Temperature rise prediction model training method, prediction method and control method
Through neural network model training and real-time data update, the problem of inaccurate temperature rise prediction of power lithium batteries is solved, accurate monitoring and control of battery temperature is achieved, and the safety and stability of the battery is improved.
Patent Information
- Application Number
- CN202410908685.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-08
- Publication Date
- 2024-11-01
AI Technical Summary
In the prior art, there is inaccurate prediction of temperature rise of power lithium batteries, which may lead to safety risks of battery overheating, especially the empirical formulas are difficult to adapt to complex situations of different battery structures and types.
Using neural network model training method, by obtaining the historical environmental data of the power battery, calculating historical temperature appreciation, adjusting model parameters until the loss value is lower than the threshold, obtaining a temperature rise prediction model, and combining real-time environmental data for updating and verifying, the Accurate prediction and control of battery temperature.
It improves the accuracy of battery temperature rise prediction, ensures real-time monitoring and control of battery temperature, improves battery safety and stability, and avoids the risk of overheating.
Smart Images

Figure CN118886540A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of power batteries, and in particular to a temperature rise prediction model training method, a prediction method and a control method. Background Art
[0002] Currently, new energy vehicles mostly use power lithium batteries as energy storage devices. The performance of power lithium batteries is greatly affected by temperature, and if the temperature of the power battery rises too quickly, there will be safety hazards. Therefore, it is necessary to predict the temperature of the power battery so that timely feedback and correction can be made when problems arise.
[0003] In the prior art, the temperature rise of the battery is calculated using an empirical formula. However, the empirical formula has certain limitations. It is difficult to accurately calculate the temperature rise of the battery for different battery structures and types. Once the temperature rise is too high and is not accurately calculated using the empirical formula, the battery current may not be controlled in time, resulting in battery overheating and a safety hazard. Summary of the invention
[0004] In view of the above-mentioned shortcomings of the prior art, an object of the present invention is to propose a temperature rise prediction model training method, a prediction method and a control method to improve the accuracy of battery temperature rise.
[0005] To achieve the above-mentioned and other related purposes, the present invention proposes a temperature rise prediction model training method, which is applied to power batteries and includes the following steps:
[0006] Acquiring historical environmental data of the power battery;
[0007] Calculating historical temperature rise values based on the historical environmental data;
[0008] Obtaining a predicted temperature rise value based on the historical environmental data and the neural network model;
[0009] Calculating a loss value based on the predicted temperature rise value and the historical temperature rise value corresponding to the predicted temperature rise value;
[0010] The model parameters of the neural network model are adjusted by the loss value, and the above steps are repeated to obtain the predicted temperature rise value and calculate the loss value until the training is completed when the loss value is lower than a preset threshold, thereby obtaining a temperature rise prediction model.
[0011] In a specific embodiment of the present invention, obtaining the predicted temperature rise value based on the historical environmental data and the neural network model includes:
[0012] Arrange the historical environmental data in chronological order;
[0013] Performing sliding window processing on the historical environmental data;
[0014] Extracting data from each sliding window to obtain multiple groups of sliding window data, each group of sliding window data constituting a group of preprocessed data;
[0015] A predicted temperature rise value is obtained based on the preprocessed data and the neural network model.
[0016] In a specific embodiment of the present invention, after the step of obtaining the temperature rise prediction model, the following steps are included:
[0017] Acquiring the current environmental data of the power battery in real time;
[0018] The historical environment data is updated according to the current environment data.
[0019] In a specific embodiment of the present invention, updating the historical environment data according to the current environment data includes:
[0020] Clear the historical environmental data corresponding to the first N moments in the data storage area in chronological order;
[0021] The current environment data corresponding to the latest N moments are added to the data storage area as the historical environment data.
[0022] In a specific embodiment of the present invention, the step of obtaining the temperature rise prediction model includes:
[0023] The temperature rise prediction model is obtained by training according to the updated historical environmental data.
[0024] In a specific embodiment of the present invention, after the step of obtaining the temperature rise prediction model, the following steps are included:
[0025] The temperature rise prediction model is verified based on the historical environmental data and the historical temperature rise values.
[0026] In a specific embodiment of the present invention, the verifying the temperature rise prediction model according to the historical environmental data and the historical temperature rise value includes:
[0027] Inputting the historical environmental data into the temperature rise prediction model to obtain a predicted temperature rise value;
[0028] Calculating the difference between the predicted temperature rise value and the corresponding historical temperature rise value;
[0029] The temperature rise prediction model is verified according to the difference.
[0030] In a specific embodiment of the present invention, the verifying the temperature rise prediction model according to the difference includes:
[0031] The determination coefficient and / or the root mean square error of the temperature rise prediction model are calculated according to the difference. When the determination coefficient and / or the root mean square error are within a preset error threshold, the current temperature rise prediction model is verified to be available, otherwise it is verified to be unavailable.
[0032] In a specific embodiment of the present invention, the calculation of the coefficient of determination and / or the root mean square error of the temperature rise prediction model according to the difference includes:
[0033] assigning a weight to each of the differences;
[0034] The coefficient of determination and / or root mean square error of the temperature rise prediction model is calculated based on the weighted difference.
[0035] In a specific embodiment of the present invention, the step of assigning a weight to each difference includes:
[0036] When the historical temperature corresponding to the historical temperature rise value is greater than a preset first temperature threshold, the difference is multiplied by a first weight, and the first weight is greater than 1;
[0037] When the historical temperature corresponding to the historical temperature rise value is less than or equal to a preset second temperature threshold, the difference is multiplied by a second weight, the second weight is less than 1, and the second temperature threshold is less than or equal to the first temperature threshold.
[0038] The present invention also proposes a battery temperature rise prediction method, comprising the following steps:
[0039] Get current environment information;
[0040] The current environmental information is input into the temperature rise prediction model as described above to obtain a predicted temperature rise value.
[0041] The present invention also provides a battery control method, comprising the following steps:
[0042] Obtaining a predicted temperature rise value by the battery temperature rise prediction method as described above;
[0043] When the predicted temperature rise value exceeds a preset temperature rise threshold, battery discharge is restricted or interrupted.
