Wind power short-term power forecasting method and device based on nonlinear weighted combination

Abstract

The present invention provides a method and device for predicting short-term wind power based on nonlinear weighted combination. The method comprises obtaining the historical output power of a wind farm; decomposing the historical output power into a number of intrinsic mode functions based on an empirical mode decomposition method; establishing a low-frequency prediction model using an LSTM neural network according to the several intrinsic mode functions; establishing a high-frequency prediction model using an improved sparrow search algorithm through a deep belief network according to the several intrinsic mode functions; integrating the prediction results of the low-frequency prediction model and the prediction results of the high-frequency prediction model using a neural network established based on an improved sparrow search algorithm-deep belief network to obtain a final prediction model for wind power. The present invention provides an effective method for extracting wind power characteristics, which can effectively improve the accuracy of short-term wind power prediction and achieve better prediction accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind power prediction, and in particular relates to a method and a device for predicting short-term wind power based on nonlinear weighted combination. Background Art

[0002] With the development of the energy revolution, renewable energy has attracted the attention of the world with its low-carbon, pollution-free and inexhaustible characteristics. Among them, wind energy has become one of the most promising alternative energy sources in the world with its clean, abundant and renewable advantages. However, the inherent characteristics of wind energy such as randomness, volatility and intermittency make the wind power sequence highly nonlinear and non-stationary, which may cause fluctuations in voltage and frequency in the power system, especially when large-scale wind power is integrated into the power system. Therefore, accurate wind power forecasting is crucial to improving the security of the power system and making full use of wind energy resources.

[0003] To date, researchers have developed a variety of wind power prediction models, mainly including two categories: physical-driven models and data-driven models. Among them, the physical-driven model relies on the numerical weather prediction system (NWP) and requires detailed modeling of the process of converting wind energy into electrical energy, which not only increases the computational burden but also makes it difficult to apply to short-term wind power prediction. Unlike the physical-driven model, the data-driven model only considers the mapping between input variables and target variables, which is very suitable for short-term wind speed / power prediction. However, how to further improve the accuracy of wind power prediction by the data-driven model has become an urgent problem to be solved. Summary of the invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art and provides a method and device for short-term wind power prediction based on nonlinear weighted combination.

[0005] In one aspect of the present invention, a method for short-term wind power forecasting based on nonlinear weighted combination is provided, the method comprising:

[0006] Get the historical output power of the wind farm;

[0007] Based on an empirical mode decomposition method, decomposing the historical output power into a number of intrinsic mode functions;

[0008] According to the several intrinsic mode functions, a low-frequency prediction model is established using an LSTM neural network;

[0009] According to the plurality of intrinsic mode functions, a high-frequency prediction model is established through a deep belief network using an improved sparrow search algorithm;

[0010] The prediction results of the low-frequency prediction model and the high-frequency prediction model are integrated by using a neural network established based on an improved sparrow search algorithm-deep belief network to obtain a final prediction model for wind power.

[0011] In some embodiments, the historical output power is decomposed into a number of intrinsic mode functions based on the empirical mode decomposition method, including:

[0012] The historical output power is used as a training sample to determine the original sequence X(t), obtain all local maximum and local minimum values ​​of X(t), and interpolate using a cubic spline function to obtain the upper envelope e max (t) and the lower envelope e min (t), and the average of the two is obtained to obtain the envelope average curve m1(t);

[0013] Subtract the envelope mean curve m1(t) from the original sequence X(t) to obtain the average value sequence h1(t);

[0014] Determine whether h1(t) satisfies the conditions for the establishment of the intrinsic mode function. If not, take h1(t) as the new X(t) and repeat the above steps until h1(t) satisfies the conditions for the establishment of the intrinsic mode function. At this time, take h1(t) as the first-order intrinsic mode function decomposed, and separate h1(t) from X(t) to obtain r1(t), where r1(t) = X(t) - h1(t);

[0015] Replace X(t) with r1(t) and decompose the new intrinsic mode function. The intrinsic mode functions obtained after each decomposition are: r2(t)=r1(t)-h2(t),...,r n (t) = r n-1 (t)-h n (t);

[0016] The decomposition ends when any of the following conditions are met: 1)r n (t) or h n (t) is less than the preset threshold, 2)r n (t) is a monotonic function, and it is impossible to screen out the intrinsic mode function from it;

[0017] After the decomposition is completed, all the intrinsic mode functions and residuals are accumulated and added to obtain

[0018] In some embodiments, the method of establishing a low-frequency prediction model using an LSTM neural network based on the plurality of intrinsic mode functions includes:

[0019] The LSTM neural network is trained using the low-frequency intrinsic mode function among the plurality of intrinsic mode functions to obtain the low-frequency prediction model, wherein the mathematical expression of the LSTM neural network is:

[0020]

[0021] Among them, f t ,i t ,o t are the outputs of the forget gate, input gate, and output gate respectively. represents the candidate cell state, c t ,h t and c t-1 ,h t-1 Represents the cell state and cell output at the current time t and the previous time t-1, respectively, t is the input of the LSTM unit, is the predicted output, w f ,w i ,w c ,w o ,W and b f ,b i ,b c ,b o , b is the weight matrix and bias vector, ο is the scalar multiplication, and σ is the sigmoid activation function.

