Optimization design method of speed-increasing wind driven generator based on OLHS-BPNN-IGWO algorithm

Abstract

The invention relates to a speed-increasing type wind driven generator optimization design method based on an OLHS-BPNN-IGWO algorithm, and belongs to the field of wind power. The speed-increasing wind driven generator comprises a stator, an outer rotor and an inner rotor, the method comprises the following steps: selecting six sizes of thicknesses hpin and hpout of inner and outer rotor permanent magnets, the number N of turns of a stator winding, the ratio b of an I-shaped clamping groove and a stator groove as to-be-optimized design variables, taking the amplitude VE of no-load back electromotive force and the distortion rate VTHD thereof as well as an electromagnetic torque ripple Trip as optimization targets, determining an optimization target function, performing sensitivity analysis on the to-be-optimized design variables, performing an OLHS experiment, establishing a BPNN proxy model, and establishing a BPNN model; the IGWO is applied to the training process of the BPNN to optimize the training of the BPNN to obtain a group of optimal design variables, so that VE, VTHD and Trip are optimal. The design precision and efficiency of the generator can be remarkably improved, and the purposes of reducing the cost and improving the performance are achieved.

Description

Technical Field

[0001] The present invention relates to an optimization design method for a speed-increasing wind turbine generator, in particular to an optimization design method for a speed-increasing wind turbine generator based on an OLHS-BPNN-IGWO algorithm, and belongs to the technical field of wind power generation. Background Art

[0002] Semi-direct-drive wind turbines, combining the advantages of permanent magnet direct-drive and doubly-fed wind turbines, have become a mainstream model. Currently, these turbines all utilize a medium-speed gearbox—a generator + mechanical gearbox structure—for speed. Mechanical gearboxes, however, utilize mechanical gear meshing and require a complex lubrication system. This results in large size, high mechanical wear, and a high failure rate. Failures can lead to high repair costs and prolonged downtime, resulting in significant economic losses.

[0003] As a new transmission method, magnetic gears, due to their lack of mechanical contact, offer advantages such as zero wear, maintenance, noise, lubrication, and overload protection. They have gained increasing attention and application in recent years. Magnetic gearboxes offer physical isolation, require no lubrication, are compact and lightweight, have low energy consumption, and transmit high torque. This significantly reduces the operating and maintenance costs of large wind turbines, particularly semi-direct drive and doubly-fed wind turbines, while improving system performance and reliability.

[0004] Therefore, there is an urgent need to develop a small, high-efficiency, low-cost speed-increasing wind turbine that integrates a speed-increasing magnetic gearbox and a generator.

[0005] However, the optimization design of a speed-increasing wind turbine is a nonlinear programming problem with complex constraints and multiple objectives. It requires finding a design solution that optimizes all performance indicators while simultaneously meeting industry standards, user requirements, and application constraints. A speed-increasing wind turbine integrates a speed-increasing magnetic gearbox with a generator and motor. The performance of the magnetic gear and generator is often coupled and conflicting, making it impossible to achieve an optimal solution that simultaneously meets multiple performance indicators. A speed-increasing wind turbine has numerous design variables, each of which is strongly coupled. The objective function optimization algorithm selects the factors that influence the design of the speed-increasing wind turbine and, in turn, determines its operating performance.

[0006] Traditional multi-objective optimization algorithms (such as NSGA-II and MOPSO) are prone to falling into local optimal solutions for high-dimensional, nonlinear, and multi-modal problems, making it difficult to obtain the optimal design solution and increasing computational costs. When neural networks are used as surrogate models, improper sample selection can lead to insufficient model accuracy. Back-propagation neural networks (BPNNs) are powerful nonlinear modeling tools capable of approximating complex functional relationships. The Grey Wolf Optimizer (GWO) is an emerging swarm intelligence algorithm with advantages such as simple structure and fast convergence. However, when solving multi-objective optimization problems, GWO is prone to falling into local optimal solutions, has slow convergence, and has poor diversity preservation capabilities. Therefore, achieving the multi-objective comprehensive optimization design of speed-increasing wind turbines faces enormous challenges. Summary of the Invention

[0007] The main purpose of the present invention is to address the deficiencies and gaps in the existing technology. The present invention provides an optimization design method for a speed-increasing wind turbine based on the OLHS-BPNN-IGWO algorithm. The multi-objective optimization method combines OLHS (Optimal Latin Hypercube Sampling), BPNN and IGWO (Improved Grey Wolf Optimization Algorithm) to improve the efficiency and accuracy of solving multi-objective optimization problems, while enhancing the global search capability to avoid falling into local optimal solutions, ensuring that the output performance of the speed-increasing wind turbine meets the design requirements, and improving the operating performance and reliability of the system.

[0008] In order to achieve the above objectives, the speed-increasing wind turbine generator described in the present invention includes: a stator, an outer rotor, an inner rotor, a rotating shaft, etc.; the stator includes a stator winding and a stator iron core; the outer rotor includes an outer rotor iron block and an outer rotor permanent magnet; the inner rotor includes an inner rotor iron core and an inner rotor permanent magnet; one end of the rotating shaft is fixed to the outer rotor, and the other end is connected to the wind wheel main shaft.