[0044] The present invention proposes a temperature rise prediction model training method, prediction method and control method. In the above scheme, the battery temperature is predicted by using the trained model after model training. Based on the analysis of historical data and related features, the prediction model can more accurately capture the law of battery temperature change. The battery temperature change can be monitored in real time, and the future temperature trend can be predicted in time, which is conducive to timely adjustment of the control strategy. Combining the prediction model and the control method, the battery temperature can be automatically controlled to improve the safety and stability of the battery. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for describing the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0046] Figure 1 It is a flow chart of a temperature rise prediction model training method in a specific embodiment of the present invention;
[0047] Figure 2 It is a schematic diagram of a convolutional neural network in a specific embodiment of the present invention;
[0048] Figure 3 This is a schematic diagram of the LSTM (Long Short-Term Memory Model) in a specific embodiment of the present invention;
[0049] Figure 4 A comparison diagram of the predicted temperature rise value and the historical temperature rise value during the model training process in a specific embodiment of the present invention;
[0050] Figure 5 A comparison diagram of the predicted temperature rise value and the historical temperature rise value of the temperature rise prediction model in a specific embodiment of the present invention;
[0051] Figure 6 A flow chart of obtaining a predicted temperature rise value in a temperature rise prediction model training method in a specific embodiment of the present invention;
[0052] Figure 7 It is a flow chart of the verification step in the temperature rise prediction model training method in a specific embodiment of the present invention;
[0053] Figure 8 A flowchart of verifying the temperature rise prediction model according to the difference in a specific embodiment of the present invention;
[0054] Fig. 9 It is a flow chart of a battery temperature rise prediction method in a specific embodiment of the present invention;
[0055] Fig.10 A flowchart of a battery control method in a specific embodiment of the present invention;
[0056] Fig.11 The present invention is a flowchart of the preprocessing steps of the temperature rise prediction model training method in a specific embodiment of the present invention. DETAILED DESCRIPTION
[0057] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.
[0058] It should be noted that the illustrations provided in the following embodiments are only used to illustrate the basic concept of the present invention in a schematic manner, and thus the illustrations only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.
[0059] The battery busbar will generate heat due to the influence of current during operation. Excessive temperature rise may cause overheating of the battery system, thus affecting its safety performance. By predicting the temperature rise of the BUSBAR (copper busbar / aluminum busbar), the creepage distance, electrical clearance and other parameters around the BUSBAR can be inferred, which directly affect the safety of the battery system. When designing the BDU (battery management unit), spine (connection between battery modules) and cell bridge connection, the design can be optimized based on the predicted temperature rise, thereby increasing the safety attributes of the entire battery system.
[0060] Predicting the temperature rise of BUSBAR helps to find the best BUSBAR structural design. By optimizing the shape, material, cross-sectional area and other parameters of BUSBAR, its resistance can be reduced, the temperature rise can be reduced, and the loss of copper / aluminum can be effectively reduced. In addition, optimizing the BUSBAR structure can also improve the thermal conductivity of the battery system, help dissipate heat, and further reduce the temperature rise. These optimization measures can not only save a lot of costs, but also improve resource utilization and enhance the performance and reliability of the entire battery system.
[0061] The empirical formulas in the prior art are usually derived based on experience and simplified models, and may not accurately capture the various influencing factors and interactions in complex systems, so their calculation results may deviate from the actual situation. Empirical formulas are usually derived and verified for specific conditions or specific types of systems, so their scope of application is limited. When the operating conditions or structural parameters of the system change, the empirical formula may no longer be applicable. The temperature rise in complex systems is often affected by multiple factors, and there may be nonlinear interactions between these factors. Empirical formulas often only consider linear relationships and cannot accurately describe nonlinear effects.
[0062] Therefore, Figure 1As shown, in order to solve the above problems, the present invention proposes a temperature rise prediction model training method, which is applied to power batteries. The temperature rise prediction model training method in the present invention can not only be used to train the temperature rise prediction model for battery BUSBAR (copper busbar / aluminum busbar), but also can be used to train models of other temperature rise environments, such as engine temperature rise, motor temperature rise, etc.
[0063] The steps include:
[0064] S1. Obtain the historical environmental data of the power battery, which includes the historical temperature, historical ambient temperature, historical contact resistance and historical current of the power battery. Collect BUSBAR (copper busbar / aluminum busbar) temperature data, ambient temperature, contact resistance and current at multiple points, collect them every 1s, store these data and then use them as historical environmental data.
[0065] S2. Calculate the historical temperature rise value according to the historical environmental data. The historical temperature rise value refers to the difference in temperature rise, which can be obtained based on the starting temperature difference within a certain time period. For example, at time t, the historical environmental temperature in the historical environmental data is 20°C, and at time t+Δt, the historical environmental temperature is 25°C, then the historical temperature rise value at time t+Δt is = 25°C-20°C = 5°C. That is, the historical temperature rise value after Δt time in the future corresponding to the historical environmental data at time t is 5°C. Where t can be any time, and Δt is a preset time interval of 1s.
[0066] S3, obtaining a predicted temperature rise value based on the historical environmental data and the neural network model, that is, inputting the historical environmental data at time t into the neural network model to obtain the predicted temperature rise value at time t+Δt.
[0067] like Figure 6 As shown, step S3 includes:
[0068] S31, preprocessing the historical environmental data to obtain preprocessed data. The historical environmental data can be preprocessed using a pandas tool. Pandas is a Python library that provides high-performance, easy-to-use data structures and data analysis tools.
[0069] like Fig.11 As shown, step S31 includes:
[0070] S311, arranging the historical environmental data in chronological order. If the historical environmental data itself is acquired in chronological order, this step can be omitted. The historical environmental data contains multiple types of data, and the historical environmental data can be divided into a data matrix, where each row in the matrix represents a type of data, such as ambient temperature, contact resistance, current, etc., and each column in the matrix represents data at a certain moment.
[0071] S312, remove the noise in the historical environmental data. Once the ambient temperature, contact resistance, current and other data suddenly change at a certain moment, it can be judged as noise, and then all the data in the column at the corresponding moment can be cleared. The location where the data occurs can be judged manually or by an algorithm. Manual judgment can be made through visualization tools (such as charts) to observe the trends and mutations of the data and determine which data points may be noise. Using algorithms to automatically detect and remove noise data can use filter techniques (such as moving average, median filter, etc.) to smooth the data and identify mutation points. When a data mutation is detected, it can be marked as noise and removed from the data set.
