Energy storage battery end-to-end capacity estimation method based on hybrid deep neural network

Abstract

The application particularly relates to an energy storage battery end-to-end capacity estimation method based on a hybrid deep neural network, which comprises the following steps: obtaining battery charging voltage data of an energy storage battery; inputting the battery charging voltage data of the energy storage battery into a trained battery capacity estimation model to output corresponding battery capacity estimation values; the battery capacity estimation model comprises: a convolutional neural network layer, which is used for extracting battery aging features of input data and generating corresponding initial feature maps; an attention mechanism layer, which is used for sequentially extracting channel attention weights and spatial attention weights of the initial feature maps and generating corresponding weighted feature maps; and a recurrent neural network layer, which is used for taking the weighted feature maps generated by the attention mechanism layer as input to be trained and outputting corresponding battery capacity estimation values. The application can automatically extract battery aging features to realize end-to-end energy storage battery capacity estimation and relieve the problems that a deep neural network is prone to fall into local optimization and gradient vanishing (dispersion).

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electrochemical energy storage science management, and particularly relates to an energy storage battery end-to-end capacity estimation method based on a hybrid deep neural network. BACKGROUND

[0002] Lithium-ion energy storage batteries are widely used in the field of electrochemical energy storage due to their high energy ratio and wide temperature working interval characteristics. However, due to the influence of external environment and internal factors, lithium batteries will inevitably age during use, showing capacity decline and performance degradation. Therefore, accurate estimation of battery capacity is helpful for the health state evaluation and remaining service life prediction of the battery, and is crucial for battery state monitoring, fault diagnosis and safety warning in the field of electrochemical energy storage.

[0003] The existing technology for energy storage battery capacity estimation often uses data-driven (physical model) estimation methods. The data-driven estimation method is to model a certain type of energy storage battery through specific health factors. For example, the Chinese patent with the publication number CN105891715A discloses a lithium-ion battery health state estimation method, which includes: establishing an equivalent physical model of the battery; obtaining the battery SOC-OCV curve through battery pulse discharge and calculating the relationship between battery SOC and open circuit voltage; collecting battery voltage, current and temperature parameters; obtaining the battery ohmic resistance based on recursive least squares method for online identification of model parameters; and calculating the battery health state using the estimated battery ohmic resistance. The existing scheme realizes the capacity estimation of the energy storage battery based on the data-driven estimation method.

[0004] The existing physical model-based estimation method can bring high accuracy, but it has the problems of complex parameters and high calculation cost. Moreover, the actual operating conditions of energy storage batteries are complex, and it is difficult to obtain complete charge and discharge data. Battery capacity estimation needs to span multiple charge and discharge cycles, which makes it difficult to apply the estimation method based on the physical model. Therefore, the existing technology begins to use the estimation method based on the deep learning model. However, the existing deep learning model-based estimation method mostly needs to manually extract battery aging feature data, and then completes battery capacity estimation based on the preprocessed data, that is, the existing technology cannot realize end-to-end estimation, which limits the practical application scenarios of energy storage battery capacity estimation. At the same time, due to the complexity of the actual operating conditions of energy storage batteries, the existing deep learning model based on the related data of energy storage batteries for battery capacity estimation is prone to local optimum and gradient disappearance (dispersion) problems, which further leads to low accuracy of energy storage battery capacity estimation. Therefore, how to design a method that can improve the practicality and accuracy of energy storage battery capacity estimation is a technical problem that needs to be solved. SUMMARY

[0005] In view of the above problems of the prior art, the technical problem to be solved by the present application is how to design a hybrid deep neural network-based energy storage battery end-to-end capacity estimation method, which can automatically extract battery aging features to realize end-to-end energy storage battery capacity estimation and alleviate the problems of deep neural network easily falling into local optimum and gradient vanishing (dispersion), so as to improve the practicability and accuracy of energy storage battery capacity estimation.

[0006] To solve the above technical problems, the present application adopts the following technical solutions:

[0007] The hybrid deep neural network-based energy storage battery end-to-end capacity estimation method comprises:

[0008] S1: obtaining battery charging voltage data of an energy storage battery;

[0009] S2: inputting the battery charging voltage data of the energy storage battery into a trained battery capacity estimation model to output a corresponding battery capacity estimation value;

[0010] The battery capacity estimation model comprises:

[0011] a convolutional neural network layer for extracting battery aging features of input data and generating a corresponding initial feature map;

[0012] an attention mechanism layer for sequentially extracting channel attention weights and spatial attention weights of the initial feature map and generating a corresponding weighted feature map;

[0013] a recurrent neural network layer for training the weighted feature map generated by the attention mechanism layer as input and outputting a corresponding battery capacity estimation value;

[0014] S3: taking the battery capacity estimation value of the energy storage battery as the battery capacity estimation result thereof.

