A permanent magnet synchronous motor heat dissipation system optimization method based on error compensation

Abstract

The application discloses a kind of permanent magnet synchronous motor heat dissipation system optimization method based on error compensation, comprising the following steps: determining the optimization parameter of motor water jacket and selecting prediction model;Error compensation method is used to improve prediction accuracy;After establishing prediction model, multi-objective optimization is carried out to improve global optimization effect.The above technical scheme first parameterization modeling is carried out to the key parameter of motor, and the key parameter is screened by designing Plackett-Burman experiment;Second, the experimental design is carried out to design variable, and the corresponding data set of factor is obtained, and machine learning model is constructed based on sample data, and error compensation strategy is used to improve machine learning model, to improve model accuracy;Then, the improved multi-objective optimization algorithm is used to improve the running speed of optimization algorithm and the ability of global search, the model is optimized, and the best parameter combination is obtained;Finally, the effectiveness of the method is verified by experiment.

Description

Technical Field

[0001] This invention relates to the field of motor optimization design technology, and in particular to an optimization method for the heat dissipation system of a permanent magnet synchronous motor based on error compensation. Background Technology

[0002] Electric vehicles are gaining increasing popularity due to their eco-friendly characteristics. As a core component of electric vehicles, the motor has a crucial impact on their performance. Induction motors are widely used in electric vehicles due to their simple structure, reliable performance, low cost, and easy maintenance. However, the high-speed rotation of the motor causes a series of power losses, such as iron and copper losses. These losses are ultimately converted into heat, causing the motor to overheat. Excessive motor temperature can damage the motor's insulation materials and weaken its magnetic properties. Therefore, effectively controlling the motor's temperature is crucial. Optimizing the cooling system design can achieve excellent heat dissipation, significantly improving internal heat dissipation without altering the motor's original operating principle, while ensuring motor safety and stable operation.

[0003] Data shows that current applications of multi-objective optimization algorithms in the field of permanent magnet motors primarily focus on electromagnetic structure optimization, such as optimizing lamination slot shape, air gap size, pole-to-slot ratio, permanent magnet size, and rotor dynamic characteristics. Research on multi-objective optimization of motor water jacket structures is relatively limited, and the accuracy of surrogate models and multi-objective optimization algorithms used in related studies still has significant room for improvement.

[0004] Chinese patent document CN115081328A discloses a "multi-objective optimization method for motors based on an improved particle swarm optimization algorithm." First, an experimental scheme is designed based on statistical experimental principles. Then, the target motor is parametrically modeled, and the corresponding experimental results are obtained through simulation. A corresponding mathematical model is generated using the response surface methodology. Next, an improved particle swarm optimization algorithm with an added mutation library is used to generate a Pareto chart to find the optimal structure of the motor. Finally, the effectiveness of the optimization is verified through simulation. The above technical solution uses the response surface methodology to simulate the mathematical model. Low-order response surface methods are difficult to simulate complex models, while high-order response surface methods are prone to overfitting. Summary of the Invention

[0005] This invention primarily addresses the technical problems of existing solutions using response surface methodology (RSM) to simulate mathematical models. Low-order RSM struggles to simulate complex models, while high-order RSM is prone to overfitting. The invention provides an error-compensated optimization method for the cooling system of a permanent magnet synchronous motor. First, key motor parameters are parametrically modeled, and key parameters are selected through Plackett-Burman experiments. Second, experiments are designed for the design variables to obtain factors and corresponding datasets. A machine learning model is constructed based on the sample data, and an error compensation strategy is used to improve the model's accuracy. Then, an improved multi-objective optimization algorithm is used to enhance the algorithm's running speed and global search capability, optimizing the model to obtain the optimal parameter combination. Finally, experiments verify the effectiveness of the proposed method.