[0072] S313, the historical environmental data with noise removed is subjected to sliding window processing, wherein the width of the sliding window is w and the interval is s, and w is greater than s. Wherein w can be 100 and s can be 50. A sliding window refers to a fixed-size window that slides on a time series or a data series. The size of the window is usually determined by a parameter, called the window width. The sliding window starts from the starting point of the data series and slides backward in sequence. The step size of each slide is determined by another parameter, called the interval. Sliding window processing can smooth the data, reduce the impact of sudden changes and noise, and make the data more stable and predictable. By setting the width of the sliding window to be greater than the interval, it is ensured that there are overlapping parts between adjacent sliding windows, and the continuity of the data can be guaranteed. In order to ensure the continuity of the data and make full use of the information of the data, it is usually set that there are overlapping parts between adjacent sliding windows. The overlapping part means that the starting position of the sliding window is not strictly incremented according to the interval, but partially overlapped, which can be a part or the entire window overlap. For example: for a sliding window with a width of 100 and an interval of 50, there are 50 overlapping data between adjacent sliding windows.
[0073] S314, extract the data in each sliding window to obtain multiple groups of sliding window data, each group of sliding window data constitutes a group of preprocessed data. Traverse each sliding window, extract sliding window data from the historical environment data, and the data in each sliding window can constitute a group of preprocessed data. The historical environment data is processed in segments according to the sliding window, and the original long sequence of historical environment data can be decomposed into multiple shorter sequences, each short sequence corresponding to the data in a sliding window. This helps to reduce the complexity of data processing and makes the analysis of each sliding window more focused and effective. The data in each sliding window can be regarded as a local area, so the features in the area can be extracted more attentively. This can reduce the complex relationship across the entire data sequence, making feature extraction more targeted and effective.
[0074] The purpose of preprocessing historical environmental data is to prepare the data for subsequent analysis, modeling, or visualization.
[0075] S32, normalize the pre-processed data to obtain training data. StandardScaler (standardization), RobustScaler (robust standardization), MaxAbsScaler (absolute maximum standardization), Normalization (regularization), etc. can be used to pre-process the data, preferably MinMaxScaler (absolute maximum standardization) pre-processed data for normalization. Normalization can eliminate the dimensional differences between different features, so that each feature is compared on the same scale. This helps to avoid the dominant role of certain features in model training, thereby improving the stability and performance of the model. Normalization can accelerate the convergence speed of the model, especially for gradient-based optimization algorithms (such as gradient descent methods), which can make it converge to the optimal solution faster.
[0076] At the same time, the training data can be divided into a training set and a test set, for example, the ratio of the training set to the test set is 0.8:0.2. Therefore, by dividing the data set, the training set can be used when training the model, and the test set can be used to evaluate the performance and generalization ability of the model after the model training is completed.
[0077] S33, inputting the training data into the neural network model to obtain the predicted temperature rise value. Figure 2 Shown is the schematic diagram of the convolutional neural network.
[0078] The neural network model can be a Transformer model, a GRU model, a CNNs model, a CNN-RNN model, a CNN LSTM model, etc., preferably a CNN LSTM model, that is, a CNN-LSTM model is built using conv1d (one-dimensional convolutional neural network) and LSTM (long short-term memory network), such as Figure 2 The CNN LSTM model shown in the figure consists of 1 convolutional layer, 1 classification layer, 5 LSTM layers, and 1 linear layer. The LSTM layer can effectively capture the time series information in the sequence data, which is very important for tasks involving time series such as temperature rise prediction. By combining CNN with LSTM, the convolutional layer can be used to extract local features of the data, and then these feature sequences are input into the LSTM layer to better process the time series information. CNN is suitable for processing multidimensional features, while LSTM is suitable for processing sequence data, such as time series data. Combining CNN with LSTM can process data with multidimensional features and time series properties, making the model more suitable for complex data structures. By introducing convolutional layers in the model, the number of parameters of the model can be effectively reduced, thereby reducing the risk of overfitting. In addition, the long short-term memory unit in LSTM helps the model learn long-term dependencies in the data, further improving the generalization ability of the model. Among them, the predicted temperature rise value is equivalent to the predicted value, and the historical temperature rise value is equivalent to the actual value.
[0079] You can also use LSTM, Transformer+LSTM, CNN-LSTM-Attention and other algorithms to predict temperature rise.
[0080] In step S3, a one-dimensional CNN can be used to extract time series features in the training data.
[0081] Scale the time series data to a similar range to ensure stability and convergence of model training. Divide the time series data into fixed-length windows or sequence segments as input to the CNN.
[0082] Determine the length of each time series window, which will become the dimension of the CNN model input. Design the size and number of convolution kernels, as well as the type and parameters of pooling layers to progressively extract and reduce features. Choose an activation function that is appropriate for the problem, such as ReLU. Optionally add regularization terms or batch normalization layers to improve the generalization ability and training speed of the model.
[0083] Select appropriate loss functions such as mean squared error (MSE) or categorical cross entropy loss, and optimizers such as Adam or SGD according to the specific task. Input the processed time series data into the CNN model for training, and optimize the model parameters through back propagation.
[0084] In the convolutional and pooling layers, the key features of the time series are extracted by gradually reducing the size and number of feature maps. A global average pooling or global maximum pooling layer is added after the last convolution layer to aggregate all values of each feature map into a single value for final prediction or classification.
[0085] The one-dimensional convolution structure is as follows Figure 2 Shown
[0086] The one-dimensional convolution formula corresponding to the convolution layer is:
[0087]
[0088] Where Hi is the input feature of the i-th layer; is a convolution operation, Wi and bi represent the weight and corresponding bias of the i-th convolution kernel, and f represents the activation function.
[0089] Use the validation set to evaluate the performance of the model and adjust hyperparameters such as convolution kernel size, number of layers, etc. Add regularization techniques as needed, or adjust the model structure to reduce the risk of overfitting.
[0090] like Figure 3The calculation order of the LSTM layer is forget gate, input gate, update memory unit, output gate and hidden state. Each gated unit has its own specific calculation method. Through the operation of these gates, LSTM can effectively process and memorize long-term dependencies in time series data.
