Apparatus and method for estimating state of charge of battery using neural network

[0001] 1. Field of the Invention

[0002] The present invention relates generally to an apparatus and a method for estimating a state of charge (SOC) of a battery, and more particularly to an apparatus and a method for estimating a state of charge (SOC) of a battery representing a non-linear characteristic by using a neural network.

[0003] 2. Description of the Related Art

[0004] In general, a state of charge (SOC) of a battery represents a non-linear characteristic, so it is difficult to precisely detect the SOC of the battery. Thus, a value of SOC of the battery is not detected but estimated. Particularly, the SOC of batteries used for hybrid electrical vehicles (HEVs) or electric vehicles (EVs) having high C-rates may represent strong non-linear characteristic, so it is almost impossible to precisely detect the SOC of such batteries.

[0005] Conventionally, an Ah-counting (ampere hour counting) scheme, an OCV (open circuit voltage) measurement scheme, or a battery impedance measurement scheme has been used for estimating the SOC of the battery.

[0006] First, the Ah-counting scheme estimates the SOC of the battery by detecting an actual capacity of the battery. In this case, an estimation value of the SOC of the battery may depend on a state of a sensor used for detecting the actual capacity of the battery. Thus, the estimation value of the SOC of the battery may vary depending on the degree of preciseness and error of the sensor.

[0007] The OCV measurement scheme estimates the SOC of the battery based on an open circuit voltage of the battery. In this case, however, the SOC of the battery can be estimated only under a resting state of the battery. In addition, the OCV measurement scheme is influenced by external environment, such as an external temperature.

[0008] The battery impedance measurement scheme estimates the SOC of the battery based on an impedance value of the battery. However, the battery impedance measurement scheme is significantly influenced by an external temperature, so that reliability for the estimation value of the SOC of the battery may be lowered.

[0009] Therefore, it is necessary to provide a method for precisely estimating the SOC of the battery while minimizing an estimation error thereof regardless of an external temperature.

[0010] The present invention is directed to an apparatus and a method for estimating an SOC of a battery that substantially obviate one or more problems due to limitations and disadvantages of the related art.

[0011] It is an object of the present invention to provide an apparatus and a method for precisely estimating an SOC of a battery by using a neural network.

[0012] Another object of the present invention is to provide an apparatus and a method capable of dynamically estimating an SOC of a battery by using a minimum of data under various environments including various temperatures and C-rates.

[0013] To achieve these objects and other advantages in accordance with the purpose of the invention, as embodied and broadly described herein, there is provided an apparatus for estimating a state of charge (SOC) of a battery, the apparatus comprising: a sensing section for detecting current, voltage and a temperature from a battery cell; a neural network performing a neural network algorithm and a learning algorithm based on data of the current, voltage and temperature transmitted thereto from the sensing section and present time data, thereby outputting the SOC of the battery estimated through a final learning algorithm; and a comparator for comparing an output value of the neural network with a predetermined target value and making the neural network iteratively perform the learning algorithm if a difference between the output value of the neural network and the predetermined target value is out of a predetermined allowable limit, and update the learning algorithm to generate the final learning algorithm.

[0014] According to the preferred embodiment of the present invention, the predetermined target value may be a “true” SOC of the battery. However, it is difficult to precisely calculate the predetermined target value, so a reference value obtained through performing an experiment under specific conditions is used as a target value. For example, the reference value is obtained based on a value of subtracting ampere-hour data of a charger/discharger from rated capacity of the battery, or an open circuit voltage value of the battery, respectively or mathematically compensating each other.

[0015] Although the present invention is described in relation to the neural network in the form of a dynamic multi-dimensional wavelet neural network algorithm, it is also possible to use a dynamic wavelet neural network algorithm or a static wavelet neural network algorithm. In addition, although the present invention is described in relation to a back-propagation learning algorithm, it is also possible to use a Kalman filter scheme, GA, or a fuzzy learning algorithm.

[0016] To achieve the above objects of the present invention, according to another aspect of the present invention, there is provided a method for estimating a state of charge (SOC) of a battery by using a neural network, the method comprising the steps of: (a) performing a learning algorithm based on data of the current, voltage and a temperature detected from the battery and present time data; and (b) outputting the SOC of the battery estimated through a final learning algorithm generated by performing the learning algorithm.

[0017] According to the preferred embodiment of the present invention, step (a) includes the substeps of detecting the current, voltage and temperature from the battery, performing a neural network algorithm by utilizing data of the current, voltage and temperature detected from the battery and present time data as learning data, checking whether a difference between an output value outputted through the neural network algorithm and a predetermined target value is within a predetermined allowable limit, and iteratively performing a learning algorithm if the difference is out of the predetermined allowable limit, thereby updating the learning algorithm to generate the final learning algorithm.

[0018] According to the preferred embodiment of the present invention, step (b) includes the substeps of detecting the current, voltage and temperature from the battery, and outputting the SOC of the battery by estimating the SOC of the battery through the final learning algorithm obtained through step (a) based on data of the current, voltage and temperature detected from the battery and present time data.

[0019] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows and in part will become apparent to those having ordinary skill in the art upon examination of the following or may be learned by practicing the invention. The objectives and other advantages of the invention may be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings.

[0025] Reference will now be made in detail to the preferred embodiment of the present invention, examples of which are illustrated in the accompanying drawings. Wherever possible, the same reference numbers will be used throughout the drawings to refer to the same or like parts.

[0026] Hereinafter, a preferred embodiment of the present invention will be explained in detail with reference to the accompanying drawings.

[0027] FIG. 1 is a block view schematically illustrating an apparatus 100 for estimating an SOC of a battery according to a preferred embodiment of the present invention.

[0028] Referring to FIG. 1, the apparatus 100 for estimating an SOC of a battery by using a neural network includes a current detecting unit 12 for detecting current (i) from a battery cell 10, a voltage detecting unit 14 for detecting voltage (V) from the battery cell 10, a temperature detecting unit 16 for detecting a temperature (T) from the battery cell 10, a neural network 20 for performing a neural network algorithm and a learning algorithm by utilizing learning data including the current (i), voltage (V) and temperature (T) detected by the above detecting units 12, 14 and 16 and time (K) and for outputting an estimated SOC based on a finally learned algorithm formula generated by completing the neural network algorithm and the learning algorithm, a charger/discharger 30 for feeding charging/discharging current to the battery cell 10, and a comparator 40 which compares an output SOC (g_o) of the battery outputted from the neural network 20 with a target SOC (g_T), checks whether a difference between the output SOC (g_o) and the target SOC (g_T) is within a predetermined allowable limit, and makes the neural network 20 iteratively perform the learning algorithm if the difference is out of the predetermined allowable limit, thereby update the learning algorithm to generate the final learning algorithm.

[0029] Preferably, the target SOC (g_T) is obtained through performing an experiment under specific conditions. For example, the target SOC (g_T) is obtained by subtracting Ah (ampere hour) data (g_r) of the charger/discharger 30 from rated capacity (g_N) of a battery, that is, g_T=g_N−g_r. This is because an ideal target SOC can be obtained by subtracting the Ah data corresponding to capacity that has already been used in the battery from the rated capacity of the battery, if the error is within the acceptable limit. Alternatively the target SOC (g_T) is obtained based on an open circuit voltage value of the battery. It is preferable to mathematically compensate the value of Ah-counting scheme and the value of OCV measurement scheme each other.

[0030] FIG. 2 is a view showing a structure of a neural network used for the present invention.

[0031] According to the preferred embodiment of the present invention, the neural network includes a dynamic multi-dimensional wavelet neural network.

[0032] Referring to FIG. 2, the dynamic multi-dimensional wavelet neural network includes an input field, a hidden layer and an output layer.

[0033] In order to apply a battery SOC estimation algorithm to the dynamic multi-dimensional wavelet neural network, an arbitrary function ƒ(x)εL²(R) can be resolved as follows based on a wavelet theory.

[0034] $\begin{array}{cc}f\left(x\right)=\sum _n{a}_n\phi \left(2^{m}x-n\right)& \mathrm{Equation}1\end{array}$

[0035] Herein, α_n is a coefficient of a basis function φ(•) for resolving the arbitrary function ƒ(x). In addition, 2^m and n are dilation and translation parameters for the basis function φ(•), respectively.

[0036] In addition, an approximation function g(x) for the arbitrary function ƒ(x) is presented as Equation 2.

[0037] $\begin{array}{cc}g\left(x\right)\approx \sum _n{a}_n\phi \left(2^{m}x-n\right)& \mathrm{Equation}2\end{array}$

[0038] Equation 2 can be applied to the dynamic multi-dimensional wavelet neural network shown in FIG. 2 as follows:

[0039] In FIG. 2, x_d(k) is an input data vector inputted into the dynamic multi-dimensional wavelet neural network. According to the present embodiment, x_d(k) is a vector including input data consisting of the current, voltage, and temperature inputted into the dynamic multi-dimensional wavelet neural network during a predetermined period of time (k). That is, x_d(k)=(i, v, T, k). In addition, g_o(x_d(k)) is an output value, which is estimated through the dynamic multi-dimensional wavelet neural network based on the input data. The g_o(x_d(k)) is presented as Equation 3.

[0040] $\begin{array}{cc}\begin{array}{c}{g}_o\left(x_d\left(k\right)\right)=\sum _n{W}_{\mathrm{no}}^o\phi \left[\sum _d{2}^m\left(x_d\left(k\right)-n\right)+W_n^D{X}_n\left(k-1\right)\right]\\ =\sum _n{W}_{\mathrm{no}}^o\phi [\sum _d{2}^m\left(x_d\left(k\right)-n\right)+\\ W_n^D\phi \left(\sum _d{2}^m\left(x_d\left(k-1\right)-n\right)+W_n^D{X}_n\left(k-2\right)\right)]\end{array}& \mathrm{Equation}3\end{array}$

[0041] Herein, W_n^D and W_no^o are coefficients representing connection weight, which is updated in every time (k) based on back-propagation (BP) learning algorithm so that the arbitrary function is approximately identified with a non-linear function.

[0042] If the difference between the output value g_o and the target value g_T detected by the comparator 40 of the dynamic multi-dimensional wavelet neural network does not exist within a predetermined allowable limit (for example, 3%), the following back-propagation learning algorithm is iteratively carried out by means of the neural network 20.

[0043] Prior to explaining the back-propagation (BP) learning algorithm, it is noted that an error function is defined as Equation 4.

[0044] $\begin{array}{cc}E=\frac{1}{2}{\left(g_T\left(k\right)-g_o\left(k\right)\right)}^2& \mathrm{Equation}4\end{array}$

[0045] Herein, g_T(k) is a desired output value, that is, the target value, and g_o(k) is an actual output value of the neural network 20. When replacing Equation 4 with g_T(k)−g_o(k)=e(k), an error gradient is obtained as represented in Equation 5.

[0046] $\begin{array}{cc}\frac{\partial E}{\partial W}=e\left(k\right)\frac{\partial g\left(k\right)}{\partial W}& \mathrm{Equation}5\end{array}$

[0047] Also, an error gradient of W_n^D, W_no^o is obtained as represented in Equation 6.

[0048] $\begin{array}{cc}\begin{array}{cc}\left(1\right)& \frac{\partial g_o\left(k\right)}{\partial W_{\mathrm{no}}^o}=X_n\left(k\right)\end{array}\begin{array}{cc}\left(2\right)& \frac{\partial g_o\left(k\right)}{\partial W_n^D}=\sum _o{W}_{\mathrm{no}}^o{P}_n\left(k\right)\end{array}\begin{array}{cc}\left(3\right)& P_n\left(k\right)=\phi \left(S_n\left(k\right)\right)\left(X_n\left(k-1\right)+W_n^D{P}_n\left(k-1\right)\right),P_n\left(0\right)=0\end{array}& \mathrm{Equation}6\end{array}$

[0049] Therefore, a final update for the connection weight is carried out as follows:

[0050] $\begin{array}{cc}W\left(t+1\right)=W\left(t\right)+\eta \left(-\frac{\partial E}{\partial W}\right),\mathrm{herein},\eta \mathrm{is}a\mathrm{learning}\mathrm{rate}.& \mathrm{Equation}7\end{array}$

[0051] In this way, the neural network 20 iteratively sends a new output value g_o, which is newly generated based on updated values of W_n^D, W_no^o, and X_n(k), to the comparator 40 while iteratively performing the back-propagation learning algorithm. In addition, such a procedure is iteratively carried out until the difference between the output value g_o and the target value g_T is defined within the predetermined allowable limit.

[0052] When the difference between the output value g_o and the target value g_T is defined within the predetermined allowable limit, the learning algorithm by using the neural network is completed and the estimated SOC is outputted by using the final estimation algorithm formula (that is, Equation 3) obtained through the learning algorithm.

[0053] Hereinafter, a method for estimating the SOC of the battery will be described with reference to FIGS. 3 and 4.

[0054] The method for estimating the SOC of the battery according to the present invention includes a learning step by using the neural network, that is, a step of obtaining a final estimation algorithm by performing the neural network algorithm and the learning algorithm, and an SOC outputting step for outputting the SOC of the battery by using the final estimation algorithm.

[0055] FIG. 3 is a flowchart sequentially illustrating learning steps by using the neural network according to the preferred embodiment of the present invention.

[0056] Referring to FIG. 3, the current (i), voltage (v) and temperature (T) are detected from the battery cell 10 (step 10). Then, the dynamic multi-dimensional wavelet neural network algorithm is carried out by utilizing the detected current (i), voltage (v) and temperature (T) and time data (k) thereof as an input data vector, that is, x_d(k)=(i, v, T, k). As a result, an output value g_o can be obtained (step 12).

[0057] After that, the output value g_o is compared with the target value g_T, and it is checked whether the difference between the output value g_o and the target g_T is within the predetermined error limit of 3% (step 14). Although the predetermined allowable error limit is set to 3% in the present embodiment, the predetermined error limit can be changed, if necessary. The SOC of the battery can be precisely estimated as the allowable error limit becomes reduced. In contrast, the SOC of the battery may be inaccurately estimated as the allowable error limit becomes enlarged.

[0058] If it is determined in step 14 that the difference exceeds the predetermined allowable error limit, the back-propagation learning algorithm is carried out, thereby obtaining an updated output value g_o (step 16). Then, the procedure returns to step 14.

[0059] In the meantime, if it is determined in step 14 that the difference is equal to or less than the predetermined allowable error limit, the neural network learning algorithm is completed (step 18). As a result, the final estimation algorithm formula (that is, Equation 3) can be obtained through the learning algorithm.

[0060] FIG. 4 is a flowchart illustrating a procedure for outputting the SOC of the battery by utilizing the final estimation algorithm obtained through the learning steps shown in FIG. 3.

[0061] Referring to FIG. 4, the current (i), voltage (v) and temperature (T) are detected from the battery cell 10 (step 20). Then, the SOC of the battery is outputted by utilizing the final estimation algorithm obtained through the learning algorithm shown in FIG. 3, while employing the detected current (i), voltage (v) and temperature (T) and time data (k) thereof as an input data vector, that is, x_d(k)=(i, v, T, k) (step 22).

[0062] Although the present invention has been described in relation to the neural network in the form of the dynamic multi-dimensional wavelet neural network, the present invention is not exclusively limited thereto. That is, the present invention is applicable for a feedforward neural network, recurrent neural network, a wavelet neural network, etc.

[0063] In addition, although the present invention employs the back-propagation learning algorithm as a learning algorithm, the present invention is not exclusively limited thereto. For instance, the present invention can employ a conventional Kalman filter scheme, GA, and a fuzzy learning algorithm.

[0064] As described above, the present invention can dynamically estimate the SOC of the battery through the neural network algorithm and the learning algorithm. Particularly, the SOC of the battery can be precisely estimated even under various external environments including various temperatures and C-rates. The present invention is effectively available in a field of hybrid electrical vehicles, in which the SOC of the battery must be precisely estimated.

[0065] The forgoing embodiments are merely exemplary and are not to be construed as limiting the present invention. The present teachings can be readily applied to other types of apparatuses. The description of the present invention is intended to be illustrative, and not to limit the scope of the claims. Many alternatives, modifications, and variations will be apparent to those skilled in the art.