[0009] The invention provides an optimization design method for a speed-increasing wind turbine generator based on the OLHS-BPNN-IGWO algorithm, comprising the following steps:

[0010] Step 1: Select the amplitude V of the no-load back electromotive force of the speed-increasing wind turbine E , distortion rate of no-load back electromotive force V THD , electromagnetic torque pulsation T rip To optimize the target, the optimization target vector O is defined as: O = [O1, O2, O3] = [V E ,V THD ,T rip ]; Determine the design variables to be optimized for the speed-increasing wind turbine generator: the thickness h of the inner rotor permanent magnet pin , the thickness of the outer rotor permanent magnet h pout , the number of turns N of the stator winding, the proportion b of the I-shaped slot of the outer rotor permanent magnet, and the six dimensions of the stator slot, namely: hs0 、h s1 、h s2 、b s0 、b s1 、b s2 , then the vector x of the design variables to be optimized is defined as: x=[x1,x2,x3,x4,x5,x6,x7,x8,x9,x 10 ]=[h pin ,h pout ,N,b,h s0 ,h s1 ,h s2 ,b s0 ,b s1 ,b s2 ];

[0011] Step 2: Determine the optimization objective function f(x):

[0012] f(x)={min|V E (x)-V N |,min V THD (x),min T rip (x)} (1)

[0013] Where V N is the rated voltage peak of the speed-increasing wind turbine;

[0014] Step 3: Determine the constraints of the design variable x to be optimized:

[0015]

[0016] In the formula, l and u in the subscripts of each design variable x to be optimized represent the lower bound and upper bound of the variable, respectively.

[0017] Step 4: Perform sensitivity analysis on the design variable x to be optimized:

[0018] Define the design variable x to be optimized i For the optimization goal O m The local sensitivity coefficient is:

[0019]

[0020] Where S m (x i ) is the i-th design variable to be optimized x i For the mth optimization objective O m Local sensitivity coefficient, i=1,2,…,10, m=1,2,3; ΔO m is the change of the mth optimization objective, O m(x0) is the initial value of the mth optimization objective, Δx i is the i-th design variable to be optimized x i The change in x i0 is the i-th design variable to be optimized x i x0 is the initial value of the vector x of the design variables to be optimized.

[0021] Define the global sensitivity function S g (x i )for:

[0022]

[0023] Where k1, k2, and k3 are the optimization targets V E 、V THD and T rip The weight coefficient of the three is equal to 1, that is: k1+k2+k3=1; S1(x i )、S2(x i )、S3(x i ) are the i-th design variable to be optimized x i Local sensitivity coefficients to the optimization targets O1, O2, and O3.

[0024] Set the criterion for whether the design variable to be optimized is a highly sensitive parameter:

[0025] If |S g (x i )|≥δ, then the design variable to be optimized x i It is a highly sensitive parameter and has a significant impact on the optimization target;

[0026] If |S g (x i )|<δ, then the design variable to be optimized x i It is a weakly sensitive parameter with limited impact;

[0027] Where δ is the global sensitivity analysis threshold.

[0028] Step 5: Perform the Optimal Latin Hypercube Sampling (OLHS) experiment:

[0029] 51) Based on the Latin Hypercube Sampling (LHS) experiment, the initial sample space is established:

[0030] For each design variable x to be optimized i , according to formula (2), its value interval is divided into K equal subintervals, and a value is randomly selected as a sample in each subinterval to ensure that there is only one sample in each subinterval; the 10 variables to be optimized x iThe samples of each subinterval of form a 10×K Latin hypercube sampling sample matrix Φ as the initial sample space:

[0031]

[0032] 52) Optimal Latin Hypercube Sampling (OLHS) rearranges and optimizes the initial samples using the maximum distance criterion:

[0033] A1. Set the maximum number of cycles F of OLHS;

[0034] A2. Calculate the straight-line distance between all sample points: For any two sample points x in formula (5) p =[x 1j …x ij …x 10j ] and x q =[x 1k …x ik …x 10k ], i = 1, 2, ..., 10, j = 1, 2, ..., K, k = 1, 2, ..., K, the distance is:

[0035]

[0036] A3. Determine the minimum distance D between all sample points min :

[0037]

[0038] A4. Adjust the sample values in formula (5), recalculate the straight-line distance between all sample points according to formula (6), and re-determine the minimum distance according to formula (7) to obtain the new minimum distance D min_new ;

[0039] A5. If D min_new >D min , then update the sample point value; otherwise restore the original sample point configuration;

[0040] A6. Determine whether the maximum number of cycles F has been reached. If so, the training ends and the optimized samples are output; otherwise, return to step A4;

[0041] Step 6: Establish a BPNN agent model: The BPNN includes 1 input layer, 1 hidden layer, and 1 output layer. The input layer has 10 neurons, corresponding to 10 design variables to be optimized x1~x 10 The hidden layer has 150 neurons, and the activation function f(·) of the hidden layer adopts Sigmoid function; the output layer has 3 neurons, corresponding to are the amplitudes V of the no-load back electromotive force of the speed-increasing wind turbine respectively. E , distortion rate of no-load back electromotive force V THD , electromagnetic torque pulsation T rip The predicted value of

[0042] Step 7: Apply IGWO to the BPNN training process to optimize the BPNN training and obtain the optimal solution for the design variables to be optimized. The training steps are as follows:

[0043] 71) Initialize BPNN: Set the connection weights between the input layer and the output layer, and between the hidden layer and the output layer to be w il and w lm , the thresholds of the hidden layer and the output layer are a l and b m , l=1,2,…,150, m=1,2,3;

[0044] 72) Initialize IGWO: Set the size of the gray wolf group M and the maximum number of iterations T of IGWO max And the gray wolf position X, the gray wolf position is the connection weight and threshold of BPNN, expressed as X = [w il ,a l ,w lm ,b m ];

[0045] 73) Determine the three alpha wolves in IGWO: α, β, and δ. The specific steps are:

[0046] B1. Perform feedforward calculation to obtain the input and output of the hidden layer of the BPNN, as well as the output of the output layer of the BPNN:

[0047] The current input H of the lth neuron in the hidden layer l , output HO l They are:

[0048]

[0049] Where l = 1, 2, ..., 150 is the number of neurons in the hidden layer, w il is the connection weight between the i-th neuron in the input layer and the l-th neuron in the hidden layer, a l is the threshold of the lth neuron in the hidden layer, f(·) is the activation function of the hidden layer, and is the Sigmoid function:

[0050] The current output of the mth neuron in the output layer for:

[0051]

[0052] Where w lm is the connection weight between the lth neuron in the hidden layer and the mth neuron in the output layer, b m is the threshold of the mth neuron in the output layer, m=1, 2, 3;

[0053] B2. Calculate the prediction error e m :

[0054]

[0055] Where Y m is the expected output.

[0056] B3. Update the connection weights of each node in the BPNN based on the prediction error obtained in step B2;

[0057]

[0058] Where η is the learning rate.

[0059] B4. Update the threshold of each node in the network based on the prediction error obtained in step B2;

[0060]

[0061] B5. Evaluate fitness: Evaluate the fitness of each wolf based on the BPNN training results. The smaller the error, the greater the fitness. Calculate the MSE error of the BPNN on the training set corresponding to each wolf's position:

[0062]

[0063] Among them, Y mk is the target output of the k-th gray wolf, is the predicted output of the BPNN for the kth gray wolf, k = 1, 2, …, M.

[0064] The three gray wolves with the highest fitness are called α, β, and δ wolves, and the other gray wolves are called other gray wolves.

[0065] 74) Use the α wolf obtained in step 73) to perform BPNN optimization training. The specific process is as follows:

[0066] C1. The position X of the wolf α obtained in step 73) α Indicated as X α =[w ilα ,a lα ,w lmα ,b mα ], use this parameter as the initial connection weight and threshold of BPNN, where wilα 、a lα 、w lmα 、b mα They are the connection weight between the input layer and the output layer, the threshold of the hidden layer, the connection weight between the hidden layer and the output layer, and the threshold of the output layer of the α wolf BPNN;

[0067] C2. Perform BPNN forward propagation to obtain the input and output of the hidden layer of BPNN, as well as the output of the output layer of BPNN: At this time, the current input H of the lth neuron in the hidden layer is lα , output HO lα They are:

[0068]

[0069] Where l = 1, 2, ..., 150 is the number of neurons in the hidden layer, w ilα is the connection weight between the i-th neuron in the input layer and the l-th neuron in the hidden layer, a lα is the threshold of the lth neuron in the hidden layer, f(·) is the activation function of the hidden layer, and the Sigmoid function is used;

[0070] The current output of the mth neuron in the output layer for:

[0071]

[0072] Where w lmα is the connection weight between the lth neuron in the hidden layer and the mth neuron in the output layer, b mα is the threshold of the mth neuron in the output layer, m=1, 2, 3;

[0073] C3. Calculate the prediction error e mα :

[0074]

[0075] Where Y mα is the expected output.

[0076] C4. Update the connection weights of each node in the BPNN based on the prediction error obtained in step C3;

[0077]

[0078] Where η is the learning rate.

[0079] C5. Update the thresholds of each node in the network based on the prediction error obtained in step C3;

[0080]

[0081] 75) Update the positions of other gray wolves: According to the update mechanism of IGWO, update the positions of other gray wolves X d , d=1,2,…,M-3, the update formula is:

[0082] X d =X d +(A1·D αd )+(A2·D βd )+(A3·D δd ) (twenty one)

[0083] Where A1, A2, and A3 are convergence factors; D αd 、D βd 、D δd are the distances between other gray wolves and α, β, and δ wolves, respectively;

[0084] A1, A2, and A3 are calculated as follows:

[0085] A1=(2r1-1)a (22)

[0086] A2=(2r2-1)a (23)

[0087] A3=(2r3-1)a (24)

[0088] Where r1, r2, and r3 are random numbers in the interval [0-1]; a is the decreasing coefficient of (2-0), which is calculated as follows:

[0089]

[0090] Where t is the current iteration number, T max is the maximum number of iterations.

[0091] D αd 、D βd 、D δd Calculate as follows:

[0092] D αd =|C×X α -X d | (26)

[0093] D βd =|C×X β -X d | (27)

[0094] D δd =|C×X δ -X d | (28)

[0095] Where, X α 、X β 、X δ are the positions of α, β, and δ wolves respectively; C is the convergence factor, C=2r, and r is a random number in the interval [0-1].

[0096] 76) Determine whether the maximum number of iterations T has been reached max If yes, the training ends, otherwise returns to step 74) C2.

[0097] 77) Determine whether the fitting accuracy is greater than 0.9. If so, the optimal solution of the design variable to be optimized is obtained; otherwise, return to step 71).

[0098] Step 8: Import the optimal solution of the design variables to be optimized obtained in step 7 into the finite element simulation model, conduct simulation experiments, and obtain three optimization target values, namely: the amplitude V of the no-load back electromotive force of the speed-increasing wind turbine E , distortion rate of no-load back electromotive force V THD , electromagnetic torque pulsation T rip .

[0099] Step 9: Evaluate whether the three optimization target values obtained in step 8 all meet the design requirements. If so, end the optimization; otherwise, return to step 5.

[0100] The beneficial effects of the present invention are:

[0101] 1) The BP neural network (BPNN) can establish the complex nonlinear relationship between the generator design parameters and performance indicators. Combined with the high-quality initial samples generated by the optimal Latin hypercube sampling (OLHS), the BPNN model can be trained more accurately, thereby improving the optimization accuracy. The improved grey wolf optimization algorithm (IGWO) effectively avoids the problem of the BPNN algorithm easily falling into the local optimum by introducing the memory-guided position update equation and nonlinear convergence factor. This enables the optimization process to more comprehensively explore the design space and find a solution closer to the global optimum.

[0102] 2) The OLHS-BPNN-IGWO multi-objective optimization method can significantly improve the accuracy and efficiency of generator design, while reducing the optimization cost, providing strong technical support for the high-performance design of generators. BRIEF DESCRIPTION OF THE DRAWINGS

[0103] Figure 1 The figure is a schematic diagram of the topological structure of the speed-increasing vertical axis wind turbine according to the present invention.

[0104] Figure 2 The figure is a schematic diagram of the stator and rotor topological structure of the speed-increasing vertical axis wind turbine of the present invention.

[0105] Figure 3 Schematic diagram of the process of optimizing the design method of the present invention.

[0106] Figure 4 Schematic diagram of some parameters to be optimized for the speed-increasing vertical axis wind turbine of the present invention.

[0107] Figure 5 This is the BPNN agent model for the optimization design method of the present invention.

[0108] Figure 6 This is a comparison chart of the effects of the two sampling methods OLHS and LHS of the present invention.

[0109] Figure 7 Schematic diagram of the BPNN proxy model fitting accuracy of the optimization design method of the present invention. Figure 8 Comparison of electromagnetic torque curves of speed-increasing vertical axis wind turbines before and after optimization. Figure 9 This is a comparison diagram of the no-load back electromotive force of the speed-increasing vertical axis wind turbine before and after optimization.

[0110] Among them, 1-stator; 2-outer rotor; 3-inner rotor; 4-rotating shaft; 5-housing; 6-wind wheel; 11-stator core; 12-stator winding; 21-outer rotor iron block; 22-outer rotor permanent magnet; 31-inner rotor core; 32-inner rotor permanent magnet. DETAILED DESCRIPTION

[0111] The present invention will be described in further detail below with reference to the accompanying drawings.

[0112] like Figure 1 、 Figure 2 As shown, the speed-increasing wind turbine generator described in the present invention includes: a stator 1, an outer rotor 2, an inner rotor 3, a rotating shaft 4, etc.; the stator 1 includes a stator winding 12 and a stator core 11, the outer rotor 2 includes an outer rotor iron block 21 and an outer rotor permanent magnet 22; the inner rotor 3 includes an inner rotor core 31 and an inner rotor permanent magnet 32; one end of the rotating shaft 4 is fixed to the outer rotor 2, and the other end is connected to the wind wheel main shaft.

[0113] like Figure 3 、 Figure 4 As shown, the optimization design method of the speed-increasing wind turbine based on the OLHS-BPNN-IGWO algorithm of the present invention comprises the following steps:

[0114] Step 1: Select the amplitude V of the no-load back electromotive force of the speed-increasing wind turbine E , distortion rate of no-load back electromotive force V THD , electromagnetic torque pulsation T rip To optimize the target, the optimization target vector O is defined as: O = [O1, O2, O3] = [VE ,V THD ,T rip ]; Determine the design variables to be optimized for the speed-increasing wind turbine: the thickness h of the inner rotor permanent magnet pin , the thickness of the outer rotor permanent magnet h pout (See Figure 4 a), the number of turns N of the stator winding, the proportion b of the I-shaped slots of the outer rotor permanent magnets (see Figure 4 b) and the six dimensions of the stator slots (see Figure 4 c), namely: h s0 、h s1 、h s2 、b s0 、b s1 、b s2 , then the vector x of the design variables to be optimized is defined as: x=[x1,x2,x3,x4,x5,x6,x7,x8,x9,x 10 ]=[h pin ,h pout ,N,b,h s0 ,h s1 ,h s2 ,b s0 ,b s1 ,b s2 ];

[0115] Step 2: Determine the optimization objective function f(x):

[0116] f(x)={min|V E (x)-V N |,min V THD (x),min T rip (x)} (1)

[0117] Where V N is the rated voltage peak; V E (x), V THD (x), T rip (x) represents the amplitude V of the no-load back electromotive force of the speed-increasing wind turbine corresponding to the vector x of the design variable to be optimized. E , distortion rate V THD and electromagnetic torque ripple T rip ;

[0118] Step 3: Determine the constraints of the design variable x to be optimized:

[0119]

[0120] In the formula, l and u in the subscripts of each design variable x to be optimized represent the lower bound and upper bound of the variable, respectively.

[0121] Step 4: Perform sensitivity analysis on the design variable x to be optimized:

[0122] Define the design variable x to be optimized i For the optimization goal O m The local sensitivity coefficient is:

[0123]

[0124] Where S m (x i ) is the i-th design variable to be optimized x i For the mth optimization objective O m Local sensitivity coefficient, i=1,2,…,10, m=1,2,3; ΔO m is the change of the mth optimization target, O(x0) is the initial value of the mth optimization target, Δx i is the i-th design variable to be optimized x i The change in x i0 is the i-th design variable to be optimized x i x0 is the initial value of the vector x of the design variables to be optimized.

[0125] Define the global sensitivity function S g (x i )for:

[0126]

[0127] Where k1, k2, and k3 are the optimization targets V E 、V THD and T rip The weight coefficient of the three is equal to 1, that is: k1+k2+k3=1; S1(x i )、S2(x i )、S3(x i ) are the i-th design variable to be optimized x i Local sensitivity coefficients to the optimization targets O1, O2, and O3.

[0128] In order to achieve refined management of structural parameters, the design variables to be optimized x are set i Whether it is a highly sensitive parameter is determined by:

[0129] If |S g (x i )|≥δ, then the design variable to be optimized x i It is a highly sensitive parameter and has a significant impact on the optimization target;

[0130] If |S g (x i)|<δ, then the design variable to be optimized x i It is a weakly sensitive parameter with limited impact;

[0131] Wherein, δ is the global sensitivity analysis threshold, for example, δ = 0.05.

[0132] Classifying the optimization variables according to this criterion helps to prioritize key parameters during optimization and improve efficiency and accuracy.

[0133] In addition, the weight coefficients (k1, k2, k3) reflect the degree of influence of the design parameters on the target characteristics, and the larger the weight coefficient, the more obvious the influence. By analyzing the size of the weight coefficients, the key factors affecting the performance of the generator can be determined.

