Magnetic core loss prediction model based on SA-SMA-GRNN

Abstract

The invention discloses a magnetic core loss prediction model based on SA-SMA-GRNN, and belongs to the technical field of signal prediction and optimization. According to the method, firstly, a training set containing characteristics of a magnetic core material, temperature, frequency, a magnetic flux density peak value and the like is constructed, and smooth factors of the GRNN are optimized through an SMA myxomycete optimization algorithm, so that the prediction precision and generalization ability of a model are improved. In order to prevent the colistia algorithm from falling into local optimum, a simulated annealing algorithm is further introduced, and by simulating a physical annealing process, the temperature in a search space is gradually reduced, the global search capability of the model is enhanced, and the convergence speed is improved. According to the method, the optimal model parameters are optimized by training the high-frequency and high-energy-consumption magnetic core loss data, loss prediction is performed through the GRNN, high-precision prediction of the magnetic core loss is realized, the global search capability and the convergence speed of the GRNN model are improved, and the method has wide applicability and relatively high prediction accuracy.

Description

Technical Field

[0001] The present invention relates to the field of signal prediction technology, and in particular to a magnetic core loss prediction model based on SA-SMA-GRNN. Background Art

[0002] Magnetic components are an indispensable part of modern electronic components, and their function is to realize the transmission and storage of electrical energy. As the core part of magnetic components, the core loss is also a non-negligible part in some high-frequency and high-energy consumption application scenarios. In addition, core loss will also reduce the service life of electronic components and reduce the reliability and safety of components. How to reduce the loss of magnetic cores under working conditions has become the mainstream research direction in academia.

[0003] In the field of core loss research, the industry lacks a core loss model that is both widely applicable and can provide high-precision prediction results. This directly restricts the accurate evaluation of losses in magnetic component design, which in turn affects the effective estimation of overall power converter efficiency. In view of this situation, the industry has high hopes for building a more convenient and accurate data-driven model, aiming to develop a core loss prediction model that can span different material types and operating conditions. Such a model will greatly improve the accuracy and efficiency of magnetic component design and lay a solid foundation for the further development of power electronics technology.

[0004] In recent years, with the development of artificial intelligence technology, data-driven modeling methods have gradually become a new trend in core loss research. Data-driven methods collect a large amount of experimental data under different conditions, and use machine learning algorithms to analyze and model the data, which can solve the problem that traditional models cannot cover complex working conditions to a certain extent. This type of method no longer relies on a specific physical model, but predicts losses by learning patterns in existing data. At present, there are some methods for data prediction, such as linear regression, decision trees, support vector machines, and other algorithms that predict data based on data learning. Although these methods can achieve the purpose of data prediction to a certain extent, they each have the disadvantages of high computational complexity, long training time, easy overfitting, and poor stability. Therefore, they cannot fully meet actual needs when predicting core loss.

[0005] In summary, finding a model that can accurately predict core loss has important practical significance. Summary of the invention

[0006] In view of the above-mentioned deficiencies in the prior art, the present invention provides a prediction model based on SA-SMA-GRNN.

[0007] The technical solution adopted by the present invention is a prediction model based on SA-SMA-GRNN, and the overall process includes the following 6 steps.

[0008] Step 1: Construct the discrete function E[u] of the original training set data measured by the core material work, and use E 1 [u],E 2 [u],E 3 [u],E 4 [u] represents material one, material two, material three, and material four respectively. E[u] is used to represent the original characteristic sequence obtained by directly sampling the radiation signal acquisition devices such as infrared radiation meters and electromagnetic radiation meters during the magnetic core material work.

[0009] Furthermore, the data of the original data discrete function E[u] in the above step 1 is composed of the core material temperature E1[u], the core material excitation frequency E2[u], the core loss data E3[u], the core excitation waveform E4[u] and the magnetic flux density data E5[u]…E 1029 [u], and its specific structural relationship is shown in formula (1).

