Optimized Informer model design method for electric vehicle SOC prediction

Abstract

The invention discloses an optimized Informer model design method for electric vehicle SOC prediction. The method comprises the steps that S1, an Informer SOC prediction model is established; s2, based on the electric vehicle operation data, a health matrix is obtained through a health assessment weight optimization algorithm; and S3, optimizing the Informer SOC prediction model from three aspects of an encoder, a decoder and a control logic of the Informer SOC prediction model by using the health matrix to obtain an optimized Informer model for predicting the SOC of the electric vehicle. According to the method, structure optimization and logic optimization are carried out on the Informer model, and the prediction performance of the Informer model is improved on the premise that the data cost is not increased; the adverse effect of truncation errors is reduced, the structure optimization effect is reserved, the cost is low, and the effect is good.

Description

Technical Field

[0001] The present invention belongs to the technical field of electric vehicle SOC prediction, and in particular to an optimized Informer model design method for electric vehicle SOC prediction. Background Art

[0002] SOC (State of Charge) affects the power output, energy management and operation of electric vehicles, so SOC prediction is very important for the state assessment of electric vehicles. At present, electric vehicle SOC prediction is roughly divided into three research directions: predicting SOC through battery state parameters; predicting SOC based on changes in the operating state of electric vehicles; and predicting SOC based on the above two combined with historical data of the same model.

[0003] In the prior art, some analyze the relationship between the air around the lithium battery and the SOC to obtain a prediction model for the content of gases such as CO2 and NOx in the circulating air of the electric vehicle and the remaining percentage of SOC. This method requires high-precision air data support. If low-precision air data is used for battery state analysis and SOC prediction, the prediction accuracy will be greatly reduced due to truncation errors. Some estimate the operating parameters of the electric vehicle and predict the SOC by the state changes of the battery compartment under different charging voltages. This method assumes that the transmission delay is fixed. When communication is blocked, the transmission speed decreases, and there is an additional lag in the characteristic data, the prediction result will be noisy. Some start from the SOC characteristic curve and use the thermal expansion and contraction phenomenon of the battery to train a model for predicting SOC through battery parameters at different temperatures. This method requires the battery temperature to rise to 400°C and then slowly decrease. In actual use, it will be difficult to collect initial data, resulting in difficulties in prediction. Some use the temperature difference of the battery surface to analyze from a thermodynamic perspective to obtain a prediction model for surface temperature SOC, but the current electric vehicle power supply module uses an integrated battery cluster, and it is difficult to collect the surface temperature data of a single battery. Some use the impact of battery group charging and discharging operations on the operating state of electric vehicles to obtain a prediction model for battery charging voltage, battery charging current, power supply module voltage, ambient temperature and SOC remaining percentage under different operating conditions. The prediction error of this model is large near the junction of the charging-discharging state. Some divide the operating state into several sub-intervals according to the speed change trends such as acceleration, deceleration, and uniform speed. Each sub-interval adopts a different SOC prediction strategy to improve the prediction accuracy. However, when integrating all sub-interval strategies and predicting SOC for a new complete interval, the prediction accuracy is not high due to the noise and multi-value prediction caused by marginal effects. Some analyze the impact of AC modules on SOC and provide a prediction model for estimating SOC based on AC module characteristic data under different operating conditions. Because the reaction of the mechanical part of the electric vehicle to the AC module is not considered, the model prediction error is large in a specific SOC interval. Some analyze the change of motor torque under different operating states, and use the relationship between motor torque, electric vehicle driving speed and SOC to propose a torque-speed-SOC prediction model. This model has fewer prediction steps and lower model complexity. The disadvantage is that the operating state is frequently switched and the motor torque changes sharply, so the SOC prediction accuracy is poor. Some use historical data of different states of the same model to obtain a prediction model of the SOC change rate under different operating states, and indirectly predict SOC, but ignore the difference in mechanical strength and the offset of the SOC characteristic curve, which increases the system error. Some classify and train historical data of different roads of the same model to obtain a speed-SOC prediction model, but this method assumes that the SOC characteristic curve is fixed and will not change with the change of the electric vehicle state, and the prediction accuracy of the model is poor.Some methods classify and train historical data based on the difference in ambient temperature to obtain a prediction model for the SOC consumption rate at different temperatures, and predict the total SOC at the next moment based on the model and the SOC margin; some methods consider the aging probability of mechanical parts on this basis, and estimate the SOC change through the weighted average of the change rate. Both methods refer to the SOC standard characteristic curve to train the prediction model, without considering the change in operating status, resulting in the SOC characteristic curve offset and poor prediction accuracy of the model. Some methods train the SOC prediction model through historical data at different altitudes to directly predict SOC, and some use historical data with different gravity constants to train the traction power prediction model, and indirectly predict SOC through mechanical related calculations. Since only the influence of gravity on driving speed is considered, and the influence of gravity on health status is not considered, both methods are difficult to handle the system errors caused by the offset of the SOC characteristic curve, and the trained SOC prediction model has poor anti-interference ability.

[0004] Optimization algorithms such as neural networks have solved some of the defects of the prediction model. For example, some have combined the Bayesian decision method with SVR, GPR, and linear regression to improve the SOC multi-value prediction problem at the state boundary, but the computational complexity has increased, and the prediction error has not been significantly reduced. Some have proposed an improved DNN (Deep Neural Network) based on the changing characteristics of the operating state of FCEVs (Fuel Cell Electric Vehicles) for SOC prediction of FCEVs. This method has fewer prediction steps and higher prediction accuracy, but the calculation is complex, the prediction speed is slow, and the hardware requirements are high, making it difficult to apply in practice. Some have introduced logical predictions of AND, OR, and NOT on the original P&O (Perturb&Observe) model to replace conventional numerical predictions. Although the frequency of multi-value predictions has been reduced, the negative impact of truncation errors has not been dealt with. The resulting P&O model has the problem of unclear changes in predicted values ​​when predicting SOC in the stable operation range of electric vehicles. Some have applied multimodal prediction to existing ANNs (Artificial Neural Networks), which temporarily improved the prediction accuracy of ANNs, but did not pay enough attention to the truncation error. When the number of significant digits after the decimal point of SOC is 1-2 digits lower than that of other features, the prediction accuracy of the model will be significantly reduced.

[0005] As mentioned above, the prediction methods in the prior art generally have the disadvantages of low prediction accuracy, slow speed, and complex calculation.