[0022] In some embodiments, the method of establishing a high-frequency prediction model through a deep belief network using an improved sparrow search algorithm based on the plurality of intrinsic mode functions includes:

[0023] Obtaining optimal weights and biases of the deep belief network using the improved sparrow search algorithm;

[0024] Substituting the optimal weights and biases into the deep belief network, training the high-frequency intrinsic mode functions among the plurality of intrinsic mode functions, and obtaining the high-frequency prediction model.

[0025] In some implementations, the improved sparrow search algorithm comprises the following steps:

[0026] Initialize the parameters of the improved sparrow search algorithm, the parameters including the population size N, the number of discoverers pNum, the number of sparrows responsible for early warning sNum, the dimension D of the objective function, the upper and lower bounds of the initial value, the maximum number of iterations and the solution accuracy;

[0027] Calculate the fitness value of each sparrow, select the current best fitness value and its corresponding position, as well as the worst fitness value and its corresponding position;

[0028] A sparrow is randomly selected from the population as sNum. After one iteration, the fitness value of each sparrow is calculated again, and the position of the discoverer is updated as follows:

[0029] In the formula, t represents the current number of iterations, j represents the dimension, and i termax Represents the maximum number of iterations, α is a random number in (0,1], It is a random number that follows a normal distribution. ST represents the safety value, and R2 represents the warning value. R2<ST means that the population is in a safe area and the discoverer can forage randomly. R2≥ST means that there are predators around the population and it is necessary to move to a safe area immediately for foraging.

[0030] Update the joiner's position as follows:

[0031] Where, X P Indicates the optimal position of the discoverer at present, X worse represents the worst position in the current population, A represents a 1×d matrix, and each element is randomly assigned 1 or -1, i>n / 2 means that the i joiners with lower fitness values ​​do not get food and need to go to other places to find food;

[0032] Update the scout's location as follows:

[0033] In the formula, β represents the step size control parameter, which is a random number that obeys a normal distribution with a mean of 0 and a variance of 1. best Indicates that the sparrow is currently in the safest position, f i represents the fitness value of the current sparrow individual, f i and f g Represent the current best and worst fitness values ​​respectively. When f i >f g When , it means that the sparrow is at the edge and is most threatened. i =f g When , it means that the sparrow in the middle is aware of the danger and tries to get closer to the nearby sparrows. γ is the minimum constant to avoid the denominator being zero.

[0034] Perform elite reverse learning on the sparrow population to obtain the latest sparrow population x i = k*(a i +b i )-x i , where x i ∈[a i ,b i ], k∈[0,1] is a random number that obeys uniform distribution, where a i 、b iare the maximum and minimum positions of the x population respectively;

[0035] According to the current state of the sparrow population, the optimal position, the worst position and the fitness value of the entire population are updated, and it is determined whether the maximum number of iterations or the solution condition is reached. If the maximum number of iterations or the solution condition is reached, the optimal value is output; if the maximum number of iterations or the solution condition is not reached, the steps after initializing the parameters of the improved sparrow search algorithm are repeated.

[0036] In some implementations, the step of obtaining the optimal weights and biases of the deep belief network using the improved sparrow search algorithm includes:

[0037] The improved sparrow search algorithm is used to optimize the weights and biases of the deep belief network, and the fitness of the optimization process is set to the root mean square error to obtain the optimal weights and biases of the deep belief network, wherein the root mean square error is expressed as: In the formula, y i is the real wind power, y' i is the wind power predicted by the extreme learning machine, and N is the predicted wind power points.

[0038] In some embodiments, the method further comprises:

[0039] A test sample is obtained from the historical output power, and the final prediction model is verified using the test sample.

[0040] Another aspect of the present invention provides a wind power short-term power prediction device based on nonlinear weighted combination, the device comprising:

[0041] An acquisition module is used to obtain the historical output power of the wind farm;

[0042] A decomposition module, used for decomposing the historical output power into a number of intrinsic mode functions based on an empirical mode decomposition method;

[0043] A first building module is used to build a low-frequency prediction model using an LSTM neural network according to the plurality of intrinsic mode functions;

[0044] A second establishment module is used to establish a high-frequency prediction model through a deep belief network using an improved sparrow search algorithm according to the plurality of intrinsic mode functions;

[0045] The integration module is used to integrate the prediction results of the low-frequency prediction model and the prediction results of the high-frequency prediction model by using a neural network established based on an improved sparrow search algorithm-deep belief network to obtain a final prediction result of wind power.