[0010] E[u]=E1[u]+E2[u]+E3[u]+… (1)

[0011] Step 2: Construct the discrete function T[u] of the original prediction set data measured by the core material work, using T 1 [u],T 2 [u],T 3 [u],T 4 [u] represents material one, material two, material three, and material four respectively, and T[u] is used to represent the original characteristic sequence obtained by directly sampling the radiation signal acquisition devices such as infrared radiation meters and electromagnetic radiation meters during the magnetic core material work.

[0012] Furthermore, the data of the original data discrete function T[u] in the above step 2 is composed of the core material serial number T1[u], the core material working temperature T2[u], the core material excitation frequency T3[u], the core material type T4[u], the core excitation waveform T5[u] and the magnetic flux density data T6[u]…T 1030 [u], and its specific structural relationship is shown in formula (2).

[0013] T[u]=T1[u]+T2[u]+T3[u]+… (2)

[0014] Step 3: Calculate the Euclidean distance D between each training sample and the input sample i , construct the operation operator x j Represents the training set input variable, x ij is the corresponding feature of the i-th training sample, and its specific structural relationship formula is shown in formula (3).

[0015]

[0016] Furthermore, the smoothing factor σ in the generalized regression network neuron is defined to output the final predicted value Its specific structural relationship formula is shown in formula (4).

[0017]

[0018] Among them, y i is the output value of the training sample, and N is the number of samples.

[0019] Step 4: Use the SMA algorithm to optimize the smoothing factor σ in the generalized regression neural network and generate random N initial smoothing factors σ i , represents the initial position of the slime mold individual, where σ min and σ max is the value range of the smoothing factor, and its structure is shown in formula (5).

[0020] σ i (0) = random(σ min ,σ max ),i=1,2,...,N (5)

[0021] Further, for each smoothing factor σ i , put it into formula (6) and calculate its corresponding prediction error MSE under GRNN.

[0022]

[0023] Step 4.1: Update the concentration according to the MSE size corresponding to each slime mold individual. The higher the concentration, the better the smoothing factor σ in the area. i , using C i represents the concentration of the i-th slime mold, as shown in formula (7).

[0024]

[0025] Depending on the concentration, the slime mold individuals will move to new locations, and t is usually used to represent the tth iteration.

[0026] Step 4.2: The smoothing factor σ of slime mold individual i at time t+1 is calculated by the smoothing factor σ at time t. i (t) plus its velocity v i (t) is used to update. In each iteration, its moving formula is shown in formula (8).

[0027] σ i (t+1)=σ i (t)+v i (t) (8)

[0028] Furthermore, v i represents the speed, which is defined as shown in formula (9), σ best represents the current global optimal smoothing factor, σ random is the number of slime mold individuals σ i Another solution randomly selected from , where r1 and r2 are random numbers between 0 and 1.

[0029] v i (t) = r1·(C i ·σ best +(1-C i )·r2·σ random ) (9)

[0030] Step 4.3: Set the slime mold individual σ i After multiple iterations, the slime mold individuals will gradually tend to the optimal smoothing factor.

[0031] Step 5: Add the simulated annealing algorithm to enhance the global search capability of the SMA-GRNN algorithm, avoid falling into the local optimal solution, and accelerate the convergence speed of the algorithm.

[0032] Step 5.1: Set the temperature variable T and generate N initial smoothing factors σ i Increasing the disturbance Δσ will increase the new solution, as shown in equation (10).

[0033] σ i ′(t+1)=σ i (t)+Δσ (10)

[0034] Step 5.2: The role of simulated annealing is to control whether the slime mold population accepts the new solution, which is determined by formula (11), where H(σ i (t)) replaces the mean square error function MSE.

[0035]

[0036] Step 5.3: When H(σ i ′(t+1)) is less than H(σ i (t)), which means that the new solution is better than the current solution, and the slime mold individual accepts the new solution. At this time, the slime mold individual σ i (t+1) is updated as shown in formula (12).