[0015] Preferably, in step S2, the convolutional neural network layer is constructed based on a one-dimensional convolutional neural network.

[0016] Preferably, the one-dimensional convolutional neural network generates the initial feature map through the following formula:

[0017]

[0018] In the formula: represents the output of the input data corresponding to the kth convolution kernel, and the outputs of each convolution kernel of the one-dimensional convolutional neural network constitute the initial feature map; w (k) represents the weight matrix of the kth convolution kernel; x c represents the sequence of the original input data; L represents the size of the convolution kernel; i represents the receptive field of the convolution kernel; and t represents the sequence length.

[0019] Preferably, in step S2, the attention mechanism layer comprises a channel attention module and a spatial attention module for extracting channel attention weights and spatial attention weights respectively;

[0020] The weighted feature map is generated by the following steps:

[0021] S201: Extracting channel attention weights of the initial feature map by the channel attention module;

[0022] S202: Multiplying the initial feature map with the channel attention weights to obtain a channel attention feature map;

[0023] S203: Extracting spatial attention weights of the channel attention feature map by the spatial attention module;

[0024] S204: Multiplying the initial feature map with the spatial attention weights to obtain a weighted feature map.

[0025] Preferably, in step S201, the channel attention module extracts the channel attention weights by the following formula:

[0026] M c (F)=softmax(Mean(F cout ));

[0027]

[0028]

[0029] In the formula: M c (F) represents the channel attention weights; F represents the initial feature map output by the convolutional neural network layer; softmax represents an activation function for normalizing the output value; Mean represents a tensor mean value operation; F cout represents the features on the output channel; z j represents the output value of the jth node; c represents the number of output nodes; z c represents the output value of the cth node; C out represents the output value of the output channel; k represents the number of convolution kernels of the convolutional neural network layer, i.e. the number of output channels.

[0030] Preferably, in step S202, the channel attention feature map is calculated by the following formula:

[0031]

[0032] In the formula: F' represents the channel attention feature map; M c (F) represents the channel attention weights; F represents the initial feature map.

[0033] Preferably, in step S202, the spatial attention module extracts the spatial attention weight by the following formula:

[0034]

[0035] wherein M s (F′) represents the spatial attention weight; f n×n represents the convolution operation with the kernel n; F′ represents the channel attention feature map; AvgPool(F′) represents the average pooling on the channel attention feature map; maxPool(F′) represents the maximum pooling on the channel attention feature map; σ represents the sigmoid activation function; [;] represents the connection of the features of two channels; and respectively represent the feature maps after the average pooling and the maximum pooling on the channel attention feature map.

[0036] Preferably, in step S204, the weighted feature map is calculated by the following formula:

[0037]

[0038] wherein F″ represents the weighted feature map; M s (F′) represents the spatial attention weight; F represents the initial feature map.

[0039] Preferably, in step S2, the recurrent neural network layer comprises a bidirectional long short-term memory neural network and a fully connected layer connected with the output of the bidirectional long short-term memory neural network.

[0040] The bidirectional long short-term memory neural network comprises a forgetting gate, an input gate and an output gate.

[0041] The forgetting gate is represented by the following formula:

[0042] f t = σ(W f h t-1 +W f x t +b f );

[0043] wherein f t represents the output of the forgetting gate; σ represents the sigmoid activation function; W f represents the forgetting gate weight; h t-1 represents the hidden layer information at t-1; x t represents the input feature at t; b f represents the forgetting gate bias.

[0044] The input gate is represented by the following formula:

[0045] it = σ(W i h t-1 + W i x t + b i ) ;

[0046]

[0047]

[0048] where i t represents the control reserved part information of the input gate; σ represents a sigmoid activation function; W i and W c represent input gate weights; h t-1 represents the hidden layer information at t-1; x t represents the input feature at t; b i and b c represent input gate biases; represents a sub-cell unit state candidate value, that is, information brought by a new input; c t represents a sub-cell unit state at t, that is, updated information; tanh represents a hyperbolic tangent function; c t-1 represents a sub-cell unit state at t-1;

[0049] The output gate is represented by the following formula:

[0050] o t = σ(W o h t-1 + W o x t + b o ) ;

[0051] h t = o t ⊙ tanh(c t ) ;

[0052] where o t represents the output feature of the output gate; h t represents the hidden layer information at t; h t-1 represents the hidden layer information at t-1; σ represents a sigmoid activation function; W o represents output gate weights; x t represents the input feature at t; b o represents output gate biases; tanh represents a hyperbolic tangent function; c t represents a sub-cell unit state at t, that is, updated information;

[0053] The output of the full connection layer is represented by the following formula:

[0054] y = aA T + b;

[0055] wherein: y represents the output of the full connection layer, i.e. the battery capacity estimation value; a represents the number of neurons; A represents the weight matrix; T represents the transpose symbol; and b represents the bias.

