A real-time prediction method for wind power

Abstract

The present invention discloses a real-time prediction method for wind power, comprising the following steps: step S1, for any wind power equipment that has been put into operation, pre-acquire the wind power and wind speed of the wind power equipment at each power generation moment after the power generation operation is completed; step S2, establish a prediction model for wind power, the prediction model for wind power is used to establish a nonlinear relationship between the wind power of the wind power equipment at a previous moment and the wind power at a later moment; step S3, collect the wind power and wind speed of the wind power equipment at the current moment in real time, and then input them into the prediction model for wind power, so as to predict and obtain the wind power of the wind power equipment at the later moment in real time. The real-time prediction method for wind power disclosed by the present invention is scientifically designed, and by establishing a prediction model for wind power, it is possible to efficiently and reliably predict the wind power of the wind power equipment in real time, and has great practical significance.

Description

Technical Field

[0001] The invention relates to the technical field of complex process system control, and in particular to a real-time prediction method for wind power. Background Art

[0002] Wind energy has the advantages of wide distribution, large reserves, low carbon emissions and cleanliness and no pollution. Vigorously developing wind power generation has become an effective means to solve problems such as energy shortages and environmental pollution.

[0003] At present, the relevant technologies of wind power generation in my country are becoming more and more mature, and wind power equipment is constantly optimized and updated. However, the intermittent, random and volatile nature of wind power generation poses huge technical challenges to the smooth operation of wind power equipment. Therefore, it is necessary to predict wind power (that is, the power that can be output by wind power equipment installed in a wind farm). Accurate wind power prediction results can guide the deployment plan of the wind farm dispatching department and ensure the safe and efficient operation of wind power equipment.

[0004] However, in order to obtain wind power forecasting results with high accuracy, most of the existing wind power forecasting methods have poor timeliness of their forecasting results, which is not conducive to the real-time deployment plan of the wind farm scheduling part.

[0005] Therefore, there is an urgent need to develop a technology to efficiently and reliably predict the wind power of wind power equipment in real time. Summary of the invention

[0006] The purpose of the present invention is to provide a real-time prediction method for wind power in view of the technical defects in the prior art.

[0007] To this end, the present invention provides a new neural network model for establishing a wind power prediction model. By using the simplified variational mode decomposition as the decomposition layer of the network, it is used to extract the features of the wind power in the input data, and then the feature learning of the decomposed wind power and the wind speed at the corresponding moment is performed through a non-incremental width learning system, so as to achieve the purpose of establishing the real-time prediction model of the wind power. Compared with other existing technologies, the neural network model provided by the present invention has a greater advantage in model calculation efficiency, can make real-time and accurate predictions of wind power, and is more conducive to helping wind farm staff to deploy wind power equipment in real time.

[0008] The present invention provides a real-time prediction method for wind power, which comprises the following steps:

[0009] Step S1, for any wind turbine that has been put into operation, the wind power and wind speed of the wind turbine at each power generation moment when the power generation operation has been completed are obtained in advance, and for any two adjacent power generation moments, the wind power and wind speed of the wind turbine at the previous moment are defined as input data of the wind turbine, and the wind power of the wind turbine at the next moment is defined as output data of the wind turbine;

[0010] Step S2, establishing a wind power prediction model, where the wind power prediction model is used to establish a nonlinear relationship between the wind power of the wind power equipment at a previous moment and the wind power at a next moment;

[0011] Step S3, collecting the wind power and wind speed of the wind power equipment at the current moment in real time, and then inputting them into the wind power prediction model, so as to predict the wind power of the wind power equipment at the next moment in real time.

[0012] It can be seen from the technical solution provided by the present invention that, compared with the prior art, the present invention provides a real-time prediction method for wind power, which is scientifically designed and can efficiently and reliably predict the wind power of wind power equipment in real time by establishing a wind power prediction model, which has great practical significance.

