TWI883920B - Method for monitoring energy storage system and energy storage monitoring system - Google Patents

Abstract

This disclosure relates to a method for monitoring an energy storage system. The system includes: acquiring multiple parameters of the energy storage system at various time points to form a training dataset, dividing the training data into independent variable vectors and dependent variable vectors; training a machine learning model based on the independent variable vectors and dependent variable vectors, and calculating corresponding similarities, setting a performance index control value based on these similarities; acquiring new operational data, inputting the new operational data into the machine learning model to obtain predicted data, calculating a similarity between the predicted data and a ground truth; and if the similarity is less than the performance index control value, then determining that the energy storage system is abnormal.

Description

Energy storage system monitoring method and energy storage monitoring system

The present disclosure relates to a monitoring method for an energy storage system, which can integrate multiple parameters of the energy storage system to calculate the performance index of the energy storage system, thereby determining whether there is an abnormality.

In order to increase renewable energy and reduce carbon emissions, countries around the world have established energy storage systems. Currently, lithium battery packs in energy storage systems are equipped with battery management systems for monitoring and protection. When an abnormality occurs in the battery pack operation, the battery pack operation can be stopped immediately. Despite this, there have been many battery energy storage system anomalies in recent years, partly because the battery management system only stops the battery pack operation when a battery parameter exceeds the operating range, which is too late for highly active lithium batteries. Take temperature parameters as an example. When an abnormality occurs in a lithium battery, the temperature rises very quickly, and stopping the battery pack operation cannot prevent the lithium battery from burning. Therefore, how to comprehensively consider various parameters of the energy storage system and detect anomalies early has become a topic of concern to technicians in this field.

The embodiment of the present disclosure provides a monitoring method for an energy storage system, which is applicable to a computer system. The monitoring method includes: obtaining a plurality of parameters of the energy storage system at m time points as a training data set, wherein m is a positive integer; obtaining the parameters corresponding to d time points as d independent variable vectors, and obtaining the parameters corresponding to the remaining (m-d) time points as (m-d) dependent variable vectors, wherein d is a positive integer, and the positive integer d is less than the positive integer m; according to the d independent variable vectors and the (m-d) dependent variable vectors, A machine learning model is trained using a variable vector, and corresponding (m-d) similarities are calculated, and a performance indicator control value is set according to the (m-d) similarities; a new operation data is obtained, and the new operation data is input into the machine learning model to obtain a predicted data, and a current similarity value between the predicted data and a real data is calculated; and if the current similarity value is less than the performance indicator control value, it is determined that the energy storage system has been abnormal.

In some embodiments, the above training dataset is represented as the following matrix.

in represents the i-th parameter at the j-th time point, i, j, n, m are positive integers and, , When training the machine learning model, the model outputs (md) predicted strain vectors, and the (md) similarities are calculated according to the following mathematical formula.

where i=1,2,…,n, k=1,2,…,(md), represents the kth similarity among (md) similarities, is the i-th element of the k-th strain vector among the (md) strain vectors, is the i-th element of the k-th predicted strain variable vector among the (md) predicted strain variable vectors.

In some embodiments, the above parameters include current, power consumption, power generation, frequency, power or voltage.

In some embodiments, the monitoring method further includes: inputting the real data into an inverse function to obtain an expected operation data; and calculating the difference between the expected operation data and the new operation data to determine whether one of the parameters is abnormal.

In some embodiments, the machine learning model is an extreme gradient boosting model, a random forest model or a deep learning model.

The embodiment of the present disclosure provides an energy storage monitoring system, including: an energy storage system; and a computer system, communicatively connected to the energy storage system, for executing the above-mentioned monitoring method.

The terms “first,” “second,” etc. used herein do not particularly refer to order or sequence, but are only used to distinguish elements or operations described with the same technical term.