[0006] As a new prediction model, the Informer model has attracted the attention of researchers due to its excellent performance. However, the actual prediction of electric vehicle SOC requires higher accuracy and faster speed, which the Informer model still lacks. Summary of the invention

[0007] In view of the above-mentioned problems existing in the prior art, the purpose of the present invention is to provide an optimized Informer model design method for electric vehicle SOC prediction, which performs structural optimization and logical optimization on the Informer model, optimizes the model structure through health assessment, and improves the prediction performance of the Informer without increasing the data cost; in terms of logical optimization, basic state variables are added, and the SOC is indirectly predicted through equivalent conversion to reduce the adverse effects of truncation errors. While retaining the structural optimization effect, the prediction performance of the Informer is further improved, with lower prediction cost and better prediction effect.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is:

[0009] The optimized Informer model design method for electric vehicle SOC prediction includes the following steps:

[0010] Step S1: Establish an Informer SOC prediction model;

[0011] Step S2: Based on the electric vehicle operation data, a health matrix is ​​obtained through a health assessment weight optimization algorithm;

[0012] Step S3: Utilizing the health matrix, the Informer SOC prediction model is optimized from three aspects, namely, the encoder, the decoder, and the control logic of the Informer SOC prediction model, to obtain an optimized Informer model for electric vehicle SOC prediction.

[0013] As a further improvement of the above technical solution:

[0014] In step S1, the input of the Informer SOC prediction model is the electric vehicle operation data, and the Informer SOC prediction model converts the electric vehicle operation data into an embedding vector of the model input.

[0015] In step S2, the health assessment weight optimization algorithm obtains the health matrix A through data collection, data cleaning, feature screening, matrix calculation, and matrix generation in sequence;

[0016] Data collection collects the operation data of the electric vehicle through the data collection module;

[0017] Data cleaning is to remove null values ​​and abnormal values ​​of running data and reduce noise fluctuations;

[0018] Feature screening is to screen out feature data from the cleaned operating data, determine the constituent elements of the health matrix and the dimension division scheme of the matrix;

[0019] The characteristic data is used to perform matrix calculation during matrix calculation;

[0020] When generating the matrix, first arrange the matrix calculation results in order to obtain the weight matrix Q n×1 , and then use factor analysis on the running data to generate the correlation matrix R 1×n , and finally multiply Q and R to get the health matrix A.

[0021] In step S3, the method for optimizing the encoder of the Informer SOC prediction model by the health matrix is:

[0022]

[0023] Where: X (i-1) is the input of the i-th encoding layer of the encoder, X (i) is the output of the i-th encoding layer of the encoder, A is the health matrix, Conv represents convolution calculation, and Attention represents probabilistic sparse self-attention.

[0024] In step S3, the method for optimizing the decoder of the Informer SOC prediction model by the health matrix is ​​as follows: i (y,c0) gets the decoded output x i , multiplied by the health factor h, is used as the input of the next decoding sublayer, where Decoder i (·) represents the decoding calculation of the i-th decoding sublayer, and 0≤i≤L; c0 and y are the initial input and feature data in the decoding process respectively; h is the health factor, which is determined by the health matrix A n×n Compression transformation.

[0025] In step S3, the method for the health matrix to optimize the control logic of the Informer SOC prediction model is: performing temperature prediction and parameter prediction, and then performing SOC calculation based on the temperature prediction and parameter prediction to obtain the SOC prediction value, wherein the parameters in the parameter prediction include voltage, current, SOC, etc.

[0026] The steps of optimizing the control logic of the Informer SOC prediction model by the health matrix are as follows:

[0027] Step a1: referring to the equipment manual or rated parameters, a certain motor temperature and the corresponding electric vehicle characteristic parameter set are set as the basic state;

[0028] Step a2: Use historical data and the Informer model that has completed encoder optimization and / or decoder optimization to generate a new health matrix in the basic state, calculate the basic SOC change rate v0, and the m, n, l values ​​corresponding to each feature;

[0029] The influence of a single feature on SOC is calculated by approximate simplification, i.e., k=m*Δx+n*Δx+l, where k is the feature influence multiplier, m is the influence coefficient of engine temperature, which is determined by the material properties of the power supply module; n is the influence coefficient of other features, which is determined by the mechanical structure of the electric vehicle; Δx is the change of the selected feature, and l is a constant. During normal operation, for each feature, m, n, and l have a corresponding fixed value, and this value does not change due to the offset of the SOC characteristic curve.

[0030] Step a3: According to the value of the motor temperature, the normal operating interval is divided into several sub-intervals. For the boundary temperature of the entire interval and the intermediate temperature of the sub-interval intersection, the data set in step a2 and the Informer model that has completed encoder optimization and / or decoder optimization are used to obtain the characteristic parameters corresponding to each motor temperature when the SOC change rate is v0, and the characteristic parameters, the basic SOC change rate and the corresponding motor temperature are marked as sub-basic states;

[0031] Step a4: When predicting, firstly, the predicted value of the motor temperature is obtained through the Informer model that has completed encoder optimization and / or decoder optimization, and the corresponding basic state or sub-basic state is found according to the predicted temperature, and then the characteristic parameter conversion is performed to obtain the corresponding SOC change rate at the prediction time; if there are p characteristics, the characteristic influence multiples k1, k2, ..., k p , relative to the basic change rate v0, the change multiple k0 of the predicted rate v1, the relevant calculation formula is:

[0032] k0=k1*k2*…*k p ;

[0033] At the prediction time, the calculation formula corresponding to the SOC prediction rate v1 is:

[0034] v1=v0*k0

[0035] Wherein, v0 is the basic change rate of SOC, obtained by step a2;

[0036] Step a5: Combine the historical SOC value and the sampling interval, and calculate the SOC prediction value through the change rate v1 obtained in step a4.

[0037] When evaluating the optimized Informer model, the determination coefficient R 2 There are three evaluation indicators: mean square error (MSE) and mean absolute error (MAE).

[0038]

[0039] In the formula, y i is the true value, is the predicted value, is the true mean, R 2 It indicates the degree of fitting. The closer the value is to 1, the better the fitting effect of the model. MSE and MAE indicate the degree of dispersion. The closer the value is to 0, the smaller the prediction fluctuation of the model.

[0040] The beneficial effects of the present invention are:

[0041] (1) The Informer model was structurally and logically optimized. The model structure was optimized through health assessment, and the prediction performance of Informer was improved without increasing data costs. In terms of logic optimization, basic state variables were added, and SOC was indirectly predicted through equivalent conversion, reducing the adverse effects of truncation errors. While retaining the structural optimization effect, the prediction performance of Informer was further improved, with lower prediction costs and better prediction effects.

[0042] (2) Optimizing the Encoder module can improve the Informer prediction speed, and optimizing the Decoder module can improve the Informer prediction accuracy. The two optimization effects do not conflict with each other. Using En-De combination optimization can simultaneously improve the Informer prediction speed and accuracy.

[0043] (3) The Informer model after prediction logic optimization can reduce the impact of truncation error and further improve prediction accuracy based on En-De optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is the flow chart of the health assessment weight optimization algorithm.

[0045] Figure 2 This is the schematic diagram of the Informer model.

[0046] Figure 3 It is a schematic diagram of the principle of optimizing the InformerSOC prediction model by using the health matrix to control the logic of the InformerSOC prediction model.

[0047] Figure 4 It is the SOC prediction diagram of the Beijing electric vehicle charging dataset in the collected dataset 1.

[0048] Figure 5 It is the SOC prediction diagram of the Shanghai electric vehicle charging dataset in the collected dataset 1.

[0049] Figure 6 This is the SOC prediction diagram of electric vehicle No. 2 in the collected data set 2.

[0050] Figure 7 This is the SOC prediction diagram of electric vehicle No. 17 in the collected data set 2.

[0051] Figure 8 This is the SOC prediction diagram of Class A electric vehicles in the collected data set 3.

[0052] Fig. 9 This is the SOC prediction diagram of Class B electric vehicles in the collected data set 3.

[0053] Fig.10 This is the SOC prediction diagram of Class A electric vehicle No. 2 in the collected data set 4.

[0054] Fig.11 This is the SOC prediction diagram of Class B electric vehicle No. 4 in the collected data set 4

[0055] Fig.12 This is the SOC prediction diagram of Class A electric vehicle No. 3 in the collected data set 4 before logic optimization in July.

[0056] Fig.13 This is the SOC prediction diagram of Class A electric vehicle No. 3 in the collected data set 4 after logic optimization in July.

[0057] Fig.14 This is the motor temperature prediction chart for Class A electric vehicles No. 3 in the collected data set 4 in July.

[0058] Fig.15 This is the SOC prediction diagram of Class A electric vehicle No. 3 in the collected data set 4 before logic optimization in August.

[0059] Fig.16 This is the SOC prediction diagram of Class A electric vehicle No. 3 in the collected data set 4 after logic optimization in August.

[0060] Fig.17 This is the motor temperature prediction chart of Class A electric vehicle No. 3 in August in the collected data set 4.

[0061] Fig.18 This is the SOC prediction diagram of Class A electric vehicle No. 3 in the collected data set 4 before logic optimization in September.

[0062] Fig.19 This is the SOC prediction diagram of Class A electric vehicle No. 3 in the collected data set 4 after logic optimization in September.

[0063] Fig. 20 This is the motor temperature prediction chart of Class A electric vehicle No. 3 in September in the collected data set 4.

[0064] Fig.21 This is the SOC prediction diagram of Class A electric vehicle No. 3 in the collected data set 4 before logic optimization in October.

[0065] Fig. 22 This is the SOC prediction diagram of Class A electric vehicle No. 3 in the collected data set 4 after logic optimization in October.

[0066] Fig.23 This is the motor temperature prediction chart for Class A electric vehicles No. 3 in the collected data set 4 in October.

[0067] Fig.24 This is the SOC prediction diagram of Class B electric vehicle No. 2 in the collected data set 4 before logic optimization in July.

[0068] Fig.25 This is the SOC prediction diagram of Class B electric vehicle No. 2 in the collected data set 4 after logic optimization in July.

[0069] Fig.26 This is the motor temperature prediction chart for Class B electric vehicles No. 2 in July in the collected data set 4.

[0070] Fig. 27 This is the SOC prediction diagram of Class B electric vehicle No. 2 in the collected data set 4 before logic optimization in August.

[0071] Fig.28 This is the SOC prediction diagram of Class B electric vehicle No. 2 in the collected data set 4 after logic optimization in August.

[0072] Fig.29 This is the motor temperature prediction chart for Class B electric vehicles No. 2 in August in the collected data set 4.

[0073] Fig.30 This is the SOC prediction diagram of Class B electric vehicle No. 2 in the collected data set 4 before logic optimization in September.

[0074] Fig.31 This is the SOC prediction diagram of Class B electric vehicle No. 2 in the collected data set 4 after logic optimization in September.

[0075] Fig.32 This is the motor temperature prediction chart for Class B electric vehicles No. 2 in the collected data set 4 in September.

[0076] Fig.33 This is the SOC prediction diagram of Class B electric vehicle No. 2 in the collected data set 4 before logic optimization in October.

[0077] Fig.34 This is the SOC prediction diagram of Class B electric vehicle No. 2 in the collected data set 4 after logic optimization in October.

[0078] Fig.35 This is the motor temperature prediction chart for Class B electric vehicles No. 2 in October in the collected data set 4. DETAILED DESCRIPTION

[0079] The specific implementation of the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the present invention, and is not used to limit the present invention.

[0080] For ease of description, spatially relative terms such as "above", "above", "on the upper surface of", "above", etc. may be used here to describe the spatial positional relationship between a device or feature and other devices or features as shown in the figure. It should be understood that spatially relative terms are intended to include different orientations of the device in use or operation in addition to the orientation described in the figure. For example, if the device in the accompanying drawings is inverted, the device described as "above other devices or structures" or "above other devices or structures" will be positioned as "below other devices or structures" or "below other devices or structures". Thus, the exemplary term "above" can include both "above" and "below". The device can also be positioned in other different ways (rotated 90 degrees or in other orientations), and the spatially relative descriptions used here are interpreted accordingly.

[0081] The optimized Informer model design method for electric vehicle SOC prediction includes the following steps:

[0082] Step S1: Establish an Informer SOC prediction model;

[0083] Step S2: Based on the electric vehicle operation data, a health matrix is ​​obtained through a health assessment weight optimization algorithm;

[0084] Step S3: Utilizing the health matrix, the Informer SOC prediction model is optimized from three aspects, namely, the encoder, the decoder, and the control logic of the Informer SOC prediction model, to obtain an optimized Informer model for electric vehicle SOC prediction.

[0085] Each step is described below.

[0086] 1. Health Assessment

[0087] Health assessment is a weight optimization algorithm that converts the correlation weights between features into a health matrix based on the characteristics of the statistical probability distribution of the model output. Figure 1 As shown, where:

[0088] (1) Data acquisition: The data acquisition module collects the total voltage, total current, ambient temperature, vehicle speed, operating mode and other operating data of the electric vehicle;

[0089] (2) Data cleaning: Eliminate empty values, abnormal values, and noise fluctuations in operating data to improve data reliability;

[0090] (3) Feature screening: Screen the feature data from the cleaned operation data to determine the constituent elements of the health matrix (i.e. Figure 1 ) and the matrix dimension partitioning scheme;

[0091] (4) Matrix calculation: Matrix calculation is performed using feature data, including the weight calculation of the impact of each feature on SOC under different operating conditions, and the dimension calculation of the mutual impact between different features;

[0092] The brief formula for matrix calculation is as follows:

[0093]

[0094] Where i is the feature number, from 1 to n; μ is the mean vector; Σ is the covariance matrix.

[0095] (5) Matrix generation: Arrange the matrix calculation results in sequence to obtain the weight matrix Q n×1 Each element of Q represents the weight factor of a specified feature under different operating conditions. Then, factor analysis is used on the operating data to generate the correlation matrix R 1×n Finally, multiply Q and R to get the health matrix A, which is:

[0096] A n×n =Q n×1 R 1×n (2)

[0097] Where n is the matrix dimension, usually 4 to 12. When n is greater than 12, it is necessary to perform dimensionality reduction on the running data.

[0098] The optimization analysis of health assessment is as follows:

[0099] Assume that the first layer input z of the prediction model (0) , a coefficient matrix with the main diagonal elements being 1 and the remaining elements being between (0,1) Right D 0 Sort so that D 0 The order of row and column labels remains consistent.

[0100] During the prediction process, the i-th layer input z (i-1) The coefficient matrix D i , which can be regarded as i D 0 Multiplication:

[0101] D i =[D 0 ] i =D 0 ×D0 ×…×D 0 (3)

[0102] Health assessment optimization can be considered as input z (0) Multiplying with the health matrix A, we get Right now but The coefficient matrix of The following calculations yield:

[0103]

[0104] In formula (4), A reduces D 0 The weak correlation weight of D is based on maintaining the dominance and symmetry of the main diagonal. 0 The weakly correlated main elements of are adjusted to between (0,1). After optimization, all D 0 quilt Instead, each time the model's forward calculation is performed, the weakly correlated weight decreases, thereby reducing the impact of weakly correlated features and improving the model's prediction performance.

[0105] Due to different prediction processes, the impact of input z and coefficient matrix D is also different, and after optimization, the degree of improvement in the prediction performance of the model is also different.

[0106] 2. Informer Optimization Model

[0107] 2.1 SOC prediction model

[0108] The Informer model is a self-attention prediction model. Different from other prediction models, the Informer adopts an encoder-decoder separation structure. During the prediction process, there is no other connection between the encoder and the decoder except for the transmission of the encoded output. There is no feedback mechanism in all prediction steps. See the Informer SOC prediction model diagram. Figure 2 .

[0109] In the Informer model, there are three key structures:

[0110] (1) Self-attention Distilling Operation: The Encoder consists of multiple encoding layers. The input data performs feature encoding operations at each layer to reduce the sequence length and passes through the last layer to output strongly correlated features. The features are then concatenated to obtain the encoded output for subsequent use.

[0111] (2) Generative Style Decoder: The decoding operation of the decoder is based on the masked multi-head attention of the decoding sublayer. Only forward calculation is required to obtain the decoding output of the specified length, preventing the diffusion of cumulative errors and reducing the probability of multi-value prediction.

[0112] (3) ProbSparse Self-attention Mechanism: Based on the conclusion that a small part of the dot product contributes the main attention, Informer increases the attention to strongly correlated features. During the prediction process, the self-attention mechanism is improved by increasing the weight of strongly correlated features.

[0113] In step S1, when establishing the Informer SOC prediction model, the electric vehicle operation data is preprocessed, and the Informer model parameters, including the number of encoder and decoder layers, are configured according to the data characteristics. In this step, the initial input is the electric vehicle operation data. The Informer SOC prediction model converts the input electric vehicle operation data into the model's embedding vector. The Informer model maps the discrete input data into a continuous high-dimensional feature vector for subsequent processing, which is equivalent to a lookup table. The establishment of the Informer SOC prediction model is a technology well known to those skilled in the art and will not be repeated here.

[0114] The existing Informer model has the following shortcomings:

[0115] 1. Frequent calculations: In order to increase the attention of strongly correlated features, the Encoder module needs to perform two dot product calculations, which reduces the prediction speed of the Informer model;

[0116] 2. Decoding defects: The decoding input consists of 0 placeholders and the encoded output. The weight of the placeholder is too high, which affects the decoding reliability of the Decoder module and reduces the prediction accuracy of the Informer model.

[0117] 3. High complexity: The sampling query of the sparse self-attention mechanism increases the computational complexity and reduces the lightweight advantage of the Informer model.

[0118] In order to make up for the above shortcomings and improve the prediction performance of the Informer model, this scheme obtains the health matrix through health assessment, and uses this matrix to optimize the Informer model from three aspects: encoder, decoder, and control logic.

[0119] 2.2Encoder Optimization

[0120] Encoder is an encoder. Encoder has multiple encoding layers, each of which has a self-attention sublayer and a convolution sublayer. Let the input of the i-th encoding layer be X (i-1) , the output is X (i) , in each encoding layer, X (i-1) We advance forward through Multi-Head Self-Attention and then use convolution to reduce the sequence length to obtain X (i) ,Right now:

[0121] X (i) =Conv(Attention(X (i-1) )) (5)

[0122] In formula (5), Conv represents convolution calculation, Attention represents probabilistic sparse self-attention, and Attention(X (i-1) ) operation, the self-attention sub-layer output p can be obtained, that is:

[0123]

[0124] In formula (6), softmax represents the normalization function, Indicates X (i-1) After sampling, the probability sparse matrix obtained is represents the variance, K, K T , V represents the adjustment matrix of the sampling key value, and p represents the output of the self-attention sub-layer.

[0125] In X (i-1) In the example, all features have equal weights. Now, we use the health matrix A to improve X. (i-1) In , the weight of strongly correlated features is:

[0126]

[0127] In formula (5), Replace X (i-1) , which can improve the self-attention mechanism and accelerate sparse sampling and The generation of ,improves the encoding speed of Encoder and the prediction speed of Informer.

[0128] The optimization analysis of Encoder is as follows:

[0129] In formula (6), The i-th query element q i , obtained through the following operations:

[0130]

[0131] In the formula, q i , k i , v i represent K, the i-th row in V, p(k j |q i )=k(q i ,k j ) / ∑ l k(q i ,k l ), k(q i ,k j ) represents an asymmetric exponential kernel Represents variance.

[0132] In equations (6) and (8), the sparse metric M(q i ,K)'s Top-u query, that is:

[0133]

[0134] Where, L K Represents the sequence length of matrix K.

[0135] In sparse self-attention, the length of query and key is equal, i.e., L Q =L K =L, the computational complexity is O(L ln L). (i-1) All features of the query have equal weights. In order to increase the weights of strongly correlated features, the Top-u query needs to perform two dot product calculations, which affects the encoding speed of the Encoder.

[0136] by Replace X (i-1) , as the input of each encoding layer, can increase the weight of strongly correlated features. Top-u query only needs one dot product calculation, and the computational complexity of sparse self-attention is reduced from O(L·ln L) to Improve the encoding speed of the Encoder, thereby speeding up the prediction speed of the Informer model.

[0137] 2.3Decoder Optimization

[0138] Decoder is a decoder. Let Decoder be composed of L overlapping decoding sub-layers. During the decoding process, the initial input c0 and feature data y need to undergo L decoding operations to obtain the decoded output c L ,Right now:

[0139] c L =Decoder(y,c L-1 )=Decoder L(Decoder L-1 (…Decoder1(y,c0)…)) (10)

[0140] In the formula, Decoder i (·) represents the decoding calculation of the i-th decoding sublayer, and 0≤i≤L. c0 is the input of the first decoding sublayer, c i-1 is the input of the i-th decoding sublayer, and y represents the feature data at the prediction time.

[0141] In c L-1 In the example, all features are equally weighted. Health assessment can be used to improve c L-1 In , the weight of strongly correlated features is:

[0142] c L =Decoder(y,c L-1 )=Decoder L (Decoder L-1 (…(Decoder1(y,c0)·h)…)) (11)

[0143] In the formula, h is the health factor, which is composed of the health matrix A n×n Compression transformation.

[0144] In each decoding sublayer, (Decoder i (y,c0)) to get the decoded output x i , after being multiplied by the health factor h, it is used as the input of the next decoding sublayer, which can improve the weight of strongly correlated features, the decoding accuracy of the Decoder and the prediction accuracy of the Informer model.

[0145] The optimization analysis of Decoder is as follows:

[0146] The input c0 of the first decoding sublayer can be expressed as a sequence Right now:

[0147]

[0148] In the formula, yes Figure 2 The encoded output in is a sequence placeholder with a scalar value of 0 (corresponding to Figure 2 Medium X de 0 part), L token Indicates the length of the sequence output by the encoding, L y Indicates the length of the placeholder sequence, d model Represents the feature dimension, and Concat represents vector concatenation.

[0149] By using 0 placeholders and -∞ masked dot products, the probabilistic sparse self-attention mechanism is applied to the decoding process of the Decoder.

[0150] Although the use of masked multi-head attention can prevent In , each feature focuses on its own autoregression phenomenon at the next moment, but the encoding output and placeholders The weights are equal, and each time the Decoder performs forward decoding, The accumulated error caused will spread once. The diffusion of the accumulated error reduces the decoding accuracy of the Decoder.

[0151] The optimization scheme of formula (11) is equivalent to improving The weight of The weight of is known from equations (3) and (4). Each time the forward decoding is performed, The accumulated error caused is reduced once, which improves the decoding reliability of the Decoder and thus improves the prediction accuracy of the Informer.

[0152] As can be seen from the above, Informer adopts an encoder-decoder separation structure, and there is no feedback mechanism. Optimizing the encoder module can improve the prediction speed of Informer, and optimizing the decoder module can improve the prediction accuracy of Informer. The two optimization effects do not conflict. Using En-De combined optimization, that is, using both encoder optimization and decoder optimization, can simultaneously improve the prediction speed and accuracy of Informer.

[0153] 3. Prediction logic optimization

[0154] 3.1 Truncation Error

[0155] Many SOC prediction models and methods in the prior art have reduced prediction accuracy due to truncation errors, and this problem occurs frequently under certain circumstances. In fact, all prediction models have truncation errors. The reason is that due to the accuracy limitations of the data acquisition module, the effective digits of SOC are mostly only 0 or 1 decimal place, which is much lower than the 3-4 effective digits of other features. When predicting, the model will unify the data accuracy, that is, make the effective digits of the predicted value consistent with the feature with the most effective digits in the training data. The greater the difference in effective digits between the predicted value and the true value, the greater the SOC truncation error.

[0156] Assume that at a certain moment, the predicted value of SOC is 28.5273. When the true value is recorded as 28.5, the truncation error is {(28.5273-28.5) / 28.5273}×100%=0.096%, and when the true value is recorded as 28, the truncation error is {(28.5273-28) / 28.5273}×100%=1.848%, and the truncation error increases by about 20 times. From the above analysis, it can be seen that the truncation error will continue to accumulate as the forward calculation proceeds. When the truncation error accumulates to a certain extent, the SOC prediction accuracy will drop significantly. In addition to the limitation of acquisition accuracy, the state change of the power supply module, resulting in the virtual power phenomenon that the actual SOC value is lower than the measured value, and the long-term low-speed operation of electric vehicles, the state change is not obvious, resulting in the SOC fixed value measurement phenomenon, all of which will produce truncation errors.

[0157] In order to reduce the negative impact of truncation error, an Informer prediction logic optimization scheme is designed to replace direct SOC prediction with indirect SOC prediction through health assessment, optimization of model prediction logic, addition of basic state variables, and equivalent conversion.

[0158] 3.2 Logic Optimization Solution

[0159] Informer's prediction logic optimization solution is based on the following four assumptions:

[0160] a. Each motor temperature corresponds to a unique SOC characteristic curve, which is a smooth and continuous single-value curve within the normal operating range, and the characteristic curves of the same model and temperature are consistent;

[0161] b. All characteristics are smooth and continuous within the normal operating range, and are affected by SOC, engine temperature and other characteristics. Moreover, when other characteristics remain unchanged and a certain characteristic changes alone, the SOC change rate will also change accordingly;

[0162] c. Within the normal operating range, the SOC characteristic curves of the same model at different temperatures have the same curve trend, with only horizontal and vertical shifts causing differences in values;

[0163] d. The impact of a single feature on SOC is calculated by approximate simplification, i.e. k = m*Δx + n*Δx + l. Among them, k is the feature impact multiplier, m is the engine temperature impact coefficient, which is determined by the material properties of the power supply module; n is the impact coefficient of other features, which is determined by the mechanical structure of the electric vehicle; Δx is the change in the selected feature, and l is a constant. In normal operation, for each feature, m, n, l has a corresponding fixed value, and the value does not change due to the offset of the SOC characteristic curve.

[0164] The prediction logic optimization scheme based on the above assumptions is as follows:

[0165] Step a1: Referring to the equipment manual or rated parameters, a certain motor temperature and the corresponding electric vehicle characteristic parameters are set as the basic state.

[0166] Step a2: Use historical data and the Informer model that has completed encoder optimization and / or decoder optimization to generate a new health matrix in the basic state, that is, a new feature matrix, and obtain the SOC basic state change rate v0, as well as the m, n, l values ​​corresponding to each feature in the assumption d. In this step, the SOC basic state change rate v0 is calculated from the new health matrix.

[0167] Step a3: According to the numerical value of the motor temperature, the normal operating range is divided into several sub-ranges. For the boundary temperature of the entire range and the intermediate temperature where the sub-ranges intersect, the data set in step a2 and the Informer model that has completed encoder optimization and / or decoder optimization are used to obtain the characteristic parameters corresponding to each motor temperature when the SOC change rate is v0, and the characteristic parameters, the basic SOC change rate and the corresponding motor temperature are marked as sub-basic states.

[0168] In this step, the division of the entire interval and sub-intervals refers to the equipment manual and industry standards. Different temperatures have different state parameters, so the entire normal operating interval is divided according to temperature.

[0169] The characteristic parameters generally include the voltage, current, speed and other parameters of the electric vehicle, in accordance with the equipment manual or industry standard parameters.

[0170] Step a4: When predicting, first obtain the predicted value of the motor temperature through the Informer model that has completed encoder optimization and / or decoder optimization, and find the corresponding basic state or sub-basic state based on the predicted temperature, and then convert the characteristic parameters to obtain the corresponding SOC change rate at the prediction moment.

[0171] If there are p features, then the feature influence ratios k1, k2, ..., k can be calculated. p Relative to the basic change rate v0, the change multiple k0 of the predicted rate v1 is calculated by the following formula:

[0172] k0=k1*k2*…*k p (13)

[0173] At the time of prediction, the calculation formula corresponding to the SOC prediction rate v1 is

[0174] v1=v0*k0 (14)

[0175] Where v0 is the basic change rate of SOC, which is obtained by step 2.

[0176] Step a5: Combine the historical SOC value and the sampling interval, and calculate the SOC prediction value through the change rate v1 obtained in step a4.

[0177] Forecast logic optimization diagram, see Figure 3 , where the input “prediction data” of the “temperature prediction” block is the temperature of the motor.

[0178] 4 Model testing

[0179] 4.1 Evaluation Metrics

[0180] Using R 2 There are three evaluation indicators: coefficient of determination, MSE (mean square error), and MAE (mean absolute error). The calculation of each indicator is shown in formulas (15)-(17):

[0181]

[0182] In the formula, y i is the true value, is the predicted value, is the true mean. 2 It indicates the degree of fitting. The closer the value is to 1, the better the fitting effect of the model. MSE and MAE indicate the degree of dispersion. The closer the value is to 0, the smaller the prediction fluctuation of the model.

[0183] Improved prediction accuracy through R 2 Improve judgment. When MSE and MAE remain unchanged or decrease, R 2 From 0.8 to 0.85, the prediction accuracy increased by 25%. 2 From 0.9 to 0.95, the prediction accuracy increased by 50%. The prediction speed was improved, and the judgment time was reduced. When MSE, MAE, R 2 When the time remains unchanged or improves, for a single prediction, it takes 30 seconds before optimization and 20 seconds after optimization, and the prediction speed is increased by 50%; it takes 30 seconds before optimization and 10 seconds after optimization, and the prediction speed is increased by 200%.

[0184] If the optimization is effective, R 2 As the MSE and MAE increase, they will also decrease. However, due to the different calculation methods of the prediction models, there are differences in the degree of prediction fluctuation, and in order to increase the overall fit, reducing the fit of specific points will increase the processing cost. Therefore, when the two indicators increase by more than 10% and the remaining one decreases by no more than 2%, it is also considered to be effective.

[0185] 4.2 Simulation Environment

[0186] Four different public data sets are used for simulation tests. The data sets are as follows:

[0187] Dataset 1: China Urban Electric Vehicle Charging Dataset: The dataset contains the current, voltage, SOC and energy recorded in the electric vehicle charging piles in Beijing, Shanghai and Shenzhen, a total of 4 features. The Beijing dataset numbered 1 and the Shanghai dataset numbered 2 were selected for simulation testing;

[0188] Dataset 2: 20 new energy vehicle battery module charging and discharging data set: The data set contains the charging and discharging data of 20 electric vehicles, and the operating data of each electric vehicle, including 9 features such as voltage, current, temperature, energy, SOC, etc. The operating data of vehicle No. 2 and vehicle No. 17 are selected for simulation testing;

[0189] Dataset 3: New energy vehicle power battery health data set: The data set contains the charging and discharging data of two electric vehicles, and the operating data of each electric vehicle, including 6 features such as speed, voltage, current, and SOC. The operating data of the two electric vehicles are simulated and tested separately;

[0190] Dataset 4: Guangdong-Hong Kong-Macao Greater Bay Area New Energy Vehicle Health Dataset; The dataset contains the operating data of 5 vehicles of 2 types of new energy vehicles. The operating conditions contained in the dataset are relatively complex, including not only charging and discharging conditions, but also high and low speed operation, emergency braking, continuous acceleration or deceleration, and other operating conditions. The operating data of each electric vehicle includes multiple features such as SOC, speed, voltage, current, and temperature. 12 features including SOC were selected for simulation testing.

[0191] In the structural optimization stage, four models were selected for simulation testing, including non-optimization, encoder optimization, decoder optimization, and encoder-decoder optimization. In the logic optimization stage, the prediction logic of the four models was optimized in turn, and the eight models before and after logic optimization were used for simulation testing.

[0192] Each data set has about 20,000 pieces of operating data, and the first 80% of the data is the training set, and the last 20% of the data is the test set. All data sets are derived from the collected electric vehicle operating data.

[0193] In each prediction effect graph, V1 represents unoptimized, V2 represents encoder optimization, V3 represents decoder optimization, V4 represents En-De optimization, true represents the true value, the horizontal axis represents the sequence number of the data, and the vertical axis represents the standardized prediction value. Each prediction effect graph has a corresponding table that records the relevant evaluation indicators.

[0194] The simulation test platform is: Intel Core i7-7700HQ CPU, 8GB DDR4 RAM, 1TB WDC HDD, and the operating system is Windows 10.

[0195] In order to reduce the impact of dimensional differences, all prediction results are standardized as follows:

[0196]

[0197] In the formula, z0 represents the unstandardized prediction result, z1 represents the standardized prediction result, and z s represents the variance of the prediction results, Represents the average of all prediction results.

[0198] Informer model parameters are shown in Table 1

[0199] Table 1 Informer model parameters

[0200]

[0201] 4.3 Model Optimization and Verification

[0202] The prediction results of the above dataset 1 are as follows Figure 4 and Figure 5 , the prediction effect of data set 2 is as follows Figure 6 and Figure 7 , the prediction effect of data set 3 is as follows Figure 8 and Fig. 9 , the prediction results of data set 4 are as follows Fig.10 and Fig.11 ,The prediction indicators of each data set are shown in Tables 2 to 9.

[0203] Table 2 Beijing electric vehicle SOC prediction indicators

[0204]

[0205]

[0206] Table 3 Shanghai Electric Vehicle SOC Prediction Index

[0207]

[0208] Table 4 SOC prediction index of electric vehicle No. 2

[0209]

[0210] Table 5 SOC prediction index of No. 17 electric vehicle

[0211]

[0212] Table 6 SOC prediction indicators for Class A electric vehicles

[0213]

[0214] Table 7 SOC prediction indicators for Class B electric vehicles

[0215]

[0216]

[0217] Table 8 SOC prediction index of Class A electric vehicle No. 2

[0218]

[0219] Table 9 SOC prediction index of No. 4 Class B electric vehicles

[0220]

[0221] Depend on Figures 4 to 11 It can be seen that after En-De optimization, the prediction fit of Informer has been improved. From Table 2 to Table 9, it can be seen that after En-De optimization, R 2 The index has been improved by 15%, MSE and MAE have been reduced, and the prediction time has been reduced by 50%. After En-De optimization, Informer's prediction accuracy has been improved by 15%, and the prediction speed has been improved by 100%. In summary, based on health assessment, optimizing the En-De module can improve the prediction performance of Informer.

[0222] 4.4 Logic Optimization Verification

[0223] contrast Fig.10 , Fig.11 and Figures 4 to 7 ,Table 8, Table 9 and Table 2 to Table 7, it is not difficult to see that when the operation of the electric vehicle changes from a single charging and discharging state to a complex operating state of multi-state switching, the prediction performance of Informer decreases significantly. From the above analysis, it can be seen that the decrease in prediction performance is caused by the truncation error of SOC, and the prediction logic of Informer needs to be optimized.

[0224] The above analysis provides a prediction logic optimization scheme. In order to verify whether the proposed scheme can reduce the impact of truncation error and improve the SOC prediction performance of Informer, the operating data of 3 Class A vehicles and 2 Class B vehicles from July to October in Dataset 4 are selected for simulation testing.

[0225] The single-month simulation test of each vehicle not only includes the SOC prediction effect diagram before and after logic optimization, but also includes the prediction effect diagram of the motor temperature after logic optimization, as well as the corresponding table to record the evaluation indicators.

[0226] In order to avoid overfitting caused by multiple training on the same data, for the same model before and after logic optimization, data from the same vehicle in the same month and different time periods were selected for simulation verification.

[0227] for Figure 3 In terms of the system, parameter prediction is a pre-calculation and is not performed during the formal prediction. The SOC calculation is a simple arithmetic operation and does not involve the Informer model. The time taken is less than 0.1% of the temperature prediction. Therefore, after the logic is optimized, the motor temperature prediction time is added to the SOC prediction time. Only in the SOC prediction table, the prediction time item is added, and in the motor temperature prediction table, the prediction time item is left blank.

[0228] The prediction results using the July data of No. 3 Class A vehicles are as follows: Figure 12 to Figure 14 , July forecast indicators for Class A vehicles No. 3 are shown in Tables 10 to 12.

[0229] Table 10 SOC prediction index before 3A-7 logic optimization

[0230]

[0231] Table 11 SOC prediction index after 3A-7 logic optimization

[0232]

[0233] Table 12 3A-7 Motor Temperature Prediction Index

[0234]

[0235] The prediction results using the August data of Class 3 A vehicles are as follows: Figures 15 to 17 , August forecast indicators for Class A vehicles No. 3 are shown in Tables 13 to 15.

[0236] Table 13 SOC prediction index before 3A-8 logic optimization

[0237]

[0238] Table 14 SOC prediction index after 3A-8 logic optimization

[0239]

[0240] Table 15 3A-8 Motor Temperature Prediction Index

[0241]

[0242] The prediction results using the September data of Class 3 A vehicles are as follows: Figures 18 to 20 The September forecast indicators for Class 3A vehicles are shown in Tables 16 to 18.

[0243] Table 16 SOC prediction index before 3A-9 logic optimization

[0244]

[0245] Table 17 SOC prediction index after 3A-9 logic optimization

[0246]

[0247]

[0248] Table 18 3A-9 Motor Temperature Prediction Index

[0249]

[0250] The prediction results using the October data of Class 3 A vehicles are as follows: Figure 21 to Figure 23 , October forecast indicators for Class A vehicles No. 3, see Tables 19 to 21.

[0251] Table 19 SOC prediction index before 3A-10 logic optimization

[0252]

[0253] Table 20 SOC prediction index after 3A-10 logic optimization

[0254]

[0255] Table 21 3A-10 Motor Temperature Prediction Index

[0256]

[0257] The prediction results using the July data of Class 2 B vehicles are as follows: Figure 24 to Figure 26 , July forecast indicators for Class 2B vehicles are shown in Tables 22 to 24.

[0258] Table 22 SOC prediction index before 2B-7 logic optimization

[0259]

[0260] Table 23 SOC prediction index after 2B-7 logic optimization

[0261]

[0262] Table 24 2B-7 Motor Temperature Prediction Index

[0263]

[0264] The prediction results using the August data of Class 2 B vehicles are as follows: Figure 27 to Figure 29 , August forecast indicators for Class 2B vehicles are shown in Tables 25 to 27.

[0265] Table 25 2B-8 SOC prediction indicators before logic optimization

[0266]

[0267] Table 26 SOC prediction index after 2B-8 logic optimization

[0268]

[0269]

[0270] Table 27 2B-8 Motor Temperature Prediction Index

[0271]

[0272] The prediction results using the September data of Class 2 B vehicles are as follows: Figure 30 to Figure 32 , September forecast indicators for Class 2B vehicles are shown in Tables 28 to 30.

[0273] Table 28 2B-9 SOC prediction indicators before logic optimization

[0274]

[0275] Table 29 SOC prediction index after 2B-9 logic optimization

[0276]

[0277] Table 30 2B-9 Motor temperature prediction index

[0278]

[0279] The prediction results using the October data of Class 2 B vehicles are as follows: Figure 33 to Figure 35 , October forecast indicators for Class 2B vehicles are shown in Tables 31 to 33.

[0280] Table 31 2B-10 SOC prediction index before logic optimization

[0281]

[0282] Table 32 SOC prediction index after 2B-10 logic optimization

[0283]

[0284] Table 33 2B-October Motor Temperature Prediction Index

[0285]

[0286] Depend on Fig.13 , Fig.16 , Fig.19 , Fig. 22 , Fig.25 , Fig.28 , Fig.31 , Fig.34 It can be seen that the prediction logic optimization does not conflict with En-De optimization. From Tables 11, 14, 17, 20, 23, 26, 29, and 32, it can be seen that after the prediction logic is optimized, the En-De optimization effect of improving the prediction accuracy by 15% and the prediction speed by 100% is retained. Figure 12-Figure 35 , Tables 10-33 show that after the prediction logic is optimized, R 2 The indicator has been improved by about 20%. After the prediction logic is optimized, the prediction accuracy of the Informer model has been improved by about 20%. The accuracy improvement of the logic optimization coexists with the accuracy improvement of the En-De optimization. By optimizing the Informer prediction logic through health assessment, the impact of truncation error can be reduced and the prediction performance of Informer can be improved.