[0046] Another aspect of the present invention provides an electronic device, comprising:

[0047] one or more processors;

[0048] A storage unit is used to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors can implement the method described above.

[0049] Another aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program can implement the method described above when executed by a processor.

[0050] The wind power short-term power prediction method and device based on nonlinear weighted combination of the present invention, taking into account the strong volatility and non-stationary characteristics of wind power sequences, proposes to use empirical mode decomposition method to extract data features of wind power and decompose it into several intrinsic mode functions, and respectively uses the decomposed data to establish a prediction model using LSTM neural network and improved sparrow search algorithm-deep belief network as subsequence predictors, thereby providing an effective method for extracting wind power features; in view of the limitations of a single deep learning method, the sparrow search algorithm is improved to enhance and optimize the performance of the algorithm, and the improved sparrow search algorithm is used to optimize the weights and bias of the deep belief network, which can effectively improve the accuracy of wind power short-term power prediction and achieve better prediction accuracy; and in order to overcome the shortcomings of linear combination, the present invention proposes a hybrid prediction model based on the nonlinear combination mechanism of the improved sparrow search algorithm-deep belief network, thereby further improving the accuracy of wind power short-term power prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a flow chart of the wind power short-term power prediction method based on nonlinear weighted combination proposed by the present invention;

[0052] Figure 2 A flow chart of the empirical mode decomposition method adopted by the present invention;

[0053] Figure 3 This is a schematic diagram of the structure of the LSTM neural network used in the present invention;

[0054] Figure 4 This is a schematic diagram of the structure of the deep belief network used in the present invention;

[0055] Figure 5 It is a structural schematic diagram of the nonlinear combination mechanism based on the ESSA-DBN network adopted in the present invention;

[0056] Figure 6This is a schematic diagram of the structure of the wind power short-term power forecast based on nonlinear weighted combination proposed by the present invention. DETAILED DESCRIPTION

[0057] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0058] In one aspect of this embodiment, Figure 1 As shown, it relates to a method S100 for short-term wind power prediction based on nonlinear weighted combination, and the method S100 includes:

[0059] S110. Obtain historical output power of the wind farm.

[0060] Specifically, in this step, the historical output power data of a 50MW wind farm can be selected, where the data interval is 15 minutes, and the annual data is divided into four quarters: spring, summer, autumn, and winter. The original wind power data of the four quarters are preprocessed, and the interpolation method is used to remove abnormal data and fill in the missing data. The first 70% of the data in each quarter are used as training samples, and the remaining 30% of the data are used as test samples.

[0061] S120. Decomposing the historical output power into a plurality of intrinsic mode functions based on an empirical mode decomposition method.

[0062] Specifically, the Empirical Mode Decomposition (EMD) method was proposed by a Chinese American in 1998. This method can be used to analyze nonlinear and non-stationary signal sequences.

[0063] The EMD process is as follows Figure 2 As shown in the figure, as a signal processing method, the empirical mode decomposition method decomposes multi-component signals into a series of quasi-single signal components, namely intrinsic mode functions (IMFs), based on the characteristics of different time scales. The decomposed IMF should meet two conditions: 1) The time series of zero-crossing points in the entire signal is the same as or differs by one in the number of maximum and minimum points. 2) The mean of the upper envelope fitted by the local maximum point and the lower envelope fitted by the local minimum point is zero. These two conditions are introduced to overcome the unnecessary fluctuation phenomenon introduced by the asymmetric conditions, so that the instantaneous frequency defined by the Hilbert spectrum has practical significance everywhere.

[0064] Step S120 may specifically include the following steps:

[0065] The historical output power is used as a training sample to determine the original sequence X(t), obtain all local maximum and local minimum values ​​of X(t), and interpolate using a cubic spline function to obtain the upper envelope emax (t) and the lower envelope e min (t), and the average of the two is obtained to obtain the envelope average curve m1(t);

[0066] Subtract the envelope mean curve m1(t) from the original sequence X(t) to obtain the average value sequence h1(t);

[0067] Determine whether h1(t) satisfies the conditions for the establishment of the intrinsic mode function. If not, take h1(t) as the new X(t) and repeat the above steps until h1(t) satisfies the conditions for the establishment of the intrinsic mode function. At this time, take h1(t) as the first-order intrinsic mode function decomposed, and separate h1(t) from X(t) to obtain r1(t), where r1(t) = X(t) - h1(t);

[0068] Replace X(t) with r1(t) and decompose the new intrinsic mode function. The intrinsic mode functions obtained after each decomposition are: r2(t)=r1(t)-h2(t),...,r n (t) = r n-1 (t)-h n (t);

[0069] The decomposition ends when any of the following conditions are met: 1)r n (t) or h n (t) is less than the preset threshold, 2)r n (t) is a monotonic function, and it is impossible to screen out the intrinsic mode function from it;

[0070] After the decomposition is completed, all the intrinsic mode functions and residuals are accumulated and added to obtain

[0071] S130. Establish a low-frequency prediction model using an LSTM neural network according to the plurality of intrinsic mode functions.