FIG1 is a schematic diagram of a monitoring system according to an embodiment. Referring to FIG1 , the energy storage monitoring system 100 includes a computer system 110 and an energy storage system 120, which are communicatively connected to each other. The computer system 110 is, for example, a desktop computer, a server, an industrial computer, an embedded system, or any electronic device with computing capabilities. The energy storage system 120 may include a battery pack and related circuits, such as an inverter, a transformer, a protection circuit, and a power-off device, and the battery pack may be a nickel-cadmium (NiCd) battery, a nickel-metal hydride (NiMH) battery, a lithium-ion (Li-ion) battery, a lithium-polymer (LiPo) battery, etc., but the present disclosure is not limited thereto. In some embodiments, the energy storage system 120 may also include a renewable energy module, such as a solar module, a wind power module, etc.

The energy storage system 120 has a plurality of sensors, such as ammeters, voltmeters, thermometers, power meters, etc., for measuring a plurality of parameters, which may include current, power consumption, power generation, frequency, power or voltage, etc. For example, if the energy storage system 120 has a three-phase circuit, the above parameters may include current, voltage, power, etc. on each phase. In some embodiments, the above parameters may also include power consumption and power generation accumulated over a period of time (e.g., 15 minutes). The above parameters may also include instantaneous system frequency, instantaneous total power factor, output (output) virtual work, and output (input) real work. In some embodiments, the above parameters may also include the highest (lowest) battery cell voltage, the highest (lowest) battery cell temperature, etc. These parameters can be transmitted to the computer system 110 via wired or wireless communication means.

The computer system 110 will execute a monitoring method based on these parameters to determine whether there is an abnormality in the energy storage system 120. Here, a machine learning method is used to determine the abnormality, so it can be divided into a training phase and a testing phase. The following will be divided into different phases for detailed description.

FIG2 is a flowchart illustrating a training phase according to an embodiment. Referring to FIG2, first in step 201, multiple parameters generated by the energy storage system at multiple time points are obtained, and these parameters have been described above. These parameters are used as a training data set and are represented as the following matrix X.

in represents the i-th parameter at the j-th time point. There are m time points in total, each with n parameters, i, j, n, m are positive integers and , . The size of the matrix X is , which means that the matrix X contains m vectors.

In step 202, the parameters corresponding to d time points are taken from the training data set as d independent variable vectors, where d is a positive integer and d < m. These independent variable vectors form a Matrix D. In step 203, the parameters corresponding to the remaining (md) time points are obtained as (md) strain vectors. These strain vectors form a matrix of size Matrix L. In some embodiments, multiple independent variable vectors are obtained at intervals and matched with a strain vector. For example, for every 10 time points, the parameters corresponding to the first 9 time points can be set as 9 independent variable vectors, and the parameters corresponding to the last time point are set as 1 strain vector. If m=1000, there are a total of d=900 independent variable vectors forming the matrix D , and (md)=100 strain vectors forming the matrix L. However, the present disclosure does not limit which time points are to be obtained as independent variable vectors (or strain vectors).

In step 204, a machine learning model is trained based on d independent variable vectors and (md) strain variable vectors. The machine learning model can be any suitable machine learning model such as a linear matrix, an extreme gradient boosting (XGBooste) model, a random forest model, a deep learning model, a neural network, etc., and the present disclosure is not limited thereto. Here, the matrix D is used as the input of the machine learning model, and the matrix L is used as the output of the machine learning model, but the machine learning model only predicts one vector at a time. For example, continuing the above example, every 9 independent variable vectors can be used to predict 1 strain variable vector, which is regarded as a training sample. In the above example, there are a total of (md) = 100 training samples. Therefore, there is a temporal relationship between the input and output of the machine learning model. In this embodiment, parameters that occurred earlier are used to predict parameters that have not yet occurred.

The above matrix L can be used as the ground truth, and the machine learning model outputs the predicted strain vector. Next, in step 205, the similarity between the ground truth and the predicted strain vector can be calculated for each training sample, as shown in the following mathematical formula 1. [Mathematical formula 1]

Where i=1,2,…,n, k=1,2,…,(md). represents the kth similarity, is the i-th element of the k-th strain variable vector in the matrix L , is the i-th element of the k-th predicted strain vector. In other words, the inverse of the L2 norm of the difference between the strain vector and the predicted strain vector is used as the similarity. The greater the similarity, the more accurate the prediction. Since each training sample can calculate the corresponding similarity, a total of (md) similarities are calculated here. .

