Multi-step wind speed prediction method based on variable decomposition depth projection coding echo state network

Abstract

The invention provides a multi-step wind speed prediction method based on a variable decomposition depth projection coding echo state network, and the method comprises the steps: 1, obtaining a data set, and carrying out the preprocessing of data; decomposing the wind speed original sequence into a plurality of subsequences with different characteristics by adopting variational mode decomposition (VMD); step 2, establishing a DEESN model; the DEESN model is composed of a plurality of echo state network ESN modules and an extreme learning machine ELM encoder, each ESN module is responsible for asynchronous advanced forecasting, and the ELM encoder between two sub-reservoirs is used for reducing dimensions and optimizing reservoir states; for Q-step advanced wind speed prediction, combining a Qth ESN module and a (Q-1) th ELM module by the DEESN model, and determining parameters of the DEESN model through a trial and error method; and step 3, performing multi-step wind speed prediction by using the DEESN model to obtain a prediction result. According to the method, the VMD-DEESN model is established, the mapping capability is improved, the time dependence between future time step length learning is improved, and then the modeling precision and efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wind speed prediction, and in particular to a multi-step wind speed prediction method based on a variational deep projection coded echo state network. Background Art

[0002] As the demand for energy increases year by year, the world's non-renewable resources are gradually withering, the difficulty of resource mining is increasing, and the energy crisis is expanding. In order to reduce the consumption of traditional energy and the emission of harmful substances, it is necessary to develop renewable energy. As a renewable energy, wind energy has the advantages of being clean, environmentally friendly, easy to obtain, and abundant in reserves, and the use of wind energy is gradually increasing around the world. However, due to the influence of time-varying meteorological factors, wind energy is intermittent, unstable, and volatile, resulting in unstable wind power generation and reducing the conversion rate of wind energy. In addition, the violent fluctuation of wind speed will also affect the stability of the grid voltage. Therefore, it is of great significance to accurately predict the wind speed.

[0003] At present, there are many methods for wind speed prediction, and patents have been applied for, such as the application publication number [CN110414045A], the invention name is a short-term wind speed prediction method based on VMD-GRU; the application publication number is [CN111950759A], the invention name is a short-term wind speed prediction method based on two-level decomposition, LSTM and AT; the application publication number is [CN114897260A], the invention name is a short-term wind speed prediction model modeling method and prediction method based on LSTM neural network; the application publication number is [CN115689039A], the invention name is a CNN+GRU fusion ultra-short-term wind speed prediction method and system; the application publication number is [CN115859816A], the invention name is a minute-level wind power prediction method and system based on CNN-LSTM algorithm.

[0004] These are mainly focused on traditional recursive neural network modeling methods and single-step prediction tasks. However, in deep learning, as the number of layers increases, traditional recursive neural networks use gradient descent to update parameters, which is prone to gradient explosion or gradient vanishing problems. At the same time, the increase in the number of layers will also increase the amount of calculation and the training time, which cannot meet the real-time prediction requirements. Summary of the invention

[0005] The purpose of the present invention is to provide a multi-step wind speed prediction method based on variational deep projection coded echo state network, which mainly solves the accuracy, efficiency and generalization problems of the prediction model in wind speed multi-step prediction. Specifically, it includes:

[0006] 1) Establish a new prediction algorithm of deep projected coded echo state network DEESN;

[0007] 2) Study the time dependency between future time steps to reduce the cumulative error in multi-step prediction and improve prediction accuracy; study the dimensionality reduction encoding method of high-dimensional state to improve modeling efficiency;

[0008] 3) The VMD data preprocessing method is used to decompose the original wind speed data into several subsequences with low complexity to improve the generalization ability of DEESN. Then, a new wind speed multi-step prediction model based on VMD-DEESN is established.

[0009] A multi-step wind speed prediction method based on variational deep projection coded echo state network VMD-DEESN specifically includes the following steps:

[0010] 1. Obtain a data set and preprocess the data. The original wind speed data has the characteristics of high nonlinearity and non-stationarity, which affects the modeling accuracy to a certain extent. To solve this problem, the present invention adopts VMD data preprocessing technology to decompose the original data sequence into several subsequences with different characteristics.