[0091] Forget Gate:
[0092] The function of the forget gate is to decide which information in the memory cell of the previous time step should be discarded in order to update the memory cell in the current time step. Its calculation formula is:
[0093] ft=σ(Wf*[ht-1,Xt]+bf)
[0094] Input Gate:
[0095] The input gate determines how new memory information is added to the memory cell. It consists of two parts: a Sigmoid activation function to determine which values should be updated, and a Tanh activation function to create new candidate values. Its calculation formula is:
[0096] it=σ(Wi*[ht-1,Xt]+bi)
[0097] Update Cell:
[0098] The update of the memory cell is completed by decaying the old memory cell according to the value of the forget gate and adding the new candidate value. The update process is as follows:
[0099] Ct=tanh(Wc*[ht-1,Xt]+bc)
[0100] This step enables LSTM to maintain long-term dependencies and selectively forget or remember new information.
[0101] Output Gate:
[0102] The output gate determines which information in the hidden state will be passed to the next time step or output layer of the network. The calculation formula of the output gate is:
[0103] Ot=σ(WO*[ht-1,Xt]+bO)
[0104] The hidden state is calculated as:
[0105] ht=Ot*tanh(Ct)
[0106] Among them, ft, it, Ot represent three gated calculation results, Xt is the input data at time t (that is, the historical environmental data in this application, the ambient temperature, contact resistance, current, etc. in the historical environmental data at time t can all be used as a separate Xt), ht is the output at time t, Ct is the cell state at time t, Wf, Wi, WO, respectively represent the weights of the three gates, bf, bi, bo represent the corresponding biases, Wc and bc represent the weight and bias of the cell state.
[0107] Throughout the process, LSTM effectively learns and captures long-term dependencies in time series data by controlling the forget gate, input gate, and output gate, as well as the update of memory cells. It is suitable for various tasks that require processing time dynamics, such as temporal relationship analysis of temperature rise data.
[0108] You can also add a DropConnect, DropBlock, Zoneout, or Dropout layer after the LSTM layer in the CNN-LSTM model, preferably the Dropout layer. The Dropout layer helps reduce the interdependencies between neurons in the model, thereby reducing the risk of overfitting. By randomly dropping a portion of neurons during training, the Dropout layer can force the model to learn more robust and generalized feature representations. This can improve the generalization ability of the model and avoid overfitting on the training data.
[0109] S4. Calculate the loss value based on the predicted temperature rise value and the historical temperature rise value corresponding to the predicted temperature rise value. Input the predicted temperature rise value and the historical temperature rise value of the model into the loss function. The loss function will return a value representing the gap (loss) between the model prediction and the historical temperature rise value. The loss function can be calculated using Mean Squared Error (MSE), Cross-Entropy Loss or Log-Likelihood Loss.
[0110] For example, the mean squared error (MSE) is:
[0111] text{MSE}=\frac{1}{n}\sum_{i=1}^{n}(y_i-\hat{y}_i)^2
[0112] Among them, y_i is the actual target value, i.e. the historical temperature rise value, \hat{y}_i is the value predicted by the model, i.e. the predicted temperature rise value, and n is the number of samples.
[0113] like Figure 4The figure shows the comparison between the predicted temperature rise value and the historical temperature rise value during the training process, where line 1 is the historical temperature rise value and line 2 is the predicted temperature rise value.
[0114] S5. Adjust the model parameters of the neural network model according to the loss value, repeat the above steps to obtain the predicted temperature rise value and calculate the loss value, until the training is completed when the loss value is lower than the preset threshold, and obtain the temperature rise prediction model.
[0115] The specific steps are:
[0116] Initialize the model's weights and biases. Usually a random initialization method is used to start the training process.
[0117] Choose an appropriate loss function, such as mean squared error (MSE) for regression problems, cross entropy loss function for classification problems, etc. The choice of loss function depends on the specific task and the nature of the output.
[0118] Choose a suitable optimization algorithm to update the model parameters to reduce the loss function. Common optimizers include stochastic gradient descent (SGD), Adam, RMSprop, etc. Each optimizer has its own hyperparameters to set, such as learning rate.
[0119] Calculate the predicted output of the model using the current parameters. For neural networks, this usually includes the complete calculation process from input to output. For example, for LSTM, it is to calculate the hidden state and output through the gated unit.
[0120] The predicted temperature rise values and the historical temperature rise values are used to calculate the value of the loss function. The loss value is a measure of the difference between the model prediction and the historical temperature rise values.
[0121] Perform backpropagation, which is the process of updating model parameters by calculating the gradient of the loss function with respect to each parameter. This step uses the chain rule to propagate back from the loss function to each parameter to determine the degree of influence of each parameter on the loss. These parameters can be weights, biases, previous layer outputs, etc.
[0122] The specific steps of back propagation are:
[0123] Propagate the error forward from the output layer (\delta^{(l)}) to the hidden layer (l):
[0124] \delta^{(l)}=((W^{(l+1)})^T\delta^{(l+1)})\odot f'(z^{(l)})
[0125] where (\delta^{(l)}) is the error term of the hidden layer (l), which represents the error contribution of each neuron in this layer.
[0126] (W^{(l+1)}) is the weight matrix from the (l)th layer to the (l+1)th layer.
[0127] ((W^{(l+1)})^T) is the transpose of (W^{(l+1)}) and is used to back-propagate the error term (\delta^{(l+1)}) of the (l+1)th layer to the (l)th layer.
[0128] (\delta^{(l+1)}) is the error term at the (l+1)th layer.
[0129] The (((W^{(l+1)})^T\delta^{(l+1)})) part means back-propagating the error term of the (l+1)th layer to the (l)th layer, and the result is a vector.
[0130] (f'(z^{(l)})) is the derivative of the activation function (f) of the (l)th layer at the input (z^{(l)}). It measures the slope of the activation function output and is used to adjust the propagation of errors.
[0131] (\odot) represents the element-by-element product (Hadamard product), which multiplies the above two parts element-by-element to obtain the final error term (\delta^{(l)}).
[0132] Use an optimizer (such as Adam, SGD, etc.) to update each parameter according to the gradient calculated by the loss function.
