CN112395815B - Temperature prediction method of permanent magnet synchronous motor - Google Patents
Temperature prediction method of permanent magnet synchronous motor Download PDFInfo
- Publication number
- CN112395815B CN112395815B CN202011402714.7A CN202011402714A CN112395815B CN 112395815 B CN112395815 B CN 112395815B CN 202011402714 A CN202011402714 A CN 202011402714A CN 112395815 B CN112395815 B CN 112395815B
- Authority
- CN
- China
- Prior art keywords
- model
- psnlstms
- steps
- temperature
- learning rate
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 50
- 230000001360 synchronised effect Effects 0.000 title claims abstract description 30
- 238000012549 training Methods 0.000 claims abstract description 49
- 230000008859 change Effects 0.000 claims abstract description 14
- 238000013528 artificial neural network Methods 0.000 claims abstract description 10
- 238000013136 deep learning model Methods 0.000 claims abstract description 10
- 238000005457 optimization Methods 0.000 claims description 20
- 238000004804 winding Methods 0.000 claims description 18
- 230000000694 effects Effects 0.000 claims description 17
- 238000012937 correction Methods 0.000 claims description 12
- 238000004364 calculation method Methods 0.000 claims description 9
- 230000006870 function Effects 0.000 claims description 9
- 238000000137 annealing Methods 0.000 claims description 7
- 238000012360 testing method Methods 0.000 claims description 5
- 238000004422 calculation algorithm Methods 0.000 claims description 4
- 238000012544 monitoring process Methods 0.000 claims description 4
- 239000000110 cooling liquid Substances 0.000 claims description 3
- 230000008901 benefit Effects 0.000 abstract description 5
- 230000015654 memory Effects 0.000 description 21
- 239000000306 component Substances 0.000 description 13
- 230000008569 process Effects 0.000 description 9
- 238000005259 measurement Methods 0.000 description 8
- 230000004907 flux Effects 0.000 description 5
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 4
- 229940088594 vitamin Drugs 0.000 description 4
- 229930003231 vitamin Natural products 0.000 description 4
- 235000013343 vitamin Nutrition 0.000 description 4
- 239000011782 vitamin Substances 0.000 description 4
- 150000003722 vitamin derivatives Chemical class 0.000 description 4
- 230000004913 activation Effects 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 239000013598 vector Substances 0.000 description 3
- 238000012795 verification Methods 0.000 description 3
- 239000008358 core component Substances 0.000 description 2
- 238000002347 injection Methods 0.000 description 2
- 239000007924 injection Substances 0.000 description 2
- 238000009413 insulation Methods 0.000 description 2
- 238000012886 linear function Methods 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 238000003058 natural language processing Methods 0.000 description 2
- 238000010200 validation analysis Methods 0.000 description 2
- RYGMFSIKBFXOCR-UHFFFAOYSA-N Copper Chemical compound [Cu] RYGMFSIKBFXOCR-UHFFFAOYSA-N 0.000 description 1
- -1 Nadam=Nesterov+Adam Natural products 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 229910052802 copper Inorganic materials 0.000 description 1
- 239000010949 copper Substances 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 239000012530 fluid Substances 0.000 description 1
- 230000009545 invasion Effects 0.000 description 1
- 230000007787 long-term memory Effects 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 239000002245 particle Substances 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Geometry (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Computer Hardware Design (AREA)
- Control Of Electric Motors In General (AREA)
- Control Of Ac Motors In General (AREA)
Abstract
The invention discloses a temperature prediction method of a permanent magnet synchronous motor, which comprises the steps of combining a deep learning model NLSTMs with a pseudo-twin neural network to construct a PSNLSTMs model, training the PSNLSTMs model by utilizing a data set, obtaining a trend of temperature sequence change of a permanent magnet motor component in a time step by utilizing data with a longer time step as left side input of the PSNLSTMs model, obtaining details of temperature sequence change by utilizing data with a shorter time step as right side input of the PSNLSTMs model, obtaining a trained PSNLSTMs model, and finally carrying out regression prediction on the left side and the right side of the trained PSNLSTMs model according to the same weight to obtain the predicted temperature of the permanent magnet synchronous motor at the next moment. The method has the advantages of accurate prediction result and small error for the temperature prediction of the permanent magnet synchronous motor.
Description
Technical Field
The invention relates to the technical field of permanent magnet synchronous motors, in particular to a temperature prediction method of a permanent magnet synchronous motor.
Background
The permanent magnet synchronous motor is one of core components of the pure electric automobile, has higher power density, and therefore causes serious temperature rise problems, and influences the working efficiency, the load capacity and the service life of the core components of the motor. For example, when the temperature exceeds a certain limit, the insulation part of the electronic winding can be aged rapidly, the insulation performance is reduced, and the motor can be burnt down when serious, so that high maintenance cost is caused, and even the safety performance of the whole vehicle can be threatened. Therefore, on the premise of ensuring the safe operation of the permanent magnet synchronous motor, the working efficiency of the motor is improved to the maximum extent, and an effective method is needed to predict the temperature of main components so as to provide assistance for a temperature control system of the motor.