[0056] Preferably, the loss function of the recurrent neural network layer adopts the mean square error loss function as follows:

[0057]

[0058] wherein: y i represents the battery capacity target value at time i; represents the battery capacity estimation value at time i; and N represents the sample quantity.

[0059] The optimizer of the recurrent neural network layer adopts the adaptive moment estimation algorithm as follows:

[0060] m t = β1m t-1 + (1- β1)g t

[0061]

[0062]

[0063]

[0064] wherein: m t and m t-1 represent the smoothed moving average at iteration t times and t-1 times; t represents the iteration number; β1 and β2 represent the smoothing constant; g t represents the gradient of the target function; V t represents the moving average of the square of the gradient; w t and w t+1 respectively represent the updated parameters at iteration t times and t+1 times; α t represents the learning rate at iteration t times; ε = 10 -8 represents the divisor 0 is avoided. and represent the t power of the smoothing constant.

[0065] The energy storage battery end-to-end capacity estimation method based on the hybrid deep neural network has the following beneficial effects:

[0066] The application takes battery charging voltage data as input, then automatically extracts battery aging features to generate feature maps (time sequence feature maps) through a convolutional neural network layer, and further predicts battery capacity estimates based on the feature maps through a recurrent neural network layer, so that the structure of the convolutional neural network layer + the recurrent neural network layer realizes direct estimation of the battery capacity through the battery charging voltage data, that is, end-to-end energy storage battery capacity estimation can be realized, thereby reducing the processing difficulty of the energy storage battery data and being conducive to ensuring that the energy storage battery capacity estimation is optimal in the full life cycle and different aging paths of the energy storage battery, thereby improving the practicality of the energy storage battery capacity estimation.

[0067] Meanwhile, the application sets an attention mechanism layer between the convolutional neural network layer and the recurrent neural network layer, and sequentially extracts the channel attention weight and the spatial attention weight of the initial feature map through the attention mechanism layer and generates an attention feature map, on the one hand, the attention mechanism layer can weight the output channels of each convolutional layer of the convolutional neural network layer and assign a value to the initial feature map, and then obtain the importance of each channel of the initial feature map to assign a weight value to each feature, so that the subsequent deep neural network can focus on some feature channels, and then generate a weighted feature map by using the spatial relationship between the features, so that the strongly related features can be emphasized on the feature, thereby the problem that the deep neural network is easy to fall into local optimum can be relieved, thereby the effectiveness of the feature map and the accuracy of the energy storage battery capacity estimation can be improved. On the other hand, the application takes the weighted feature map generated by the attention mechanism layer as the input of the recurrent neural network layer for training, which avoids complex manual extraction of battery aging feature data as model input, and further relieves the problems of gradient vanishing and dispersion of the recurrent neural network hidden layer by weighting the extracted features, so that the optimal weight of each hidden layer of the deep neural network can be obtained, thereby the performance of the recurrent neural network layer and the accuracy of the energy storage battery capacity estimation can be improved. BRIEF DESCRIPTION OF DRAWINGS

[0068] In order to make the purpose, technical scheme and advantages of the application more clear, the application will be further described in detail below with reference to the drawings, in which:

[0069] Figure 1 The logic block diagram of the energy storage battery end-to-end capacity estimation method based on the hybrid deep neural network;

[0070] Figure 2 The modeling flowchart of the energy storage battery capacity estimation method based on the end-to-end structure;

[0071] Figure 3 The logic principle schematic diagram of the attention mechanism layer;

[0072] Figure 4The battery capacity estimation value of the battery capacity estimation model of the application is compared with the experimental value;

[0073] Figure 5 The capacity estimation error analysis result under the constant current constant voltage charging condition. DETAILED DESCRIPTION

[0074] In order to make the objects, technical solutions and advantages of the embodiments of the application clearer, the following will combine the drawings in the embodiments of the application to clearly and completely describe the technical solutions in the embodiments of the application. Obviously, the described embodiments are some of the embodiments of the application, rather than all the embodiments. The components of the embodiments of the application described and shown in the drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the application provided in the drawings is not intended to limit the scope of the claimed application, but only represents selected embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the application.