[0006] The above-mentioned technical problems of the present invention are mainly solved by the following technical solution: The present invention includes the following steps:

[0007] S1 determines the optimized parameters of the motor water jacket and selects the prediction model;

[0008] S2 employs an error compensation method to improve prediction accuracy;

[0009] After establishing the prediction model, S3 performs multi-objective optimization to improve the global optimization effect.

[0010] First, a Plackett-Burman experiment was designed to screen experimental factors, and then the sample set was determined using Latin hypercube sampling (LHS). The temperature rise and pressure loss of the motor were determined using CFD methods. An error-compensated GRNN network was established, and R... 2 The model accuracy was evaluated using RMSE. Multi-objective optimization was performed using the improved NSGA-II algorithm, and finally, experimental verification was conducted.

[0011] Preferably, step S1 specifically includes:

[0012] S1.1 Calculate the electromagnetic loss of the motor, and then preliminarily determine the optimized parameters of the motor water jacket;

[0013] S1.2 Design Plackett-Burman experiment to screen key parameters based on the highest temperature of the motor coil and the pressure loss of the water jacket;

[0014] S1.3 uses Latin hypercube sampling to obtain sample points, sets simulation analysis parameters to perform fluid dynamics calculations and simulations, and determines the motor coil temperature rise and water jacket pressure loss;

[0015] S1.4 A generalized regression neural network (GRNN) was selected for prediction through comparative experiments.

[0016] To verify the superiority of the error compensation model, extreme learning machine (ELM), generalized regressive neural network (GRNN), back regression neural network (BP), and extreme gradient boosting tree (XGboost) were selected for comparative experiments. GRNN showed better prediction performance, but all models performed poorly in predicting pressure loss.

[0017] Preferably, step S1.1 calculates the electromagnetic loss of the motor and then preliminarily determines the optimized parameters of the motor water jacket, specifically including seven variables: water channel height H, total water channel width L, water channel angle α, rib width l, inlet and outlet pipe diameter β, water channel roughness γ, and water volume V.

[0018] Preferably, in step S1.3 during the simulation, the fluid temperature is set to 25℃, the ambient temperature is set to 15℃, and the convective heat dissipation coefficient is set to 12W / (m²). 2 The solution algorithm was selected as couple, the model was selected as K-epslion, and the number of grids was 2.5 million. In this process, Latin hypercube sampling was first used to select 100 sample points, and the temperature rise of the motor coil and the pressure loss of the water jacket were determined by CFD method. The sample set was divided into training set 1, training set 2 and test set, with a ratio of 4:4:2.

[0019] Preferably, the prediction model is a generalized regression neural network (GRNN), which is represented as follows:

[0020]

[0021] f(X,y) is the joint probability density of the observed value X and the maximum probability value y of the non-independent variable Y relative to the independent variable X. Y is the predicted output of Y given the input X.

[0022] As a preferred approach, Parzen nonparametric estimation is applied based on the sample dataset. Estimating the density function

[0023]

[0024] In the formula, X i ,Y i Let be the sample observations of random variables x and y; n be the sample size; p be the dimension of random variable x; and σ be the width coefficient of the Gaussian function, referred to here as the smoothing factor.

[0025] use Substitute f(X,y) into the above equation and swap the order of integration and summation:

[0026]

[0027] because The network output can be obtained by calculating the two integrals. for

[0028]

[0029] estimated value For all sample observations Y i The weighted average of each observation Y i The weighting factor is the corresponding sample X i The exponent of the squared Euclid distance between X and X.

[0030] R 2 Coefficient of determination

[0031]

[0032] To further improve the accuracy of the GRNN neural network, an error compensation method was adopted to improve the model. The improved temperature rise R... 2 The pressure loss R is 0.9883. 2 A value of 0.9910 indicates a significant improvement in accuracy.

[0033] Preferably, step S3 specifically includes:

[0034] S3.1 First, a more uniform initial population is generated using the Sobol sequence during population initialization;

[0035] S3.2 Replace the binomial crossover operator (SBX) used by NSGA-II with the normally distributed crossover operator (NDX); S3.3 Add the metropolis criterion of the annealing algorithm (SA) to the elite selection strategy to increase the uncertainty of the algorithm.