Next, in step 206, a performance index control value is set according to the (m-d) similarities calculated above. For example, the minimum value, the minimum tenth value, or the minimum 1% value among the (m-d) similarities can be set as the performance index control value. This performance index control value can be used as an indicator for determining whether the energy storage system 120 is abnormal.

FIG3 is a flow chart illustrating the test phase according to an embodiment. Referring to FIG3, in step 301, new operation data is obtained, which is represented as a vector , which also contains the n parameters mentioned above. In step 302, the new operation data is input into the trained machine learning model to obtain the prediction data, which is represented by the vector (including n predicted values). Since the machine learning model predicts the parameters of the next time point, the corresponding real data (also with n parameters) can be collected after a period of time. In step 303, the predicted data is calculated according to the above mathematical formula 1 The similarity between the current data and the corresponding real data (referred to as the current similarity). In step 304, it is determined whether the current similarity is less than the performance indicator control value calculated above. If the current similarity is greater than or equal to the performance indicator control value, then return to step 301 to obtain the next new operation data. If the current similarity is less than the performance indicator control value, in step 305, it is determined that the energy storage system 120 is abnormal.

FIG4 is a schematic diagram showing experimental data according to an embodiment. The horizontal axis of FIG4 represents time, and the vertical axis represents predicted data. When the similarity is less than the performance index control value 410, it indicates that the energy storage system 120 is abnormal. The reason for this is that when the energy storage system 120 is normal (or healthy), it is easy to predict the parameters at the next time point. When the system is abnormal, the accuracy of the prediction will decrease, and the calculated similarity will decrease.

Next, it can be determined which device has the abnormality. In step 306, the real data corresponding to the predicted data is input into the inverse function to obtain the expected operation data. Specifically, the operation during the training phase can be expressed as ,in is the machine learning model mentioned above, and the inverse function is expressed as For example, if the machine learning model is a linear matrix, the inverse function can be the inverse matrix of the linear matrix, or the machine learning model can be a cycle generative adversarial network (cycle GAN), then there are two generators in the network that can serve as and . Assume that the forecast data The corresponding real data is represented as y, then the above operation can be expressed as ,in The expected operation data.

In step 307, the expected operation data is calculated With new operation data To determine which parameter is abnormal. Specifically, the expected operation data With new operation data Both contain parameters at multiple time points, where the difference between corresponding parameters can be calculated, for example, the difference between the i-th parameter at the t-th time point in two data is calculated. If this difference is greater than a critical value, it means that the corresponding device is abnormal, where t is a positive integer. Please refer to the following Table 1 for experimental data. New operation data Expected operation data Parameter 1 2328 1904 Parameter 2 2377 1951 Parameter 3 2304 1847 Table 1

For example, the above parameters 1 to 3 are the current of each phase in the three-phase circuit. The actual measured values (new operation data) are 2328, 2377, and 2304 mA, respectively, but the expected values (expected operation data) calculated by the inverse function are 1904, 1951, and 1847 mA, respectively. When the difference between the actual measured value and the expected value is greater than a critical value, it means that the corresponding three-phase circuit has an abnormality.

Through the above-mentioned method, the parameters of the energy storage system 120 can be comprehensively considered, and the possible parameters where abnormalities may occur can be found through the inverse function.

Although the present invention has been disclosed as above by the embodiments, they are not intended to limit the present invention. Any person with ordinary knowledge in the relevant technical field can make some changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be defined by the scope of the attached patent application.

100: Energy storage monitoring system 110: Computer system 120: Energy storage system 201-206, 301-307: Steps 410: Performance index control value

In order to make the above features and advantages of the present invention more clearly understandable, the following examples are specifically cited and detailed descriptions are given in conjunction with the attached figures. FIG. 1 is a schematic diagram of a monitoring system according to an embodiment. FIG. 2 is a flow chart of a training phase according to an embodiment. FIG. 3 is a flow chart of a test phase according to an embodiment. FIG. 4 is a schematic diagram of experimental data according to an embodiment.