[0075] It should be noted that: similar reference numerals and letters represent similar items in the following drawings, therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. In the description of the application, it needs to be explained that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship when the product of the application is usually placed, which is only for the convenience of describing the application and simplifying the description, and does not indicate or imply that the indicated device or element must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the application. In addition, the terms "first", "second", "third" and the like are only used for differentiation in description, and cannot be understood as indicating or implying relative importance. In addition, the terms "horizontal", "vertical" and the like do not mean that the components must be absolutely horizontal or vertical, but can be slightly inclined. For example, "horizontal" only means that its direction is relatively more horizontal than "vertical", and does not mean that the structure must be completely horizontal, but can be slightly inclined. In the description of the application, it also needs to be explained that, unless otherwise explicitly specified and limited, the terms "arrangement", "installation", "connection", "connection" should be understood broadly, for example, it can be fixedly connected, or detachably connected, or integrally connected; it can be mechanically connected, or electrically connected; it can be directly connected, or indirectly connected through an intermediate medium, or the communication inside two elements. For those of ordinary skill in the art, the specific meaning of the above terms in the application can be understood according to the specific circumstances.

[0076] The specific embodiments are further described in detail below:

[0077] Embodiments:

[0078] The embodiment discloses an energy storage battery end-to-end capacity estimation method based on a hybrid deep neural network.

[0079] As shown in Figure 1 and Figure 2 The energy storage battery end-to-end capacity estimation method based on a hybrid deep neural network comprises:

[0080] S1: Obtain battery charging voltage data of the energy storage battery;

[0081] S2: Input the battery charging voltage data of the energy storage battery into the trained battery capacity estimation model, and output the corresponding battery capacity estimation value;

[0082] The battery capacity estimation model comprises:

[0083] A convolutional neural network layer is used to extract battery aging features of input data and generate corresponding initial feature maps;

[0084] An attention mechanism layer is used to sequentially extract channel attention weights and spatial attention weights of the initial feature maps and generate corresponding weighted feature maps;

[0085] A recurrent neural network layer is used to train the weighted feature maps generated by the attention mechanism layer as input and output the corresponding battery capacity estimation value;

[0086] In the embodiment, when training the battery capacity estimation model, the energy storage battery is selected to carry out a cycle aging life experiment, and the battery test end voltage, capacity and other data are collected as training samples to construct corresponding training sets, test sets and validation sets, and the existing training means is used for training.

[0087] S3: The battery capacity estimation value of the energy storage battery is used as the battery capacity estimation result of the energy storage battery, and the health state of the energy storage battery can be analyzed based on the battery capacity estimation result of the energy storage battery, which can be used as a reference index for maintenance and replacement of the energy storage battery.

[0088] This invention takes battery charging voltage data as input, then automatically extracts battery aging features through a convolutional neural network layer to generate a feature map (time-series feature map), and then predicts the battery capacity estimate based on the feature map through a recurrent neural network layer. The structure of convolutional neural network layer + recurrent neural network layer realizes the direct estimation of battery capacity through battery charging voltage data, that is, it can realize end-to-end energy storage battery capacity estimation, thereby reducing the processing difficulty of energy storage battery data, and helping to ensure that the energy storage battery capacity estimation reaches the optimal level throughout the entire life cycle of the energy storage battery and under different aging paths, thus improving the practicality of energy storage battery capacity estimation.

[0089] Meanwhile, this invention sets up an attention mechanism layer between the convolutional neural network (CNN) layer and the recurrent neural network (RNN) layer. This layer sequentially extracts channel attention weights and spatial attention weights from the initial feature map and generates an attention feature map. On one hand, the attention mechanism layer weights the output channels of each convolutional layer in the CNN and assigns weights to the initial feature map. By obtaining the importance of each channel in the initial feature map, it assigns weights to each feature, allowing subsequent deep neural networks to focus on certain feature channels. Furthermore, it utilizes the spatial relationships between features to generate a weighted feature map, emphasizing strongly correlated features and mitigating the problem of deep neural networks easily getting trapped in local optima. This improves the effectiveness of the feature map and the accuracy of energy storage battery capacity estimation. On the other hand, this invention uses the weighted feature map generated by the attention mechanism layer as input to the RNN layer for training, avoiding the complex manual extraction of battery aging feature data as model input. By weighting the extracted features, it further alleviates the vanishing and diffusion problems of gradients in the hidden layers of the RNN, obtaining the optimal weights for each hidden layer of the deep neural network. This improves the performance of the RNN layer and the accuracy of energy storage battery capacity estimation.