[0133] Taking Adam as an example, the steps of adjusting the parameters of the neural network model according to the loss value are:
[0134] Use the loss function to calculate the gradient of the model parameters with respect to the loss function:
[0135] (g_t=\nabla_{\theta}\text{Loss}(\theta_{t-1})),
[0136] Where (\nabla_{\theta}) represents the gradient of the parameter (\theta). The gradient is a vector whose components are the partial derivatives of the loss function with respect to the corresponding parameters. (\text{Loss}(\theta_{t-1})) represents the loss function (\text{Loss}), which is a function used to measure the difference between the model prediction and the true value. (\theta_{t-1}) represents the model parameters at the (t-1)th iteration.
[0137] (g_t) is a gradient vector, which represents the gradient of the loss function (\text{Loss}) with respect to the parameter (\theta) at the current iteration step (t). That is, (g_t) is a vector whose (i)th element ((g_t)_i) represents the partial derivative of the loss function with respect to (\theta_i).
[0138] Compute the exponential moving average of the gradient:
[0139] (m_t=\beta_1\cdot m_{t-1}+(1-\beta_1)\cdot g_t),
[0140] Where (m_t) is the first moment estimate of momentum. In the Adam algorithm, it is used to store the exponentially weighted average of past gradients in order to maintain momentum when the parameters are updated.
[0141] (\beta_1) is the exponential decay rate of momentum, usually set to 0.9. It controls the contribution of past gradients to the current momentum, that is, the influence of past gradients on the current update.
[0142] (m_{t-1}) is the first-order moment estimate of momentum at the previous iteration.
[0143] (g_t) is the gradient vector of the loss function (\text{Loss}) with respect to the parameter (\theta) at the current iteration step (t).
[0144] Compute the exponential moving average of the squared gradients:
[0145] (v_t=\beta_2\cdot v_{t-1}+(1-\beta_2)\cdot g_t^2),
[0146] Where (v_t) is the second-order moment estimate in the Adam algorithm, which is used to store the exponentially weighted average of the square of the gradient.
[0147] (\beta_2) is the exponential decay rate used to estimate the second-order moment, and its value is usually 0.999.
[0148] (v_{t-1}) is the second-order moment estimate at the previous iteration.
[0149] (g_t^2) is the square of the gradient of the loss function (\text{Loss}) with respect to the parameter (\theta) at the current iteration step (t). That is, ((g_t)_i^2) represents the square of the (i)th element in the gradient vector (g_t).
[0150] Update the model parameters using first-order and second-order moment estimates: (\theta_t=\theta_{t-1}-\frac{\eta}{\sqrt{\hat{v}_t}+\epsilon}\cdot\hat{m}_t),
[0151] where (\theta_t) is the updated model parameter, representing the new parameter value after iteration step (t).
[0152] (\theta_{t-1}) is the model parameter at the previous iteration.
[0153] (\eta): This is the learning rate, which is used to control the step size or rate of each parameter update, which can be 10 -2 .
[0154] (\sqrt{\hat{v}_t}) is the square root of the second-order moment estimate (\hat{v}_t). (\hat{v}_t) is the exponentially weighted moving average of the squared gradient, which is used to adaptively adjust the learning rate of each parameter.
[0155] (\epsilon) is a small constant (usually (10^{-8})) used to ensure that the denominator is not zero to enhance numerical stability.
[0156] (\hat{m}_t) is a correction to the first-order moment estimate (\hat{m}_t). (\hat{m}_t) is an exponentially weighted moving average of the first-order moment of the gradient, used to maintain momentum.
[0157] The above steps are repeated, and each iteration calculates the loss, back-propagates the gradient, and updates the parameters according to the current parameters until the loss value is lower than the preset threshold or the maximum number of iterations is reached.
[0158] After each iteration or epoch, check whether the loss value reaches the preset threshold. If it reaches or is close enough, stop the training process.
[0159] Determine a preset loss value threshold as the condition for stopping training. This threshold can be determined based on the requirements of the task and the expectations of the performance indicators. Using the given training data set, start training the model. In each training iteration, the model calculates the loss value according to the loss function, and adjusts the model parameters according to the optimization algorithm (such as Adam) to reduce the loss value. During the training process, continue to monitor the loss value. You can use the validation set to evaluate the performance of the model and check whether the loss value reaches or is lower than the preset threshold. If the loss value is higher than the preset threshold, continue to train the model and adjust the parameters. You can try to adjust the learning rate, increase the number of training iterations, try different optimization algorithms, etc. to try to reduce the loss value. When the loss value is lower than the preset threshold, stop training. The parameters of the model have been adjusted to a better state at this time, and the loss value has reached the expected target.
[0160] The temperature rise prediction model is obtained by adjusting the neural network model as the trained temperature rise prediction model. Once the loss value of the model is reduced to below a preset threshold and the model has converged, the trained model can be regarded as the temperature rise prediction model.
[0161] S6. Acquire the current environmental data of the power battery in real time. The sensor data can be collected regularly through the vehicle electronic system or data acquisition unit. The sampling period can be set according to demand, usually in milliseconds, seconds or minutes. Shorter sampling periods can provide more frequent data updates, but will also increase the computing and communication burden of the system. The processed data can be transmitted to the vehicle's central processing unit or cloud server through the vehicle communication system for further processing and storage. Data transmission can be carried out through wired or wireless communication, such as CAN bus, Bluetooth, Wi-Fi or mobile network.
[0162] Specifically:
[0163] Determine the data sampling cycle based on actual needs and system design. The sampling cycle should fully meet the monitoring needs while avoiding frequent sampling that leads to resource waste or data overload.
[0164] Set a timer or trigger in the system to periodically trigger the data acquisition process. This may be achieved through software programming, such as using a timer interrupt or a timer task in an embedded system to perform data acquisition.
[0165] When the timer is triggered, the system starts reading the current environmental data from the connected sensors. This data may include information such as battery temperature, pressure, humidity, etc., depending on the type of sensors installed and the sensor location.
[0166] The sensor data read often needs to be processed and stored. Processing may include data calibration, unit conversion, or simple data cleaning to ensure data quality and usability. The processed data is usually stored in memory or written to a persistent storage device (such as a solid-state drive or flash memory) for subsequent analysis and use.