At present, two main methods for predicting the temperature of a permanent magnet synchronous motor are a direct method and an indirect method. Direct methods mainly include Finite Element Analysis (FEA) and Computational Fluid Dynamics (CFD). Their common advantage is that any shape of device can be modeled to obtain the temperature value of the motor local unit. However, modeling and computing of these two methods is highly computer performance demanding and requires a significant amount of time. In practical applications, indirect methods are often used. Indirect methods mainly include flux observation, invasion and thermal simulation. When the motor temperature is predicted by adopting a magnetic flux observer method, magnetic flux deviation (model error and measurement error) can cause serious errors of a predicted result. Therefore, the flux bias must be reduced when using this method. Invasive methods are mainly achieved by injecting a current, the recognition result being dependent on the kinetic parameters. It is the signal injection that causes output torque ripple, increasing copper consumption, and thus affecting motor performance. Optimization of signal injection has been one of the hot spots of research. Lumped parameter axial flux permanent magnet machine thermal models (LTPNs) reduce the order of the model by simplifying the complex thermal behavior inside the system. Although many studies optimize the original model to achieve higher accuracy, the complexity of the model is still higher and the estimation of the loss distribution is a difficulty.
With the rapid development of artificial intelligence technology, deep learning has achieved good results in Computer Vision (CV) and Natural Language Processing (NLP), and has been applied in many fields of temperature sequence prediction. In 2017, some students first studied long-term memory artificial neural networks (Long Short Term Memory Networks, LSTMs) in the field of permanent magnet synchronous motor temperature prediction, and demonstrated the applicability of LSTMs in this field. In this study, authors processed the data using appropriate methods and used particle swarm optimization to determine model appropriate hyper-parameters. Although the prediction accuracy obtained by the method is not low, a large amount of training time is consumed, and the performance requirement on a computer is high. Therefore, the temperature prediction of the permanent magnet synchronous motor has the problems of complex model, high prediction cost, insufficient precision, complex training process and the like.
Disclosure of Invention
The invention aims to provide a temperature prediction method of a permanent magnet synchronous motor. The method can be used for predicting the temperature of the permanent magnet synchronous motor, and has the advantages of accurate prediction result and small error.
The technical scheme of the invention is as follows: a temperature prediction method of a permanent magnet synchronous motor is characterized by comprising the following steps of: combining a deep learning model NLSTMs with a pseudo-twin neural network to construct a PSNLSTMs model, then training the PSNLSTMs model by utilizing a data set, and using a time step as followsThe data of each moment is used as the left side input of the PSNLSTMs model to obtain the trend of the temperature sequence change of the permanent magnet motor part in the time step, and the time step is used as +.>The data of each moment is used as the right side input of the PSNLSTMs model to obtain the details of temperature sequence change, wherein +.>Thereby obtaining a trained PSNLSTMs model; and finally, inputting the left side and the right side of the trained PSNLSTMs model to carry out regression prediction according to the same weight to obtain the predicted temperature of the permanent magnet synchronous motor at the next moment.
In the temperature prediction method of the permanent magnet synchronous motor, the deep learning model NLSTMs are expressed as follows:
；
；
；
；
；
；
；
；
；
；
；
；
wherein the method comprises the steps ofFor the dot product operator, [,]representing merging the two vectors in brackets; />The input, forgetting gate, output gate, input gate, cell state, internal memory cell state and output at the current moment of NLSTMs are sequentially respectively; />The cell state, the output, the internal memory cell state and the internal memory output are respectively carried out at the last moment in sequence;the internal memory input, the internal memory forget gate, the internal memory output gate, the internal memory input gate and the internal memory output at the current moment are sequentially respectively; />Respectively corresponding to the weight matrix in each gate, < ->Respectively toBias terms in each gate; />Are sigmoid activation functions; wherein->For the tanh function, +.>Is a linear function.
According to the temperature prediction method of the permanent magnet synchronous motor, the PSNLSTMs model is trained by utilizing the data set, namely, the data set is divided into a training set and a testing set, wherein input data of the training set comprise continuous data of environment temperature, cooling liquid temperature, voltage d coordinate component, voltage q coordinate component, motor rotating speed, motor torque, current d coordinate component, current q coordinate component, permanent magnet temperature, stator tooth temperature and winding temperature at s times, and the characteristics of 11 dimensions are shared;
the left side input is:
；
the right side input is:
；
in the method, in the process of the invention,、/>respectively representing input data of different time steps of left input and right input of PSNLSTMs model, +.>Indicate->No. 5 of the time>Vitamin characteristics (I)>Indicate->No. 5 of the time>Vitamin characteristics (I)>，。
In the temperature prediction method of the permanent magnet synchronous motor, in the training of the PSNLSTMs model, the mean square error is used as a loss function:
，
wherein:is mean square error>For data volume, +.>For the moment of->For the real data at time t, +.>The predicted data at the time t;
obtaining a determination coefficient by using the mean square error:
：
wherein:is the average of the data;
in the above formula, the determining coefficient is to use the mean value as an error reference for monitoring the overall fitting effect of the PSNLSTMs model.