[0090] In this embodiment, the battery capacity estimation model is a two-dimensional attention mechanism one-dimensional convolutional-bidirectional long short-term memory neural network model (hereinafter referred to as 1DCNN-CBAM-BiLSTM).

[0091] In practice, the convolutional neural network layer is built based on a one-dimensional convolutional neural network (1DCNN). The convolutional neural network layer uses a one-dimensional convolution kernel of a specified size to perform one-dimensional convolution operations on the multi-channel one-dimensional input data.

[0092] A one-dimensional convolutional neural network generates an initial feature map using the following formula:

[0093]

[0094] In the formula: represents the output of the input data corresponding to the kth convolution kernel, and the output of each convolution kernel of the one-dimensional convolutional neural network constitutes an initial feature map; w (k) represents the weight matrix of the kth convolution kernel; x c represents the sequence of the original input data; L represents the size of the convolution kernel; i represents the receptive field of the convolution kernel; and t represents the length of the sequence.

[0095] In the implementation process, the attention mechanism layer includes a channel attention module (CAM) and a spatial attention module (SAM) for extracting channel attention weights and spatial attention weights, respectively; wherein the channel attention module and the spatial attention module can be embedded in different convolution layers of the one-dimensional convolutional neural network.

[0096] As shown in Figure 3 , the weighted feature map is generated by the following steps:

[0097] S201: extracting channel attention weights of the initial feature map through the channel attention module;

[0098] S202: multiplying the initial feature map by the channel attention weights to obtain a channel attention feature map;

[0099] S203: extracting spatial attention weights of the channel attention feature map through the spatial attention module;

[0100] S204: multiplying the initial feature map by the spatial attention weights to obtain a weighted feature map.

[0101] Specifically, the channel attention module performs a tensor mean operation on the output values of the output channels, and then normalizes the output values obtained by the tensor mean operation through an activation function, normalizes each channel attention weight of the feature map to between 0 and 1, and the normalized weight is the channel attention weight.

[0102] The channel attention module extracts the channel attention weights by the following formula:

[0103] M c (F)=softmax(Mean(F cout ));

[0104]

[0105]

[0106] In the formula, M c (F) represents the channel attention weight; F represents the initial feature map output by the convolutional neural network layer; softmax represents an activation function for normalizing the output value; Mean represents a tensor mean operation; and F coutrepresents the feature on the output channel; z j represents the output value of the jth node; c represents the number of output nodes; z c represents the output value of the cth node; c out represents the output value of the output channel; k represents the number of convolution kernels of the convolutional neural network layer, i.e., the number of output channels.

[0107] The channel attention feature map is calculated by the following formula:

[0108]

[0109] In the formula: F' represents the channel attention feature map; M c (F) represents the channel attention weight; F represents the initial feature map.

[0110] The spatial attention module performs spatial domain processing on the channel attention feature map. First, the input feature map is subjected to maximum pooling and average pooling in the channel dimension, the two feature maps after pooling are stacked in the channel dimension, then the channel information is extracted by the convolution kernel, the channel number of the feature map is fused to 1, finally, the result after convolution is subjected to sigmoid function to normalize the spatial weight of the feature map. The normalized weight is the spatial attention weight.

[0111] The spatial attention module extracts the spatial attention weight by the following formula:

[0112]

[0113] In the formula: M s (F') represents the spatial attention weight; F represents the initial feature map. n×n represents the convolution operation with the kernel n; F' represents the channel attention feature map; AvgPool(F') represents the average pooling of the channel attention feature map; maxPool(F') represents the maximum pooling of the channel attention feature map; σ represents the sigmoid activation function; [;] represents the connection of the features of two channels; and represent the feature maps after average pooling and maximum pooling of the channel attention feature map, respectively.

[0114] The weighted feature map is calculated by the following formula:

[0115]

[0116] In the formula: F'' represents the weighted feature map; M s (F') represents the spatial attention weight; F represents the initial feature map.

[0117] The attention mechanism layer of the application can weight the output channels of each convolution layer of the convolutional neural network layer and assign a value to the initial feature map, and then assign a weight value to each feature by obtaining the importance of each channel of the initial feature map, so that the subsequent deep neural network can focus on certain feature channels, and the spatial relationship between the features is used to generate a weighted feature map, so that the strong correlation features can be emphasized on the feature, thereby the problem that the deep neural network is easy to fall into local optimum can be relieved, and the effectiveness of the feature map can be further improved.

[0118] In the implementation process, the recurrent neural network layer includes a bidirectional long short-term memory neural network (BiLSTM), and a full connection layer connected with the output of the bidirectional long short-term memory neural network;

[0119] The bidirectional long short-term memory neural network includes a forgetting gate, an input gate and an output gate.