[0167] The processed data can be transmitted via the vehicle’s communication system to other locations, such as the vehicle bus, cloud storage, or on-board displays. In the cloud or on-board system, the data can be further analyzed and utilized for real-time monitoring of battery status, predictive maintenance, or other applications.
[0168] S7. Update the historical environment data according to the current environment data.
[0169] The specific steps are:
[0170] Collect environmental data at the current time point, such as current temperature, humidity, air pressure, etc.
[0171] Obtain previously recorded historical environmental data and historical temperature rise data. These data are usually stored in the form of time series, such as records in a database or logs in a file.
[0172] The historical environment data corresponding to the first N moments in the data storage area are cleared in chronological order.
[0173] The current environment data corresponding to the latest N moments are added to the data storage area as the historical environment data. N can be preset according to the actual application scenario, and can be specifically 1. Taking N=1 as an example, for example, there are 1-100 groups of historical environment data in the data storage area. When a group of the latest current environment data is obtained, the first group of historical environment data can be cleared, and the historical environment data of the 2nd to 100th groups can be retained. Then the obtained group of current environment data is used as the 101st group of historical environment data to obtain the same 100 groups of data.
[0174] Add the current environment data to the history. This is done by updating a database entry or appending to a log file.
[0175] Ensure that the updated historical environmental data and historical temperature rise data are saved in the system. This can be done by updating the database, files or other data storage mechanisms.
[0176] The temperature rise prediction model is obtained by training according to the updated historical environmental data. A new temperature rise prediction model is obtained by repeating the above-mentioned temperature rise prediction model training method according to the updated historical environmental data.
[0177] Since the state of the power battery will change with the use of the vehicle, the data of the use process can be updated in real time into the historical data as the data for training the model, so that the model can be trained in real time so that the temperature rise prediction model is closer to the current battery working condition. Through online learning, the model can continuously learn from the newly collected data and adjust and update according to the latest information to more accurately predict the temperature rise of the battery. This method helps the model better adapt to the changes in the power battery state and improve the accuracy and reliability of the prediction.
[0178] S8. Verify the temperature rise prediction model according to the historical environmental data and the historical temperature rise values.
[0179] like Figure 7 As shown, step S8 includes the following steps:
[0180] S81, input the historical environmental data into the temperature rise prediction model to obtain a predicted temperature rise value. The model processes the input historical environmental data and outputs a corresponding predicted temperature rise value. This predicted temperature rise value represents the battery temperature rise estimated by the model based on the historical environmental data.
[0181] S82, calculate the difference between the predicted temperature rise value and the corresponding historical temperature rise value. Obtain the historical temperature rise value data corresponding to the time point involved in the prediction from the historical records. Compare the predicted temperature rise value with the corresponding historical temperature rise value data and calculate the difference between them. This can be done by simply subtracting the historical temperature rise value from the predicted temperature rise value.
[0182] S83. Verify the temperature rise prediction model according to the difference. Use statistical indicators such as root mean square error (RMSE), mean absolute error (MAE) or mean absolute percentage error (MAPE) to evaluate the prediction accuracy of the model. Lower error values usually indicate better model prediction capabilities. In actual applications, monitor the prediction performance of the model in real time to check whether the model can maintain accuracy under different working conditions. Rolling prediction and sliding window techniques can be used to continuously update the model and promptly detect potential prediction deviations.
[0183] like Figure 8 As shown, step S83 includes:
[0184] S831, assigning a weight to each difference. At different temperatures, the temperature rise will have different effects on the safety performance of the battery, so a larger weight is set in the area with a greater impact on the safety performance of the battery, and a smaller weight is set in the area with a smaller impact on the safety performance of the battery.
[0185] When the historical temperature corresponding to the historical temperature rise value is greater than the preset first temperature threshold, the difference is multiplied by the first weight, and the first weight is greater than 1. When the historical temperature corresponding to the historical temperature rise value is less than or equal to the preset second temperature threshold, the difference is multiplied by the second weight, the second weight is less than 1, and the second temperature threshold is less than or equal to the first temperature threshold. The historical temperature corresponding to the historical temperature rise value refers to the temperature value corresponding to the end of the historical temperature rise value. The higher the temperature, the greater the impact on the safety performance of the battery. Therefore, a weight greater than 1 is set in a higher temperature area, and a weight less than 1 is set in a lower temperature area. The first weight and the second weight can change linearly with the temperature. Specifically, the higher the temperature, the greater the weight. By considering the dynamically adjusted weight in the difference calculation, the safety performance of the battery at different temperatures can be more accurately evaluated.
[0186] Specifically:
[0187] For each time point, the difference between the current temperature and a reference temperature (eg, standard operating temperature) is calculated.
[0188] According to the actual situation that battery safety performance is affected by temperature, the weight of each temperature difference range is determined. Generally, a high temperature difference may have a greater impact on battery safety performance, while a small temperature difference has a smaller impact.
[0189] Assign a weight to each temperature difference range. For example, a rule or rule of thumb may be used to determine the weights so that a higher weight is assigned to a temperature change with a greater impact, and a lower weight is assigned to a temperature change with a smaller impact.
[0190] When calculating the impact of temperature differences on battery performance, each temperature difference is multiplied by the corresponding weight. This can more accurately reflect the actual impact of different temperature changes on battery safety performance.
[0191] Ensure that the assigned weights are in line with expectations and verify them in actual applications. Based on actual data and feedback, the weights need to be adjusted to better suit different working conditions and environmental changes.
[0192] S832. Calculate the coefficient of determination and / or root mean square error of the temperature rise prediction model according to the weighted difference.
[0193] R 2 It is the coefficient of determination, which is used to evaluate the degree of explanation of the model for the sample observation results. The closer its value is to 1, the higher the degree of explanation of the model and the better the model.
[0194] R 2 The calculation formula is:
[0195]
[0196] RMSE is the root mean square error, which is the square of the difference between the predicted temperature rise and the historical temperature rise (weighted here), then averaged and finally took the square root.
[0197] The RMSE calculation formula is:
[0198]
[0199] In the above two formulas, y i is the historical temperature rise value, f i is the predicted temperature rise, is the average of the historical temperature rise values. FVU is the proportion of unexplained variance, RSS is the residual sum of squares, TSS is the total sum of squares, and n is the number of samples.
[0200] The temperature rise prediction model obtained in this application was subjected to the above algorithm and the results were R2=0.9762 and RMSE=1.9793.
[0201] S833. Verify the current temperature rise prediction model according to the determination coefficient and / or the root mean square error. When the determination coefficient and / or the root mean square error are within the preset error threshold, verify that the current temperature rise prediction model is available, otherwise verify that it is unavailable. If the determination coefficient is higher than the preset threshold, it means that the model can well explain the variation of the observed data and has a strong prediction ability. If the root mean square error is lower than the preset threshold, it means that the difference between the predicted temperature rise value of the model and the actual observed value is small, and the prediction accuracy is high.
[0202] like Fig. 9 As shown, the present invention also proposes a battery temperature rise prediction method, comprising the following steps:
[0203] B1. Obtain current environmental information. The BUSBAR (copper busbar / aluminum busbar) ambient temperature, contact resistance, and current can be collected at multiple points through on-board sensors. By real-time monitoring and collecting this environmental information, potential problems can be identified in a timely manner and appropriate measures can be taken to ensure the safety performance and stability of the battery system.
[0204] B2. Input the current environmental information into the temperature rise prediction model to obtain the predicted temperature rise value. Deploy the trained model to the actual vehicle system for real-time temperature rise prediction. The deployment process can integrate the model into the vehicle's software or hardware platform and ensure that the model can run stably and reliably. Set up a data interface in the vehicle system to obtain BUSBAR ambient temperature, contact resistance and current data in real time. It is also necessary to pre-process the input data to ensure that the data format and range match the input requirements of the model. Figure 5The figure shows the comparison between the predicted temperature rise value of the trained temperature rise prediction model and the historical temperature rise value, where line 3 is the historical temperature rise value and line 4 is the predicted temperature rise value.
[0205] like Fig.10 As shown, the present invention also proposes a battery control method, comprising the following steps:
[0206] C1. Obtain the predicted temperature rise value through the battery temperature rise prediction method. Input the collected environmental information (BUSBAR ambient temperature, contact resistance and current) into the pre-trained temperature rise prediction model to obtain the predicted temperature rise value. The predicted temperature rise value is used to monitor the status of the battery system in real time, and take appropriate measures to adjust the operation of the system to ensure the safety performance and stability of the battery.
[0207] C2. When the predicted temperature rise value exceeds the preset temperature rise threshold, the battery discharge is limited or interrupted. Use a real-time temperature rise prediction system to monitor the temperature changes of the battery system and obtain the predicted temperature rise value. Compare the predicted temperature rise value with the preset temperature rise threshold. If the predicted temperature rise value exceeds the threshold, it means that the battery system is at a potential risk of overheating. When the predicted temperature rise value exceeds the preset threshold, the corresponding safety mechanism in the system is triggered to limit or interrupt the discharge process of the battery. This can be achieved by controlling the battery management system (BMS), such as pausing or adjusting the discharge current, or completely interrupting the discharge of the battery. Before executing the restriction or interruption of discharge, a safety alarm or notification can be triggered to remind the vehicle operator or maintenance personnel that there is a risk of excessive battery temperature. Such an alarm can help take timely countermeasures to avoid potential safety issues. Once the battery temperature drops below a safe level, the system can automatically resume the discharge process, or resume the discharge after manual authorization by the operator.
[0208] In addition, the solution in this application can not only be used for the BUSBAR (copper busbar / aluminum busbar) of the battery, but can also be applied to the temperature prediction of PACK (battery pack) high and low voltage power connectors, liquid-cooled laminated all-aluminum busbar / copper busbar, super-charging DCU busbar (copper busbar / aluminum busbar), DC charging port and other equipment.
[0209] In summary, the present invention proposes a temperature rise prediction model training method, prediction method and control method. In the above scheme, the battery temperature is predicted by using the trained model after model training. Based on the analysis of historical data and related features, the prediction model can more accurately capture the law of battery temperature change. The battery temperature change can be monitored in real time, and the future temperature trend can be predicted in time, which is conducive to timely adjustment of the control strategy. Combined with the prediction model and the control method, the battery temperature can be automatically controlled to improve the safety and stability of the battery.
[0210] The above embodiments are only illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with this technology may modify or change the above embodiments without violating the spirit and scope of the present invention. Therefore, all equivalent modifications or changes made by those with ordinary knowledge in the relevant technical field without departing from the spirit and technical ideas disclosed by the present invention should still be covered by the claims of the present invention. Efficiently realize the connection between the high voltage of the motor and the high voltage of the controller, small size, oil-resistant sealing design, reliable sealing of the controller cavity and the motor cavity, temperature rise prediction model training method, prediction method and control method oil cooling design, effectively reduce the temperature, achieve greater flow and increase product life.
[0211] In the description herein, many specific details, such as examples of components and / or methods, are provided to provide a complete understanding of embodiments of the present invention. However, those skilled in the art will recognize that embodiments of the present invention may be practiced without one or more of the specific details or with other devices, systems, components, methods, components, materials, parts, etc. In other cases, well-known structures, materials, or operations are not specifically shown or described in detail to avoid obscuring aspects of embodiments of the present invention.
[0212] References throughout this specification to "one embodiment," "an embodiment," or "a specific embodiment" mean that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the invention, and not necessarily in all embodiments. Thus, various appearances of the phrases "in one embodiment," "in an embodiment," or "in a specific embodiment" in different places throughout this specification do not necessarily refer to the same embodiment. In addition, the particular features, structures, or characteristics of any specific embodiment of the invention may be combined with one or more other embodiments in any suitable manner. It should be understood that other variations and modifications of the embodiments of the invention described and illustrated herein may be possible in light of the teachings herein and are to be considered part of the spirit and scope of the invention.
[0213] It should also be understood that one or more of the elements shown in the figures may also be implemented in a more separate or more integrated manner, or even removed because they are inoperable in certain circumstances or provided because they may be useful depending on the application.