[0037] σ i (t+1)=σ i ′(t+1) (12)

[0038] When H(σ i ′(t+1)) is greater than or equal to H(σ i(t)), the probability P is used to decide whether to accept the worse new solution. This probability is determined by the current temperature function T(t). As the temperature decreases, the probability of accepting the worse solution decreases.

[0039] Step 5.4: As the temperature function T(t) is iterated, the temperature gradually decreases and eventually approaches 0. As the temperature decreases, the probability of accepting a worse solution decreases, and the slime mold individuals are more inclined to choose a better solution. The cooling coefficient α is set, usually with a value of 0.9≤α≤0.99. The temperature iteration formula is shown in formula (13).

[0040] T(t+1)=αT(t) (13)

[0041] Step 6: In each iteration, the slime mold individuals update their positions according to the global optimal solution and random factors. This step is combined with the simulated annealing mechanism, which enables the individuals to accept a larger range of disturbances in the early stage and gradually converge in the later stage.

[0042] After each iteration, the global optimal smoothing factor σ is updated best , until the maximum number of iterations is reached or the accuracy of the solution meets the requirements. Its specific structure is shown in formula (14).

[0043] σ best =min(MSE(σ1),MSE(σ2),...,MSE(σ N )) (14)

[0044] Further through the above steps, the optimal smoothing factor σ can be obtained best , bring it into GRNN, so that the prediction error of GRNN is minimized.

[0045] The beneficial effects of the above technical solution are as follows: the present invention provides a SA-SMA-GRNN prediction model to construct the original training set and prediction set data of the magnetic core material; the discrete functions E[u] and T[u] are used to represent the original radiation signal data collected by the device, which are used in the training and prediction stages respectively; the data includes multiple characteristics such as the temperature, excitation frequency, and loss of the magnetic core material, and for the training samples and the input samples, their Euclidean distance D is calculated. i , so as to find the similarity between the training samples and the input samples, introduce the SMA optimized by simulated annealing to optimize the smoothing factor σ in GRNN, set the temperature variable T to control the updating process of the smoothing factor, and gradually update the global optimal smoothing factor through multiple iterations until the maximum number of iterations or the target accuracy is reached; by combining simulated annealing with the slime mold algorithm, the accurate prediction of the core loss is achieved, and the global search capability and convergence speed of the GRNN model are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 The overall step flow chart of the prediction model based on SA-SMA-GRNN of the present invention is as follows:

[0047] Figure 2 This is a flow chart of the steps of optimizing the smoothing factor σ using the SMA algorithm in the present invention.

[0048] Figure 3 This is a flowchart of the decomposition steps of the SMA algorithm optimized by the simulated annealing algorithm in the present invention.

[0049] Figure 4 This is a comparison chart of the predicted results and actual results of the training set in the embodiment of the present invention.

[0050] Figure 5 This is a diagram of the prediction results of the prediction set according to an embodiment of the present invention. DETAILED DESCRIPTION

[0051] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0052] This embodiment takes the infrared radiation signal and the original data directly obtained by electromagnetic signal when the magnetic element is working as an example. Figure 1 The six overall steps S1 to S6 shown in the figure successively complete the construction of data discrete function, model construction, model training, and feature optimization, thereby realizing the prediction of the core loss of the magnetic component when it is working, as described in detail as follows.

[0053] Step 1: Construct the discrete function E[u] of the original training set data measured by the core material work, and use E 1 [u],E 2 [u],E 3 [u],E 4 [u] represents the categories of magnetic materials, such as material one, material two, material three, material four, etc. E[u] is used to represent the original characteristic sequence obtained by direct sampling of radiation signal acquisition devices such as infrared radiation meters and electromagnetic radiation meters during the work of magnetic core materials.

[0054] Furthermore, the data of the original data discrete function E[u] in the above step 1 is composed of the core material temperature E1[u], the core material excitation frequency E2[u], the core loss data E3[u], the core excitation waveform E4[u] and the magnetic flux density data E5[u]…E 1029 [u], and its specific structural relationship is shown in formula (1).