[0072] Specifically, in order to alleviate the shortcomings of shallow neural networks, such as local optimization, overfitting, and weak generalization performance, deep learning networks have been applied to the field of wind power prediction as machine learning. Among them, LSTM neural network is a special recurrent neural network (RNN) that effectively solves the problems of gradient explosion and gradient disappearance in classic RNN networks. LSTM neural network consists of a "gated structure" (input gate, forget gate, output gate) with short-term memory and a unit state with long-term memory. Its structure is as follows Figure 3 shown.

[0073] Step S130 may specifically include the following steps:

[0074] The LSTM neural network is trained using the low-frequency intrinsic mode function among the plurality of intrinsic mode functions to obtain the low-frequency prediction model, wherein the mathematical expression of the LSTM neural network is:

[0075]

[0076] Among them, f t ,i t ,o t are the outputs of the forget gate, input gate, and output gate respectively. represents the candidate cell state, c t ,h t and c t-1 ,h t-1 Represents the cell state and cell output at the current time t and the previous time t-1, respectively, t is the input of the LSTM unit, is the predicted output, w f ,w i ,w c ,w o ,W and b f ,b i ,b c ,b o , b is the weight matrix and bias vector, ο is the scalar multiplication, and σ is the sigmoid activation function.

[0077] S140. According to the plurality of intrinsic mode functions, a high-frequency prediction model is established through a deep belief network using an improved sparrow search algorithm.

[0078] Specifically, step S140 may include the following steps:

[0079] Obtaining optimal weights and biases of the deep belief network using the improved sparrow search algorithm;

[0080] Substituting the optimal weights and biases into the deep belief network, training the high-frequency intrinsic mode functions among the plurality of intrinsic mode functions, and obtaining the high-frequency prediction model.

[0081] Deep Belief Networks (DBN) is a probabilistic generative model that consists of a series of restricted Boltzmann machines (RBM) and logistic regression layers. The RMB network consists of a visualization layer trained with data and a hidden layer with feature detection capabilities. More specifically, the underlying DBN prediction model uses a greedy algorithm for unsupervised learning training, specifically, the parameters obtained from the first layer of meta-structure training are used as the input of the second layer of meta-structure. The purpose of this process is to extract sample features more efficiently. In addition, the top layer of DBN uses the classic BP (Back Propagation) neural network structure for regression prediction. The structure of DBN is as follows: Figure 3 As shown. Among them, x t , are the input and predicted output respectively. {w1,w2,...,w n} is the weight matrix, v1 and {h1,h2,...,h n} represent the weight matrix input to the hidden layer and the bias of the hidden layer respectively.

[0082] The Sparrow Search Algorithm (SSA) is a new type of swarm intelligence optimization algorithm proposed by Xue Jiankai et al. in 2019, inspired by the foraging and anti-predation behaviors of sparrows. In the SSA algorithm, sparrows in the population are usually divided into discoverers and joiners, and their identities change dynamically. Whenever a sparrow becomes a discoverer, another sparrow must become a joiner. In addition, each sparrow will be given an initial position and a fitness determined by a fitness function. In the SSA algorithm, the size of the fitness value indicates the ability of the discoverer to search for food. The position of the discoverer in the iteration process is updated as follows:

[0083]

[0084] Where: t represents the current number of iterations; j represents the dimension; i termax Indicates the maximum number of iterations; α is a random number in (0,1); It is a random number that obeys the normal distribution; ST represents the safety value; R2 represents the warning value. R2<ST means that the population is in a safe area and the discoverer can forage randomly. R2≥ST means that there are predators around the population and it is necessary to immediately move to a safe area for foraging.

[0085] If the joiners in the population notice that the discoverer has found better food, they will immediately compete for the food. If the joiners succeed in the competition, they can get the food of the discoverer, otherwise they will continue to monitor the discoverer's foraging. The position of the joiners is updated as follows:

[0086]

[0087] In the above formula: X P Indicates the optimal position of the discoverer at present; X worse Represents the worst position in the current population; A represents a 1×d matrix, and each element is randomly assigned 1 or -1; i>n / 2 means that the i participants with lower fitness values ​​do not get food and need to go to other places to find food.

[0088] In addition, there are predators in the sparrow algorithm. The number of sparrows that find predators in the entire population accounts for 10% to 20%. The initial positions of these sparrows are randomly generated, and the expression is as follows:

[0089]

[0090] In the above formula: β represents the step size control parameter, which is a random number that follows a normal distribution with a mean of 0 and a variance of 1; X best Indicates that the sparrow is currently in the safest position; f i Indicates the fitness value of the current sparrow individual; f i and f g Represent the current best and worst fitness values ​​respectively. When f i >f g When , it means that the sparrow is at the edge and is under the greatest threat; i =f g This means that the sparrow in the middle is aware of the danger and tries to get closer to the nearby sparrows; γ is the minimum constant to avoid the denominator being zero.