In the temperature prediction method of the permanent magnet synchronous motor, in the PSNLSTMs model training, the learning rate is selected to preheat, so that the PSNLSTMs model reaches a stable state, and then the learning is performed by using the preset learning rate; the learning rate preheating calculation formula is as follows:
；
in the method, in the process of the invention,for the current learning rate, < >>Respectively, a set maximum learning rate and a set minimum learning rate, < >>The number of steps to be executed at present and the total number of steps to be preheated are respectively;
after the learning rate is preheated, the decay of the preset learning rate is realized by adopting cosine annealing, and the formula is as follows:
；
； />；
in the aboveFor the number of steps currently performed +.>For the total number of steps to be performed, +.>For the number of steps taken from the start of training, < + >>Step number preheated for learning rate, +.>For the number of steps remaining unchanged after the end of learning rate warm-up,/->For learning rate after attenuation, +.>For a set minimum learning rate, +.>The maximum learning rate is set;
meanwhile, nadam algorithm is combined for improving training effect;
the Nadam calculation procedure is as follows:
；
；
；
；
；
；
wherein,the gradient at the current t moment, the first moment estimation, the second moment estimation and the updated gradient are respectively; />The first moment estimation and the second moment estimation respectively correspond to the t-1 moment; />The correction gradient, the correction first moment estimation and the correction second moment estimation at the current t moment after correction calculation are respectively carried out; />Estimating an average first moment at the current t moment; />Momentum factor estimated for the second moment, +.>First moment momentum factors corresponding to time t, time t-1 and any time i respectively,/->Is a positive number close to 0 but not 0.
Compared with the prior art, the method comprises the steps of combining a deep learning model NLSTMs with a pseudo-twin neural network to construct a PSNLSTMs model, training the PSNLSTMs model by utilizing a data set, obtaining a trend of temperature sequence change of a permanent magnet motor part in a time step by utilizing data with a longer time step as the left side input of the PSNLSTMs model, obtaining details of temperature sequence change by utilizing data with a shorter time step as the right side input of the PSNLSTMs model, obtaining a trained PSNLSTMs model, and finally inputting the left side and the right side of the trained PSNLSTMs model to carry out regression prediction according to the same weight to obtain the predicted temperature of the permanent magnet synchronous motor at the next moment; the invention utilizes the deep learning model NLSTMs to process the internal memory more flexibly, can grasp more details of the temperature change of the permanent magnet synchronous motor, combines the characteristics of the pseudo-twin neural network to obtain the PSNLSTMs, trains the PSNLSTMs, and has the advantages of accurate prediction result and small error after training. In addition, the invention carries out optimization design in the training process of the PSNLSTMs model, combines two learning rate optimization methods of learning rate preheating and cosine annealing, and uses the Nadam momentum self-adaptive optimizer to accelerate the training of the PSNLSTMs model and promote the training effect of the model.
Drawings
FIG. 1 is a schematic diagram of the structure of the present invention;
FIG. 2 is a schematic structural diagram of the deep learning model NLSTMs;
FIG. 3 is a graph of learning rate decay using cosine annealing;
FIG. 4 is a graph of learning rate variation;
FIG. 5 (a) is a graph showing the loss of a training set of NLSTMs models using a learning rate optimization strategy;
FIG. 5 (b) is a graph of the loss of the NLSTMs training set without employing a learning rate optimization strategy;
FIG. 6 (a) is a graph of loss for an NLSTMs model validation set employing a learning rate optimization strategy;
FIG. 6 (b) is a graph of loss of an NLSTMs model validation set that does not employ a learning rate optimization strategy;
FIG. 7 is a graph of a predicted temperature fit of NLSTMs models to stator yokes;
FIG. 8 is a graph of predicted temperature bias of NLSTMs model versus stator yoke;
FIG. 9 is a graph of a predicted temperature fit of the PSNLSTMs model to the stator yoke;
FIG. 10 is a graph of predicted temperature bias of the PSNLSTMs model versus stator yoke;
FIG. 11 is a graph of predicted temperature fit of NLSTMs models to stator teeth;
FIG. 12 is a graph of predicted temperature deviation of LSTM model versus stator teeth;
FIG. 13 is a graph of a predicted temperature fit of the PSNLSTMs model to stator teeth;
FIG. 14 is a graph of predicted temperature deviation of the PSNLSTMs model versus stator teeth;
FIG. 15 is a graph of predicted temperature fits of NLSTMs models to stator windings;
FIG. 16 is a graph of predicted temperature deviation of the LSTM model from stator windings;
FIG. 17 is a graph of a predicted temperature fit of the PSNLSTMs model to the stator windings
FIG. 18 is a graph of predicted temperature deviation of the PSNLSTMs model versus stator winding.