[0011] 2. Establish a DEESN prediction algorithm. In order to improve the mapping ability and learn the time dependency between future time steps, the present invention proposes a new DEESN model, which consists of multiple echo state networks (ESNs)

[0012] The kth ESN learner is responsible for the prediction of the kth step and connects the prediction output of the previous ESN module with the input variables and the encoding state to form a new input signal for the next prediction. Secondly, the ELM encoder is used to optimize and reduce the dimension of the reserve pool state information to reduce time consumption.

[0013] 3. Establish a new VMD-DEESN multi-step wind speed prediction model and conduct performance evaluation. On the basis of the above, a VMD-DEESN multi-step wind speed prediction model is established, and the effectiveness and superiority of the model are verified. First, experiments are carried out on the DEESN model with different decomposition methods and without decomposition methods to prove the effectiveness of the variational mode decomposition VMD method. Secondly, under the same decomposition method VMD, the DEESN model is experimentally compared with other models to verify the superiority of DEESN. Finally, the same two comparative experiments are carried out using wind speed data on different time scales to verify that the proposed VMD-DEESN is superior to other models in handling multi-step wind speed prediction.

[0014] The beneficial effects of the present invention are:

[0015] 1. Establish a DEESN model to improve the mapping ability and learn the temporal dependencies between future time steps, thereby improving modeling accuracy and efficiency.

[0016] 2. The variational mode decomposition (VMD) method is used to decompose the wind speed series into several subsequences with different characteristics to improve the accuracy and generalization ability of the prediction model.

[0017] 3. By conducting wind speed prediction experiments at different time scales, it is proved that the proposed VMD-DEESN prediction model is superior to other methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flowchart of the multi-step wind speed prediction method based on the variational deep projection coded echo state network VMD-DEESN;

[0019] Figure 2 It is a schematic diagram of VMD decomposition;

[0020] Figure 3 It is the structural diagram of the ESN model;

[0021] Figure 4 This is the structural diagram of the DEESN model. DETAILED DESCRIPTION

[0022] The present invention is further described below in conjunction with specific embodiments, but the protection scope of the present invention is not limited thereto.

[0023] This embodiment provides a multi-step wind speed prediction method based on variational deep projection coded echo state network (VMD-DEESN) to improve modeling accuracy and efficiency. In addition, in order to deal with the nonlinear problem of wind speed data, the variational mode decomposition (VMD) method is used to pre-process the data to further improve the generalization ability of the model. Finally, a multi-step wind speed prediction model based on variational deep projection coded echo state network VMD-DEESN is established, and the effectiveness and superiority of the prediction results are verified.

[0024] Figure 1 The flowchart of the wind speed prediction method based on VMD-DEESN is shown. The flowchart includes the following steps:

[0025] Step 1. Obtain the data set. Collect two data sets at different time scales, namely wind speed records every 10 minutes and every hour, and split them into training set and test set in a ratio of 4:1. Then, preprocess the data and use VMD to decompose the original wind speed sequence into several subsequences with different characteristics.

[0026] Step 2. Establish the DEESN model. The DEESN model consists of multiple echo state network (ESN) modules and extreme learning machine (ELM) encoders. Each ESN module is responsible for asynchronous advance forecasts, and the ELM encoder between two sub-reservoirs is used to reduce the dimension and optimize the reservoir state. For Q-step ahead wind speed forecast, the DEESN model combines the Qth ESN module and the Q-1th ELM module, and the parameters of the DEESN model are determined by trial and error.

[0027] Step 3. Use the DEESN model to perform multi-step wind speed forecasting and obtain the forecast results.

[0028] Step 4. Quantitative evaluation. Three different evaluation indicators were used to evaluate the performance of VMD-DEESN, including mean absolute error (MAE), root mean square error (RMSE), and normalized root mean square error (NRMSE), and experiments were conducted. The effectiveness and superiority of introducing VMD decomposition in DEESN were verified by comparing with EEMD, CEEMDAN decomposition technology and no decomposition technology. The overall performance of VMD-DEESN was verified by comparing with LSTM, CBLM and classic recursive ESN models. Wind speed prediction experiments on different time scales of every hour and every ten minutes were used to prove the generalization ability of the model.

