A method for estimating the state of charge of a power lithium battery based on a multi-layer perceptron algorithm

Abstract

The present invention discloses a method for estimating the state of charge of a power lithium battery based on a multi-layer perceptron algorithm. The improvement lies in the following steps: Step 1, data splitting; Step 2, data preprocessing; Step 3, designing a multi-layer perceptron network; Step 4, selection of hyperparameters; Step 5, parameter adjustment; Step 6, model verification. The estimation method disclosed by the present invention is more accurate in estimation and prediction compared with the relatively simple look-up table method and model estimation method, is insensitive to factors such as the environment, and has good stability. Compared with the LSTM method of the data-driven method, the training speed is faster.

Description

Technical Field

[0001] The present invention belongs to the field of power lithium battery state of charge estimation methods, and particularly relates to an estimation method for the state of charge of a power lithium battery based on a multi-layer perceptron algorithm in this field. Background Art

[0002] Since the 21st century, problems such as the energy crisis and automobile exhaust pollution have become increasingly prominent, and reducing carbon emissions has become a consensus. Power lithium batteries are an important direction for the development of new energy, and they can be widely used in new energy vehicles and energy storage and other fields. The state of charge (SOC) of a battery is an important factor in measuring the performance of the battery and is one of the key parameters of a lithium-ion battery. The more accurate the battery SOC estimation is, the more superior the performance of the battery management system is. On the contrary, if the battery SOC estimation is inaccurate, it will not only waste energy but also damage the battery, and in severe cases, it will even cause safety hazards.

[0003] The conventional battery SOC refers to the total value obtained by dividing the amount of electricity that the battery can discharge under a certain discharge current by its rated capacity under this condition. Lithium-ion batteries have typical non-linear characteristics, and it is difficult to measure the total amount of electricity released by the battery through existing means or methods. According to the ampere-hour integration method theory, the battery SOC is particularly crucial, and it can accurately reflect the energy status. The current estimation methods for battery SOC mainly include: simple look-up table method, model estimation method, and data-driven estimation method.

[0004] The simple look-up table method generally uses the open-circuit voltage method to perform static tests on the open-circuit voltage and the battery SOC, fits the OCVSOC curve as a static correction. Then, the ampere-hour integration method is used for dynamic estimation, and finally, the value of the battery state of charge is estimated. The biggest disadvantage of the simple look-up table method is that it takes a long time, wastes energy, cannot perform real-time estimation, and more importantly, it uses an open-loop estimation method, resulting in cumulative errors.

[0005] The model estimation method generally establishes a battery model, such as an electrochemical impedance model (EIM) and an equivalent circuit model (ECM), etc. The state of charge of the power lithium battery is estimated through the state of charge of the model. The main disadvantages are difficult modeling. The model-based estimation method requires a very in-depth understanding of the battery in order to model using parameters such as electrochemistry and material properties, and there are also disadvantages such as large parameter identification difficulty and large computational amount.

[0006] The data-driven estimation method is based on the data accumulated during the application of lithium-ion batteries. Through machine learning methods, it learns the internal mechanism of the state of charge and uses supervised learning methods to estimate the state of charge. The data-driven estimation method describes the battery performance from the perspective of the whole-process test data, and then analyzes the battery SOC. It requires the use of professional lithium battery test equipment and high-precision data acquisition circuits to simulate the real working conditions of lithium batteries through the testing of lithium batteries. Its main disadvantage is that it highly depends on the integrity of the data and the speed of computer processing, has high requirements for data, and the discrete training takes a long time. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide an estimation method for the state of charge of a power lithium battery based on a multi-layer perceptron algorithm.

[0008] The present invention adopts the following technical solutions:

[0009] An estimation method for the state of charge of a power lithium battery based on a multi-layer perceptron algorithm, the improvement lies in that it includes the following steps:

[0010] Step 1, data splitting:

[0011] Divide the total battery test data set T ot into two parts, 75% is the training data set T ra , 25% is the test data set T es ;

[0012] Step 2, data preprocessing:

[0013] Normalize the data in the training data set T ra , and the formula is as follows:

[0014]

[0015] Among them, x represents a certain feature, x min represents the minimum value in this feature, x max represents the maximum value in this feature;

[0016] Step 3, design a multi-layer perceptron network:

[0017] Design a fully connected neural network:

[0018]

[0019] m = f(z)

[0020]

[0021] h = f(s)

[0022]

[0023] In the above formula, D represents the scale, and x1, x2, …… x d indicates that there are d feature inputs, and m1, m2, m3, …… m d represent features with d hidden layers, and h1, h2, ……, h d represents the number of features in the second hidden layer, o1 represents the output value of the multi-layer perceptron network, a represents the weight of each feature, b is the offset, W, X, and M are all matrices, and f(·) represents the activation function;

[0024] Step 4: Selection of hyperparameters:

[0025] The loss function J(a) uses the root mean square error RMSE and the mean absolute error MAE.

