Apparatus and method for estimating a state of a battery

1 . An apparatus for estimating a state of a battery, the apparatus comprising:

a memory configured to store a program of a neural network including a plurality of pre-trained predictive models and an adaptive hidden layer; and

at least one processor configured to execute the program,

wherein the program includes an instruction for:

receiving battery data of a target battery,

inputting the battery data to the adaptive hidden layer, wherein the adaptive hidden layer is configured to select one predictive model from the plurality of pre-trained predictive models,

inputting the battery data to the selected predictive model and outputting prediction data for a remaining useful life of the target battery through the selected predictive model,

wherein each of the plurality of pre-trained predictive models is trained, in response to input data corresponding to a first cycle, to predict result data for a second cycle that is spaced apart from the first cycle by skipping a predetermined number of non-adjacent cycles (P), instead of predicting data for a cycle immediately following the first cycle, and

wherein the predetermined number of cycles (P) to be skipped is determined based on a correlation coefficient, the correlation coefficient being calculated based on the formula Correlation(Q i+T −Q i /T,Q i+T+P ), where Q is a capacity of data, i is a starting point, T is a total length of entire input data, and P is a difference between a cycle of the input data and a cycle of output data.

2 . The apparatus of claim 1 , wherein the adaptive hidden layer comprises:

identifying a similarity between the battery data and data corresponding to each of the plurality of pre-trained predictive models; and

learning so that a single predictive model to which data with a highest similarity to the battery data corresponds applies to the battery data, based on the identified similarity.

3 . The apparatus of claim 2 , wherein the adaptive hidden layer comprises applying a weight proportional to a similarity corresponding to each of the plurality of pre-trained predictive models, based on the identified similarity.

4 . The apparatus of claim 3 , wherein the applying of the weight proportional to the similarity corresponding to each of the plurality of pre-trained predictive models comprises:

calculating an average weight of at least one weight applied in advance and a matched weight, with respect to each of the plurality of pre-trained predictive models; and

applying the calculated average weight to each of the plurality of pre-trained predictive models.

5 . The apparatus of claim 4 , wherein the adaptive hidden layer comprises selecting the single predictive model having a highest average weight among the average weights with respect to each of the plurality of pre-trained predictive models and learning so that the single predictive model and the battery data match to each other.

6 . The apparatus of claim 1 , wherein the battery data includes battery data of a first cycle from charging to discharging of the target battery.

7 . The apparatus of claim 6 , wherein each of the plurality of pre-trained predictive models is trained to predict result data for a remaining useful life of a corresponding battery, based on an input of first battery data of a first cycle from charging to discharging of the corresponding battery, and

wherein predicting the result data for the remaining useful life of the corresponding battery comprises:

predicting, by each of the plurality of pre-trained predictive models, first data for a remaining useful life corresponding to a second cycle after skipping the predetermined number of cycles from the first cycle, based on the input of the first battery data of the first cycle from charging to discharging of the corresponding battery to each of the plurality of pre-trained predictive models; and

minimizing, by each of the plurality of pre-trained predictive models, a difference between the predicted first data and battery data of the second cycle of the corresponding battery.

8 . The apparatus of claim 1 , wherein the program further includes an instruction for predicting result data for a remaining useful life of a corresponding battery, and

wherein the predicting of the result data for the remaining useful life of the corresponding battery comprises predicting, by each of the plurality of pre-trained predictive models, second data for a remaining useful life corresponding to a third cycle after skipping the predetermined number of cycles from a second cycle, based on an input of first data to each of the plurality of pre-trained predictive models.

9 . The apparatus of claim 1 , wherein each of the plurality of pre-trained predictive models includes a long short-term memory (LSTM) model.

10 . A method for estimating a state of a battery, the method comprising:

receiving battery data of a target battery;

inputting the battery data to an adaptive hidden layer included in a neural network, wherein the adaptive hidden layer is configured to select a single predictive model from a plurality of pre-trained predictive models included in the neural network;

inputting the battery data to the selected single predictive model; and

outputting prediction data for a remaining useful life of the target battery through the selected single predictive model,

wherein each of the plurality of pre-trained predictive models is configured, in response to input data corresponding to a first cycle, to predict result data for a second cycle that is spaced apart from the first cycle by skipping a predetermined number of non-adjacent cycles (P), instead of predicting data for a cycle immediately following the first cycle, and

wherein the predetermined number of cycles (P) to be skipped is determined based on a correlation coefficient, the correlation coefficient being calculated based on the formula Correlation(Q i+T −Q i /T,Q i+T+P ), where Q is a capacity of data, i is a starting point, T is a total length of entire input data, and P is a difference between a cycle of the input data and a cycle of output data.

11 . The method of claim 10 , wherein the selecting of the single predictive model from the plurality of pre-trained predictive models included in the neural network comprises:

identifying, by the adaptive hidden layer, a similarity between the battery data and data corresponding to each of the plurality of pre-trained predictive models; and

learning so that the single predictive model to which data with a highest similarity to the battery data corresponds applies to the battery data, based on the identified similarity.

12 . The method of claim 11 , wherein the selecting of the single predictive model from the plurality of pre-trained predictive models included in the neural network comprises:

applying, by the adaptive hidden layer, a weight proportional to a similarity corresponding to each of the plurality of pre-trained predictive models, based on the identified similarity.

13 . The method of claim 12 , wherein the applying of the weight proportional to the similarity corresponding to each of the plurality of pre-trained predictive models comprises:

calculating an average weight of at least one weight applied in advance and a matched weight, with respect to each of the plurality of pre-trained predictive models; and

applying the calculated average weight with respect to each of the plurality of pre-trained predictive models.

14 . The method of claim 13 , wherein the selecting of the single predictive model from the plurality of pre-trained predictive models included in the neural network comprises:

learning, by the adaptive hidden layer, so that the single predictive model having a highest average weight among the average weights with respect to each of the plurality of pre-trained predictive models and the battery data match to each other.

15 . The method of claim 10 , wherein the battery data includes battery data of a first cycle from charging to discharging of the target battery.

16 . The method of claim 15 , wherein each of the plurality of pre-trained predictive models is trained to predict result data for a remaining useful life of a corresponding battery, based on an input of first battery data of a first cycle from charging to discharging of the corresponding battery, and

wherein predicting the result data for the remaining useful life of the corresponding battery comprises:

predicting, by each of the plurality of pre-trained predictive models, first data for a remaining useful life corresponding to a second cycle after skipping the predetermined number of cycles from the first cycle, based on the input of the first battery data of the first cycle from charging to discharging of the corresponding battery to each of the plurality of pre-trained predictive models; and

minimizing, by each of the plurality of pre-trained predictive models, a difference between the predicted first data and battery data of the second cycle of the corresponding battery.

17 . The method of claim 10 , further comprising predicting result data for a remaining useful life of a corresponding battery,

wherein the predicting of the result data for the remaining useful life of the corresponding battery comprises predicting, by each of the plurality of pre-trained predictive models, second data for a remaining useful life corresponding to a third cycle after skipping the predetermined number of cycles from a second cycle, based on an input of first data to each of the plurality of pre-trained predictive models.

18 . The method of claim 10 , wherein each of the plurality of pre-trained predictive models includes a long short-term memory (LSTM) model.