[0055] E[u]=E1[u]+E2[u]+E3[u]+… (1)

[0056] Step 2: Construct the discrete function T[u] of the original prediction set data measured by the core material work, using T 1 [u],T 2 [u],T 3 [u],T 4 [u] represents the categories of magnetic materials, such as material one, material two, material three, and material four. T[u] is used to represent the original characteristic sequence obtained by directly sampling radiation signal acquisition devices such as infrared radiation meters and electromagnetic radiation meters during the work of magnetic core materials.

[0057] Furthermore, the data of the original data discrete function T[u] in the above step 2 is composed of the core material serial number T1[u], the core material working temperature T2[u], the core material excitation frequency T3[u], the core material type T4[u], the core excitation waveform T5[u] and the magnetic flux density data T6[u]…T 1030 [u], and its specific structural relationship is shown in formula (2).

[0058] T[u]=T1[u]+T2[u]+T3[u]+… (2)

[0059] Step 3: Calculate the Euclidean distance D between each training sample and the input sample i , construct the operation operator x j Represents the training set input variable, x ij is the corresponding feature of the i-th training sample. From the data, we can know that there are 5 feature variables. The specific structural relationship formula is shown in formula (3).

[0060]

[0061] Furthermore, the smoothing factor σ in the generalized regression network neuron is defined to output the final predicted value Its specific structural relationship formula is shown in formula (4).

[0062]

[0063] Among them, y i is the output value of the training sample, and N is the number of samples.

[0064] Step 4: Use the SMA algorithm to optimize the smoothing factor σ in the generalized regression neural network. The decomposition process is as follows: Figure 2 As shown, the specific steps are as follows.

[0065] Step 4.1: Generate N random initial smoothing factors σ i , represents the initial position of the slime mold individual, where σ min and σ max is the value range of the smoothing factor, and its structure is shown in formula (5).

[0066] σ i (0) = random(σ min ,σ max ),i=1,2,...,N (5)

[0067] Further, for each smoothing factor σ i , put it into formula (6) and calculate its corresponding prediction error MSE under GRNN.

[0068]

[0069] Step 4.2: Update the concentration according to the MSE size corresponding to each slime mold individual. The higher the concentration, the better the smoothing factor σ in the area. i , using C i represents the concentration of the i-th slime mold, as shown in formula (7).

[0070]

[0071] Depending on the concentration, the slime mold individuals will move to new locations, and t is usually used to represent the tth iteration.

[0072] Step 4.3: The smoothing factor σ of slime mold individual i at time t+1 is calculated by the smoothing factor σ at time t. i (t) plus its velocity v i (t) is used to update, and its moving formula in each iteration is shown in formula (8).

[0073] σ i (t+1)=σ i (t)+v i (t) (8)

[0074] Furthermore, v i represents the speed, which is defined as shown in formula (9), σ best represents the current global optimal smoothing factor, σ random is the number of slime mold individuals σ i Another solution randomly selected from , where r1 and r2 are random numbers between 0 and 1.

[0075] v i (t) = r1·(C i ·σ best +(1-C i )·r2·σ random ) (9)

[0076] Step 4.4: Set the slime mold individual σ i After 10 iterations, the slime mold individuals will gradually tend to the optimal smoothing factor.

[0077] Step 5: Add simulated annealing algorithm to enhance the global search capability of SMA-GRNN algorithm, avoid falling into local optimal solution, and accelerate the convergence speed of the algorithm. The decomposition process is as follows: Figure 3 shown.

[0078] Step 5.1: Set the temperature variable function T(t) and generate N initial smoothing factors σ i Increasing the disturbance Δσ will increase the new solution, as shown in equation (10).

[0079] σ i ′(t+1)=σ i (t)+Δσ (10)

[0080] Step 5.2: The role of simulated annealing is to control whether the slime mold population accepts the new solution, which is determined by formula (11), where H(σ i (t)) replaces the mean square error function MSE.

[0081]

[0082] Step 5.3: When H(σ i ′(t+1)) is less than H(σ i (t)), which means that the new solution is better than the current solution, and the slime mold individual accepts the new solution. At this time, the slime mold individual σ i (t+1) is updated as shown in formula (12).