[0120] The forgetting gate is represented by the following formula:

[0121] f t =σ(W f h t-1 +W f x t +b f );

[0122] In the formula, f t represents the output of the forgetting gate; sigma represents a sigmoid activation function; W f represents the forgetting gate weight; h t-1 represents the hidden layer information at t-1; x t represents the input feature at t; and b f represents the forgetting gate bias.

[0123] The input gate is represented by the following formula:

[0124] i t =σ(W i h t-1 +W i x t +b i );

[0125]

[0126]

[0127] In the formula, i t represents the control reserved part of information of the input gate; sigma represents a sigmoid activation function; W i and W c represent the input gate weight; h t-1represents the hidden layer information at time t-1; x t represents the input feature at time t; b i , b c represents the input gate bias; represents the sub-cell unit state candidate value, i.e. the information brought by the new input; c t represents the sub-cell unit state at time t, i.e. the updated information; tanh represents the hyperbolic tangent function; c t-1 represents the sub-cell unit state at time t-1;

[0128] The output gate is represented by the following formula:

[0129] o t = σ(W o h t-1 + W o x t + b o );

[0130] h t = o t ⊙ tanh(c t );

[0131] In the formula: o t represents the output feature of the output gate; h t represents the hidden layer information at time t; h t-1 represents the hidden layer information at time t-1; σ represents the sigmoid activation function; W o represents the output gate weight; x t represents the input feature at time t; b o represents the output gate bias; tanh represents the hyperbolic tangent function; c t represents the sub-cell unit state at time t, i.e. the updated information;

[0132] The activation functions of the bidirectional long short-term memory neural network are the sigmoid function and the tanh function, and the formulas are respectively:

[0133]

[0134]

[0135] The output of the fully connected layer is represented by the following formula:

[0136] y = aA T + b;

[0137] In the formula: y represents the output of the fully connected layer, i.e. the battery capacity estimation value; a represents the number of neurons; A represents the weight matrix; T represents the transpose symbol; and b represents the bias.

[0138] In the implementation process, the loss function of the recurrent neural network layer adopts the following mean square error loss function (MSE);

[0139]

[0140] In the formula: y i represents the battery capacity target value at time i; represents the battery capacity estimation value at time i; N represents the sample quantity;

[0141] The optimizer of the recurrent neural network layer adopts the following adaptive moment estimation (Adam) algorithm;

[0142] m t = beta1m t-1 + (1-beta1)g t

[0143]

[0144]

[0145]

[0146] In the formula: m t and m t-1 represent the smoothed moving average at iteration t times and t-1 times; t represents the iteration number; beta1 and beta2 represent the smoothing constant; g t represents the gradient of the target function; V t represents the moving average of the square of the gradient; w t and w t+1 respectively represent the updated parameters at iteration t times and t+1 times; alpha t represents the learning rate at iteration t times; epsilon = 10 -8 represents the divisor 0; and represent the t power of the smoothing constant.

[0147] The application constructs a recurrent neural network layer based on a long short-term memory neural network and a full connection layer, adopts a mean square error loss function as a loss function, and adopts an adaptive moment estimation algorithm as an optimizer, so that the battery capacity estimation model can realize accurate estimation of the battery capacity under the whole life cycle and different aging paths, thereby further improving the practicality of the energy storage battery capacity estimation. At the same time, the long short-term memory neural network takes the weighted feature map as the input for training, avoids complex manual extraction of battery aging feature data as the model input, and further alleviates the problems of gradient disappearance and dispersion of the recurrent neural network hidden layer by weighting the extracted features, can obtain the optimal weight of each hidden layer of the deep neural network, and thereby can further improve the accuracy of the energy storage battery capacity estimation.

[0148] In order to better illustrate the advantages of the technical scheme of the present application, the following experiments are disclosed in this embodiment.

[0149] Figure 4 The battery capacity estimation model proposed by the present application, i.e. the battery capacity estimation value of the two-dimensional attention mechanism one-dimensional convolution-bidirectional long short-term memory neural network model (1DCNN-CBAM-BiLSTM), is compared with the battery capacity estimation value of the existing convolution-long short-term memory neural network (CNN-LSTM) and the experimental value.

[0150] Through Figure 4 It can be seen that the battery capacity estimation model (1DCNN-CBAM-BiLSTM) of the present application has a smoother estimation result compared with the existing convolution-long short-term memory neural network (CNN-LSTM). Under the early use cycle of the battery, the capacity estimation results of the two models have no great difference, but when the battery has experienced a certain number of cycles, i.e. after the battery is aged, the capacity degradation is more nonlinear, the battery capacity estimation model (1DCNN-CBAM-BiLSTM) can deeply extract the battery capacity features, and ensure high capacity estimation accuracy.