Detailed Description
The invention is further illustrated by the following figures and examples, which are not intended to be limiting.
Example 1: a temperature prediction method of a permanent magnet synchronous motor combines a deep learning model NLSTMs (Nested long-short-term memory artificial neural Networks (NLSTMs)) with a pseudo-twin neural Network to construct a PSNLSTMs model shown in figure 1 (Siamese Network is a Network structure, which is a conjoined neural Network, the Siamese Network realizes 'conjoined' through a weight sharing mode, when weights of the neural networks on the left side and the right side are not shared or the models on the two sides are not the same, the model is called pseudo-Siamese networks, in the embodiment, the NLSTMs and the pseudo-Siamese networks are combined to obtain a PSNLSTMs model), then a data set is utilized to train the PSNLSTMs model, and a baseline data set used in the embodiment is from a Kaggele data science online competition platform, and data measurement and collection are carried out by power electronics and electric drive trains of the Pade university. The data set has been normalized to include temperatures of the main components of the permanent magnet motor, such as the stator yoke, stator teeth, windings, etc., and the meaning and sign of the specific data label are shown in table 1 below.
TABLE 1
The whole data set is totally 99 pieces of data, the measurement frequency is 2Hz, 52 measurement sessions are included, the measurement sessions are distinguished through session ids, the measurement sessions are independent, and the measurement data in the same session are continuously measured. Each measurement session in the dataset may represent substantially the entire electrothermal change process of the permanent magnet synchronous motor, and in order to reduce the time and cost of training the model, the present embodiment uses 51 sessions as the training set, 1 session (id=32) as the test set, downsamples at an appropriate frequency, and performs an appropriate data cleansing operation. The number of the final training sets is 32263, and the number of the test sets is 412. Using time steps asThe data of each moment is used as the left side input of the PSNLSTMs model to obtain the trend of the temperature sequence change of the permanent magnet motor part in the time step, and the time step is used as +.>The data of each moment is used as the right side input of the PSNLSTMs model to obtain the details of temperature sequence change, wherein +.>(the temperature at the next time is often close to the temperature at this timeAnd finally, inputting the left side and the right side of the trained PSNLSTMs to carry out regression prediction according to the same weight to obtain the predicted temperature of the permanent magnet synchronous motor at the next moment.
Example 2: on the basis of embodiment 1, the structure of the deep learning model NLSTMs is shown in fig. 2, expressed as:
；
；
；
；
；
；
；
；
；
；
；
；
wherein the method comprises the steps ofFor the dot product operator, [,]representing merging the two vectors in brackets; />The input, forgetting gate, output gate, input gate, cell state, internal memory cell state and output at the current moment of NLSTMs are sequentially respectively; />The cell state, the output, the internal memory cell state and the internal memory output are respectively carried out at the last moment in sequence; />The internal memory input, the internal memory forget gate, the internal memory output gate, the internal memory input gate and the internal memory output at the current moment are sequentially respectively; />Respectively corresponding to the weight matrix in each gate, < ->Respectively corresponding to bias items in the gates; the input and output of NLSTMs are unchanged relative to LSTMs, but the temporary cell state is increased, which is relative to the transmission of cell states between internal memories of cells at different times, which is originally in LSTMCalculate->While the two vectors for addition are +.>And->Become->And->As an input to the internal memory. />The function is a sigmoid activation function, and aims to realize selective memory and forgetting and realize longer learning; />Are sigmoid activation functions; wherein->As a function of the tanh,the linear function, here distinguished from a standard LSTM, is to better describe the temporary cell state of the external LSTM input.
The training of the PSNLSTMs by using the data set is specifically that the deep learning models related to the embodiment are all supervised learning, and the prediction task of the PSNLSTMs is the temperature of the stator yoke, the stator teeth and the windings of the important component motor of the permanent magnet motor at the next moment. In order to ensure the accuracy of temperature prediction of three main components, training is performed to divide a data set into a training set and a test set, wherein input data of the training set comprises continuous environmental temperature, cooling liquid temperature and voltage d coordinate division at s momentsThe data of the quantity, the voltage q coordinate component, the motor rotating speed, the motor torque, the current d coordinate component, the current q coordinate component, the permanent magnet temperature, the stator tooth temperature and the winding temperature are 11 dimensional features; acquiring time step length from training set as followsThe individual moments and time steps are +.>Input data at each instant, and->Left and right inputs as PSNLSTMs model, respectively:
the left side input is:
；
the right side input is:
；
in the method, in the process of the invention,、/>respectively representing input data of different time steps of left input and right input of PSNLSTMs model, +.>Indicate->No. 5 of the time>Vitamin characteristics (I)>Indicate->No. 5 of the time>Vitamin characteristics (I)>，。
In the training of the PSNLSTMs model, the mean square error is taken as a loss function:
，
wherein:is mean square error>For data volume, +.>For the moment of->For the real data at time t, +.>The predicted data at the time t;
obtaining a determination coefficient by using the mean square error:
：
wherein:is the average of the data;
in the above formula, the determining coefficient uses the mean value as an error reference for monitoring the fitting effect of the PSNLSTMs model.