[0029] Specifically, the variational mode decomposition (VMD) in step 1 is described as follows:

[0030] VMD can capture the complex features of the original sequence and then decompose it into K subsequences with different sparsity features. Figure 2 This is a schematic diagram of decomposing the original sequence into three subsequences. Assume that each decomposed subsequence is u k , then the corresponding decomposition process is shown in formula (1):

[0031]

[0032] Among them, ω k Indicates u k is the center frequency of , δ(t) represents the Dirac distribution, (*) represents the convolution operation, and z is the original signal.

[0033] In order to solve the above optimization problem, the quadratic penalty function factor a and the Lagrange multiplier λ are introduced to ensure the accuracy and constraints of the reconstruction. The expanded equation is rewritten as follows:

[0034]

[0035] By iterating the alternating direction method of multipliers (ADMM), each subsequence and center frequency are updated to find the optimal y k ,ω kand λ. Update as follows:

[0036]

[0037] Where n is the number of iterations, γ is the noise tolerance, and Respectively represent the Fourier transform of z(t), u(t), λ(t). The iteration terminates when the following conditions are met.

[0038]

[0039] Among them, ∈ is and The limit error between .

[0040] Specifically, in step 2, ESN and DEESN are described as follows:

[0041] (1) Echo State Network (ESN)

[0042] ESN is a variant RNN with a three-layer structure, namely, input layer, hidden layer and output layer, such as Figure 3 As shown in Figure 1, the hidden layer is also called a reserve pool, and the connection weights between neurons are randomly and sparsely initialized. The reserve pool provides a complex nonlinear mapping that can encode the input signal from a low-dimensional input space to a high-dimensional state space.

[0043] Assume that the number of neurons in each layer is K, M, and N respectively. Therefore, the input signal can be set as The status of the reserve pool is l train represents the length of the training sequence; the output weight is expressed as The training process mainly includes state collection and output weight calculation. When the input signal enters the reservoir, the state information can be updated and obtained. According to the reservoir state, the output weight matrix is ​​adjusted using the regression method. The updated reservoir state equation is shown below:

[0044]

[0045] d(t) represents the reservoir state at time t, α is the leakage rate in the range [0, 1], and Represent the connection weights between input and reservoir and between reservoir and reservoir respectively, ranging from [-1, 1]. The final feature in the output layer contains the input signal and the reservoir state, that is,

[0046] x(t)=[U(t),d(t)] (8)

[0047] Finally, the ridge regression method is used to find Wout , expressed as:

[0048] W out =YX(X T X+δE) -1 (9)

[0049] Where Y is the target value, δ represents the regularization parameter, E is the identity matrix, and X is collected from x(t) as follows:

[0050] X=[x(1),x(2),…,x(l train )] T (10)

[0051] (2)DEESN algorithm

[0052] In order to improve the mapping ability and learn the temporal dependencies between future time steps, this paper proposes a new DEESN model for multi-step wind speed prediction, such as Figure 4 As shown. DEESN consists of multiple ESN modules and ELM encoders, each ESN module is responsible for the asynchronous advance forecast, and the ELM encoder between the two sub-reservoirs is used to reduce the dimension and optimize the reservoir state. Starting from the second module, the total input includes not only the input variables, but also the predicted output of the previous module and the encoded state information. Therefore, DEESN integrates the predicted output of the previous step in each future step prediction, which can learn time dependencies and reduce cumulative errors. For Q-step ahead wind speed forecast, DEESN combines the Qth ESN module and the Q-1th ELM module. The reservoir state of the first ESN module at time step t is calculated by the following formula.

[0053]

[0054] U1(t) represents the input signal at time t, and denote the connection weights from input to reservoir and from reservoir to reservoir, respectively.

[0055] Then, the state d1(t) of the above storage pool is cascaded with the input signal U1(t) to form the following new state information.