[0013] The present invention aims at the intermittent, random and volatile problems of wind power generation. By combining simplified variational mode decomposition (VMD) and non-incremental broad learning system (BLS), a method such as Figure 1 The new neural network model shown is used to establish a wind power prediction model and achieve the purpose of real-time monitoring and online prediction of wind power.

[0014] For the present invention, the new neural network model proposed in the present invention can greatly shorten the time for predicting the wind power at the next moment in the future while ensuring the accuracy of the model, compared with other existing wind power modeling methods. The prediction results obtained by this model are more timely, thereby achieving the purpose of real-time prediction of wind power, which is conducive to helping wind farm managers to deploy wind power equipment operation plans in real time.

[0015] Compared with the prior art, the novel neural network model proposed in the present invention has the following beneficial technical effects:

[0016] First, the variational mode decomposition method is used to transform the one-dimensional time domain characteristics of wind power into multi-dimensional frequency domain characteristics of wind power. The preliminary feature extraction of wind power is carried out through the decomposition layer, which effectively increases the useful information that the model can learn.

[0017] Second, a new type of neural network architecture, the width learning system, is used to perform feature learning on multidimensional data. The decomposed multidimensional data is sparsely represented, linearly transformed, and nonlinearly enhanced through the sparse layer, mapping layer, and enhancement layer respectively. Through the joint action of multiple methods, further feature extraction of wind power is achieved, effectively improving the feature learning ability of the model.

[0018] Third, in the entire process of modeling wind power, the neural network model proposed in the present invention does not learn the sequence data of wind power by capturing time characteristics or spatial characteristics, but directly establishes the nonlinear relationship between the wind power at the previous moment and the wind power at the next moment in the future by combining multiple methods, thereby effectively improving the calculation speed of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A flow chart of a real-time prediction method for wind power provided by the present invention;

[0020] Figure 2 A schematic diagram of the topological structure of a novel neural network model adopted in a real-time prediction method for wind power provided by the present invention;

[0021] Figure 3 A real-time prediction method for wind power provided by the present invention, a schematic diagram of a training set fitting curve of a wind power prediction model in an embodiment;

[0022] Figure 4 A real-time prediction method for wind power provided by the present invention is a schematic diagram of a test set fitting curve of a wind power prediction model in an embodiment. DETAILED DESCRIPTION

[0023] In order to enable those skilled in the art to better understand the scheme of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and implementation modes.

[0024] See also Figures 1 to 4 The present invention provides a real-time prediction method for wind power, comprising the following steps:

[0025] Step S1, for any wind power equipment that has been put into operation, the wind power and wind speed of the wind power equipment at each power generation moment after the power generation operation is completed are obtained in advance, and for any two adjacent power generation moments, the wind power and wind speed of the wind power equipment at the previous moment (i.e., the previous moment) are defined as input data of the wind power equipment, and the wind power of the wind power equipment at the next moment (i.e., the next moment) is defined as output data of the wind power equipment;

[0026] In the present invention, it should be noted that wind power refers to the power that can be output by wind power equipment installed in a wind farm.

[0027] In the present invention, it should be noted that each power generation moment at which the power generation operation has been completed is each moment at which power generation has been completed in history (early period).

[0028] It should be noted that the wind power and wind speed of the wind power equipment at each power generation moment after the power generation operation is completed can be obtained from the existing wind farm SCADA system (i.e., data acquisition and monitoring control system). The wind farm SCADA system is used for wind turbine control and data monitoring of the wind farm. The system monitors the operating status of the wind turbine, real-time alarm data, and remote start, stop or reset operations to achieve automatic monitoring of the wind farm. The wind power and wind speed data used in the present invention are all from this system.

[0029] Step S2, establishing a wind power prediction model, where the wind power prediction model is used to establish a nonlinear relationship between the wind power of the wind power equipment at a previous moment and the wind power at a next moment;

[0030] Step S3, collecting the wind power and wind speed of the wind power equipment at the current moment in real time, and then inputting them into the wind power prediction model, so as to predict the wind power of the wind power equipment at the next moment (ie, the next moment of the current moment) in real time.