[0026]

[0027]

[0028]

[0029] In the above formula, a represents the weight of each feature, x represents a certain feature, y represents the value of SOC, N represents the total number of samples in the training data set, and Y n represents the true SOC value of each training data sample, and O n (x) represents the SOC value of each training data sample calculated by the multi-layer perceptron algorithm;

[0030] During the model training process, the learning rate adopts the cosine annealing algorithm, and the initial learning rate is 0.00015;

[0031] Step 5: Parameter adjustment:

[0032] First, adjust the depth of the network, fix the maximum network width at 80. After training the network depth, then fix two network depths to optimize the network width. The number of training loops is randomly selected from 30 to 200 for measurement. Finally, select the optimal model with the best accuracy and precision;

[0033] Step 6, Model verification:

[0034] Use the optimal model selected in Step 5 as the verification model to view its training error and test error. When the training error E tr is less than one-thousandth and the test error E te is less than five-thousandths, then select this model as the final algorithm model. If the requirements are not met, repeat the process of Step 5.

[0035] Furthermore, layer normalization operations are added to each layer in step 3, and DropOut pruning operations are added after each layer. The DropOut pruning operation is to randomly cut off network connections with a certain probability P d randomly, and P d = 0.3.

[0036] Furthermore, the activation functions in step 3 use the ReLU activation function and the Mish activation function, and the ReLU activation function and the Mish activation function are used alternately between layers. The formulas are as follows:

[0037] f(v) = ReLU(v) = MAX(0, v)

[0038] f(v) = Mish(v) = v * tanh(1 + e v )

[0039] In the above formula, v represents the eigenvalue trained by the previous layer of the neural network, and tanh is the hyperbolic tangent function.

[0040] Furthermore, the fully connected neural network in step 3 is a nine-layer fully connected neural network. The first layer has 15 input features and 80 output features, and the activation function used is the Mish activation function. The second layer has 80 input feature parameters and 100 output parameters, and the activation function used is still ReLU. The third layer has 100 input features and 100 output features, and the activation function used is the Mish activation function. The fourth layer has 100 input features and 70 output features, and the activation function is ReLU. The fifth layer has 70 input features and 70 output features, and the activation function is Mish. The sixth layer has 70 input features and 50 output features, and the activation function is ReLU. The seventh layer has 50 input features and 50 output features, and the activation function is Mish. The eighth layer has 50 input features and 20 output features, and the activation function is ReLU. The ninth layer is the prediction layer with 20 input features and 1 output feature, and the activation function is Mish.

[0041] Furthermore, in the loss function of step 4, the training error uses the RMSE root mean square error as the basis, and the RMSE error E tr should be less than one-thousandth. The test error uses the absolute value error MAE as the evaluation basis, and the MAE error E te should be less than five-thousandths.

[0042] The beneficial effects of the present invention are:

[0043] The estimation method disclosed by the present invention is more accurate in estimation and prediction compared with the relatively simple look-up table method and the model estimation method, is insensitive to factors such as the environment, and has good stability. Compared with the LSTM method of the data-driven method, the training speed is faster.

[0044] The estimation method disclosed by the present invention can more accurately estimate the value of the state of charge, and the training error MSE is less than 2%. The more accurate the state of charge is, the more accurate driving mileage can be provided, and the safety of lithium batteries and new energy vehicles can be better guaranteed. The implementation cost is low, it is simple and easy to use, is more suitable for engineering practice, and has broad market prospects and promotion value. Description of the Drawings

[0045] Figure 1 It is a schematic diagram of the fully connected neural network in step 3 of the estimation method of the present invention. Detailed Embodiments

[0046] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0047] Embodiment 1. This embodiment discloses an estimation method for the state of charge of a power lithium battery based on a multi-layer perceptron algorithm. By deeply analyzing each parameter in the publicly available test data set of lithium battery 18650, a nine-layer neural network is used to learn the internal mechanism of the state of charge. The activation function uses the cross-use of Mish and ReLU to meet non-linearity. 70% of the data set is used to train the model parameters of this method, and 30% of the data set is used to test and verify the generalization of the model. The finally designed and trained model method can very accurately estimate the state of charge (SOC). The comprehensive performance of the lithium battery is judged by the state of charge, and finally the state of charge is used to determine the driving mileage of the new energy vehicle. The specific steps are as follows:

[0048] Step 1, data splitting:

[0049] The total battery test data set T ot (Total DataSet) is divided into two parts, 75% is the training data set T ra (Training DataSet), and 25% is the test data set T es (Testing DataSet);

[0050] Step 2, data preprocessing:

[0051] In the dataset, since the value ranges of some features are relatively large and those of some features are relatively small, it may lead to the features with larger value ranges being considered more important by the neural network algorithm. Moreover, the large variation range of feature values will also cause difficulties in training the weights. In this case, the normalization technique is mainly used. Note that the normalization operation is only used in the training dataset.