[0151] Figure 5 The battery capacity estimation error analysis of the battery capacity estimation model (1DCNN-CBAM-BiLSTM) and the existing convolution-long short-term memory neural network (CNN-LSTM) under the condition of the cycle aging experiment of the energy storage battery is given.

[0152] Among them, Figure 5 (a) is the root mean square error statistics of the capacity estimation of the two models, the root mean square error of the battery capacity estimation of the battery capacity estimation model (1DCNN-CBAM-BiLSTM) is 0.0015, which is 36% lower than the root mean square error of the battery capacity estimation of the existing convolution-long short-term memory neural network (CNN-LSTM).

[0153] Figure 5 (b) is the average absolute error statistics of the capacity estimation of the two models, the root mean square error of the battery capacity estimation of the battery capacity estimation model (1DCNN-CBAM-BiLSTM) is 0.0011, which is 41.6% lower than the root mean square error of the battery capacity estimation of the existing convolution-long short-term memory neural network (CNN-LSTM).

[0154] The simulation results show that under the condition of the cycle aging experiment of the energy storage battery, the battery capacity estimation model (1DCNN-CBAM-BiLSTM) of the present application is superior to the existing neural network model in overall estimation accuracy.

[0155] Finally, it needs to be explained that the above examples are only used to illustrate the technical solutions of the present application but not to limit the technical solutions, and those of ordinary skill in the art should understand that the technical solutions of the present application are modified or equivalently replaced without departing from the purpose and scope of the technical solutions, which should be covered in the scope of claims of the present application.

Claims

1. An end-to-end capacity estimation method for energy storage batteries based on hybrid deep neural networks, characterized in that, include: S1: Obtain the battery charging voltage data of the energy storage battery; S2: Input the battery charging voltage data of the energy storage battery into the trained battery capacity estimation model, and output the corresponding battery capacity estimate. Battery capacity estimation models include: Convolutional neural network layers are used to extract battery aging features from the input data and generate corresponding initial feature maps; The attention mechanism layer is used to sequentially extract the channel attention weights and spatial attention weights of the initial feature map and generate the corresponding weighted feature map; The attention mechanism layer includes a channel attention module and a spatial attention module, which are used to extract channel attention weights and spatial attention weights, respectively. The weighted feature map is generated using the following steps: S201: Extract the channel attention weights of the initial feature map through the channel attention module; S202: Multiply the initial feature map by the channel attention weights to obtain the channel attention feature map; S203: Extract spatial attention weights from the channel attention feature map using the spatial attention module; S204: Multiply the initial feature map by the spatial attention weights to obtain a weighted feature map; The attention mechanism layer can assign weights to the output channels of each convolutional layer in the convolutional neural network and assign weights to the initial feature map. By obtaining the importance of each channel in the initial feature map, weights are assigned to each feature, allowing the subsequent deep neural network to focus on certain feature channels. Then, the spatial relationship between features is used to generate a weighted feature map, which can emphasize strongly related features and alleviate the problem that deep neural networks are prone to getting trapped in local optima. The recurrent neural network layer is used to train the system by taking the weighted feature map generated by the attention mechanism layer as input and outputting the corresponding battery capacity estimate. S3: Use the estimated battery capacity of the energy storage battery as its battery capacity estimation result.

2. The end-to-end capacity estimation method for energy storage batteries based on hybrid deep neural networks as described in claim 1, characterized in that: In step S2, the convolutional neural network layer is constructed based on a one-dimensional convolutional neural network.

3. The end-to-end capacity estimation method for energy storage batteries based on hybrid deep neural networks as described in claim 2, characterized in that: A one-dimensional convolutional neural network generates an initial feature map using the following formula: In the formula: This represents the output corresponding to the k-th convolutional kernel of the input data. The outputs of each convolutional kernel in a one-dimensional convolutional neural network constitute the initial feature map; w (k) x represents the weight matrix of the k-th convolutional kernel; c L represents the sequence of the original input data; L represents the kernel size; i represents the receptive field of the kernel; t represents the sequence length.

