Vehicle starting lithium battery heating duration prediction method and system

Abstract

The invention discloses a method and a system for predicting heating duration of a starting lithium battery for a vehicle, and the method comprises the steps: carrying out the feature extraction of a plurality of historical temperatures, and obtaining historical temperature data features; the battery heating duration prediction model is adopted to predict the battery heating duration, the time is reserved in advance to heat the battery, and vehicle parking and timely starting and battery protection are achieved. The heating system comprises a battery pack and a remote control module; the battery pack comprises a starting battery pack, a heating module and a battery management system; the battery management system comprises an acquisition module, a main control module and a wireless communication module; the acquisition module acquires and starts temperature data and residual electric quantity of the battery pack, and the main control module transmits the temperature data to the remote control module through the wireless communication module to form a historical temperature data set. According to the invention, preheating of the battery in a low-temperature environment can be realized, timely starting of vehicle parking and protection of the battery are realized, and the preparation efficiency and the intelligent level of the battery are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium battery thermal management, and in particular to a method and system for predicting the heating time of a vehicle starting lithium battery. Background Art

[0002] With the popularization of automobiles and the rapid development of new energy vehicles, the pain points of using automotive starting lithium batteries have also become a topic of concern for car owners. In low temperature environments, the performance of lithium batteries will be affected, resulting in a decrease in the battery's discharge capacity, thereby affecting the vehicle's starting and driving performance. The life of lithium batteries will also be quickly consumed during frequent sudden starts at low temperatures, which will not only cause losses to car owners, but the frequently replaced discarded lithium batteries will also cause pollution to the variable environment.

[0003] In order to solve this pain point, the existing technology will choose to alleviate the performance loss of the vehicle starting lithium battery and prolong the battery life by preheating. The Chinese invention patent with application number 201810553042.6 discloses a power battery preheating method and system. The preheating signal is sent through the telematics processor, and the car owner preheats the power battery before departure, thereby avoiding the problem of the vehicle being unable to start due to the battery temperature being too low when using the car. However, in this process, the preheating of the battery needs to rely on the car owner to start it manually, and estimate the charging time and arrange the trip by himself. How to help car owners predict the lithium battery heating time point and the shortest heating time in advance based on the current temperature and power, so as to further facilitate the car owner's travel schedule, is a new direction worthy of research. Summary of the invention

[0004] Purpose of the invention:

[0005] The present invention proposes a method and system for predicting the heating time of a vehicle starting lithium battery. By acquiring the user's historical travel data and historical power consumption, a travel data set is formed, and a linear regression prediction model is input to predict the user's travel situation; the battery heating time is predicted by constructing a battery heating time prediction model, and the heating time is reserved according to the user's travel needs and the lithium battery is heated to wait for travel. The above method uses the battery pack and the remote control module as carriers, provides parameter visualization and operable functions, and then assists the user to complete the travel plan under the premise of fully utilizing the performance of the starting lithium battery and prolonging the battery life.

[0006] Technical solution:

[0007] The present invention provides a method for predicting the heating time of a vehicle starting lithium battery, the method comprising the following steps:

[0008] S1 collects the user's vehicle travel history data and current remaining power to form a user's historical travel data set.

[0009] S2 inputs the user's historical travel data set into the linear regression model, uses the least squares method to solve the regression coefficient, and predicts whether the user will travel today. If the user does not travel, it returns to S1, and if the user travels, it enters S3.

[0010] S3 predicts the user's travel time point according to the user's habits, sets the predicted time point based on the user's travel time point, and executes S4 when the predicted time point is reached, otherwise S3 is repeated.

[0011] S4 collects the current battery temperature T0 in real time, and determines whether the battery temperature T0 is in the optimal operating temperature range [T1, T2] of the lithium battery. If yes, S4 is repeatedly executed.

[0012] If the battery temperature T0 is not within the optimal operating temperature range [T1, T2] of the lithium battery, execute S5.