[0083] σ i (t+1)=σ i ′(t+1) (12)

[0084] When H(σ i ′(t+1)) is greater than or equal to H(σ i (t)), the probability P is used to decide whether to accept the worse new solution. This probability is determined by the current temperature function T(t). As the temperature decreases, the probability of accepting the worse solution decreases.

[0085] Step 5.4: As the temperature function T(t) is iterated, the temperature gradually decreases and eventually approaches 0. As the temperature decreases, the probability of accepting a worse solution decreases, and the slime mold individuals are more inclined to choose a better solution. The cooling coefficient α is set, usually with a value of 0.9≤α≤0.99. The temperature iteration formula is shown in formula (13).

[0086] T(t+1)=αT(t) (13)

[0087] Step 6: In each iteration, the slime mold individuals update their positions according to the global optimal solution and random factors. This step is combined with the simulated annealing mechanism, which enables the individuals to accept a larger range of disturbances in the early stage and gradually converge in the later stage.

[0088] After each iteration, the global optimal smoothing factor σ is updated best , until it reaches 10 iterations. Its specific structure is shown in formula (14).

[0089] σ best =min(MSE(σ1),MSE(σ2),...,MSE(σ N )) (14)

[0090] Further through the above steps, the optimal smoothing factor σ can be obtained best is 0.017742, which is brought into the GRNN model to minimize the prediction error of the GRNN model; the training set data is put in, and the prediction results are compared with the actual data. Figure 4 As shown; bring the prediction set into the trained model, and the prediction results are as follows Figure 5 As shown, it can be seen that the prediction accuracy of the core loss prediction using this method is high.

[0091] Although the present invention has been disclosed as above in terms of preferred embodiments, they are not intended to limit the present invention. Anyone skilled in the art can make various changes or modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection defined by the claims of this application.

Claims

1. A core loss prediction model based on SA-SMA-GRNN, characterized in that: The method includes the following 6 steps: Step 1: Construct the discrete function E[u] of the original training set data measured by the core material work, and use E 1 [u],E 2 [u],E 3 [u],E 4 [u] represents material 1, material 2, material 3, and material 4 respectively, and E[u] is used to represent the original characteristic sequence obtained by direct sampling of radiation signal acquisition devices such as infrared radiation meters and electromagnetic radiation meters during the work of magnetic core materials; Furthermore, the data of the original data discrete function E[u] in the above step 1 is composed of the core material temperature E1[u], the core material excitation frequency E2[u], the core loss data E3[u], the core excitation waveform E4[u] and the magnetic flux density data E5[u]…E 1029 [u], and its specific structural relationship is shown in formula (1): E[u]=E1[u]+E2[u]+E3[u]+… (1) Step 2: Construct the discrete function T[u] of the original prediction set data measured by the core material work, using T 1 [u],T 2 [u],T 3 [u],T 4 [u] represents material 1, material 2, material 3, and material 4 respectively, and T[u] is used to represent the original characteristic sequence obtained by directly sampling radiation signal acquisition devices such as infrared radiation meters and electromagnetic radiation meters during the work of magnetic core materials; Furthermore, the data of the original data discrete function T[u] in the above step 2 is composed of the core material serial number T1[u], the core material working temperature T2[u], the core material excitation frequency T3[u], the core material type T4[u], the core excitation waveform T5[u] and the magnetic flux density data T6[u]…T 1030 [u], and its specific structural relationship is shown in formula (2): T[u]=T1[u]+T2[u]+T3[u]+… (2) Step 3: Calculate the Euclidean distance D between each training sample and the input sample i , construct the operation operator x j Represents the training set input variable, x ij is the corresponding feature of the i-th training sample. Its specific structural relationship formula is shown in formula (3): Furthermore, the smoothing factor σ in the generalized regression network neuron is defined to output the final predicted value Its specific structural relationship formula is shown in formula (4): Among them, y i is the output value of the training sample, N is the number of samples; Step 4: Use the SMA algorithm to optimize the smoothing factor σ in the generalized regression neural network and generate random N initial smoothing factors σ i , represents the initial position of the slime mold individual, where σ min and σ max is the value range of the smoothing factor, and its structure is shown in formula (5); s i (0)=random(σ min ,s max ),i=1,2,...,N (5) Further, for each smoothing factor σ i , put it into formula (6) and calculate its corresponding prediction error MSE under GRNN: Step 5: Add the simulated annealing algorithm to enhance the global search capability of the SMA-GRNN algorithm, avoid falling into the local optimal solution, and accelerate the convergence speed of the algorithm; Step 6: In each iteration, the slime mold individuals update their positions according to the global optimal solution and random factors. This step combines the simulated annealing mechanism, which enables the individuals to accept a larger range of disturbances in the early stage and gradually converge in the later stage. After each iteration, the global optimal smoothing factor σ is updated best , until the maximum number of iterations is reached or the accuracy of the solution meets the requirements. Its specific structure is shown in formula (7): s best =min(MSE(σ1),MSE(σ2),...,MSE(σ N )) (7) Further through the above steps, the optimal smoothing factor σ can be obtained best , bring it into GRNN to minimize the prediction error of GRNN, and bring the prediction set data into it to get the final core loss prediction data.

