as:

[0082]

[0083] The larger the quantile skill score S Q , the better the sharpness and overall performance of the prediction quantile.

[0084] c. The quantile crossing can be measured by the crossing degree index (CDI), which is defined as:

[0085]

[0086] In the formula: determines whether the quantile crossing occurs, (α w -α v) to characterize the degree of quantile crossing. The smaller the quantile crossing degree, the less the crossing of predicted quantiles. The quantile crossing degree of zero means that there is no quantile crossing.

[0087] The wind power data of the Barrow Offshore Wind Farm in East Irish Sea, Cambria, UK in 2021 is selected to verify the effectiveness of the method, the time resolution of the data set is 30 minutes, the ratio of the training set, the validation set and the test set is about 6:2:2, and the prediction target is the prediction quantile with a nominal quantile level from 5% to 95% with an interval of 5% in advance for 1 hour.

[0088] Table 1 compares the average proportional bias indicators of the method, the bootstrap neural network, the classical quantile regression and the quantile regression neural network, and the results show that the predicted quantiles obtained by the method exhibit outstanding reliability at different quantile levels.

[0089] Table 1

[0090]

[0091] Unit: %

[0092] Table 2 compares the quantile skill score and the quantile crossing degree indicator, and the results show that the method has higher sharpness and comprehensive performance. Since the bootstrap neural network generates a prediction distribution based on the Gaussian distribution assumption and then calculates the prediction quantile, the quantile crossing degree is 0. Compared with the quantile regression method, the quantile crossing degree indicator of the predicted quantile obtained by the method is 0, and the method has strict monotonicity.

[0093] Table 2

[0094]

[0095] In summary, the method constructs a non-crossing quantile wind power probability prediction method based on integrated deep learning, uses the Huber loss function to construct a deep learning quantile regression model, ensures the non-crossing property of the predicted quantile through the exponential stack mapping method, determines the integration weight of the homogeneous integrated deep learning and the heterogeneous integrated deep learning model by using the quantile loss indicator, realizes the adaptive integration of different types of deep learning models, avoids the complex optimization process of the integration weight, and the two-stage integrated deep learning framework improves the overall prediction performance and generalization ability of the method.

[0096] The specific embodiments of the application are described above with reference to the drawings, which are not a limitation on the scope of protection of the application, and all equivalent models or equivalent algorithm processes made by using the content of the specification and drawings of the application are within the scope of patent protection of the application.

Claims

1. An integrated deep learning-based non-crossing quantile wind power probability prediction method, characterized in that, The method is: a deep learning quantile regression model based on Huber loss is established; on this basis, a non-cross quantile regression probability prediction model is proposed; a two-stage integration method is proposed to realize adaptive integration of homogeneous deep learning models and heterogeneous deep learning models; The deep learning quantile regression model based on Huber loss is as follows: ， In the formula: For the input vector of the prediction model, This is the actual power output value. These represent three heterogeneous ensemble deep learning models: multilayer perceptron, long short-term memory neural network, and deep belief neural network. To predict the output quantile vector, and These are the deep learning model and its parameters; For a given quantile level Hubell loss function: ， where: is the predicted quantile at a given quantile level , is the set of predicted quantiles at a given set of quantile levels , , is an arbitrarily small positive number, is defined as: ； The method for constructing the non-cross quantile regression probability prediction model comprises a non-cross quantile generation strategy based on an exponential stack mapping, first generating non-negative quantile elements by using a deep learning quantile regression model and then generating a predicted quantile: ， wherein: is the output layer result of the deep learning quantile regression model, the activation function of the output layer of the deep learning quantile regression model adopts a ReLU function to ensure non-negativity, the ReLU function is defined as: , The predicted quantile output by the non-cross quantile generation strategy based on the exponential stack mapping satisfies the monotone increasing characteristic, that is: ； The two-stage integrated method is to build a homogeneous integrated deep learning model in the first stage and build a heterogeneous integrated deep learning model in the second stage : ， ， wherein: is a set of indices of homogeneous deep learning models, and denote ensemble weights of homogeneous ensemble deep learning and heterogeneous ensemble deep learning models, respectively, denotes an ensemble method of homogeneous ensemble deep learning; The ensemble weights of the homogeneous ensemble deep learning model and the heterogeneous ensemble deep learning model are determined according to the proposed quantile loss index for evaluating the overall performance of quantile prediction determined, defined as: ， In the formula: ， ， ； The integration method of the homogeneous deep learning model is average integration, and the calculation formula of the integration weight is: ； The integration method of the heterogeneous deep learning model is average integration, and the calculation formula of the integration weight is: ， In the formula: and respectively, are the quantile loss indicators calculated using the validation set in the homogeneous ensemble deep learning model and the heterogeneous ensemble deep learning model.

Images