301~307: Steps

Claims (10)

A monitoring method for an energy storage system is applicable to a computer system, and the monitoring method includes: Obtaining multiple parameters of the energy storage system at m time points as a training data set, where m is a positive integer; Obtaining the parameters corresponding to d time points among the time points as d independent variable vectors, and obtaining the parameters corresponding to the remaining (m-d) time points as (m-d) strain variable vectors, where d is a positive integer and d is less than m; Training a machine learning model based on the d independent variable vectors and the (m-d) strain variable vectors, and calculating the corresponding (m-d) similarities, and setting a performance indicator control value based on the (m-d) similarities; Obtain a new operating data, input the new operating data into the machine learning model to obtain a predicted data, calculate a current similarity value between the predicted data and a real data; and If the current similarity value is less than the performance indicator control value, it is determined that the energy storage system has been abnormal. The monitoring method as claimed in claim 1, wherein the training data set is represented by the following matrix , in represents the i-th parameter at the j-th time point, i, j, n, m are positive integers and , , wherein when training the machine learning model, the machine learning model outputs (md) predicted strain vectors, and the (md) similarities are calculated according to the following mathematical formula, where i=1,2,…,n, k=1,2,…,(md), represents the kth similarity among the (md) similarities, is the i-th element of the k-th strain vector in the (md) strain vectors, is the i-th element of the k-th predicted strain variable vector among the (md) predicted strain variable vectors. A monitoring method as described in claim 1, wherein the parameters include current, power consumption, power generation, frequency, power or voltage. The monitoring method as described in claim 3 further includes: Inputting the real data into an inverse function to obtain an expected operation data; and Calculating the difference between the expected operation data and the new operation data to determine whether one of the parameters is abnormal. The monitoring method as described in claim 1, wherein the machine learning model is an extreme gradient boosting model, a random forest model or a deep learning model. An energy storage monitoring system comprises: an energy storage system; and a computer system, communicatively connected to the energy storage system, for executing a plurality of steps: obtaining a plurality of parameters of the energy storage system at m time points as a training data set, wherein m is a positive integer; obtaining the parameters corresponding to d time points among the time points as d independent variable vectors, and obtaining the parameters corresponding to the remaining (m-d) time points as (m-d) dependent variable vectors, wherein d is a positive integer and d is less than m; A machine learning model is trained based on the d independent variable vectors and the (m-d) dependent variable vectors, and the corresponding (m-d) similarities are calculated, and a performance indicator control value is set based on the (m-d) similarities; A new operation data is obtained, and the new operation data is input into the machine learning model to obtain a predicted data, and a current similarity value between the predicted data and a real data is calculated; and If the current similarity value is less than the performance indicator control value, it is determined that the energy storage system has been abnormal. The energy storage monitoring system as described in claim 6, wherein the training data set is represented by the following matrix , in represents the i-th parameter at the j-th time point, i, j, n, m are positive integers and , , wherein when training the machine learning model, the machine learning model outputs (md) predicted strain vectors, and the (md) similarities are calculated according to the following mathematical formula, where i=1,2,…,n, k=1,2,…,(md), represents the kth similarity among the (md) similarities, is the i-th element of the k-th strain vector in the (md) strain vectors, is the i-th element of the k-th predicted strain variable vector among the (md) predicted strain variable vectors. An energy storage monitoring system as described in claim 6, wherein the parameters include current, power consumption, power generation, frequency, power or voltage. The energy storage monitoring system as described in claim 8, wherein the steps further include: Inputting the real data into an inverse function to obtain an expected operating data; and Calculating the difference between the expected operating data and the new operating data to determine whether one of the parameters is abnormal. An energy storage monitoring system as described in claim 6, wherein the machine learning model is an extreme gradient boosting model, a random forest model or a deep learning model.