[0013] S5 builds a battery heating time prediction model to calculate the time required for the battery cell temperature T0 to heat to the optimal operating temperature of the lithium battery, and reserves time in advance to complete the battery's self-heating and wait for the user to travel;

[0014] The process of building a battery heating time prediction model is as follows:

[0015] S5.1: Optimize and construct the objective function with the shortest heating time as the goal.

[0016] S5.2: Based on the RVFLNN algorithm, the discrete wavelet transform (DWT) algorithm is used to preprocess and train the input data set, which includes heating power, initial cell temperature, ambient temperature, battery voltage and current, and current battery SOC value.

[0017] S5.3 uses the V-TST algorithm to perform fine search and optimization on the learning parameters in the RVFLNN algorithm to further reduce the prediction error.

[0018] Furthermore, the objective function described in S5.1 is constructed with the shortest heating time as the goal, and the objective function is as follows:

[0019]

[0020] Among them, t heat Heating time for the battery; E total is the total heating energy during the battery heating period; η is the efficiency factor, which indicates the energy conversion efficiency during the battery heating process, usually 0≤η≤1, δ η is the efficiency error correction term, P load is the load power during battery heating, I(t) is the battery heating current, is the correction term for ambient temperature change, V OC(SOC(t)) is the open circuit voltage of the battery at the current SOC, and R is the internal resistance of the battery.

[0021] Furthermore, the P load The load power calculation formula during battery heating is as follows:

[0022]

[0023] Where T b is the temperature to which heating is required, T a is the initial battery temperature, R load is the battery thermal resistance.

[0024] Furthermore, the DWT data set processing calculation method in S5.2 is as follows:

[0025] The input data x[n] is decomposed through a pair of low-pass h[n] and high-pass g[n] filters as follows:

[0026]

[0027] Where k is the index after downsampling, A1[k] and D1[k] are the approximate coefficient and detail coefficient of the first layer decomposition, respectively.

[0028] The low-pass filter output A1[k] is recursively decomposed. Each layer uses the approximate coefficient A1[k] of the previous layer as the new input data, and passes through the low-pass and high-pass filters again to obtain the approximate coefficients and detail coefficients of the next layer. The above process is repeated until the required number of decomposition layers is reached, as shown in the following formula:

[0029]

[0030] Here, i is the number of decomposition levels.

[0031] Furthermore, the optimized RVFLNN algorithm in S5.2 processes the data set as follows:

[0032] The RVFLNN algorithm includes the original input data x = [x1, x2, ···, x n ] and y enhancement nodes, passing the original input data through a randomly generated weight matrix And the bias vector β is transformed nonlinearly to obtain the output of the hidden layer, as follows:

[0033]

[0034] Among them, ξ is the activation function, N is the number of input samples, I N is an N×1 vector of all ones, used to expand the bias vector β to match the dimension of N, Represents element-wise multiplication.

[0035] Use the transformed data together with the original input data as the input layer of the new Directly connected to the output layer, the output of the output layer is obtained as follows:

[0036] ο=Wv T

[0037] Among them, ο is the output matrix, which contains the predicted outputs of all training samples.

[0038] Conjugate gradient training is used to iteratively update the weight matrix of RVFLNN, and the initial weight vector W0 is selected, and the target value vector a=[a1,a2,···,a i ], the input vectors form the input matrix V = [v1,v2,···,v i ], the steps to obtain the optimal weight matrix are as follows:

[0039] Calculate the initial gradient: z0 = W0V-a

[0040] Set search direction: g0 = -z0

[0041] Along k Direction calculates the step size along the search direction:

[0042] Update weight vector: W k+1 =W k +α k g k

[0043] Compute the new gradient:

[0044] Compute the conjugate direction update:

[0045] Where i is the training sample, k is the number of iterations, z k is the gradient, z0 is the initial gradient, z k+1 is the new gradient, g k is the search direction, g0 is the initial search direction, g k+1 is the new search direction, α k is the step length along the search direction, W k+1 is the updated weight.