[0031] In the present invention, in step S2, the prediction model of wind power is specifically a neural network model;

[0032] See also Figure 2 As shown, the neural network model includes an input layer, a decomposition layer, a sparse layer, a mapping layer, an enhancement layer and an output layer for processing input data in sequence;

[0033] Among them, after the input data of the wind power equipment is input from the input layer, the decomposition layer, sparse layer, mapping layer and enhancement layer respectively perform feature decomposition, sparse representation, linear transformation and nonlinear enhancement processing operations on the input data, and then the connection weight between the processed input data and the output data is calculated in the output layer, so as to establish a nonlinear relationship between the wind power of the wind power equipment at the previous moment and the wind power at the next moment.

[0034] It should be noted that, in the present invention, the decomposition layer is used to convert the single-dimensional time domain data into multi-dimensional frequency domain data, and the wind power data is subjected to preliminary feature extraction, which increases the useful information that the model can learn;

[0035] The sparse layer is used to perform sparse representation on the decomposed data, so that the model has better noise resistance and higher computational efficiency;

[0036] The mapping layer is used to perform linear transformation on the sparsely represented data and map the data to the interval of 0-1 for model training;

[0037] The enhancement layer is used to perform nonlinear enhancement on the mapped data, and use the activation function to increase the nonlinear features in the model to improve the learning ability of the model.

[0038] See also Figure 2 As shown, in step S2, in specific implementation, the neural network model is a neural network model with preset calculation rules;

[0039] The preset calculation rules of the neural network model specifically include the following five stages of calculation and processing operations performed in sequence according to the number of network layers it includes:

[0040] Step S21, performing calculation processing operations from the input layer to the decomposition layer;

[0041] In specific implementation, step S21 specifically includes the following sub-steps:

[0042] Step S211: a set of wind power time series data with a sample size of t is given in advance. Represents the input data of the neural network model, Y = [F 2 , F 3 , ...F t ] represents the output data of the neural network model;

[0043] Among them, S is wind speed, F is wind power;

[0044] It should be noted that step S21 performs feature decomposition on the wind power in the input data, and uses the decomposed multi-dimensional frequency domain data and the wind speed at the corresponding time as the decomposition layer P k The input data.

[0045] Step S212, by expression Establish the connection between the single-dimensional time domain characteristics and the multi-dimensional frequency domain characteristics of wind power in the input data;

[0046] Assume ω k is the frequency domain sequence f k The center frequency of the frequency domain sequence fk It is obtained by solving the constrained variational problem shown in the following formula (1);

[0047]

[0048] Where: K is the frequency domain sequence f k dimension; δ(t) is the Dirac distribution; * represents the convolution operator; j is an imaginary number;

[0049] Step S213, introducing a quadratic penalty factor α and a Lagrange multiplier operator θ into formula (1), transforming the constrained variational problem shown in the above formula (2) into an unconstrained variational problem, and obtaining an expression as shown in the following formula (2);

[0050]

[0051] Step S214, using the alternating direction multiplier method, solve the variational problem shown in the above formula (2), then the frequency domain sequence f k , center frequency ω k The update iterative mechanism of the Lagrange multiplier operator θ is shown in the following formulas (3), (4) and (5) respectively;

[0052]

[0053]

[0054]

[0055] in: and They are h(ω), f k (ω), f p (ω) and θ(ω), τ is the update parameter; s is the number of iterations.

[0056] Step S215, by running the output data P of the decomposition layer k The expression of the decomposition layer is calculated to obtain the output data P k ;

[0057] The output data P of the decomposition layer k The expression of is shown in the following formula (6);

[0058] P k =[f 1 , f 2 , ..., f k , S]., Formula (6);

[0059] Among them, f kIt is the frequency domain sequence obtained by variational mode decomposition of the input data X of the neural network model, and k is the number of decompositions of the input data X through the network model.

[0060] It should be noted that, for the real-time prediction method of wind power of the present invention, the input data X is the wind power F and wind speed S at the previous moment, and the output data is the wind power at the next moment; wherein, the decomposition layer only decomposes F, and the decomposed F and S are merged as the input of the sparse layer.

[0061] Step S22, performing a calculation process from the decomposition layer to the sparse layer;

[0062] In specific implementation, step S22 specifically includes the following sub-steps:

[0063] Step S221, by expression Create the output data Q of the sparse layer and the decomposition layer P k the relationship between;

[0064] Among them, the optimal weight It is obtained by solving the constraint problem shown in the following formula (7);

[0065]

[0066] where λ is the regularization coefficient in the sparse layer; ||·|| 2 represents the L2 norm; ||·|| 1 represents the L1 norm;

[0067] Step S222, using the alternating direction multiplier method to solve the convex optimization problem shown in the above formula (7), thereby converting the formula (7) into the form shown in the formula (8);

[0068]

[0069] in,

[0070] Step S223, further transform formula (8) to obtain the following formula (9);

[0071]

[0072] Among them: w and o are both a set of optimal solutions;

[0073] Step S224, using the iterative mechanism shown in the following formula (10), To solve;

[0074]

[0075] Where r is the number of iterations; Sλ is the symbol of the soft threshold operator, S λ The definition of is shown in the following formula (11);

[0076]

[0077] Then, the output data Q of the sparse layer is ln The expressions of are shown in the following formulas (12) and (13);

[0078]

[0079]

[0080] in, is the optimal weight obtained by sparse autoencoding of the randomly generated weights of the sparse layer, l is the number of windows in the mapping layer, and n is the number of nodes in each window in the mapping layer.

[0081] Step S23, performing a calculation process from the sparse layer to the mapping layer;

[0082] In specific implementation, step S23 specifically includes the following sub-steps:

[0083] Output data Z of running the mapping layer ln The expressions are as shown in the following formulas (14) and (15);

[0084]

[0085]

[0086] Among them, φ(·) is the feature mapping function, that is, the features are uniformly mapped to the interval [-1, 1]; β ei is the randomly generated bias of the mapping layer.

[0087] Step S24, performing a calculation processing operation from the mapping layer to the enhancement layer;

[0088] In specific implementation, step S24 specifically includes the following sub-steps:

[0089] Run Enhancement Layer H m The expression of is as shown in the following formula (16);

[0090] H m =ζ(Z ln W h +β h )., formula (16);

[0091] Among them, ζ(·) is a nonlinear activation function, W h and β hare the randomly generated weights and biases of the enhancement layer.

[0092] Step S25, performing calculation processing operations from the enhancement layer to the output layer;

[0093] In specific implementation, step S25 specifically includes the following sub-steps:

[0094] The expression of the output data Y of the output layer is shown in the following formula (17);

[0095] Y=[Z ln |H m ]W., formula (17);

[0096] Where W is the connection weight output by the neural network model, and its expression is shown in the following formula (18);

[0097] W=[Z ln |H m ] -1 Y., formula (18);

[0098] Among them, the regularization coefficient η of the enhancement layer is introduced to solve the connection weight W of the above formula (18), and the obtained expressions are shown in formulas (19), (20) and (21);

[0099]

[0100]

[0101]

[0102] Where: ||·|| 2 represents the L2 norm, and I is an identity matrix.

[0103] In the calculation and processing operation of the neural network model described in step S21 to step S25, the description of each parameter is shown in Table 1 below.

[0104] Table 1:

[0105] parameter describe k The number of feature nodes in the decomposition layer α Penalty factor in decomposition layer <![CDATA[s max ]]> Maximum number of iterations in a decomposition layer λ Regularization coefficients in sparse layers <![CDATA[r max ]]> Maximum number of iterations in a sparse layer l Number of windows in the mapping layer n The number of feature nodes per window in the mapping layer m The number of feature nodes in the enhancement layer η Regularization coefficients in the enhancement layer

[0106] In order to more clearly understand the technical solution of the present invention, the technical solution of the present invention is described below through specific embodiments.

[0107] Example

[0108] First, the experimental samples were taken from 50,530 measured data (i.e., wind power and wind speed of the wind turbine at 50,530 power generation moments) of a single wind turbine (i.e., wind power equipment) in a wind farm in 2018. The recording interval was 10 minutes, including the actual wind speed and actual power. The cut-in wind speed of the wind turbine was 3m / s, and the cut-out wind speed was 25m / s.

[0109] Then, the neural network model proposed in the present invention is used to model the wind power. The parameters of the model are set as follows: the number of feature nodes, the penalty factor and the maximum number of iterations in the decomposition layer are set to 11, 100 and 50 respectively, the regularization coefficient and the maximum number of iterations in the sparse layer are set to 0.01 and 50 respectively, the number of windows in the mapping layer and the number of feature nodes in each window are set to 19 and 1 respectively, and the number of feature nodes and the regularization coefficient in the enhancement layer are set to 6 and 2 respectively. -30 , and the nonlinear activation function in the enhancement layer is the tanh function.

[0110] Then, the mean of the root mean square error (RMSE) of 30 independent experiments is used to evaluate the model accuracy, and the standard deviation is used to evaluate the model stability. The specific results of the established wind power prediction model (i.e., the wind power prediction model) are shown in Table 2 below.

[0111] Table 2:

[0112]

[0113] In Table 2, AVG is the average of 30 root mean square errors (RMSE) obtained after the model is run independently for 30 times, which is used to evaluate the accuracy of the model; SD is the standard deviation of 30 root mean square errors (RMSE) obtained after the model is run independently for 30 times, which is used to evaluate the stability of the model.

[0114] The fitting curves of the neural network model proposed in the present invention for the training sample and the test sample are as follows: Figure 3 and Figure 4 As shown, the fitting result of the neural network model on the training set is as follows Figure 3 As shown, the fitting results on the test set are as follows Figure 4 shown.

[0115] Depend on Figure 3 and Figure 4 It can be seen that the output value of the wind power prediction model established based on the neural network model proposed in the present invention (i.e. Figure 3 and Figure 4 The predicted value in the calculation result can be well approximated to the target value (ie, the true value), which verifies that the real-time prediction method for wind power proposed in the present invention is an effective, relatively accurate and reliable calculation scheme.

[0116] Based on the above technical solutions, it can be known that for the real-time prediction method of wind power provided by the present invention, first, the historical wind power of the wind power equipment and the wind speed data at the corresponding moment are retrieved, and the wind power data at the next moment is recorded as output data, and the wind power and wind speed at the previous moment are recorded as input data; then, the single-dimensional time domain features of the wind power in the input data are transformed into multi-dimensional frequency domain features using simplified variational mode decomposition, and are modeled together with the wind speed at the corresponding moment as new network inputs, and the new network inputs are sparsely represented, linearly transformed and nonlinearly enhanced respectively through the sparse layer, mapping layer and enhancement layer of the incremental width learning system, so as to establish a connection (nonlinear relationship) between the wind power and wind speed at the previous moment (i.e. the previous moment) and the wind power at the next moment (i.e. the next moment); then, the wind power and wind speed at the current moment are continuously collected, and based on the prediction model established above, the wind power at the next moment is predicted in real time.

[0117] In summary, compared with the prior art, the real-time prediction method for wind power provided by the present invention is scientifically designed. By establishing a wind power prediction model, the wind power of wind power equipment can be efficiently and reliably predicted in real time, which has great practical significance.