By the above formula, it can be observed whether the prediction error is greater or less than the mean reference error,the normal value of (1) is [0,1 ]]The closer to 1 the better the model fitting effect is, and vice versa, due to +.>Is a dimensionless index, which can bring convenience to solve the problems of different metrics.
In the training of the PSNLSTMs model, a smaller learning rate is set first, and after a part of rounds or steps are trained, the training is carried out by modifying the training rate into a preset learning rate. Because the PSNLSTMs model is randomly initialized with its weight just after training, if a larger learning rate is selected, the model is unstable (e.g., the model is over-fitted at the beginning of training, and then it takes more training rounds to change). The learning rate is selected for preheating, the model can be learned with a smaller learning rate in the set number of rounds or steps, so that the model reaches a stable state, and the model is learned with the preset learning rate, so that the convergence rate of the model is higher, the learning rate is better to be selected for preheating, the PSNLSTMs model reaches a stable state, and the model is learned with the preset learning rate;
the learning rate preheating calculation formula is as follows:
；
in the method, in the process of the invention,for the current learning rate, < >>Respectively, a set maximum learning rate and a set minimum learning rate, < >>The number of steps to be executed at present and the total number of steps to be preheated are respectively;
after the learning rate is preheated, the state of the PSNLSTMs model is relatively stable, and the model learning vibration can be caused by continuously using a larger learning rate. Therefore, in order to achieve the global optimum, training is performed with a smaller learning rate when approaching the global optimum, and the present embodiment adopts cosine annealing as shown in fig. 3 to achieve attenuation of the learning rate, where the formula is as follows:
；
； />；
in the aboveFor the number of steps currently performed +.>For the total number of steps to be performed, +.>For the number of steps taken from the start of training, < + >>Step number preheated for learning rate, +.>For the number of steps remaining unchanged after the end of learning rate warm-up,/->For learning rate after attenuation, +.>Setting 0 for the set minimum learning rate; />The maximum learning rate is set;
meanwhile, the cosine annealing and random gradient descent are combined together to achieve a better training effect. Nadam and Adam are optimization algorithms which use momentum and self-adaptive learning rate to accelerate convergence speed, adam is expansion of SGD, and stability and great training speed are improved compared with SGD. Nadam integrates Nesterov on the basis of Adam, namely Nadam=Nesterov+Adam, and cosine annealing is combined with Nadam algorithm to improve training effect;
the Nadam calculation procedure is as follows:
；
；
；
；
；
；
wherein,the gradient at the current t moment, the first moment estimation, the second moment estimation and the updated gradient are respectively; />The first moment estimation and the second moment estimation respectively correspond to the t-1 moment; />The correction gradient, the correction first moment estimation and the correction second moment estimation at the current t moment after correction calculation are respectively carried out; />Estimating an average first moment at the current t moment; />Momentum factor estimated for the second moment, +.>First moment momentum factors corresponding to time t, time t-1 and any time i respectively,/->Is a positive number close to 0 but not 0.
In combination with the above learning rate optimization strategy, the learning rate changes during the training process are shown in fig. 4. The learning rate in the figure is set to 0.008, the total number of steps is 25200, the first 2500 steps are the learning rate preheating part, and after the learning rate is linearly increased to the set learning rate, the learning rate enters the attenuation part.
Example 3: in example 2, in the training of the PSNLSTMs model, the applicant uses the same data set and sets the same super parameters when the training rounds are the same, and the comparison experiment is performed on the NLSTMs model, and the results are shown in fig. 5 and 6 (in the figure), where fig. 5 (a) is a graph of loss of the training set of the NLSTMs model using the learning rate optimization strategy, fig. 5 (b) is a graph of loss of the training set of the NLSTMs model not using the learning rate optimization strategy, fig. 6 (a) is a graph of loss of the verification set of the NLSTMs model using the learning rate optimization strategy, and fig. 6 (b) is a graph of loss of the verification set of the NLSTMs model not using the learning rate optimization strategy; as can be seen from fig. 5, the NLSTMs model using the learning rate optimization strategy converges significantly slower in the first 5 rounds than the NLSTMs model not using the learning rate optimization strategy, because the learning rate is smaller at the beginning, but the degree of convergence is similar in the later stages as the learning rate steadily increases; as can be seen from fig. 6, the loss graph of the NLSTMs model verification set using the learning rate optimization strategy is significantly more stable and more convergent in the initial stage and the later stage, which indicates that the learning rate optimization strategy used in the present invention can improve the fitting effect of the PSNLSTMs model.