[0046] When k<x+y, the optimal weight matrix W is obtained.

[0047] Furthermore, the V-TST algorithm in S5.3 performs a fine search and optimization process for the learning parameters in the RVFLNN algorithm as follows:

[0048] V-TST utilizes multiple encoder-decoder blocks, multi-head attention mechanism and position encoding PE to process the time series dynamics composed of state information. V-TST takes an input tensor [X t-W ,X t-W+1 ,…X t ], and produces a predicted point and an output tensor [y t+1 ,y t+2 ,…y t+H ], after using PE, the input tensor is processed by the scaled dot product attention mechanism, the dot product between the query and each key is calculated, and the dot product is converted into a probability distribution using the softmax function to determine the importance of each value, and finally the weighted sum is used to get the output, as shown in the following formula:

[0049]

[0050] Among them, A P,n is the attention value of head n in encoder block P, Q, K, V represent the query matrix, key matrix and value matrix respectively, D k is the dimension of the key vector.

[0051] In V-TST, N encoder blocks are stacked together, using the self-attention mechanism, and N decoding blocks use the cross-attention mechanism. The entire encoder-decoder device is connected to the final linear layer that predicts the output of the system. In V-TST with n attention heads, each attention head calculates the attention score of each encoder and decoder block, which is then combined using weights cj and the input / output relationship is processed by RVFLNN. The FFN in V-TST consists of a linear neuron network, while the FFN is replaced by the RVFLNN layer. The RVFLNN network puts the hidden layer into the input layer, performs a nonlinear transformation on the input vector, and obtains the optimal weight matrix of the output layer.

[0052] Furthermore, the present invention also provides a vehicle starting lithium battery heating time prediction system, the system comprising a heating module, a collection module, a main control module, a wireless communication module and a remote control module;

[0053] The heating module includes a vehicle lithium battery pack and a heating device, which is used to receive the heating time instruction issued by the main control module and take heating measures.

[0054] The acquisition module is used to collect the current remaining power of the automotive lithium battery pack, real-time battery temperature T0, heating power, battery cell initial temperature, ambient temperature, battery voltage and current, and current battery SOC value.

[0055] The main control module loads the battery heating time prediction model, receives the heating instruction from the wireless communication module, predicts the shortest heating time by receiving the data obtained by the acquisition module, and sends a heating time instruction to the heating module.

[0056] The wireless communication module and the remote control module are connected through the Internet. The remote control module integrates a linear regression model to determine whether the user is traveling based on the user's vehicle travel history data and the current remaining power. If the user needs to travel, the preheating time point is set based on the user's travel time point, and a heating instruction is generated and transmitted to the wireless communication module.

[0057] Beneficial effects:

[0058] 1. The present invention uses a linear regression prediction model to take the user's historical travel behavior as the original data to predict the user's travel time, so that the starting battery can automatically start the preheating system within a set range, assist the user in traveling, and avoid battery damage caused by the user's forgetfulness.

[0059] 2. The present invention constructs a battery heating time prediction model, which can optimize the scheduling according to many parameters such as current power, temperature, travel time, energy conversion efficiency, etc., and predict the optimal heating time data, so as to maintain the battery in a healthy state for a long time, further extend the battery life and improve the battery energy utilization efficiency.

[0060] 3. The present invention adopts a modular design of the battery adaptive heating system. While maintaining the travel prediction, heating time prediction, low temperature and low power warning, and battery heating functions, it also ensures the maintainability and scalability of the system, and has a faster speed in subsequent fault repairs and new function addition iterations, further reducing the possibility of delaying the owner's daily travel. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a schematic diagram of the specific process of the present invention;

[0062] Figure 2 It is a schematic diagram of the system structure of the present invention;

[0063] Figure 3 This is a schematic diagram of the DWT algorithm processing of the present invention;

[0064] Figure 4 This is a schematic diagram of data processing of the RVFLNN algorithm of the present invention; DETAILED DESCRIPTION

[0065] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. The following embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent or modification made according to the spirit of the present invention should be included in the protection scope of the present invention.