[0091] The improved sparrow search algorithm (ESSA) introduces a reverse solution into the standard SSA algorithm, which greatly expands the search range of the algorithm compared with before.

[0092] In order to effectively improve the convergence speed of the algorithm, the reverse solution of the current solution is first obtained, and then the individuals whose original solution fitness value is greater than the reverse solution fitness value are found according to the fitness value, and they are formed into an elite group, and then a new search space is generated in the elite group to find the reverse solution of the individuals whose original solution fitness value is less than the reverse solution fitness value. When the algorithm finds the optimal solution, it will inevitably find the area where the optimal solution is located, and then generate a reverse solution in the interval defined by the elite group, so as to search for the optimal solution.

[0093] Elite Opposite Solution (EOS) In N-dimensional space, the elite individual x of the current group best =(x1,x2,...,x N ) can be expressed as x′ best =(x′1,x′2,...,x′ N ), which can be defined as xi = k*(a i +b i )-x i , where x i ∈[a i ,b i ], k∈[0,1] is a random number that obeys uniform distribution, where a i 、b i They are the maximum and minimum positions of the x population respectively. Using k can generate multiple reverse solutions of elite individuals, which can effectively improve the mining ability of the algorithm.

[0094] Based on the above content, the improved sparrow search algorithm includes the following steps:

[0095] Initialize the parameters of the improved sparrow search algorithm, the parameters including the population size N, the number of discoverers pNum, the number of sparrows responsible for early warning sNum, the dimension D of the objective function, the upper and lower bounds of the initial value, the maximum number of iterations and the solution accuracy;

[0096] Calculate the fitness value of each sparrow, select the current best fitness value and its corresponding position, as well as the worst fitness value and its corresponding position;

[0097] A sparrow is randomly selected from the population as sNum. After one iteration, the fitness value of each sparrow is calculated again, and the position of the discoverer is updated as follows:

[0098] In the formula, t represents the current number of iterations, j represents the dimension, and i termax Represents the maximum number of iterations, α is a random number in (0,1], It is a random number that follows a normal distribution. ST represents the safety value, and R2 represents the warning value. R2<ST means that the population is in a safe area and the discoverer can forage randomly. R2≥ST means that there are predators around the population and it is necessary to move to a safe area immediately for foraging.

[0099] Update the joiner's position as follows:

[0100] Where, X P Indicates the optimal position of the discoverer at present, X worse represents the worst position in the current population, A represents a 1×d matrix, and each element is randomly assigned 1 or -1, i>n / 2 means that the i joiners with lower fitness values ​​do not get food and need to go to other places to find food;

[0101] Update the scout's location as follows:

[0102] In the formula, β represents the step size control parameter, which is a random number that obeys a normal distribution with a mean of 0 and a variance of 1. best Indicates that the sparrow is currently in the safest position, f i represents the fitness value of the current sparrow individual, f i and f g Represent the current best and worst fitness values ​​respectively. When f i >f g When , it means that the sparrow is at the edge and is most threatened. i =f g When , it means that the sparrow in the middle is aware of the danger and tries to get closer to the nearby sparrows. γ is the minimum constant to avoid the denominator being zero.

[0103] Perform elite reverse learning on the sparrow population to obtain the latest sparrow population x i = k*(a i +b i )-x i , where x i ∈[a i ,b i ], k∈[0,1] is a random number that obeys uniform distribution, where a i 、b i are the maximum and minimum positions of the x population respectively;

[0104] According to the current state of the sparrow population, the optimal position, the worst position and the fitness value of the entire population are updated, and it is determined whether the maximum number of iterations or the solution condition is reached. If the maximum number of iterations or the solution condition is reached, the optimal value is output; if the maximum number of iterations or the solution condition is not reached, the steps after initializing the parameters of the improved sparrow search algorithm are repeated.

[0105] Furthermore, the step of obtaining the optimal weights and biases of the deep belief network by using the improved sparrow search algorithm includes:

[0106] The improved sparrow search algorithm is used to optimize the weights and biases of the deep belief network, and the fitness of the optimization process is set to the root mean square error to obtain the optimal weights and biases of the deep belief network, wherein the root mean square error is expressed as: In the formula, y i is the real wind power, y' i is the wind power predicted by the extreme learning machine, and N is the predicted wind power points.

[0107] Specifically, in this step, the weights and biases of DBN are optimized using the improved sparrow search algorithm in DBN. The fitness selected in the optimization process is the root mean square error (RMSE), where the formula of RMSE is as follows:

[0108]

[0109] Among them, y i is the real wind power, y' i is the wind power predicted by ELM (Extreme Learning Machine), and N is the predicted wind power points.