[0287] In summary, through health assessment, the health matrix is ​​obtained, and the encoder and decoder are improved, the prediction process can be optimized, and the prediction accuracy and speed of Informer can be improved. On this basis, through health assessment, the prediction logic is optimized, basic state variables are added, and equivalent conversion is used to indirectly predict SOC, which can reduce the impact of truncation error and further improve the prediction accuracy of Informer. Through simulation tests, the En-De optimized Informer model is verified, and the prediction accuracy is improved by about 15% and the prediction speed is improved by about 100%. The Informer model after the prediction logic is optimized can reduce the impact of truncation error on the basis of En-De optimization and further improve the prediction accuracy by about 20%. Compared with the unoptimized Informer model, the Informer optimization model based on health assessment has lower prediction cost and better prediction effect when predicting the SOC of electric vehicles.

[0288] Finally, it is necessary to explain here that the above embodiments are only used to further illustrate the technical solution of the present invention in detail and cannot be understood as limiting the scope of protection of the present invention. Some non-essential improvements and adjustments made by technicians in this field based on the above content of the present invention all fall within the scope of protection of the present invention.

Claims

1. An optimized informer model design method for electric vehicle SOC prediction, characterized in that: The steps include: Step S1: Establish an Informer SOC prediction model; Step S2: Based on the electric vehicle operation data, a health matrix is ​​obtained through a health assessment weight optimization algorithm; Step S3: Utilizing the health matrix, the Informer SOC prediction model is optimized from three aspects, namely, the encoder, the decoder, and the control logic of the Informer SOC prediction model, to obtain an optimized Informer model for electric vehicle SOC prediction.

2. The design method according to claim 1, characterized in that: In step S1, the input of the Informer SOC prediction model is the electric vehicle operation data, and the Informer SOC prediction model converts the electric vehicle operation data into an embedding vector of the model input.

3. The design method according to claim 1, characterized in that: In step S2, the health assessment weight optimization algorithm obtains the health matrix A through data collection, data cleaning, feature screening, matrix calculation, and matrix generation in sequence; Data collection collects the operation data of the electric vehicle through the data collection module; Data cleaning is to remove null values ​​and abnormal values ​​of running data and reduce noise fluctuations; Feature screening is to screen out feature data from the cleaned operating data, determine the constituent elements of the health matrix and the dimension division scheme of the matrix; The characteristic data is used to perform matrix calculation during matrix calculation; When generating the matrix, first arrange the matrix calculation results in order to obtain the weight matrix Q n×1 , and then use factor analysis on the running data to generate the correlation matrix R 1×n , and finally multiply Q by R to get the health matrix A.

4. The design method according to claim 1, characterized in that: In step S3, the method for optimizing the encoder of the InformerSOC prediction model by the health matrix is: Where: X (i-1) is the input of the i-th encoding layer of the encoder, X (i) is the output of the i-th encoding layer of the encoder, A is the health matrix, Conv represents convolution calculation, and Attention represents probabilistic sparse self-attention.

5. The design method according to claim 1, characterized in that: In step S3, the method for optimizing the decoder of the InformerSOC prediction model by the health matrix is ​​as follows: i (y,c0) gets the decoded output x i , multiplied by the health factor h, is used as the input of the next decoding sublayer, where Decoder i (·) represents the decoding calculation of the i-th decoding sublayer, and 0≤i≤L; c0 and y are the initial input and feature data in the decoding process respectively; h is the health factor, which is determined by the health matrix A n×n Compression transformation.

6. The design method according to claim 1, characterized in that: In step S3, the method for the health matrix to optimize the control logic of the InformerSOC prediction model is: performing temperature prediction and parameter prediction, and then performing SOC calculation based on the temperature prediction and parameter prediction to obtain the SOC prediction value, wherein the parameters in the parameter prediction include voltage, current, and SOC.

7. The design method according to claim 6, characterized in that: The steps of optimizing the control logic of the Informer SOC prediction model by the health matrix are as follows: Step a1: referring to the equipment manual or rated parameters, set a motor temperature and the corresponding electric vehicle characteristic parameter set as the basic state; Step a2: Use historical data and the Informer model that has completed encoder optimization and / or decoder optimization to generate a new health matrix in the basic state, calculate the basic SOC change rate v0, and the m, n, l values ​​corresponding to each feature; The influence of a single feature on SOC is calculated by approximate simplification, i.e., k = m*Δx + n*Δx + l, where k is the feature influence multiplier, m is the influence coefficient of engine temperature, which is determined by the material properties of the power supply module; n is the influence coefficient of other features, which is determined by the mechanical structure of the electric vehicle; Δx is the change of the selected feature, and l is a constant. During normal operation, for each feature, m, n, and l have a corresponding fixed value, and this value does not change due to the offset of the SOC characteristic curve. Step a3: According to the value of the motor temperature, the normal operating interval is divided into several sub-intervals. For the boundary temperature of the entire interval and the intermediate temperature of the sub-interval intersection, the data set in step a2 and the Informer model that has completed encoder optimization and / or decoder optimization are used to obtain the characteristic parameters corresponding to each motor temperature when the SOC change rate is v0, and the characteristic parameters, the basic SOC change rate and the corresponding motor temperature are marked as sub-basic states; Step a4: When predicting, firstly, the predicted value of the motor temperature is obtained through the Informer model that has completed encoder optimization and / or decoder optimization, and the corresponding basic state or sub-basic state is found according to the predicted temperature, and then the characteristic parameter conversion is performed to obtain the corresponding SOC change rate at the prediction time; if there are p characteristics, the characteristic influence multiples k1, k2, ···, k p , relative to the basic change rate v0, the change multiple k0 of the predicted rate v1, the relevant calculation formula is: k0=k1*k2* · · · *k p ; At the prediction time, the calculation formula corresponding to the SOC prediction rate v1 is: v1=v0*k0 Wherein, v0 is the basic change rate of SOC, obtained by step a2; Step a5: Combine the historical SOC value and the sampling interval, and calculate the SOC prediction value through the change rate v1 obtained in step a4.

8. The design method according to any one of claims 1 to 6, characterized in that: When evaluating the optimized Informer model, the determination coefficient R 2 There are three evaluation indicators: mean square error (MSE) and mean absolute error (MAE).

9. The design method according to claim 1, characterized in that: In the formula, y i is the true value, is the predicted value, is the true mean, R 2 It indicates the degree of fitting. The closer the value is to 1, the better the fitting effect of the model. MSE and MAE indicate the degree of dispersion. The closer the value is to 0, the smaller the prediction fluctuation of the model.

Images