Further, the applicant compared the fitting performance of LSTM model, NLSTMs model and PSNLSTMs model on stator yoke, stator tooth and winding temperature predictions, and evaluated the prediction effect of three models on different important components with four technical indicators including Mean Square Error (MSE), mean square error (RMSE), mean Absolute Error (MAE), and decision coefficient (R-Squared, R), finally obtaining the prediction effect as shown in tables 2-4.
Table 2 stator yoke
| Model | MSE | MAE | RMSE | R² |
| LSTM model | 0.001304 | 0.028133 | 0.036106 | 0.996349 |
| NLSTMs model | 0.000859 | 0.022817 | 0.029305 | 0.997589 |
| PSNLSTMs model | 0.000807 | 0.021904 | 0.028406 | 0.997734 |
Table 3 stator teeth
| Model | MSE | MAE | RMSE | R² |
| LSTM model | 0.002125 | 0.035742 | 0.046103 | 0.997504 |
| NLSTMs model | 0.001454 | 0.026653 | 0.038132 | 0.998290 |
| PSNLSTMs model | 0.001137 | 0.024393 | 0.033723 | 0.998663 |
Table 4 stator winding
| Model | MSE | MAE | RMSE | R² |
| LSTM model | 0.004858 | 0.051114 | 0.069701 | 0.996553 |
| NLSTMs model | 0.002834 | 0.036052 | 0.053237 | 0.997988 |
| PSNLSTMs model | 0.002387 | 0.035798 | 0.048857 | 0.998305 |
From the R perspective, the results of the LSTM model, the NLSTMs model and the PSNLSTMs model on the temperature predictions of three important components are very close to 1, and good effects are obtained. However, according to the other three evaluation indexes MSE, MAE, RMSE, the deviations between the predicted values and the true values are intuitively evaluated, and the NLSTM and the PSNLSTM have better effects than the LSTM. The PSNLSTM is close to the NLSTM prediction effect, especially on the stator yoke temperature prediction. But the temperature prediction of important parts of the permanent magnet motor has great significance for temperature monitoring and temperature control with higher accuracy.
To better compare the PSNLSTMs model and the NLSTMs model, fig. 7 shows a predicted temperature fit curve of the NLSTMs model to a stator yoke (stator yoke) (in the figure, curve 1 is a true temperature, curve 2 is a predicted temperature, the ordinate is a temperature, and the abscissa is a time, and the same applies below), fig. 8 shows a predicted temperature deviation curve of the NLSTMs model to the stator yoke, fig. 9 shows a predicted temperature fit curve of the PSNLSTMs model to the stator yoke, and fig. 10 shows a predicted temperature deviation curve of the PSNLSTMs model to the stator yoke. As can be seen from fig. 7 and 10, the PSNLSTMs model and the NLSTMs model are quite close to each other simply from the curve fitting effect; however, from the predicted temperature deviation curves of fig. 8 and 10, at the point of the temperature abrupt change, the deviation of the PSNLSTMs model is slightly smaller than that of the LSTM model, and the predicted temperature deviation PSNLSTMs model at the start and end ends is significantly smaller than that of the NLSTMs model, and at the non-abrupt temperature point, the error distribution is more uniform, so that it can be explained that the PSNLSTMs model of the present invention is more accurate in the effect of temperature prediction and the error is smaller.
Still further, the PSNLSTMs model and the NLSTMs model are compared to the predicted temperatures of the stator teeth and the stator windings, fig. 11 shows a predicted temperature fitting curve of the NLSTMs model to the stator teeth (stator tooth), fig. 12 shows a predicted temperature deviation curve of the LSTM model to the stator teeth, fig. 13 shows a predicted temperature fitting curve of the PSNLSTMs model to the stator teeth, fig. 14 shows a predicted temperature deviation curve of the PSNLSTMs model to the stator teeth, fig. 15 shows a predicted temperature fitting curve of the NLSTMs model to the stator windings (stator winding), fig. 16 shows a predicted temperature deviation curve of the LSTM model to the stator windings, fig. 17 shows a predicted temperature deviation curve of the PSNLSTMs model to the stator windings, fig. 18 shows a predicted temperature deviation curve of the PSNLSTMs model to the stator windings; as can be seen from the predicted temperature fitting curves in the diagrams, the fitting results of the PSNLSTMs and the NLSTMs are very similar, but on the predicted temperature deviation curves, the PSNLSTMs are superior to the NLSTMs on the temperature predictions of the two important parts, the deviation variation amplitude is smaller, and the prediction performance is more stable and accurate.
In conclusion, the temperature prediction method can be used for predicting the temperature of the permanent magnet synchronous motor, and has the advantages of accurate prediction result and small error.