[0036] Traditional multi-objective optimization algorithm NSGA-II is prone to getting trapped in local optima and failing to find the global optimum. This invention uses an improved NSGA-II algorithm to enhance the global optimization effect.

[0037] Preferably, the NDX operator formula in step S3.2 is as follows:

[0038]

[0039] Where u is a random number uniformly distributed between the interval (0,1), and |N(0,1)| is a random number following a normal distribution of N(0,1).

[0040] Preferably, the Metropolis criterion in step S3.3 is as follows:

[0041]

[0042] Where df < 0 represents accepting the old choice with a probability of 1, and df ≥ 0 represents accepting the new choice with a probability of P. The optimal solution set obtained through screening is defined as having a pressure loss of less than 0.3 bar.

[0043] The beneficial effects of this invention are: it uses an error-compensated GRNN neural network to improve prediction accuracy, and uses an improved NSGA-II algorithm to improve global search capability to achieve multi-objective optimization of the water jacket of permanent magnet motor with temperature rise and pressure loss as the objectives. This solves the problems of insufficient accuracy of surrogate models and easy trapping of multi-objective optimization algorithms in current multi-objective optimization of motor water jackets. Attached Figure Description

[0044] Figure 1 This is a flowchart of the present invention.

[0045] Figure 2 This is a schematic diagram illustrating the working principle of the present invention.

[0046] Figure 3 This is a Plackett-Burman experimental data graph of the present invention.

[0047] Figure 4 This is a generalized regression network structure diagram of the present invention.

[0048] Figure 5 This is a Pareto front plot of the NSGA-II algorithm of the present invention.

[0049] Figure 6 This is a Pareto front plot of an improved NSGA-II algorithm according to the present invention.

[0050] Figure 7 This is a diagram showing the structural parameters of a motor according to the present invention. Detailed Implementation

[0051] The technical solution of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0052] Example: This example describes an optimization method for the heat dissipation system of a permanent magnet synchronous motor based on error compensation. Figure 1 As shown, it includes the following steps:

[0053] S1 determines the optimized parameters of the motor water jacket and selects the prediction model;

[0054] S2 employs an error compensation method to improve prediction accuracy;

[0055] After establishing the prediction model, S3 performs multi-objective optimization to improve the global optimization effect.

[0056] First, the electromagnetic losses of the motor were calculated. Then, the optimal parameters for the motor's water jacket were initially determined: seven variables, including water channel height H, total water channel width L, water channel angle α, rib width l, inlet and outlet pipe diameter β, water channel roughness γ, and water flow rate V. A Plackett-Burman experiment was designed, using the highest temperature of the motor coil and the pressure loss of the water jacket as responses to screen key parameters. Experimental data are as follows: Figure 3 As shown.

[0057] The optimal parameters obtained through Plackett-Burman experiments are H, L, and V. The multi-objective optimization model can be described as follows:

[0058] min T max ,P

[0059]

[0060] Where T max Where is the highest temperature, and P is the pressure loss.

[0061] Then, Latin hypercube sampling was used to obtain sample points, and simulation analysis parameters were set to perform fluid dynamics calculations. During the simulation, the fluid temperature was set to 25℃, and the ambient temperature was set to 15℃. The convective heat dissipation coefficient was set to 12 W / (m²). 2 The algorithm chosen was a coupled approach, and the model was K-epslion with a grid size of 2.5 million. First, Latin hypercube sampling was used to select 100 sample points, and CFD was used to determine the motor coil temperature rise and water jacket pressure loss. The sample set was divided into training set 1, training set 2, and test set in a ratio of 4:4:2. To verify the superiority of the error compensation model, Extreme Learning Machine (ELM), Generalized Regression Neural Network (GRNN), Backward Regression Neural Network (BP), and Extreme Gradient Boosting Tree (XGboost) were selected for comparative experiments. The experimental results are shown in Table 1.