4. The end-to-end capacity estimation method for energy storage batteries based on hybrid deep neural networks as described in claim 1, characterized in that: In step S201, the channel attention module extracts the channel attention weights using the following formula: M c (F)=softmax(Mean(F cout )); Where: M c (F) represents the channel attention weight; F represents the initial feature map of the convolutional neural network layer output; softmax represents the normalization operation of the activation function on the output value; Mean represents the tensor mean operation; F cout Indicates the characteristics on the output channel; z j represents the output value of the j-th node; c represents the number of output nodes; z c This represents the output value of the c-th node; C out represents the output value of the output channel; k represents the number of convolutional kernels in the convolutional neural network layer, i.e., the number of output channels.

5. The end-to-end capacity estimation method for energy storage batteries based on hybrid deep neural networks as described in claim 1, characterized in that: In step S202, the channel attention feature map is calculated using the following formula: In the formula: F′ represents the channel attention feature map; M c (F) represents the channel attention weight; F represents the initial feature map.

6. The end-to-end capacity estimation method for energy storage batteries based on hybrid deep neural networks as described in claim 1, characterized in that: In step S202, the spatial attention module extracts the spatial attention weights using the following formula: Where: M s (F′) represents the spatial attention weight; f n×n represents a convolution operation with kernel n; F′ represents the channel attention feature map; AvgPool(F′) represents average pooling of the channel attention feature map; maxPool(F′) represents max pooling of the channel attention feature map; σ represents the sigmoid activation function; [;] represents concatenating the features of two channels; and These represent the feature maps after average pooling and max pooling, respectively, applied to the channel attention feature maps.

7. The end-to-end capacity estimation method for energy storage batteries based on hybrid deep neural networks as described in claim 1, characterized in that: In step S204, the weighted feature map is calculated using the following formula: In the formula: F″ represents the weighted feature map; M s (F′) represents the spatial attention weight; F represents the initial feature map.

8. The end-to-end capacity estimation method for energy storage batteries based on hybrid deep neural networks as described in claim 1, characterized in that: In step S2, the recurrent neural network layer includes a bidirectional long short-term memory neural network and a fully connected layer connected to the output of the bidirectional long short-term memory neural network. Bidirectional long short-term memory (LSSM) neural networks include a forget gate, an input gate, and an output gate. The forget gate is represented by the following formula: f t =σ(W f h t-1 +W f x t +b f ); In the formula: f t The output of the forget gate is represented by σ; the sigmoid activation function is represented by W. f Indicates the forget gate weight; h t-1 Represents the hidden layer information at time t-1; x t b represents the input features at time t; f This indicates forgetting gate bias; The input gate is represented by the following formula: i t =σ(W i h t-1 +W i x t +b i ); In the formula: i t This indicates that the input gate retains some control information; σ represents the sigmoid activation function; W i and W c Indicates the input gate weights; h t-1 Represents the hidden layer information at time t-1; x t This represents the input features at time t; b i b c Indicates input gate deviation; This represents the candidate state value of the sub-cell unit, i.e., the information brought by the new input; c t The expression represents the state of the sub-cell unit at time t, i.e., the updated information; tanh represents the hyperbolic tangent function; c t-1 This indicates the state of the subcell unit at time t-1; The output gate is represented by the following formula: o t =σ(W o h t-1 +W o x t +b o ); h t =o t ⊙tanh(c t ); In the formula: o t Indicates the output characteristics of the output gate; h t Represents the hidden layer information at time t; h t-1 Represents the hidden layer information at time t-1; σ represents the sigmoid activation function; W o Indicates the output gate weight; x t This represents the input features at time t; b o Represents the output gate bias; tanh represents the hyperbolic tangent function; c t This represents the state of the sub-cell unit at time t, i.e., the updated information; The output of a fully connected layer can be expressed by the following formula: y=aA T +b; In the formula: y represents the output of the fully connected layer, i.e., the battery capacity estimate; a represents the number of neurons; A represents the weight matrix; T represents the transpose sign; b represents the bias.

9. The end-to-end capacity estimation method for energy storage batteries based on hybrid deep neural networks as described in claim 8, characterized in that: The loss function for the recurrent neural network layer is the following mean squared error loss function; In the formula: y i This represents the target battery capacity at time i; Let represent the estimated battery capacity at time i; N represents the sample size. The optimizer for the recurrent neural network layer uses the following adaptive moment estimation algorithm; m t =β1m t-1 +(1-β1)g t Where: m t and m t-1 The moving averages after smoothing for iterations t and t-1 are represented by β1 and β2, respectively; t represents the number of iterations; β1 and β2 represent smoothing constants; g t V represents the gradient of the objective function; t w represents the moving average of the squared gradient. t and w t+1 These represent the update parameters for iterations t and t+1, respectively; α t This represents the learning rate after t iterations; ε = 10 -8 This indicates that the division by zero should be avoided; and This represents the smoothing constant raised to the power of t.

Images