Claims (3)
1. A temperature prediction method of a permanent magnet synchronous motor is characterized by comprising the following steps of: constructing a PSNLSTMs model by combining a deep learning model NLSTMs with a pseudo-twin neural network, training the PSNLSTMs model by using a data set, obtaining a trend of temperature sequence change of a permanent magnet motor part in a time step by using data with the time step of m moments as left side input of the PSNLSTMs model, obtaining details of temperature sequence change by using data with the time step of n moments as right side input of the PSNLSTMs model, wherein m is more than n, and obtaining the trained PSNLSTMs model; finally, inputting the left side and the right side of the trained PSNLSTMs model to carry out regression prediction according to the same weight to obtain the predicted temperature of the permanent magnet synchronous motor at the next moment;
the PSNLSTMs model is trained by utilizing a data set, namely, the data set is divided into a training set and a testing set, input data of the training set comprises continuous data of environment temperature, cooling liquid temperature, voltage d coordinate component, voltage q coordinate component, motor rotating speed, motor torque, current d coordinate component, current q coordinate component, permanent magnet temperature, stator tooth temperature and winding temperature at s times, and the total data is 11 dimensionalities;
the left side input is:
the right side input is:
wherein S is 1 、S 2 Respectively representing input data of different time steps of left input and right input of PSNLSTMs model, a m，k A represents the kth dimension characteristic at the mth moment, a n，k The kth dimensional feature at the nth time is represented, k=1, 2,3.
2. The method for predicting the temperature of a permanent magnet synchronous motor according to claim 1, wherein: in the training of the PSNLSTMs model, the mean square error is taken as a loss function:
wherein: MSE is mean square error, n is the data amount, t is the time, a t Is the real data at the time t, p t The predicted data at the time t;
obtaining a determination coefficient by using the mean square error:
wherein:is the average of the data;
in the above formula, the determining coefficient is to use the mean value as an error reference for monitoring the overall fitting effect of the PSNLSTMs model.
3. The method for predicting the temperature of a permanent magnet synchronous motor according to claim 2, wherein: in PSNLSTMs model training, selecting learning rate to preheat, enabling the PSNLSTMs model to reach a stable state, and learning by using a preset learning rate; the learning rate preheating calculation formula is as follows:
wherein eta is the current learning rate, eta max ，η min Respectively setting maximum learning rate and minimum learning rate, T n ，T w The number of steps to be executed at present and the total number of steps to be preheated are respectively;
after the learning rate is preheated, the decay of the preset learning rate is realized by adopting cosine annealing, and the formula is as follows:
T s ＝global_steps-warmup_steps-hold_base_rate_steps；
T s ＝global_steps-warmup_steps-hold_base_rate_steps；
t in the above s T is the number of steps currently performed t For the total number of steps to be performed, global_steps is the number of steps passed from the start of training, norm_steps is the number of steps preheated by the learning rate, hold_base_rate_steps is the number of steps remaining unchanged after the end of the learning rate preheating, η is the learning rate after decay, η min For a set minimum learning rate, eta max The maximum learning rate is set;
combining the learning rate optimization strategy with a Nadam optimization algorithm to improve training effect;
the Nadam calculation procedure is as follows:
m t ＝u t ×m t-1 +(1-u t )×g t ；
wherein g t ，m t ，n t ，Δθ t The gradient at the current t moment, the first moment estimation, the second moment estimation and the updated gradient are respectively; m is m t-1 ，n t-1 The first moment estimation and the second moment estimation respectively correspond to the t-1 moment;the correction gradient, the correction first moment estimation and the correction second moment estimation at the current t moment after correction calculation are respectively carried out; />Estimating an average first moment at the current t moment; v is the momentum factor of the second moment estimation, u t ，u t-1 ，u i The first moment momentum factors corresponding to time t, time t-1 and any time i respectively, ζ is a positive number close to 0 but not 0.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011402714.7A CN112395815B (en) | 2020-12-04 | 2020-12-04 | Temperature prediction method of permanent magnet synchronous motor |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011402714.7A CN112395815B (en) | 2020-12-04 | 2020-12-04 | Temperature prediction method of permanent magnet synchronous motor |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112395815A CN112395815A (en) | 2021-02-23 |
| CN112395815B true CN112395815B (en) | 2024-03-29 |
Family
ID=74604232
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011402714.7A Active CN112395815B (en) | 2020-12-04 | 2020-12-04 | Temperature prediction method of permanent magnet synchronous motor |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112395815B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113239783A (en) * | 2021-05-11 | 2021-08-10 | 广西科学院 | Plant leaf identification system based on jigsaw disordered data disturbance mechanism |