[0134] Step 5: Perform the Optimal Latin Hypercube Sampling (OLHS) experiment:

[0135] 51) Based on the Latin Hypercube Sampling (LHS) experiment, the initial sample space is established:

[0136] For each design variable x to be optimized i , according to formula (2), its value interval is divided into K equal subintervals, and a value is randomly selected in each subinterval as a sample to ensure that there is only one sample in each subinterval; the samples of each subinterval of the 10 variables to be optimized constitute a 10×K Latin hypercube sampling sample matrix Φ as the initial sample space:

[0137]

[0138] 52) Optimal Latin Hypercube Sampling (OLHS) rearranges and optimizes the initial samples using the maximum distance criterion:

[0139] A1. Set the maximum number of cycles F of OLHS;

[0140] A2. Calculate the straight-line distance between all sample points: For any two sample points x in formula (5) p =[x 1j …x ij …x 10j ] and x q =[x 1k …x ik …x 10k ], i = 1, 2, ..., 10, j = 1, 2, ..., K, k = 1, 2, ..., K, the distance is:

[0141]

[0142] A3. Determine the minimum distance D between all sample points min :

[0143]

[0144] A4. Adjust the sample values in formula (5), recalculate the straight-line distance between all sample points according to formula (6), and re-determine the minimum distance according to formula (7) to obtain the new minimum distance D min_new ;

[0145] A5. If D min_new >D min , then update the sample point value; otherwise restore the original sample point configuration;

[0146] A6. Determine whether the maximum number of cycles F has been reached. If so, the training ends and the optimized samples are output; otherwise, return to step A4.

[0147] Step 6: Establish BPNN agent model: Figure 5 As shown in Figure 2, BPNN consists of 1 input layer, 1 hidden layer, and 1 output layer. The input layer has 10 neurons, corresponding to 10 design variables to be optimized x1~x 10 ; The hidden layer has 150 neurons; The output layer has 3 neurons They correspond to the amplitude V of the no-load back electromotive force of the speed-increasing wind turbine. E , distortion rate of no-load back electromotive force V THD , electromagnetic torque pulsation T rip The predicted value of ; the activation function f(·) of the hidden layer adopts the Sigmoid function;

[0148] Step 7: Apply IGWO to the BPNN training process to optimize the BPNN training and obtain the optimal solution for the design variables to be optimized. The specific steps are as follows:

[0149] 71) Initialize BPNN: Set the connection weights between the input layer and the output layer, and between the hidden layer and the output layer to be w il and w lm , the thresholds of the hidden layer and the output layer are a l and b m , l=1,2,…,150, m=1,2,3;

[0150] 72) Initialize IGWO: Set the size of the gray wolf group M and the maximum number of iterations T of IGWO max And the gray wolf position X, the gray wolf position is the connection weight and threshold of BPNN, expressed as X = [w il ,a l ,w lm ,b m ];

[0151] 73) Determine the three alpha wolves in IGWO: α, β, and δ. The specific steps are:

[0152] B1. Perform feedforward calculations to obtain the input and output of the hidden layer of the BPNN, as well as the output of the output layer of the BPNN:

[0153] The current input H of the lth neuron in the hidden layer l , output HO l They are:

[0154]

[0155] Where l = 1, 2, ..., 150 is the number of neurons in the hidden layer, w il is the connection weight between the i-th neuron in the input layer and the l-th neuron in the hidden layer, a l is the threshold of the lth neuron in the hidden layer,

[0156] The current output of the mth neuron in the output layer for:

[0157]

[0158] Where w lm is the connection weight between the lth neuron in the hidden layer and the mth neuron in the output layer, b m is the threshold of the mth neuron in the output layer, m = 1, 2, 3;

[0159] B2. Calculate the prediction error e m :

[0160]

[0161] Where Y m is the expected output.

[0162] B3. Update the connection weights of each node in the BPNN based on the prediction error obtained in step B2;

[0163]

[0164] Where η is the learning rate.

[0165] B4. Update the threshold of each node in the network based on the prediction error obtained in step B2;

[0166]

[0167] A5. Evaluate fitness: Evaluate the fitness of each wolf based on the BPNN training results. The smaller the error, the greater the fitness. Calculate the MSE error of the BPNN on the training set for each wolf's location:

[0168]

[0169] Among them, Y mk is the target output of the k-th gray wolf, is the predicted output of the BPNN for the kth gray wolf, k = 1, 2, …, M.

[0170] The three gray wolves with the highest fitness are called α, β, and δ wolves, and the other gray wolves are called other gray wolves.

[0171] 74) Use the α wolf obtained in step 73) to perform BPNN optimization training. The specific process is as follows:

[0172] C1. The position X of the wolf α obtained in step 73) α Indicated as X α =[w ilα ,a lα ,w lmα ,b mα ], and use this parameter as the initial connection weight and threshold of α wolf BPNN;

[0173] C2. Perform BPNN forward propagation to obtain the input and output of the hidden layer of the α wolf BPNN and the output of the output layer: At this time, the current input H of the lth neuron in the hidden layer lα , output HO lα They are:

[0174]

[0175] Where l = 1, 2, ..., 150 is the number of neurons in the hidden layer, w ilα is the connection weight between the i-th neuron in the input layer and the l-th neuron in the hidden layer, a lα is the threshold of the lth neuron in the hidden layer;

[0176] The current output of the mth neuron in the output layer for:

[0177]

[0178] Where w lmα is the connection weight between the lth neuron in the hidden layer and the mth neuron in the output layer, b mα is the threshold of the mth neuron in the output layer, m = 1, 2, 3;

[0179] C3. Calculate the prediction error e mα :

[0180]

[0181] Where Y mα is the expected output.

[0182] C4. Update the connection weights of each node in the BPNN based on the prediction error obtained in step C3;

[0183]

[0184] Where η is the learning rate.