[0062]

[0063] Table 1 Comparison of prediction accuracy of various models

[0064] It can be seen that GRNN performs well in prediction, but all models perform poorly in predicting pressure loss. This invention employs an error compensation method to improve prediction accuracy. The GRNN network structure diagram is shown below. Figure 4 As shown.

[0065] Generalized regressive neural networks can be represented as

[0066]

[0067] f(X,y) is the joint probability density of the observed value X and the maximum probability value y of the non-independent variable Y relative to the independent variable x.

[0068] Y is the predicted output of Y given the input X.

[0069] Applying Parzen nonparametric estimation, it can be derived from the sample dataset. Estimating the density function

[0070]

[0071] In the formula, X i ,Y i Let be the sample observations of random variables x and y; n be the sample size; p be the dimension of random variable x; and σ be the width coefficient of the Gaussian function, referred to here as the smoothing factor.

[0072] use Substitute f(X,y) into the above equation and swap the order of integration and summation:

[0073]

[0074] because The network output can be obtained by calculating the two integrals. for

[0075]

[0076] estimated value For all sample observations Y i The weighted average of each observation Y i The weighting factor is the corresponding sample X i The exponent of the squared Euclid distance between X and X.

[0077] To further improve the accuracy of the GRNN neural network, an error compensation method was adopted to improve the model. The improved temperature rise R... 2 The pressure loss R is 0.9883. 2 A value of 0.9910 indicates a significant improvement in accuracy.

[0078] After establishing the prediction model, this invention begins multi-objective optimization. The traditional multi-objective optimization algorithm NSGA-II is prone to getting stuck in local optima and failing to find the global optimum. This invention uses an improved NSGA-II algorithm to enhance the global optimization effect.

[0079] First, a more uniform initial population is generated using Sobol sequences during population initialization. Second, a normally distributed crossover operator (NDX) is used instead of the binomial crossover operator (SBX) used in NSGA-II. Finally, the metropolis criterion of the annealing algorithm (SA) is added to the elite selection strategy to increase the algorithm's uncertainty. A comparison of NSGA-II and the improved NSGA-II is shown in the figure below. Figure 5 , Figure 6 As shown

[0080] The NDX operator formula is shown below:

[0081]

[0082] Where u is a random number uniformly distributed between the interval (0,1), and |N(0,1)| is a random number following a normal distribution of N(0,1).

[0083] The Metropolis guidelines are as follows:

[0084]

[0085] Where df < 0 means accepting the old choice with probability 1, and df ≥ 0 means accepting the new choice with probability P.

[0086] The optimal solution set obtained through screening was determined with a pressure loss of less than 0.3 bar. The final optimal solutions are shown in Table 2.

[0087]

[0088] Table 2 Comparison of Pareto curves between NSGA-II and improved NSGA-II

[0089]

[0090] Table 3. Parameters before and after optimization

[0091] Therefore, modifying the motor's water jacket structure can reduce the maximum temperature of the motor coil by approximately 5°C and increase the pressure loss to 0.29 bar, which meets the requirements. This method effectively reduces motor heat generation and extends the motor's service life.

[0092] To verify the accuracy of the algorithm, an experimental setup was built for testing. The experimental steps are as follows:

[0093] 1. Three sets of PT1000 sensors were embedded in the coil at the motor shaft extension end, and temperature data was recorded using a temperature monitoring instrument.

[0094] 2. During the experiment, the motor was run under rated conditions. The motor speed, torque and temperature rise data were recorded every 1 minute. When the temperature fluctuation was less than 0.5℃, the coil temperature was recorded at this time.

[0095] The pressure loss value can be read directly from the water chiller inlet as 0.285 bar.

[0096] The recorded temperature curve shows that the highest temperature rise of the motor is about 95.1℃, indicating that the method is highly accurate.