[0110] S150, integrating the prediction results of the low-frequency prediction model and the prediction results of the high-frequency prediction model by using a neural network established based on an improved sparrow search algorithm-deep belief network to obtain a final prediction model for wind power.

[0111] Specifically, since most current studies have shown that hybrid prediction models based on decomposition strategies are superior to single models, but are affected by the combination mechanism, the existing linear combination mechanism ignores the inherent nonlinear relationship between the models and cannot meet the needs of high-precision prediction. Therefore, the present invention uses a hybrid model based on a nonlinear combination mechanism to aggregate the prediction results of a single model. Specifically, after obtaining the prediction results of the two individual models, the predictions of the two models are aggregated through a new nonlinear combination mechanism based on the ESSA-DBN network. The structure of the nonlinear combination mechanism is as follows: Figure 4 In this step, the essence of the nonlinear combination mechanism is the secondary prediction of the prediction results of the high-frequency component and the low-frequency component of the wind farm output power.

[0112] The wind power short-term power prediction method based on nonlinear weighted combination of the present invention takes into account the strong volatility and non-stationary characteristics of wind power sequences, proposes to use empirical mode decomposition method to extract data features of wind power and decompose it into several intrinsic mode functions, and uses the decomposed data to respectively establish prediction models with LSTM neural network and improved sparrow search algorithm-deep belief network as subsequence predictors, thereby providing an effective method for extracting wind power features; in view of the limitations of a single deep learning method, the sparrow search algorithm is improved to enhance and optimize the performance of the algorithm, and the improved sparrow search algorithm is used to optimize the weights and bias of the deep belief network, which can effectively improve the accuracy of wind power short-term power prediction and achieve better prediction accuracy; and, in order to overcome the shortcomings of linear combination, the present invention proposes a hybrid prediction model based on the nonlinear combination mechanism of the improved sparrow search algorithm-deep belief network, thereby further improving the accuracy of wind power short-term power prediction.

[0113] Furthermore, the method S100 may further include the following steps:

[0114] A test sample is obtained from the historical output power, and the final prediction model is verified using the test sample.

[0115] Another aspect of the present invention, as Figure 6 As shown, a wind power short-term power forecasting device 100 based on nonlinear weighted combination is provided, and the device 100 is applicable to the method described above. The device 100 comprises:

[0116] An acquisition module 110 is used to acquire the historical output power of the wind farm;

[0117] A decomposition module 120, configured to decompose the historical output power into a plurality of intrinsic mode functions based on an empirical mode decomposition method;

[0118] A first establishing module 130 is used to establish a low-frequency prediction model using an LSTM neural network according to the plurality of intrinsic mode functions;

[0119] A second establishing module 140 is used to establish a high-frequency prediction model through a deep belief network using an improved sparrow search algorithm according to the plurality of intrinsic mode functions;

[0120] The integration module 150 is used to integrate the prediction results of the low-frequency prediction model and the prediction results of the high-frequency prediction model by using a neural network established based on an improved sparrow search algorithm-deep belief network to obtain a final prediction result of wind power.

[0121] The wind power short-term power prediction device based on nonlinear weighted combination of the present invention takes into account the strong volatility and non-stationary characteristics of wind power sequences, proposes to use empirical mode decomposition method to extract data features of wind power and decompose it into several intrinsic mode functions, and uses the decomposed data to respectively establish prediction models with LSTM neural network and improved sparrow search algorithm-deep belief network as subsequence predictors, thereby providing an effective device for extracting wind power features; in view of the limitations of a single deep learning method, the sparrow search algorithm is improved to enhance and optimize the performance of the algorithm, and the improved sparrow search algorithm is used to optimize the weights and bias of the deep belief network, which can effectively improve the accuracy of wind power short-term power prediction and achieve better prediction accuracy; and, in order to overcome the shortcomings of linear combination, the present invention proposes a hybrid prediction model based on the nonlinear combination mechanism of the improved sparrow search algorithm-deep belief network, thereby further improving the accuracy of wind power short-term power prediction.

[0122] Another aspect of the present invention provides an electronic device, comprising:

[0123] one or more processors;

[0124] A storage unit is used to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors can implement the method described above.

[0125] Another aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program can implement the method described above when executed by a processor.

[0126] The computer-readable medium may be included in the apparatus, device, or system of the present invention, or may exist independently.

[0127] Among them, computer-readable storage media can be any tangible media that contains or stores a program, which can be an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or equipment. More specific examples include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, an optical fiber, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0128] The computer-readable storage medium may also include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code, specific examples of which include but are not limited to electromagnetic signals, optical signals, or any suitable combination thereof.