[0185] C5. Update the thresholds of each node in the network based on the prediction error obtained in step C3;

[0186]

[0187] 75) Update the positions of other gray wolves: According to the update mechanism of IGWO, update the positions of other gray wolves X d , d=1,2,…,M-3, the update formula is:

[0188] X d =X d +(A1·D αd )+(A2·D βd )+(A3·D δd ) (twenty one)

[0189] Where A1, A2, and A3 are nonlinear convergence factors, which determine the update step size of the wolf at each iteration; D αd 、D βd 、D δd are the distances between other gray wolves and α, β, and δ wolves, respectively.

[0190] A1, A2, and A3 are calculated as follows:

[0191] A1=(2r1-1)a (22)

[0192] A2=(2r2-1)a (23)

[0193] A3=(2r3-1)a (24)

[0194] Among them, r1, r2, and r3 are random numbers in the interval [0-1]; a is the decreasing coefficient of (2-0], which is calculated as follows:

[0195]

[0196] Where t is the current iteration number, T maxis the maximum number of iterations.

[0197] D αd 、D βd 、D δd Calculate as follows:

[0198] D αd =|C×X α -X d | (26)

[0199] D βd =|C×X β -X d | (27)

[0200] D δd =|C×X δ -X d | (28)

[0201] Where, X α 、X β 、X δ are the positions of the wolves α, β, and δ respectively; C is the convergence factor, C=2r, r is a random number in the interval [0-1], which adjusts the update step size of the wolf together with A1, A2, and A3.

[0202] 76) Determine whether the maximum number of iterations T has been reached max If yes, the training ends, otherwise returns to step 74) C2.

[0203] 77) Determine whether the fitting accuracy is greater than 0.9. If so, obtain the optimal design variable combination; otherwise, return to step 71).

[0204] Step 8: Import the optimal solution of the design variables to be optimized obtained in step 7 into the finite element simulation model, conduct simulation experiments, and obtain three optimization target values, namely: the amplitude V of the no-load back electromotive force of the speed-increasing wind turbine E and its distortion rate V THD , electromagnetic torque pulsation T rip .

[0205] Step 9: Evaluate whether the three optimization target values obtained in step 8 all meet the design requirements. If so, end the optimization; otherwise, return to step 5.

[0206] The present invention is further described below with reference to a preferred embodiment.

[0207] Taking a 5kW speed-increasing wind turbine generator according to the present invention as an example, its technical parameters are shown in Table 1.

[0208] Table 1 Technical parameters of speed-increasing wind turbines

[0209]

[0210] According to the design requirements in Table 1, the preliminary main size parameters and the values of the design variables to be optimized of the speed-increasing wind turbine are shown in Table 2 and Table 3 respectively.

[0211] Table 2 Main dimensional parameters preliminarily determined

[0212]

[0213] Table 3 Preliminary determined values of design variables to be optimized

[0214]

[0215] First, the constraints of the design variables to be optimized are determined, and their initial adjustment ranges are shown in Table 4.

[0216] Figure 6 For Latin hypercube sampling LHS( Figure 6 a) and the optimal Latin hypercube sampling OLHS of the present invention ( Figure 6 b) Comparison of the effects of the two sampling methods. LHS uses stratified sampling to evenly divide the range of experimental factor values into several sub-intervals for independent sampling. Compared with manual sampling, it is more flexible and the number of experiments can be controlled, but the distribution is uneven. OLHS rearranges and optimizes the initial samples through optimization criteria to achieve uniform, random, and orthogonal sampling, obtaining rich information with a small number of points, and good space filling and balance. Figure 6 It can be seen that the test points are more evenly distributed in the two-dimensional space by using OLHS.

[0217] Table 4 Design variables to be optimized and their adjustment ranges

[0218]

[0219] After multi-objective optimization is performed using the optimization design method of the present invention, random sampling experiments are conducted to verify the fitting accuracy of the surrogate model and evaluate the prediction efficiency. Figure 7 Predict graphs for three optimization objectives, Figure 7 a, 7b, 7c are the amplitudes of the no-load back electromotive force V of the speed-increasing wind turbine E , distortion rate of no-load back electromotive force V THD , electromagnetic torque pulsation T rip The prediction graph is shown in the figure. The dotted line is the ideal target line Y = T. As can be seen from the figure, most of the data points are distributed on the diagonal line of Y = T, which indicates that the difference between the predicted value and the actual observed value is small, showing that the model has excellent prediction performance. Generally speaking, when the determination coefficient R 2 When it is greater than 0.9, it means that the result is more reliable, that is, the fitting accuracy meets the requirements. Figure 7R of each subgraph 2 The values are all higher than 0.99, indicating that the BPNN proxy model has very good fitting performance for the no-load back EMF and its distortion rate, as well as the electromagnetic torque ripple.

[0220] Table 5 shows the values of the design variables to be optimized of the speed-increasing wind turbine before and after optimization using the optimization design method of the present invention.

[0221] Table 5 Design variable values to be optimized before and after optimization of speed-increasing wind turbine

[0222] Optimization parameters Value before optimization Optimized values <![CDATA[Inner rotor permanent magnet thickness h pin / mm]]> 4 6 <![CDATA[Thickness h of outer rotor permanent magnet pout / mm]]> 55 53 Number of stator winding turns N 100 110 The proportion of I-shaped slots in the outer rotor permanent magnets b 1.1 1.1 <![CDATA[h s0 ]]> 2 2 <![CDATA[h s1 ]]> 5 8.5 <![CDATA[h s2 ]]> 12 50 <![CDATA[b s0 ]]> 5 3.2 <![CDATA[b s1 ]]> 8 9.7 <![CDATA[b s2 ]]> 13 10.7

[0223] Figure 8 The following graph compares the electromagnetic torque curves before and after optimization. It clearly shows that after optimization, the electromagnetic torque of the speed-increasing wind turbine is closer to the stall torque. Based on the data in the graph, calculations show that torque ripple has dropped from 3.63% to 2.42%, a 33.3% reduction, demonstrating significant optimization results.

[0224] Figure 9 The following figure compares the no-load back EMF before and after optimization. The rated no-load back EMF phase voltage is set at 220V. The comparison shows that the no-load back EMF phase voltage has been optimized from 301V before optimization to 220.8V, which is basically consistent with the rated voltage, demonstrating a significant optimization effect.

[0225] In addition, according to the distortion rate calculation formula, it can be calculated that the distortion rate of no-load back electromotive force is reduced from 5.9% before optimization to 2.9%, and the optimization effect is significant.

[0226] In summary, the optimization design method for speed-increasing wind turbines based on the OLHS-BPNN-IGWO algorithm of the present invention effectively avoids the problem that the BPNN algorithm is easily trapped in the local optimum, so that the optimization process can more comprehensively explore the design space and find a solution that is closer to the global optimum; at the same time, it can significantly improve the accuracy and efficiency of generator design, reduce optimization costs, and provide strong technical support for the high-performance design of generators.

Claims

1. An optimization design method for a speed-increasing wind turbine generator based on an OLHS-BPNN-IGWO algorithm, wherein the speed-increasing wind turbine generator comprises: stator, outer rotor, inner rotor; the stator includes stator winding and stator core; the outer rotor includes outer rotor iron block and outer rotor permanent magnet; the inner rotor includes inner rotor core and inner rotor permanent magnet; characterized in that the following steps are adopted: Step 1: Select the amplitude V of the no-load back electromotive force of the speed-increasing wind turbine E , no-load back electromotive force distortion rate V THD , electromagnetic torque pulsation T rip To optimize the target, the vector O of the optimization target is defined as: O = [O1, O2, O3] = [V E ,V THD ,T rip ]; Determine the design variables to be optimized for the speed-increasing wind turbine generator: the thickness h of the inner rotor permanent magnet pin , the thickness of the outer rotor permanent magnet h pout , the number of turns N of the stator winding, the proportion b of the I-shaped slot of the outer rotor permanent magnet, and the six dimensions of the stator slot, namely: h s0 、h s1 、h s2 、b s0 、b s1 、b s2 , then the vector x of the design variables to be optimized is defined as: x=[x1,x2,x3,x4,x5,x6,x7,x8,x9,x 10 ]=[h pin ,h pout ,N,b,h s0 ,h s1 ,h s2 ,b s0 ,b s1 ,b s2 ]; Step 2: Determine the optimization objective function f(x): f(x)={min|V E (x)-V N |,minV THD (x),minT rip (x)} (1) Where V N is the rated voltage peak of the speed-increasing wind turbine; Step 3: Determine the constraints of the design variable x to be optimized: In the formula, l and u in the subscripts of each design variable x to be optimized represent the lower bound and upper bound of the variable respectively; Step 4: Perform sensitivity analysis on the design variable x to be optimized: Define the design variable x to be optimized i For the optimization goal O m The local sensitivity coefficient is: Where S m (x i ) is the i-th design variable to be optimized x i For the mth optimization objective O m The local sensitivity coefficient, i = 1, 2, ..., 10, m = 1, 2, 3; ΔO m is the change of the mth optimization objective, O m (x0) is the initial value of the mth optimization objective, Δx i is the i-th design variable to be optimized x i The change in x i0 is the i-th design variable to be optimized x i The initial value of ; x0 is the initial value of the vector x of the design variable to be optimized; Define the global sensitivity function S g (x i )for: Where k1, k2, and k3 are the optimization targets V E 、V THD and T rip The weight coefficient of the three is equal to 1; S1(x i )、S2(x i )、S3(x i ) are the i-th design variable to be optimized x i Local sensitivity coefficients to the optimization targets O1, O2, and O3; The criterion for whether the design variable to be optimized is a highly sensitive parameter is set as: If |S g (x i )|≥δ, then the design variable to be optimized x i It is a highly sensitive parameter and has a significant impact on the optimization target; If |S g (x i )|<δ, then the design variable to be optimized x i It is a weakly sensitive parameter with limited impact; Among them, δ is the global sensitivity analysis threshold; Step 5, perform the optimal Latin hypercube sampling OLHS experiment; Step 6: Establish a BPNN agent model: The BPNN includes 1 input layer, 1 hidden layer, and 1 output layer. The input layer has 10 neurons, corresponding to 10 design variables to be optimized x1~x 10 The hidden layer has 150 neurons, and the activation function f(·) of the hidden layer adopts Sigmoid function; the output layer has 3 neurons, corresponding to are the amplitudes V of the no-load back electromotive force of the speed-increasing wind turbine respectively. E , no-load back electromotive force distortion rate V THD , electromagnetic torque pulsation T rip The predicted value of Step 7: Apply the improved grey wolf optimization algorithm (IGWO) to the BPNN training process to optimize the BPNN training and obtain the optimal solution of the design variables to be optimized. Step 8: Import the optimal solution of the design variables to be optimized obtained in step 7 into the finite element simulation model, conduct simulation experiments, and obtain three optimization target values, namely: the amplitude V of the no-load back electromotive force of the speed-increasing wind turbine E , no-load back electromotive force distortion rate V THD , electromagnetic torque pulsation T rip ; Step 9: Evaluate whether the three optimization target values obtained in step 8 all meet the design requirements. If so, end the optimization; otherwise, return to step 5.