[0097] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. An optimization method for the heat dissipation system of a permanent magnet synchronous motor based on error compensation, characterized in that, Includes the following steps: S1 determines the optimized parameters of the motor water jacket and selects the prediction model; S2 employs an error compensation method to improve prediction accuracy; After establishing the prediction model, S3 performs multi-objective optimization to improve the global optimization effect; Step S3 specifically includes: S3.1 First, a more uniform initial population is generated using the Sobol sequence during population initialization; S3.2 Replaces the binomial crossover operator used by NSGA-II with a normally distributed crossover operator; S3.3 adds the Metropolis criterion of the annealing algorithm to the elite selection strategy, increasing the uncertainty of the algorithm.

2. The method for optimizing the heat dissipation system of a permanent magnet synchronous motor based on error compensation according to claim 1, characterized in that, Step S1 specifically includes: S1.1 Calculate the electromagnetic loss of the motor, and then preliminarily determine the optimized parameters of the motor water jacket; S1.2 Design Plackett-Burman experiment to screen key parameters based on the highest temperature of the motor coil and the pressure loss of the water jacket; S1.3 uses Latin hypercube sampling to obtain sample points, sets simulation analysis parameters to perform fluid dynamics calculations and simulations, and determines the motor coil temperature rise and water jacket pressure loss; S1.4 A generalized regression neural network (GRNN) was selected for prediction through comparative experiments.

3. The method for optimizing the heat dissipation system of a permanent magnet synchronous motor based on error compensation according to claim 2, characterized in that, Step S1.1 calculates the electromagnetic loss of the motor and then preliminarily determines the optimized parameters of the motor water jacket, specifically including seven variables: water channel height H, total water channel width L, water channel angle α, rib width l, inlet and outlet pipe diameter β, water channel roughness γ, and water volume V.

4. The method for optimizing the heat dissipation system of a permanent magnet synchronous motor based on error compensation according to claim 2, characterized in that, In step S1.3, during the simulation, the fluid temperature is set to 25℃, the ambient temperature is set to 15℃, the convective heat dissipation coefficient is set to 12W / (m²K), the solution algorithm is selected as couple, the model is selected as K-epslion, and the number of grids is 2.5 million. In this process, 100 sample points are first selected by Latin hypercube sampling, and the temperature rise of the motor coil and the pressure loss of the water jacket are determined by CFD method. The sample set is divided into training set 1, training set 2 and test set, with a ratio of 4:4:

2.

5. The method for optimizing the heat dissipation system of a permanent magnet synchronous motor based on error compensation according to claim 1, characterized in that, The prediction model selected is a generalized regression neural network (GRNN), which is represented as follows: for Observations X and dependent variable Y relative to independent variable maximum probability value The joint probability density, That is, the predicted output of Y given input X.

6. The method for optimizing the heat dissipation system of a permanent magnet synchronous motor based on error compensation according to claim 5, characterized in that, Applying Parzen nonparametric estimation from the sample dataset Estimating the density function , In the formula, For random variables and Sample observations; For sample size; For random variables dimensionality; is the width coefficient of the Gaussian function, referred to here as the smoothness factor; use replace Substitute into the above equation and change the order of integration and summation: because The network output can be obtained by calculating the two integrals. for estimated value For all sample observations The weighted average of each observation The weighting factors are the corresponding samples The exponent of the squared Euclid distance between X and X.

7. The method for optimizing the heat dissipation system of a permanent magnet synchronous motor based on error compensation according to claim 1, characterized in that, The NDX operator formula in step S3.2 is as follows: in, These are random numbers uniformly distributed between the intervals (0,1). These are random numbers that follow a normal distribution N(0,1).

8. The method for optimizing the heat dissipation system of a permanent magnet synchronous motor based on error compensation according to claim 1, characterized in that, The Metropolis criteria in step S3.3 are as follows: in, <0 means accepting the old choice with a probability of 1. Representative with The probability accepts a new choice.

Images