[0214] In addition, unless otherwise explicitly indicated, any marking arrows in the drawings should be regarded as exemplary only and not limiting. In addition, unless otherwise indicated, the term "or" used herein is generally intended to mean "and / or". In the case where the term is not clear because it is anticipated that the ability to separate or combine is provided, the combination of components or steps will also be regarded as indicated.
[0215] As used in the description herein and throughout the claims that follow, "a", "an", and "the" include plural references unless otherwise indicated. Likewise, as used in the description herein and throughout the claims that follow, the meaning of "in" includes "in" and "on" unless otherwise indicated.
[0216] The above description of the illustrated embodiments of the present invention (including the contents described in the Abstract) is not intended to be an exhaustive list or to limit the present invention to the precise form disclosed herein. Although specific embodiments of the present invention and examples of the present invention are described herein for illustrative purposes only, various equivalent modifications are possible within the spirit and scope of the present invention as will be recognized and appreciated by those skilled in the art. As noted, these modifications may be made to the present invention in accordance with the above description of the embodiments described in the present invention, and these modifications will be within the spirit and scope of the present invention.
[0217] Systems and methods have been generally described herein as details that aid in understanding the present invention. In addition, various specific details have been given to provide an overall understanding of embodiments of the present invention. However, those skilled in the relevant art will recognize that embodiments of the present invention may be practiced without one or more of the specific details, or may be practiced using other devices, systems, accessories, methods, components, materials, parts, etc. In other cases, well-known structures, materials, and / or operations are not specifically shown or described in detail to avoid confusion with various aspects of embodiments of the present invention.
[0218] Thus, although the invention has been described herein with reference to specific embodiments thereof, freedom of modification, various changes and substitutions are also within the foregoing disclosure, and it should be understood that in some cases, some features of the invention will be employed without the corresponding use of other features without departing from the scope and spirit of the proposed invention. Thus, many modifications may be made to adapt a particular environment or material to the essential scope and spirit of the invention. The invention is not intended to be limited to the specific terms used in the claims below and / or the specific embodiments disclosed as the best mode contemplated for carrying out the invention, but the invention will include any and all embodiments and equivalents falling within the scope of the appended claims. Thus, the scope of the invention will be determined solely by the appended claims.
Claims
1. A temperature rise prediction model training method, applied to power batteries, It is characterized in that The steps include: Acquiring historical environmental data of the power battery; Calculating historical temperature rise values based on the historical environmental data; Obtaining a predicted temperature rise value based on the historical environmental data and the neural network model; Calculating a loss value based on the predicted temperature rise value and the historical temperature rise value corresponding to the predicted temperature rise value; The model parameters of the neural network model are adjusted by the loss value, and the above steps are repeated to obtain the predicted temperature rise value and calculate the loss value until the training is completed when the loss value is lower than a preset threshold, thereby obtaining a temperature rise prediction model.
2. The temperature rise prediction model training method according to claim 1, It is characterized in that The obtaining of the predicted temperature rise value based on the historical environmental data and the neural network model comprises: Arrange the historical environmental data in chronological order; Performing sliding window processing on the historical environmental data; Extracting data from each sliding window to obtain multiple groups of sliding window data, each group of sliding window data constituting a group of preprocessed data; A predicted temperature rise value is obtained based on the preprocessed data and the neural network model.
3. The temperature rise prediction model training method according to claim 1, It is characterized in that After the step of obtaining the temperature rise prediction model, the method further comprises: Acquiring the current environmental data of the power battery in real time; The historical environment data is updated according to the current environment data.
4. The temperature rise prediction model training method according to claim 3, It is characterized in that Updating the historical environment data according to the current environment data includes: Clear the historical environmental data corresponding to the first N moments in the data storage area in chronological order; The current environment data corresponding to the latest N moments are added to the data storage area as the historical environment data.
5. The temperature rise prediction model training method according to claim 3, It is characterized in that The step of obtaining the temperature rise prediction model comprises: The temperature rise prediction model is obtained by training according to the updated historical environmental data.
6. The temperature rise prediction model training method according to claim 1, It is characterized in that After the step of obtaining the temperature rise prediction model, the following steps are included: The temperature rise prediction model is verified based on the historical environmental data and the historical temperature rise values.
7. The temperature rise prediction model training method according to claim 6, It is characterized in that The verifying the temperature rise prediction model according to the historical environmental data and the historical temperature rise values includes: Inputting the historical environmental data into the temperature rise prediction model to obtain a predicted temperature rise value; Calculating the difference between the predicted temperature rise value and the corresponding historical temperature rise value; The temperature rise prediction model is verified according to the difference.
8. The temperature rise prediction model training method according to claim 7, It is characterized in that The verifying the temperature rise prediction model according to the difference includes: The determination coefficient and / or the root mean square error of the temperature rise prediction model are calculated according to the difference. When the determination coefficient and / or the root mean square error are within a preset error threshold, the current temperature rise prediction model is verified to be available, otherwise it is verified to be unavailable.
9. The temperature rise prediction model training method according to claim 8, It is characterized in that The step of calculating the coefficient of determination and / or the root mean square error of the temperature rise prediction model according to the difference comprises: assigning a weight to each of the differences respectively; The coefficient of determination and / or root mean square error of the temperature rise prediction model is calculated based on the weighted difference.
10. The temperature rise prediction model training method according to claim 9, It is characterized in that The step of assigning a weight to each difference value comprises: When the historical temperature corresponding to the historical temperature rise value is greater than a preset first temperature threshold, the difference is multiplied by a first weight, and the first weight is greater than 1; When the historical temperature corresponding to the historical temperature rise value is less than or equal to a preset second temperature threshold, the difference is multiplied by a second weight, the second weight is less than 1, and the second temperature threshold is less than or equal to the first temperature threshold.
11. A battery temperature rise prediction method, It is characterized in that The steps include: Get current environment information; The current environmental information is input into the temperature rise prediction model as described in any one of claims 1 to 10 to obtain a predicted temperature rise value.
12. A battery control method, It is characterized in that The steps include: Obtaining a predicted temperature rise value by the battery temperature rise prediction method according to claim 11; When the predicted temperature rise value exceeds a preset temperature rise threshold, battery discharge is restricted or interrupted.
Citation Information
Cited By
Lithium battery multi-feature remaining service life prediction method based on regression prediction model
CN122263678A