[0056] x1(t)=[U1(t),d1(t)] (12)

[0057] Therefore, the output weight It can be expressed as follows:

[0058]

[0059] Among them, Y1 is the actual wind speed value predicted in the first step. The following reserve pool status is indicated.

[0060]

[0061] in Represents the state signal of the jth node in the ith reservoir at time t.

[0062] Finally, the first step ahead prediction value at time t is calculated as follows:

[0063]

[0064] For the second module, the total input includes not only the input variables, but also the predicted output and encoded state information of the previous module. The state information of the first reservoir is encoded by ELM according to the following formula:

[0065]

[0066] Among them, g1(t) represents the encoding information of the first ELM encoder at time t, is the weight uniformly and randomly generated by the first reserve pool, and b1 represents the bias of the compensation coding signal.

[0067] The total input signal to the second reservoir is then:

[0068]

[0069] Then, the state of the second reserve pool is updated according to the following formula:

[0070]

[0071] in, and They represent the weights of the second reservoir input to the reservoir and from reservoir to reservoir respectively.

[0072] Then, the state of the reservoir is collected and concatenated with the input signal U1(t) as follows:

[0073] x2(t)=[U1(t),d2(t)] (19)

[0074] Through the iteration steps, the state of each reservoir is calculated as

[0075]

[0076] d q (t) represents the state of the qth storage pool at time t, where and They represent the connection weights from the qth ESN module input to the reservoir and from the reservoir to the reservoir. q(t) is the input of the qth reservoir:

[0077]

[0078] Among them, g q-1 (t) represents the state signal encoded by the (q-1)th ELM encoder at time t. is the (q-1)th step ahead prediction result at time t. The state information encoded by ELM is as follows:

[0079]

[0080] Among them, W q-1 and b q-1 are the weight and regularization value of the (q-1)th ELM encoder, d q-1 (t) represents the final features that need to be trained through the (q-1)th ridge regression.

[0081] After each reservoir state update is stable, the output weight of each ESN module can be calculated as follows:

[0082]

[0083] in represents the output weight of the qth ESN module, Y q represents the target value of the qth step, X q is the final feature matrix passing through the qth ESN module, expressed as follows:

[0084]

[0085] where x q (t)=[U1(t),d q (t)].

[0086] Finally, the prediction output of each step can be obtained by the following formula:

[0087]

[0088] in, Represents the q-th step ahead prediction result at time t.

[0089] Specifically, in step 4:

[0090] (1) Quantitative evaluation indicators

[0091] Three different evaluation indicators are used to evaluate the performance of VMD-DEESN, namely, mean absolute error (MAE), root mean square error (RMSE), and normalized root mean square error (NRMSE). The calculation formula is as follows:

[0092]

[0093] Where T is the length of the test data, Y q (t) and Represents the true value and predicted value of the Qth step at time t.

[0094] (2) Performance Verification

[0095] 1) Comparison with EEMD decomposition, CEEMDAN decomposition and no decomposition method

[0096] VMD, EEMD and CEEMDAN are used to decompose the original sequence respectively, and experimental verification is carried out on the DEESN model. The experimental results are analyzed to evaluate the effectiveness and superiority of introducing VMD decomposition in DEESN.

[0097] 2) Comparison of DEESN with LSTM, CBLM and classic recursive ESN models

[0098] Under the conditions of ensuring fairness and reasonable parameter settings, the original sequences are decomposed by VMD, and then experimental comparisons are conducted on four prediction models: DEESN, LSTM, CBLM, and classic recursive ESN. Through the experimental results, it is analyzed whether the performance of the DEESN model is better than the other three models.

[0099] 3) Wind speed prediction experiments on different time scales

[0100] To prove the generalization ability of the VMD-DEESN model, wind speed data at different time scales are used for verification, including hourly wind speed prediction and ten-minute wind speed prediction. The above two comparative experiments are repeated. Through the experimental results, it is analyzed whether the VMD-DEESN model has good performance.

[0101] (3) Experimental results and analysis

[0102] To verify the performance of the VMD-DEESN model, a wind speed dataset with two time scales was collected from three stations in Tanghe County, Henan Province, China, including hourly and ten-minute wind speed records from October 2016 to September 2017. The dataset was divided into a training set and a test set according to a 4:1 ratio, and corresponding experiments were carried out.

[0103] 1) Tables 1 and 2 list the average statistical results of MAE, RMSE, and NRMSE for every ten minutes and every hour of the multi-step forecast values ​​without using the decomposition tool and using three different decomposition tools, respectively, with a prediction step of 12. The results show that the model using the decomposition tool is better than the model without the decomposition tool in terms of prediction performance. When using the VMD decomposition tool, the prediction accuracy of the DEESN model is the best.

[0104] Table 1 Average statistical results of wind speed prediction evaluation indicators every ten minutes based on different decomposition tools of DEESN model

[0105]

[0106] Table 2 Average statistical results of hourly wind speed prediction evaluation indicators of different decomposition tools based on the DEESN model

[0107]

[0108]

[0109] 2) Tables 3 and 4 list the average statistical results of MAE, RMSE, and NRMSE for every ten minutes and every hour of the multi-step forecast values ​​calculated by the four models, respectively, with a forecast step of 12. The results show that the performance of DEESN is better than the other three models, with lower prediction error and less calculation time.

[0110] Table 3 Average statistical results of wind speed prediction evaluation indicators for four models every ten minutes

[0111]

[0112]

[0113] Table 4 Average statistical results of hourly wind speed prediction evaluation indicators of four models

[0114]

[0115] 3) The results of the above comparative experiments show that the VMD-DEESN model has good performance at both the ten-minute and hourly time scales. The VMD-DEESN model has good generalization ability.

[0116] The above is only a preferred embodiment of the present invention, which is used to illustrate the technical solution of the present invention, but not to limit the present invention. It should be pointed out that for ordinary technicians in this field, without departing from the concept of the present invention, some of the technical features therein are modified and replaced, and they will not depart from the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be based on the attached claims.

Claims

1. A multi-step wind speed prediction method based on variational deep projection coded echo state network, characterized in that: The following steps are involved: Step 1: Obtain the data set and preprocess the data; Two data sets at different time scales are collected, namely wind speed records every 10 minutes and every hour, and divided into training set and test set in a ratio of 4:

1. Then, the data is preprocessed, and the variational mode decomposition (VMD) is used to decompose the original wind speed sequence into several subsequences with different characteristics. Step 2, establish a DEESN model; the DEESN model consists of multiple echo state network ESN modules and extreme learning machine ELM encoders, each ESN module is responsible for asynchronous advance forecasting, and the ELM encoder between two sub-reservoirs is used to reduce the dimension and optimize the reservoir state; for Q-step advance wind speed prediction, the DEESN model combines the Qth ESN module and the Q-1th ELM module, and determines the parameters of the DEESN model by trial and error; Step 3: Use the DEESN model to perform multi-step wind speed forecasting and obtain the forecast results.

2. The multi-step wind speed prediction method based on variational deep projection coded echo state network according to claim 1 is characterized in that: In step 1, VMD is used to capture the complex features of the original sequence, and then decompose it into K subsequences with different sparsity features. Each decomposed subsequence is u k , then the corresponding decomposition process is shown in formula (1): Among them, ω k Indicates u k The center frequency of , δ(t) represents the Dirac distribution, (*) represents the convolution operation, and z is the original signal; The quadratic penalty function factor a and the Lagrange multiplier λ are introduced to ensure the accuracy and constraints of the reconstruction; the expanded equation is rewritten as follows: Through the iteration of the alternating direction multiplier method ADMM, each subsequence and center frequency are updated to find the optimal y k ,ω k and λ; are updated as follows: Where n is the number of iterations, γ is the noise tolerance, and Respectively represent the Fourier transform of z(t), u(t), λ(t); the iteration terminates when the following conditions are met; Among them, ∈ is and The limit error between .

3. The multi-step wind speed prediction method based on variational deep projection coded echo state network according to claim 2 is characterized in that: In step 2, ESN is a variant RNN with a three-layer structure, namely, input layer, hidden layer and output layer; the hidden layer is also called a reserve pool, and the connection weights between neurons are randomly sparsely initialized; the reserve pool provides a complex nonlinear mapping, which encodes the input signal from a low-dimensional input space to a high-dimensional state space; The number of neurons in each layer is K, M, and N respectively, and the input signal is set to The status of the reserve pool is l train represents the length of the training sequence; the output weight is expressed as The training process includes state collection and output weight calculation; when the input signal enters the reservoir, the state information is updated and obtained; according to the reservoir state, the output weight matrix is ​​adjusted using the regression method; the updated reservoir state equation is as follows: d(t)=(1-α)d(t-1)+αtanh(W in U(t)+W res d(t-1)) (7) d(t) represents the reservoir state at time t, α is the leakage rate in the range [0, 1], and They represent the connection weights between the input and the reservoir and between the reservoir and the reservoir, respectively, ranging from [-1, 1]; the final feature in the output layer contains the input signal and the reservoir state, that is, x(t)=[U(t),d(t)] (8) Finally, the ridge regression method is used to find W out , expressed as: W out =YX(X T X+δE) -1 (9) Where Y is the target value, δ represents the regularization parameter, E is the identity matrix, and X is collected from x(t) as follows: X=[x(1),x(2),…,x(l train )] T (10)。 4. The multi-step wind speed prediction method based on variational deep projection coded echo state network according to claim 3 is characterized in that: In step 2, DEESN consists of multiple ESN modules and ELM encoders. Each ESN module is responsible for asynchronous advance forecasts, and the ELM encoder between two sub-reservoirs is used to reduce the dimension and optimize the reservoir state. Starting from the second module, the total input includes not only the input variables but also the predicted output and encoded state information of the previous module. For Q-step ahead wind speed forecast, DEESN combines the Q-th ESN module and the Q-1-th ELM module. The reservoir state of the first ESN module at time step t is calculated by the following formula: U1(t) represents the input signal at time t, and denote the connection weights from input to reservoir and from reservoir to reservoir, respectively; The state d1(t) of the above storage pool is cascaded with the input signal U1(t) to form the following new state information; x1(t)=[U1(t),d1(t)] (12) Output weight It is expressed as follows: Among them, Y1 is the actual wind speed value predicted in the first step. Indicates the following reserve pool status; in represents the state signal of the jth node of the i-th reservoir at time t; Finally, the first step ahead prediction value at time t is calculated as follows: For the second module, the total input includes not only the input variables but also the predicted output and encoded state information of the previous module; the state information of the first reservoir is encoded by ELM according to the following formula: Among them, g1(t) represents the encoding information of the first ELM encoder at time t, is the weight uniformly generated randomly by the first reserve pool, and b1 represents the bias of the compensation coding signal; The total input signal to the second reservoir is then: Then, the state of the second reserve pool is updated according to the following formula: in, and denote the weights of the second reservoir input to the reservoir and from reservoir to reservoir respectively; Then, the state of the reservoir is collected and concatenated with the input signal U1(t) as follows: x2(t)=[U1(t),d2(t)] (19) Through the iteration steps, the state of each reservoir is calculated as: d q (t) represents the state of the qth storage pool at time t, where and They represent the connection weights from the qth ESN module input to the reservoir and from reservoir to reservoir respectively; U q (t) is the input of the qth reservoir: Among them, g q-1 (t) represents the state signal encoded by the (q-1)th ELM encoder at time t; is the (q-1)th step ahead prediction result at time t; the state information encoded by ELM is as follows: Among them, W q-1 and b q-1 are the weight and regularization value of the (q-1)th ELM encoder, d q-1 (t) represents the final feature that needs to be trained through the (q-1)th ridge regression; After each reservoir state update is stable, the output weight of each ESN module can be calculated as follows: in represents the output weight of the qth ESN module, Y q represents the target value of the qth step, X q is the final feature matrix passing through the qth ESN module, expressed as follows: where x q (t)=[U1(t),d q (t)]; Finally, the prediction output of each step is obtained by the following formula: in, Represents the q-th step ahead prediction result at time t.

Images