2. A prediction model based on SA-SMA-GRNN according to claim 1, characterized in that: The method for optimizing the smoothing factor in step 4 is as follows: Step 4.1: Update the concentration according to the MSE size corresponding to each slime mold individual. The higher the concentration, the better the smoothing factor σ in the area. i , using C i represents the concentration of the i-th slime mold. According to the concentration, the slime mold individual will move to a new position. Usually t is used to represent the t-th iteration, as shown in formula (8): Step 4.2: The smoothing factor σ of slime mold individual i at time t+1 is calculated by the smoothing factor σ at time t. i (t) plus its velocity v i (t) is updated, and its moving formula in each iteration is shown in formula (9): σ i (t+1)=σ i (t)+v i (t) (9) Furthermore, v i represents the speed, which is defined as shown in formula (10), σ best represents the current global optimal smoothing factor, σ random is the number of slime mold individuals σ i Another solution randomly selected from , where r1 and r2 are random numbers between 0 and 1: v i (t)=r1·(C i ·σ best +(1-C i )·r2·σ random ) (10) Step 4.3: Set the slime mold individual σ i After multiple iterations, the slime mold individuals will gradually tend to the optimal smoothing factor.

3. The prediction model based on SA-SMA-GRNN according to claim 1, characterized in that: The process of step 5 is as follows; Step 5.1: Set the temperature variable T(t) and generate N initial smoothing factors σ i Adding disturbance Δσ will add new solutions, as shown in formula (11): s i ′(t+1)=σ i (t)+Δσ (11) Step 5.2: The role of simulated annealing is to control whether the slime mold population accepts the new solution, which is determined by equation (12), where H(σ i (t)) instead of the mean square error function MSE: Step 5.3: When H(σ i ′(t+1)) is less than H(σ i (t)), which means that the new solution is better than the current solution, and the slime mold individual accepts the new solution. At this time, the slime mold individual σ i (t+1) update, as shown in formula (13): s i (t+1)=σ i ′(t+1) (13) When H(σ i ′(t+1)) is greater than or equal to H(σ i (t)), the probability P is used to decide whether to accept the worse new solution. This probability is determined by the current temperature function T(t). As the temperature decreases, the probability of accepting the worse solution decreases. Step 5.4: As the temperature function T(t) is iterated, the temperature gradually decreases and eventually approaches 0. As the temperature decreases, the probability of accepting a worse solution decreases, and the slime mold individuals are more inclined to choose a better solution. The cooling coefficient α is set, usually with a value of 0.9≤α≤0.

99. The temperature iteration formula is shown in formula (14): T(t+1)=αT(t) (14).

Images