[0052] Normalize the data in the training dataset T ra using the following formula:

[0053]

[0054] where x represents a certain feature, x min represents the minimum value in this feature, and x max represents the maximum value in this feature;

[0055] Step 3, design a multi-layer perceptron network:

[0056] As Figure 1 shown, design a fully connected neural network, where all input parameters act on the hidden layer:

[0057]

[0058] m = f(z)

[0059]

[0060] h = f(s)

[0061]

[0062] In the above formula, D represents the scale, and different network layers have different d values, indicating the number of features. x1, x2,... x d represent d feature inputs, m1, m2, m3,... m d represent the features with d hidden layers, and different network layers have different feature values. h1, h2,..., h d represent the number of features in the second hidden layer, o1 represents the output value of the multi-layer perceptron network, a represents the weight of each feature, b is the offset, W, X, and M are all matrices, and f(·) represents the activation function, which is usually a non-linear function;

[0063] For the selection of the activation function, mainly the ReLU activation function and the Mish activation function are used, and the ReLU activation function and the Mish activation function are used alternately between layers. The formula is as follows:

[0064] f(v) = ReLU(v) = MAX(0, v)

[0065] f(v) = Mish(v) = v * tanh(1 + e v )

[0066] In the above formula, v represents the eigenvalue trained by the previous layer of neural network, and tanh is the double tangent function.

[0067] The non-linearity of the ReLU function comes from the fact that when the value of the input feature is greater than 0, the value of this feature is used to maintain linear transmission. When the value of the input feature is less than 0, 0 is used for substitution to maintain non-linearity. The non-linearity of the Mish activation function comes from the multiplication of the value v of the feature and the double tangent function.

[0068] Batch Normalization operation is added to each layer, and at the same time, DropOut pruning operation is added after each layer. The DropOut pruning operation is to randomly cut off the network connections with a certain probability P d randomly, P d = 0.3. By designing the depth of different multi-layer perceptron networks and the width of different multi-layer perceptron networks, and using different DropOut pruning weights, the purpose of optimizing network expression is achieved.

[0069] In this embodiment, the fully connected neural network in this step is a nine-layer fully connected neural network. The first layer has 15 input feature parameters and 80 output feature parameters, and the activation function used is the Mish activation function. The second layer has 80 input feature parameters and 100 output parameters, and the activation function used is still ReLU. The third layer has 100 input features and 100 output features, and the activation function used is the Mish activation function. The fourth layer has 100 input features and 70 output features, and the activation function is ReLU. The fifth layer has 70 input features and 70 output features, and the activation function is Mish. The sixth layer has 70 input features and 50 output features, and the activation function is ReLU. The seventh layer has 50 input features and 50 output features, and the activation function is Mish. The eighth layer has 50 input features and 20 output features, and the activation function is ReLU. The ninth layer is the prediction layer with 20 input features and 1 output feature, and the activation function is Mish.

[0070] Step 4: Selection of hyperparameters:

[0071] The loss function J(a) uses the root mean square error RMSE and the mean absolute error MAE.

[0072]

[0073]

[0074]

[0075] In the above formula, a represents the weight of each feature, x represents a certain feature, y represents the true value of the label, that is, the value of SOC, N represents the total number of samples in the training dataset, and Y n represents the true SOC value of each training data sample, O n (x) represents the SOC value of each training data sample calculated by the multi-layer perceptron algorithm;

[0076] Regarding the use of the learning rate in hyperparameter selection, after designing and improving the multi-layer perceptron, the cosine annealing algorithm is adopted for the learning rate during the process of training the multi-layer perceptron algorithm model to accelerate the convergence of the model. As the number of training times increases, the learning rate will show a trend of rising first and then falling. The initial learning rate is 0.00015.

[0077] In the loss function of this step, the training error uses the RMSE root mean square error as the basis, and the RMSE error E tr should be less than one-thousandth. The test error uses the absolute value error MAE as the evaluation basis, and the MAE error E te should be less than five-thousandths.

[0078] Step 5: Parameter adjustment:

[0079] During the training process, due to different activation functions and different hyperparameters, the width of the multi-layer perceptron network is different, and the depth of the multi-layer perceptron is also different. These will all lead to a decrease in the accuracy and accuracy of the training model.