[0129] It is to be understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of the present invention, but the present invention is not limited thereto. For those of ordinary skill in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A wind power short-term power forecasting method based on nonlinear weighted combination, characterized in that: The method comprises: Get the historical output power of the wind farm; Based on an empirical mode decomposition method, decomposing the historical output power into a number of intrinsic mode functions; According to the several intrinsic mode functions, a low-frequency prediction model is established using an LSTM neural network; According to the plurality of intrinsic mode functions, a high-frequency prediction model is established through a deep belief network using an improved sparrow search algorithm; The prediction results of the low-frequency prediction model and the high-frequency prediction model are integrated by using a neural network established based on an improved sparrow search algorithm-deep belief network to obtain a final prediction model of wind power; The method based on the empirical mode decomposition is to decompose the historical output power into a number of intrinsic mode functions, including: The historical output power is used as a training sample to determine the original sequence X(t), obtain all local maximum and local minimum values ​​of X(t), and interpolate using a cubic spline function to obtain the upper envelope e max (t) and the lower envelope e min (t), and the average of the two is obtained to obtain the envelope average curve m1(t); Subtract the envelope mean curve m1(t) from the original sequence X(t) to obtain the average value sequence h1(t); Determine whether h1(t) satisfies the conditions for the establishment of the intrinsic mode function. If not, take h1(t) as the new X(t) and repeat the above steps until h1(t) satisfies the conditions for the establishment of the intrinsic mode function. At this time, take h1(t) as the first-order intrinsic mode function decomposed, and separate h1(t) from X(t) to obtain r1(t), where r1(t) = X(t) - h1(t); Replace X(t) with r1(t) and decompose the new intrinsic mode function. The intrinsic mode functions obtained after each decomposition are: r2(t)=r1(t)-h2(t),...,r n (t) = r n-1 (t)-h n (t); The decomposition ends when any of the following conditions are met: 1)r n (t) or h n (t) is less than the preset threshold, 2)r n (t) is a monotonic function, and it is impossible to screen out the intrinsic mode function from it; After the decomposition is completed, all the intrinsic mode functions and residuals are accumulated and added to obtain The method of establishing a low-frequency prediction model using an LSTM neural network according to the plurality of intrinsic mode functions includes: The LSTM neural network is trained using the low-frequency intrinsic mode function among the plurality of intrinsic mode functions to obtain the low-frequency prediction model, wherein the mathematical expression of the LSTM neural network is: Among them, f t ,i t ,o t are the outputs of the forget gate, input gate, and output gate respectively. represents the candidate cell state, c t ,h t and c t-1 ,h t-1 Represents the cell state and cell output at the current time t and the previous time t-1, respectively, t is the input of the LSTM unit, is the predicted output, w f ,w i ,w c ,w o ,W and b f ,b i ,b c ,b o ,b is the weight matrix and bias vector, is scalar multiplication, σ is the sigmoid activation function; The method of establishing a high-frequency prediction model based on the plurality of intrinsic mode functions using an improved sparrow search algorithm through a deep belief network includes: Obtaining optimal weights and biases of the deep belief network using the improved sparrow search algorithm; Substituting the optimal weights and biases into the deep belief network, training the high-frequency intrinsic mode functions among the plurality of intrinsic mode functions, and obtaining the high-frequency prediction model.

2. The method according to claim 1, characterized in that The improved sparrow search algorithm comprises the following steps: Initialize the parameters of the improved sparrow search algorithm, the parameters including the population size N, the number of discoverers pNum, the number of sparrows responsible for early warning sNum, the dimension D of the objective function, the upper and lower bounds of the initial value, the maximum number of iterations and the solution accuracy; Calculate the fitness value of each sparrow, select the current best fitness value and its corresponding position, as well as the worst fitness value and its corresponding position; A sparrow is randomly selected from the population as sNum. After one iteration, the fitness value of each sparrow is calculated again, and the position of the discoverer is updated as follows: where t represents the current iteration number, j represents the dimension, and i termax represents the maximum iteration number, α is a random number belonging to (0, 1], belongs to a random number following a normal distribution, ST represents the safety value, R2 represents the warning value, R2 < ST indicates that the population is in a safe area and the discoverers can forage randomly, R2 ≥ ST indicates that there are predators around the population and it is necessary to immediately transfer to a safe area for foraging; Update the joiner's position as follows: Where, X P Indicates the optimal position of the discoverer at present, X worse represents the worst position in the current population, A represents a 1×d matrix, and each element is randomly assigned 1 or -1, i>n / 2 means that the i joiners with lower fitness values ​​do not get food and need to go to other places to find food; Update the scout's location as follows: In the formula, β represents the step size control parameter, which is a random number that obeys a normal distribution with a mean of 0 and a variance of 1. best Indicates that the sparrow is currently in the safest position, f i represents the fitness value of the current sparrow individual, f i and f g Represent the current best and worst fitness values ​​respectively. When f i >f g When , it means that the sparrow is at the edge and is most threatened. i =f g When , it means that the sparrow in the middle is aware of the danger and tries to get closer to the nearby sparrows. γ is the minimum constant to avoid the denominator being zero. Perform elite reverse learning on the sparrow population to obtain the latest sparrow population x i = k*(a i +b i )-x i , where x i ∈[a i ,b i ], k∈[0,1] is a random number that obeys uniform distribution, where a i 、b i are the maximum and minimum positions of the x population respectively; According to the current state of the sparrow population, the optimal position, the worst position and the fitness value of the entire population are updated, and it is determined whether the maximum number of iterations or the solution condition is reached. If the maximum number of iterations or the solution condition is reached, the optimal value is output; if the maximum number of iterations or the solution condition is not reached, the steps after initializing the parameters of the improved sparrow search algorithm are repeated.