2. The method for optimizing the design of a speed-increasing wind turbine generator based on the OLHS-BPNN-IGWO algorithm according to claim 1, characterized in that: The specific steps of step 5 are: 51) Based on the Latin Hypercube Sampling LHS experiment, the initial sample space is established: For each design variable x to be optimized i , according to formula (2), its value interval is divided into K equal subintervals, and a value is randomly selected as a sample in each subinterval to ensure that there is only one sample in each subinterval; the 10 variables to be optimized x i The samples of each subinterval of form a 10×K Latin hypercube sampling sample matrix Φ as the initial sample space: 52) Optimal Latin Hypercube Sampling (OLHS) rearranges and optimizes the initial samples using the maximum distance criterion: A1. Set the maximum number of cycles F of OLHS; A2. Calculate the straight-line distance between all sample points: For any two sample points x in formula (5) p =[x 1j …x ij …x 10j ] and x q =[x 1k …x ik …x 10k ], i = 1, 2, ..., 10, j = 1, 2, ..., K, k = 1, 2, ..., K, the distance is: A3. Determine the minimum distance D between all sample points min : A4. Adjust the sample values in formula (5), recalculate the straight-line distance between all sample points according to formula (6), and re-determine the minimum distance according to formula (7) to obtain the new minimum distance D min_new ; A5. If D min_new >D min , then update the sample point value; otherwise restore the original sample point configuration; A6. Determine whether the maximum number of cycles F has been reached. If so, training ends and the optimized samples are output. Otherwise, return to step A4.

3. The method for optimizing the design of a speed-increasing wind turbine generator based on the OLHS-BPNN-IGWO algorithm according to claim 1, characterized in that: The specific steps of step 7 are: 71) Initialize BPNN: Set the connection weights between the input layer and the output layer, and between the hidden layer and the output layer to be w il and w lm , the thresholds of the hidden layer and the output layer are a l and b m , l=1,2,…,150, m=1,2,3; 72) Initialize IGWO: Set the size of the gray wolf group M and the maximum number of iterations T of IGWO max And the gray wolf position X, the gray wolf position is the connection weight and threshold of BPNN, expressed as X = [w il ,a l ,w lm ,b m ]; 73) Determine the three alpha wolves in IGWO: α, β, and δ. The specific steps are: B1. Perform feedforward calculation to obtain the input and output of the hidden layer of the BPNN, as well as the output of the output layer of the BPNN: The current input H of the lth neuron in the hidden layer l , output HO l They are: Where l = 1, 2, ..., 150 is the number of neurons in the hidden layer, w il is the connection weight between the i-th neuron in the input layer and the l-th neuron in the hidden layer, a l is the threshold of the lth neuron in the hidden layer, f(·) is the activation function of the hidden layer, and is the Sigmoid function: The current output of the mth neuron in the output layer for: Where w lm is the connection weight between the lth neuron in the hidden layer and the mth neuron in the output layer, b m is the threshold of the mth neuron in the output layer; B2. Calculate the prediction error e m : Where Y m is the expected output; B3. Update the connection weights of each node in the BPNN based on the prediction error obtained in step B2; Where η is the learning rate; B4. Update the threshold of each node in the network based on the prediction error obtained in step B2; B5. Evaluate fitness: Evaluate the fitness of each wolf based on the BPNN training results. The smaller the error, the greater the fitness. Calculate the mean square error (MSE) of the BPNN on the training set for each wolf's location: Among them, Y mk is the target output of the k-th gray wolf, is the predicted output of the BPNN for the kth gray wolf, k = 1, 2, …, M; The three gray wolves with the highest fitness are called α, β, and δ wolves, and the other gray wolves are called other gray wolves; 74) Use the α wolf obtained in step 73) to perform BPNN optimization training. The specific process is as follows: C1. The position X of the wolf α obtained in step 73) α Indicated as X α =[w ilα ,a lα ,w lmα ,b mα ], and use this parameter as the initial connection weight and threshold of BPNN at this time, where w ilα 、a lα 、w lmα 、b mα They are the connection weight between the input layer and the output layer, the threshold of the hidden layer, the connection weight between the hidden layer and the output layer, and the threshold of the output layer of the α wolf BPNN; C2. Perform BPNN forward propagation to obtain the input and output of the hidden layer of BPNN, as well as the output of the output layer of BPNN: the current input H of the lth neuron in the hidden layer lα , output HO lα They are: Where l = 1, 2, ..., 150 is the number of neurons in the hidden layer, w ilα is the connection weight between the i-th neuron in the input layer and the l-th neuron in the hidden layer, a lα is the threshold of the lth neuron in the hidden layer, f(·) is the activation function of the hidden layer, and the Sigmoid function is used; The current output of the mth neuron in the output layer for: Where w lmα is the connection weight between the lth neuron in the hidden layer and the mth neuron in the output layer, b mα is the threshold of the mth neuron in the output layer, m=1, 2, 3; C3. Calculate the prediction error e mα : Where Y mα is the expected output; C4. Update the connection weights of each node in the BPNN based on the prediction error obtained in step C3; Where η is the learning rate; C5. Update the thresholds of each node in the network based on the prediction error obtained in step C3; 75) Update the positions of other gray wolves: According to the update mechanism of IGWO, update the positions of other gray wolves X d , d=1,2,…,M-3, the update formula is: X d =X d +(A1·D αd )+(A2·D βd )+(A3·D δd ) (21) Where A1, A2, and A3 are convergence factors; D αd 、D βd 、D δd are the distances between other gray wolves and α, β, and δ wolves, respectively; A1, A2, and A3 are calculated as follows: A1=(2r1-1)a (22) A2=(2r2-1)a (23) A3=(2r3-1)a (24) Where r1, r2, and r3 are random numbers in the interval [0-1]; a is the decreasing coefficient of (2-0), which is calculated as follows: Where t is the current iteration number, T max is the maximum number of iterations. D αd 、D βd 、D δd Calculate as follows: D αd =|C×X α -X d | (26) D βd =|C×X β -X d | (27) D δd =|C×X δ -X d | (28) Where, X α 、X β 、X δ are the positions of α, β, and δ wolves respectively; C is the convergence factor, C = 2r, r is a random number in the interval [0-1]; 76) Determine whether the maximum number of iterations T has been reached max If yes, the training ends, otherwise returns to step 74) C2; 77) Determine whether the fitting accuracy is greater than 0.

9. If so, the optimal solution of the design variable to be optimized is obtained; otherwise, return to step 71).

Images