[0080] First, adjust the depth of the network, fix the maximum network width at 80. After training a better network depth, then fix two network depths to optimize the network width. The number of training loops (EPOCH) is randomly selected from 30 to 200 for measurement. Finally, select the optimal model with the best accuracy and accuracy;

[0081] Step 6, Model verification:

[0082] Take the optimal model selected in Step 5 as the verification model to check its training error and test error. When the training error E tr is less than one-thousandth and the test error E te is less than five-thousandths, select this model as the final algorithm model. If the requirements are not met, repeat the process of Step 5.

Claims

1. A method for estimating the state of charge of a power lithium battery based on a multi-layer perceptron algorithm, characterized in that, It includes the following steps: Step 1, data splitting: Divide the total battery test dataset T ot into two parts, with 75% being the training dataset T ra and 25% being the test dataset T es ; Step 2, data preprocessing: Normalize the data in the training dataset T ra using the following formula: where x represents a certain feature, and x min represents the minimum value of this feature, and x max represents the maximum value of this feature; Step 3, designing a multi-layer perceptron network: Designing a fully-connected neural network: m = f(z) h = f(s) In the above formula, D represents the scale, and x1, x2, …… x d indicates that there are d feature inputs, and m1, m2, m3, …… m d represents the features with d hidden layers, and h1, h2 ……, h d represents the number of features in the second hidden layer, o1 represents the output value of the multi-layer perceptron network, a represents the weight of each feature, b is the offset, W, X, and M are all matrices, and f(·) represents the activation function; Step 4: Selection of hyperparameters: The loss function J(a) uses the root mean square error RMSE and the mean absolute error MAE. In the above formula, a represents the weight of each feature, x represents a certain feature, y represents the value of SOC, N represents the total number of samples in the training data set, and Y n represents the true SOC value of each training data sample, and O n (x) represents the SOC value of each training data sample calculated by the multi-layer perceptron algorithm; During the model training process, the learning rate adopts the cosine annealing algorithm, and the initial learning rate is 0.00015. Step 5: Parameter adjustment: First, adjust the depth of the network, fix the maximum network width at 80. After training the network depth, then fix two network depths to optimize the network width. The number of training loops is randomly selected from 30 to 200 for measurement. Finally, select the optimal model with the best accuracy and precision. Step 6, model verification: Take the optimal model selected in step 5 as the validation model to check its training error and test error. When the training error E tr is less than one-thousandth and the test error E te is less than five-thousandths, select this model as the final algorithm model. If the requirements are not met, repeat the process of step 5.

2. The method for estimating the state of charge of a power lithium battery based on the multi-layer perceptron algorithm according to claim 1, characterized in that: Add layer normalization operations to each layer in step 3, and at the same time add DropOut pruning operations after each layer. The DropOut pruning operation is to randomly cut off network connections with a certain probability P d randomly, where P d = 0.

3.

3. The method for estimating the state of charge of a power lithium battery based on the multi-layer perceptron algorithm according to claim 1, characterized in that: The activation functions in Step 3 use the ReLU activation function and the Mish activation function, and the ReLU activation function and the Mish activation function are used alternately between layers. The formula is as follows: f(v) = ReLU(v) = MAX(0, v) f(v) = Mish(v) = v * tanh(1 + e v ) In the above formula, v represents the eigenvalue trained by the previous layer of the neural network, and tanh is the hyperbolic tangent function.

4. The method for estimating the state of charge of a power lithium battery based on the multi-layer perceptron algorithm according to claim 3, characterized in that: The fully-connected neural network in Step 3 is a nine-layer fully-connected neural network. The first layer has 15 input features and 80 output features, and the activation function used is the Mish activation function. The second layer has 80 input feature parameters and 100 output parameters, and the activation function used is still ReLU. The third layer has 100 input features and 100 output features, and the activation function used is the Mish activation function. The fourth layer has 100 input features and 70 output features, and the activation function is ReLU. The fifth layer has 70 input features and 70 output features, and the activation function is Mish. The sixth layer has 70 input features and 50 output features, and the activation function is ReLU. The seventh layer has 50 input features and 50 output features, and the activation function is Mish. The eighth layer has 50 input features and 20 output features, and the activation function is ReLU. The ninth layer is the prediction layer with 20 input features and 1 output feature, and the activation function is Mish.

5. The method for estimating the state of charge of a power lithium battery based on the multi-layer perceptron algorithm according to claim 1, characterized in that: In the loss function of step 4, the training error is based on the RMSE (Root Mean Square Error), and the RMSE error E tr should be less than one-thousandth. The test error uses the absolute error MAE as the evaluation criterion, and the MAE error E te should be less than five-thousandths.

Images