[0066] like Figure 1As shown, this embodiment provides a method for predicting the heating time of a vehicle starting lithium battery, and the specific steps of the method are as follows:

[0067] S1 collects the user's vehicle travel history data and current remaining power to form a user's historical travel data set;

[0068] S2 inputs the user's historical travel data set into the linear regression model, uses the least squares method to solve the regression coefficient, and predicts whether the user will travel today. If the user does not travel, it returns to S1, and if the user travels, it enters S3;

[0069] S3 predicts the user's travel time point according to the user's habits, sets the predicted time point based on the user's travel time point, and executes S4 when the predicted time point is reached, otherwise repeats S3;

[0070] S4 collects the current battery temperature T0 in real time, and determines whether the battery temperature T0 is in the optimal operating temperature range [T1, T2] of the lithium battery. If yes, S4 is repeatedly executed;

[0071] If the battery temperature T0 is not within the optimal operating temperature range [T1, T2] of the lithium battery, execute S5;

[0072] S5 builds a battery heating time prediction model to calculate the time required for the battery cell temperature T0 to heat to the optimal operating temperature of the lithium battery, and reserves time in advance to complete the battery's self-heating and wait for the user to travel;

[0073] The process of building a battery heating time prediction model is as follows:

[0074] S5.1: Optimize and construct the objective function with the shortest heating time as the goal;

[0075] The objective function is as follows:

[0076]

[0077] Among them, t heat Heating time for the battery; E total is the total heating energy during the battery heating period; η is the efficiency factor, which indicates the energy conversion efficiency during the battery heating process, usually 0≤η≤1, δ η is the efficiency error correction term, P load is the load power during battery heating, I(t) is the battery heating current, δ t0 (T) is the correction term for ambient temperature change, V OC (SOC(t)) is the open circuit voltage of the battery at the current SOC, and R is the internal resistance of the battery.

[0078] The calculation formula for the load power during Pload battery heating is as follows:

[0079]

[0080] Where T b is the temperature to which heating is required, T a is the initial battery temperature, R load is the battery thermal resistance.

[0081] S5.2: Based on the RVFLNN algorithm, the discrete wavelet transform DWT algorithm is used to preprocess and train the input data set, which includes: heating power, battery cell initial temperature, ambient temperature, battery voltage and current, and current battery SOC value;

[0082] like Figure 3 As shown, the DWT data set processing and calculation method is as follows:

[0083] The input data x[n] is decomposed through a pair of low-pass h[n] and high-pass g[n] filters as follows:

[0084]

[0085] Where k is the index after downsampling, A1[k] and D1[k] are the approximate coefficient and detail coefficient of the first layer decomposition, respectively.

[0086] The low-pass filter output A1[k] is recursively decomposed. Each layer uses the approximate coefficient A1[k] of the previous layer as the new input data, and passes through the low-pass and high-pass filters again to obtain the approximate coefficients and detail coefficients of the next layer. The above process is repeated until the required number of decomposition layers is reached, as shown in the following formula:

[0087]

[0088] Here, i is the number of decomposition levels.

[0089] like Figure 4 As shown, the optimized RVFLNN algorithm processes the data set as follows:

[0090] The RVFLNN algorithm includes the original input data x = [x1, x2, ···, x n ] and y enhancement nodes, passing the original input data through a randomly generated weight matrix W h =[W1,W2,···,W x ,W x+1 ,···,W x+y ] and the bias vector β are transformed nonlinearly to obtain the output of the hidden layer, as shown in the following formula:

[0091]

[0092] Among them, ξ is the activation function, N is the number of input samples, I Nis an N×1 vector of all ones, used to expand the bias vector β to match the dimension of N, represents element-wise multiplication;

[0093] Use the transformed data together with the original input data as the input layer of the new Directly connected to the output layer, the output of the output layer is obtained as follows:

[0094] ο=Wv T

[0095] Among them, ο is the output matrix, which contains the predicted outputs of all training samples.

[0096] Conjugate gradient training is used to iteratively update the weight matrix of RVFLNN, and the initial weight vector W0 is selected, and the target value vector a=[a1,a2,···,a i ], the input vectors form the input matrix V = [v1,v2,···,v i ], the steps to obtain the optimal weight matrix are as follows:

[0097] Calculate the initial gradient: z0 = W0V-a

[0098] Set search direction: g0 = -z0

[0099] Along k Direction calculates the step size along the search direction:

[0100] Update weight vector: W k+1 =W k +α k g k

[0101] Compute the new gradient:

[0102] Compute the conjugate direction update:

[0103] Where i is the training sample, k is the number of iterations, z k is the gradient, z0 is the initial gradient, z k+1 is the new gradient, g k is the search direction, g0 is the initial search direction, g k+1 is the new search direction, α k is the step length along the search direction, W k+1 is the updated weight.

[0104] When k<x+y, the optimal weight matrix W is obtained.

[0105] S5.3 uses the V-TST algorithm to perform a fine search and optimization on the learning parameters in the RVFLNN algorithm to further reduce the prediction error;

[0106] The V-TST algorithm performs a detailed search and optimization process for the learning parameters in the RVFLNN algorithm as follows:

[0107] V-TST utilizes multiple encoder-decoder blocks, multi-head attention, and position encoding PE to process the time series dynamics composed of state information. V-TST takes an input tensor [X t-W ,X t-W+1 ,…X t ], and produces a predicted point and an output tensor [y t+1 ,y t+2 ,…y t+H ]. After using PE, the input tensor is processed by the scaled dot product attention mechanism, the dot product between the query and each key is calculated, and the dot product is converted into a probability distribution using the softmax function to determine the importance of each value, and finally the weighted sum is used to obtain the output, as shown below:

[0108]

[0109] Among them, A P,n is the attention value of head n in encoder block P, Q, K, V represent the query matrix, key matrix and value matrix respectively, D k is the dimension of the key vector.

[0110] In V-TST, N encoder blocks are stacked together and use self-attention, and N decoder blocks use cross-attention. The entire encoder-decoder device is connected to a final linear layer that predicts the output of the system. In V-TST with n attention heads, each attention head calculates the attention score for each encoder and decoder block, and then uses the weight c j Combine them, weight c j It can be obtained through model training or preset by itself, and the input / output relationship is processed by RVFLNN. The FFN in V-TST is composed of a linear neural network, and the FFN is replaced by the RVFLNN layer. The RVFLNN network puts the hidden layer into the input layer, performs a nonlinear transformation on the input vector, and obtains the optimal weight matrix of the output layer.

[0111] In addition, this embodiment also provides a vehicle starting lithium battery heating time prediction system, such as Figure 2 As shown, the system is used to implement all the steps of the above-mentioned vehicle starting lithium battery heating time prediction method, and includes: a heating module, a collection module, a main control module, a wireless communication module and a remote control module.

[0112] The heating module includes a vehicle lithium battery pack and a heating device, which is used to receive the heating time instruction issued by the main control module and take heating measures;

[0113] The acquisition module is used to collect the current remaining power of the automotive lithium battery pack, real-time battery temperature T0, heating power, battery cell initial temperature, ambient temperature, battery voltage and current, and current battery SOC value;

[0114] The main control module is loaded with a battery heating time prediction model, receives a heating instruction from the wireless communication module, predicts the shortest heating time by receiving data obtained by the acquisition module, and issues a heating time instruction to the heating module;

[0115] The wireless communication module and the remote control module are connected through the Internet. The remote control module integrates a linear regression model to determine whether the user is traveling based on the user's vehicle travel history data and the current remaining power. If the user needs to travel, the preheating time point is set based on the user's travel time point, and a heating instruction is generated and transmitted to the wireless communication module.

[0116] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent transformation or modification made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting the heating time of a vehicle starting lithium battery, characterized in that: The method comprises the following steps: S1 collects the user's vehicle travel history data and current remaining power to form a user's historical travel data set; S2 inputs the user's historical travel data set into the linear regression model, uses the least squares method to solve the regression coefficient, and predicts whether the user will travel today. If the user does not travel, it returns to S1, and if the user travels, it enters S3; S3 predicts the user's travel time point according to the user's habits, sets the predicted time point based on the user's travel time point, and executes S4 when the predicted time point is reached, otherwise repeats S3; S4 collects the current battery temperature T0 in real time, and determines whether the battery temperature T0 is in the optimal operating temperature range [T1, T2] of the lithium battery. If yes, S4 is repeatedly executed; If the battery temperature T0 is not within the optimal operating temperature range [T1, T2] of the lithium battery, execute S5; S5 builds a battery heating time prediction model to calculate the time required for the battery cell temperature T0 to heat to the optimal operating temperature of the lithium battery, and reserves time in advance to complete the battery's self-heating and wait for the user to travel; The process of building a battery heating time prediction model is as follows: S5.1: Optimize and construct the objective function with the shortest heating time as the goal; S5.2: Based on the RVFLNN algorithm, the discrete wavelet transform DWT algorithm is used to preprocess and train the input data set, which includes: heating power, battery cell initial temperature, ambient temperature, battery voltage and current, and current battery SOC value; S5.3 uses the V-TST algorithm to perform fine search and optimization on the learning parameters in the RVFLNN algorithm to further reduce the prediction error.

2. The method for predicting the heating time of a vehicle starting lithium battery according to claim 1, characterized in that: The objective function described in S5.1 is constructed with the goal of minimizing the heating time. The objective function is as follows: Among them, t heat Heating time for the battery; E total is the total heating energy during the battery heating period; η is the efficiency factor, which indicates the energy conversion efficiency during the battery heating process, usually 0≤η≤1, δ η is the efficiency error correction term, P load is the load power during battery heating, I(t) is the battery heating current, is the correction term for ambient temperature change, V OC (SOC(t)) is the open circuit voltage of the battery at the current SOC, and R is the internal resistance of the battery.

3. The method for predicting the heating time of a vehicle starting lithium battery according to claim 2, characterized in that: The P load The load power calculation formula during battery heating is as follows: Where T b is the temperature to which heating is required, T a is the initial battery temperature, R load is the battery thermal resistance.

4. The method for predicting the heating time of a vehicle starting lithium battery according to claim 1, characterized in that: The DWT data set processing and calculation method in S5.2 is as follows: The input data x[n] is decomposed through a pair of low-pass h[n] and high-pass g[n] filters as follows: Where k is the index after downsampling, A1[k] and D1[k] are the approximate coefficient and detail coefficient of the first layer decomposition respectively; The low-pass filter output A1[k] is recursively decomposed. Each layer uses the approximate coefficient A1[k] of the previous layer as the new input data, and passes through the low-pass and high-pass filters again to obtain the approximate coefficients and detail coefficients of the next layer. The above process is repeated until the required number of decomposition layers is reached, as shown in the following formula: Here, i is the number of decomposition levels.

5. The method for predicting the heating time of a vehicle starting lithium battery according to claim 1 or 4, characterized in that: The optimized RVFLNN algorithm in S5.2 processes the data set as follows: The RVFLNN algorithm includes the original input data x = [x1, x2, ···, x n ] and y enhancement nodes, passing the original input data through a randomly generated weight matrix And the bias vector β is transformed nonlinearly to obtain the output of the hidden layer, as follows: Among them, ξ is the activation function, N is the number of input samples, I N is an N×1 vector of all ones, used to expand the bias vector β to match the dimension of N, represents element-wise multiplication; Use the transformed data together with the original input data as the input layer of the new Directly connected to the output layer, the output of the output layer is obtained as follows: o=Wv T Among them, ο is the output matrix, which contains the predicted output of all training samples; Conjugate gradient training is used to iteratively update the weight matrix of RVFLNN, and the initial weight vector W0 is selected, and the target value vector a=[a1,a2,···,a i ], the input vectors form the input matrix V = [v1,v2,···,v i ], the steps to obtain the optimal weight matrix are as follows: Calculate the initial gradient: z0 = W0V-a Set search direction: g0 = -z0 Along k Direction calculates the step size along the search direction: Update weight vector: W k+1 =W k +α k g k Compute the new gradient: Compute the conjugate direction update: Where i is the training sample, k is the number of iterations, and z k is the gradient, z0 is the initial gradient, z k+1 is the new gradient, g k is the search direction, g0 is the initial search direction, g k+1 is the new search direction, α k is the step length along the search direction, W k+1 is the updated weight; When k<x+y, the optimal weight matrix W is obtained.

6. The method for predicting the heating time of a vehicle starting lithium battery according to claim 1, characterized in that: The V-TST algorithm in S5.3 performs a detailed search and optimization process for the learning parameters in the RVFLNN algorithm as follows: V-TST utilizes multiple encoder-decoder blocks, multi-head attention mechanism and position encoding PE to process the time series dynamics composed of state information. V-TST takes an input tensor [X t-W ,X t-W+1 ,…X t ], and produces a predicted point and an output tensor [y t+1 ,y t+2 ,…y t+H ], after using PE, the input tensor is processed by the scaled dot product attention mechanism, the dot product between the query and each key is calculated, and the dot product is converted into a probability distribution using the softmax function to determine the importance of each value, and finally the weighted sum is used to get the output, as shown in the following formula: Among them, A P,n is the attention value of head n in encoder block P, Q, K, V represent the query matrix, key matrix and value matrix respectively, D k is the dimension of the key vector; In V-TST, N encoder blocks are stacked together using the self-attention mechanism, and N decoder blocks use the cross-attention mechanism. The entire encoder-decoder device is connected to a final linear layer that predicts the output of the system. In V-TST with n attention heads, each attention head calculates the attention score for each encoder and decoder block, and then uses the weight c j Combine them and process the input / output relationship through RVFLNN. The FFN in V-TST consists of a linear neural network, while the FFN is replaced by the RVFLNN layer. The RVFLNN network puts the hidden layer into the input layer, performs a nonlinear transformation on the input vector, and obtains the optimal weight matrix of the output layer.

7. The vehicle starting lithium battery heating time prediction system according to any one of claims 1 to 6, characterized in that: The method performs any of the above steps through a modular system, the system comprising: a heating module, a collection module, a main control module, a wireless communication module and a remote control module; The heating module includes a vehicle lithium battery pack and a heating device, which is used to receive the heating time instruction issued by the main control module and take heating measures; The acquisition module is used to collect the current remaining power of the automotive lithium battery pack, real-time battery temperature T0, heating power, battery cell initial temperature, ambient temperature, battery voltage and current, and current battery SOC value; The main control module loads the battery heating time prediction model, receives the heating instruction from the wireless communication module, predicts the shortest heating time by receiving the data obtained by the acquisition module, and issues a heating time instruction to the heating module; The wireless communication module and the remote control module are connected through the Internet. The remote control module integrates a linear regression model to determine whether the user is traveling based on the user's vehicle travel history data and the current remaining power. If the user needs to travel, the preheating time point is set based on the user's travel time point, and a heating instruction is generated and transmitted to the wireless communication module.

Images