[0118] The present invention aims at the intermittent, random and volatile problems of wind power generation. By combining simplified variational mode decomposition (VMD) and non-incremental broad learning system (BLS), a method such as Figure 1 The new neural network model shown is used to establish a wind power prediction model and achieve the purpose of real-time monitoring and online prediction of wind power.

[0119] For the present invention, the new neural network model proposed in the present invention can greatly shorten the time for predicting the wind power at the next moment in the future while ensuring the accuracy of the model, compared with other existing wind power modeling methods. The prediction results obtained by this model are more timely, thereby achieving the purpose of real-time prediction of wind power, which is conducive to helping wind farm managers to deploy wind power equipment operation plans in real time.

[0120] Compared with the prior art, the novel neural network model proposed in the present invention has the following beneficial technical effects:

[0121] First, the variational mode decomposition method is used to transform the one-dimensional time domain characteristics of wind power into multi-dimensional frequency domain characteristics of wind power. The preliminary feature extraction of wind power is carried out through the decomposition layer, which effectively increases the useful information that the model can learn.

[0122] Second, a new type of neural network architecture, the width learning system, is used to perform feature learning on multidimensional data. The decomposed multidimensional data is sparsely represented, linearly transformed, and nonlinearly enhanced through the sparse layer, mapping layer, and enhancement layer respectively. Through the joint action of multiple methods, further feature extraction of wind power is achieved, effectively improving the feature learning ability of the model.

[0123] Third, in the entire process of modeling wind power, the neural network model proposed in the present invention does not learn the sequence data of wind power by capturing time characteristics or spatial characteristics, but directly establishes the nonlinear relationship between the wind power at the previous moment and the wind power at the next moment in the future by combining multiple methods, thereby effectively improving the calculation speed of the model.

[0124] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A real-time prediction method for wind power, characterized in that: The following steps are involved: Step S1, for any wind turbine that has been put into operation, the wind power and wind speed of the wind turbine at each power generation moment when the power generation operation has been completed are obtained in advance, and for any two adjacent power generation moments, the wind power and wind speed of the wind turbine at the previous moment are defined as input data of the wind turbine, and the wind power of the wind turbine at the next moment is defined as output data of the wind turbine; Step S2, establishing a wind power prediction model, where the wind power prediction model is used to establish a nonlinear relationship between the wind power of the wind power equipment at a previous moment and the wind power at a next moment; Step S3, collecting the wind power and wind speed of the wind power equipment at the current moment in real time, and then inputting them into the wind power prediction model, so as to predict the wind power of the wind power equipment at the next moment in real time; In step S2, the prediction model of wind power is specifically a neural network model; The neural network model includes an input layer, a decomposition layer, a sparse layer, a mapping layer, an enhancement layer and an output layer for sequentially processing input data; Among them, after the input data of the wind power equipment is input from the input layer, the decomposition layer, the sparse layer, the mapping layer and the enhancement layer respectively perform feature decomposition, sparse representation, linear transformation and nonlinear enhancement processing operations on the input data, and then the connection weight between the processed input data and the output data is calculated in the output layer, so as to establish a nonlinear relationship between the wind power of the wind power equipment at the previous moment and the wind power at the next moment; In step S2, the neural network model is a neural network model with preset calculation rules; The preset calculation rules of the neural network model specifically include the following five stages of calculation and processing operations performed in sequence according to the number of network layers it includes: Step S21, performing calculation processing operations from the input layer to the decomposition layer; Step S22, performing a calculation process from the decomposition layer to the sparse layer; Step S23, performing a calculation process from the sparse layer to the mapping layer; Step S24, performing a calculation processing operation from the mapping layer to the enhancement layer; Step S25, performing calculation processing operations from the enhancement layer to the output layer; Step S21 specifically includes the following sub-steps: Step S211: a set of wind power time series data with a sample size of t is given in advance. Represents the input data of the neural network model, Y = [F2, F3, ...F t ] represents the output data of the neural network model; Among them, S is wind speed, F is wind power; Step S212, by expression Establish the connection between the single-dimensional time domain characteristics and the multi-dimensional frequency domain characteristics of wind power in the input data; Assume ω k is the frequency domain sequence f k The center frequency of the frequency domain sequence f k It is obtained by solving the constrained variational problem shown in the following formula (1); Where: K is the frequency domain sequence f k dimension; δ(t) is the Dirac distribution; * represents the convolution operator; j is an imaginary number; Step S213, introducing a quadratic penalty factor α and a Lagrange multiplier operator θ into formula (1), transforming the constrained variational problem shown in the above formula (2) into an unconstrained variational problem, and obtaining an expression as shown in the following formula (2); Step S214, using the alternating direction multiplier method, solve the variational problem shown in the above formula (2), then the frequency domain sequence f k , center frequency ω k The update iterative mechanism of the Lagrange multiplier operator θ is shown in the following formulas (3), (4) and (5) respectively; in: and They are h(ω), f k (ω), f p The Fourier transform of (ω) and θ(ω), τ is the update parameter; s is the number of iterations; Step S215, by running the output data P of the decomposition layer k The expression of the decomposition layer is calculated to obtain the output data P k ; The output data P of the decomposition layer k The expression of is shown in the following formula (6); P k =[f1, f2, ..., f k , S]., Formula (6); Among them, f k is the frequency domain sequence obtained by variational mode decomposition of the input data X of the neural network model, and k is the number of decompositions of the input data X through the network model; Step S22 specifically includes the following sub-steps: Step S221, by expression Create the output data Q of the sparse layer and the decomposition layer P k the relationship between; Among them, the optimal weight It is obtained by solving the constraint problem shown in the following formula (7); Where λ is the regularization coefficient in the sparse layer; ||·||2 represents the L2 norm; ||·||1 represents the L1 norm; Step S222, using the alternating direction multiplier method to solve the convex optimization problem shown in the above formula (7), thereby converting the formula (7) into the form shown in the formula (8); in, Step S223, further transform formula (8) to obtain the following formula (9); stv-0=0., formula (9); Among them: w and o are both a set of optimal solutions; Step S224, using the iterative mechanism shown in the following formula (10), To solve; o r+1 :=S λ (v r+1 +u r ), u r+1 := u r +(v r+1 - o r+1 ), Equation (10); Where r is the number of iterations; S λ is the symbol of the soft threshold operator, S λ The definition of is shown in the following formula (11); Then, the output data Q of the sparse layer is ln The expressions of are shown in the following formulas (12) and (13); in, is the optimal weight obtained by sparse autoencoding of the randomly generated weights of the sparse layer, l is the number of windows in the mapping layer, and n is the number of nodes in each window in the mapping layer; Step S23 specifically includes the following sub-steps: Output data Z of running the mapping layer ln The expressions are as shown in the following formulas (14) and (15); Among them, φ(·) is the feature mapping function, that is, the features are uniformly mapped to the interval [-1, 1]; β ei is the randomly generated bias of the mapping layer; Step S24 specifically includes the following sub-steps: Run Enhancement Layer H m The expression of is as shown in the following formula (16); H m = ζ(Z ln W h + β h ), Equation (16); Among them, ζ(·) is a nonlinear activation function, W h and β h are the randomly generated weights and biases of the enhancement layer; Step S25 specifically includes the following sub-steps: The expression of the output data Y of the output layer is shown in the following formula (17); Y = [Z ln |H m ]W, Formula (17); Where W is the connection weight output by the neural network model, and its expression is shown in the following formula (18); W = [Z ln |H m ] -1 Y, formula (18); Among them, the regularization coefficient η of the enhancement layer is introduced to solve the connection weight W of the above formula (18), and the obtained expressions are shown in formulas (19), (20) and (21); Where: ||·||2 represents the L2 norm, and I is an identity matrix.

Images