3. The method according to claim 2, characterized in that The step of using the improved sparrow search algorithm to obtain the optimal weight and bias of the deep belief network includes: The improved sparrow search algorithm is used to optimize the weights and biases of the deep belief network, and the fitness of the optimization process is set to the root mean square error to obtain the optimal weights and biases of the deep belief network, wherein the root mean square error is expressed as: In the formula, y i is the real wind power, y i is the wind power predicted by the extreme learning machine, and N is the predicted wind power points.

4. The method according to any one of claims 1 to 3, characterized in that: The method further comprises: A test sample is obtained from the historical output power, and the final prediction model is verified using the test sample.

5. A wind power short-term power prediction device based on nonlinear weighted combination, characterized in that: The device comprises: An acquisition module is used to obtain the historical output power of the wind farm; A decomposition module, used for decomposing the historical output power into a number of intrinsic mode functions based on an empirical mode decomposition method; A first building module is used to build a low-frequency prediction model using an LSTM neural network according to the plurality of intrinsic mode functions; A second establishment module is used to establish a high-frequency prediction model through a deep belief network using an improved sparrow search algorithm according to the plurality of intrinsic mode functions; An integration module, used to integrate the prediction results of the low-frequency prediction model and the prediction results of the high-frequency prediction model using a neural network established based on an improved sparrow search algorithm-deep belief network, so as to obtain a final prediction result of wind power; The method based on the empirical mode decomposition is to decompose the historical output power into a number of intrinsic mode functions, including: The historical output power is used as a training sample to determine the original sequence X(t), obtain all local maximum and local minimum values ​​of X(t), and interpolate using a cubic spline function to obtain the upper envelope e max (t) and the lower envelope e min (t), and the average of the two is obtained to obtain the envelope average curve m1(t); Subtract the envelope mean curve m1(t) from the original sequence X(t) to obtain the average value sequence h1(t); Determine whether h1(t) satisfies the conditions for the establishment of the intrinsic mode function. If not, take h1(t) as the new X(t) and repeat the above steps until h1(t) satisfies the conditions for the establishment of the intrinsic mode function. At this time, take h1(t) as the first-order intrinsic mode function decomposed, and separate h1(t) from X(t) to obtain r1(t), where r1(t) = X(t) - h1(t); Replace X(t) with r1(t) and decompose the new intrinsic mode function. The intrinsic mode functions obtained after each decomposition are: r2(t)=r1(t)-h2(t),...,r n (t) = r n-1 (t)-h n (t); The decomposition ends when any of the following conditions are met: 1)r n (t) or h n (t) is less than the preset threshold, 2)r n (t) is a monotonic function, and it is impossible to screen out the intrinsic mode function from it; After the decomposition is completed, all the intrinsic mode functions and residuals are accumulated and added to obtain The method of establishing a low-frequency prediction model using an LSTM neural network according to the plurality of intrinsic mode functions includes: The LSTM neural network is trained using the low-frequency intrinsic mode function among the plurality of intrinsic mode functions to obtain the low-frequency prediction model, wherein the mathematical expression of the LSTM neural network is: Among them, f t ,i t ,o t are the outputs of the forget gate, input gate, and output gate respectively. represents the candidate cell state, c t ,h t and c t-1 ,h t-1 Represents the cell state and cell output at the current time t and the previous time t-1, respectively, t is the input of the LSTM unit, is the predicted output, w f ,w i ,w c ,w o ,W and b f ,b i ,b c ,b o ,b is the weight matrix and bias vector, is scalar multiplication, σ is the sigmoid activation function; The method of establishing a high-frequency prediction model based on the plurality of intrinsic mode functions using an improved sparrow search algorithm through a deep belief network includes: Obtaining optimal weights and biases of the deep belief network using the improved sparrow search algorithm; Substituting the optimal weights and biases into the deep belief network, training the high-frequency intrinsic mode functions among the plurality of intrinsic mode functions, and obtaining the high-frequency prediction model.

6. An electronic device, characterized in that: include: one or more processors; A storage unit, used to store one or more programs, which, when executed by the one or more processors, enable the one or more processors to implement the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 4 can be implemented.

Images