| CN113742903B (en) * | 2021-08-19 | 2023-06-23 | 安徽科技学院 | Permanent magnet synchronous motor fault detection and identification method based on digital twin technology |
| CN115808944B (en) * | 2023-02-09 | 2023-06-02 | 国能大渡河枕头坝发电有限公司 | Stator temperature rise test control method |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106788028A (en) * | 2016-12-20 | 2017-05-31 | 江苏大学 | Bearing-free permanent magnet synchronous motor intensified learning controller and its building method |
| CN109787524A (en) * | 2019-02-18 | 2019-05-21 | 北京理工大学 | A kind of permanent magnet synchronous motor on-line parameter identification method |
| CN109977837A (en) * | 2019-03-20 | 2019-07-05 | 常熟理工学院 | Pedestrian based on human body attitude invariant features recognition methods again |
| US10565498B1 (en) * | 2017-02-28 | 2020-02-18 | Amazon Technologies, Inc. | Deep neural network-based relationship analysis with multi-feature token model |
-
2020
- 2020-12-04 CN CN202011402714.7A patent/CN112395815B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106788028A (en) * | 2016-12-20 | 2017-05-31 | 江苏大学 | Bearing-free permanent magnet synchronous motor intensified learning controller and its building method |
| US10565498B1 (en) * | 2017-02-28 | 2020-02-18 | Amazon Technologies, Inc. | Deep neural network-based relationship analysis with multi-feature token model |
| CN109787524A (en) * | 2019-02-18 | 2019-05-21 | 北京理工大学 | A kind of permanent magnet synchronous motor on-line parameter identification method |
| CN109977837A (en) * | 2019-03-20 | 2019-07-05 | 常熟理工学院 | Pedestrian based on human body attitude invariant features recognition methods again |
Non-Patent Citations (5)
| Title |
|---|
| Learning Deep Neural Networks for Vehicle Re-ID with Visual-spatio-temporal Path Proposals;Yantao Shen 等;《《2017 IEEE International Conference on Computer Vision (ICCV)》》;全文 * |
| Sachin Kumar 等.Earth Mover's Distance Pooling over Siamese LSTMs for Automatic Short Answer Grading.《Twenty-Sixth International Joint Conference on Artificial Intelligence》.2017,全文. * |
| 基于LSTM网络的短期风力发电功率预测模型研究;张晋霞;《中国优秀硕士论文全文数据库 工程科技Ⅱ辑》;全文 * |
| 基于混合小波LSTM-DNN结构的光伏预测模型;罗潇;《电力电子技术》;全文 * |
| 基于混合网络的异源遥感图像变化检测;周圆 等;《北京航空航天大学学报》;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112395815A (en) | 2021-02-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112395815B (en) | Temperature prediction method of permanent magnet synchronous motor | |
| Kirchgässner et al. | Estimating electric motor temperatures with deep residual machine learning | |
| Diao et al. | Multiobjective system level optimization method for switched reluctance motor drive systems using finite-element model | |
| Chemali et al. | Long short-term memory networks for accurate state-of-charge estimation of Li-ion batteries | |
| CN114004164A (en) | Motor rotor temperature prediction method and system for control | |
| Lee et al. | Temperature estimation of PMSM using a difference-estimating feedforward neural network | |
| Xie et al. | Imitation and transfer Q-learning-based parameter identification for composite load modeling | |
| Ben Regaya et al. | Electric drive control with rotor resistance and rotor speed observers based on fuzzy logic | |
| Zhao et al. | Intelligent second‐order sliding mode control for permanent magnet linear synchronous motor servo systems with robust compensator | |
| Karami-Shahnani et al. | Online Inductance Estimation of PM-Assisted Synchronous Reluctance Motor Using Artificial Neural Network | |
| Bicer et al. | Electric Vehicle Thermal Management System Modeling with Informed Neural Networks | |
| Xiao et al. | Parameter identification of direct‐drive permanent magnet synchronous generator based on EDMPSO‐EKF | |
| Li et al. | Global attention-based encoder-decoder lstm model for temperature prediction of permanent magnet synchronous motors | |
| Maruta et al. | Identification of multiple mode models via distributed particle swarm optimization | |
| Parekh et al. | Performance analysis of electrical machines using a hybrid data-and physics-driven model | |
| CN108015761B (en) | Single-connecting-rod flexible mechanical arm control method and system | |
| Ai et al. | Optimisation of reverse salient rotor for interior permanent magnet synchronous motor and experimental validation | |
| Hinchi et al. | A deep long-short-term-memory neural network for lithium-ion battery prognostics | |
| Li et al. | A Thermal-Aware Digital Twin Model of Permanent Magnet Synchronous Motors (PMSM) Based on BP Neural Networks | |
| Yutthanawa et al. | Multi-surface Permanent Magnet Synchronous Motor Temperature Estimation based on Automate Machine Learning Approach | |
| Farah et al. | Robust Model-Free Reinforcement Learning Based Current Control of PMSM Drives | |
| Fekri et al. | The design and development of multivariable controls with the application for energy management of hybrid electric vehicles | |
| Kang et al. | An integrated parameter identification method of asynchronous motor combined with adaptive load characteristics | |
| Acikgoz et al. | Long short-term memory network-based speed estimation model of an asynchronous motor | |
| Zhao et al. | Research of maximum ratio of torque to current control method for PMSM based on least square support vector machines |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |