Abnormality detection method, power regulation system, and temperature prediction device

Abstract

The invention provides an anomaly detection method for an energy storage system. The anomaly detection method comprises the following steps: receiving a plurality of sensing data acquired from an electronic component of the energy storage system; inputting a plurality of sensing data to a temperature prediction model; receiving a plurality of sensing data of the same type through each model encoder, and respectively calculating a plurality of time sequence features associated with the type from the plurality of sensing data of each type; outputting a plurality of time sequence features of each type to a weight model; calculating a predicted temperature feature for the plurality of time series features through a weight model and outputting the predicted temperature feature to a model decoder; reconstructing, by the model decoder, the predicted temperature feature to generate a predicted temperature of the electronic component; and evaluating whether the operation of the electronic component is abnormal by calculating an error value between the predicted temperature and a current temperature. The invention further provides a power regulation system and a temperature prediction device.

Description

Technical Field

[0001] The present disclosure relates to a detection method, a power supply system and a device using the detection method, and in particular to an abnormality detection method, a power regulation system and a temperature prediction device for an energy storage system. Background Art

[0002] The current practice of predicting abnormalities in power supply devices is to monitor the sensing data of the electronic components inside the power supply device, such as voltage, current or temperature, and design the allowable range of these sensing data. When the control system connected to the power supply device determines that these sensing data exceed the allowable range, it will immediately determine that the power supply device has an abnormality and immediately reduce the load to prevent damage to the power supply device. At the same time, other power supply devices will immediately dispatch power to maintain uninterrupted power output.

[0003] One of the methods of monitoring the temperature of electronic components is to monitor the temperature change of the water cooling system. However, since there are too many factors affecting the temperature change of the water cooling system, such as the temperature of the water cooling system is related to the environmental conditions and fans of the water cooling system, the load of the power supply device cannot be represented by the temperature change of the water cooling system. If the abnormality of the power supply device caused by component wear is directly evaluated by the temperature change of the water cooling system, it is easy to cause misjudgment.

[0004] In addition, the control system receives sensing data from the power supply device, one type of sensing data is a high power signal. Since the control system includes multiple modules and is bulky, collecting high power signals to perform related operations is dangerous in operation.

[0005] Prior art has proposed training artificial intelligence models to predict anomalies in corresponding fields, but due to at least the above reasons, the prediction anomaly model is inaccurate and the implementation is unsafe. Therefore, there is still room for improvement in the anomaly prediction of the prior art. Summary of the invention

[0006] The purpose of the invention disclosed herein is to provide an abnormality detection method, a power regulation system and a temperature prediction device for an energy storage system.

[0007] According to an embodiment of the present disclosure, a method for detecting an abnormality in an energy storage system is disclosed, comprising: (a) receiving a plurality of sensing data obtained from an electronic component of an energy storage system and each sensing data is associated with a type; (b) inputting the plurality of sensing data into a temperature prediction model according to the type of each sensing data, wherein the temperature prediction model comprises a plurality of model encoders, a weight model and a model decoder; (c) receiving a plurality of sensing data of the same type through each model encoder and calculating a plurality of time series features associated with the type from the plurality of sensing data of each type respectively; (d) outputting the plurality of time series features of each type to the weight model; (e) calculating a predicted temperature feature for the plurality of time series features through the weight model and outputting the predicted temperature feature to a model decoder; (f) reconstructing the predicted temperature feature through the model decoder to generate a predicted temperature of the electronic component; and (g) evaluating whether the operation of the electronic component is abnormal by calculating an error value between the predicted temperature and a current temperature.

[0008] According to an embodiment of the present disclosure, a power regulation system is disclosed, including a sensing module, a storage medium and a computing device. The sensing module is configured to obtain a plurality of sensing data of an electronic component, wherein each sensing data is associated with a type. The storage medium is coupled to the sensing module and configured to store a temperature prediction model, wherein the temperature prediction model includes a plurality of model encoders, a weight model and a model decoder. The plurality of model encoders are configured to receive a plurality of sensing data, wherein each sensing data is associated with a type; the weight model is coupled to the plurality of model encoders. The model decoder is coupled to the weight model. The computing device is coupled to the sensing module and the storage medium, and configured to use the temperature prediction model to perform a plurality of operations, including receiving a plurality of sensing data of the same type through each model encoder and calculating a plurality of time series features associated with the type from the plurality of sensing data of each type respectively; outputting the plurality of time series features of each type to the weight model; calculating a predicted temperature feature for the plurality of time series features through the weight model and outputting the predicted temperature feature to a model decoder; reconstructing the predicted temperature feature through the model decoder to generate a predicted temperature of the electronic component; and evaluating whether the operation of the electronic component is abnormal by comparing the error value between the predicted temperature and a current temperature.

[0009] According to one embodiment of the present disclosure, a temperature prediction device is disclosed, including a storage medium and a processor. The storage medium is configured to store a plurality of program codes. The processor is coupled to the storage medium and configured to load a plurality of program codes to perform the following operations: receiving a plurality of sensing data obtained from an electronic component of an energy storage system and each sensing data is associated with a type; inputting the plurality of sensing data into a temperature prediction model according to the type of each sensing data, wherein the temperature prediction model includes a plurality of model encoders, a weight model and a model decoder; receiving a plurality of sensing data of the same type through each model encoder and calculating a plurality of time series features associated with the type from the plurality of sensing data of each type respectively; outputting a plurality of time series features of each type to the weight model; calculating a predicted temperature feature for the plurality of time series features through the weight model and outputting the predicted temperature feature to a model decoder; and reconstructing the predicted temperature feature through the model decoder to generate a predicted temperature of the electronic component.

[0010] The present disclosure can improve the accuracy of temperature prediction, avoid the problem of overall system failure caused by damage to some electronic components of the equipment, and achieve the technical effect of maintaining the stability of the overall system and uninterrupted power supply. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a block diagram of an energy storage system according to an embodiment of the present disclosure.

[0012] Figure 2 It is a block diagram of a power regulation system according to an embodiment of the present disclosure.

[0013] Figure 3 The present invention is a flowchart of an abnormality detection method for an energy storage system according to an embodiment of the present invention.

[0014] Figure 4 It is a model architecture diagram of a temperature prediction model shown in the present disclosure according to an embodiment.

[0015] Figure 5 It is a model architecture diagram of a weight model shown in the present disclosure according to one embodiment.

[0016] Figure 6 It is a block diagram of a temperature prediction device according to an embodiment of the present disclosure.

[0017] Figure 7 The figure is a flow chart of a temperature prediction method according to an embodiment of the present disclosure.

[0018] Figure 8a and Figure 8b The diagram is an architecture diagram of an automatic encoder training method according to an embodiment of the present disclosure.

[0019] Fig. 9 The present invention is a flowchart of a training model encoder according to one embodiment.

[0020] Fig.10 The present invention is a flowchart of a training weight model according to an embodiment of the present invention.

[0021] Fig.11 The present invention is a flowchart of a method for performing anomaly detection by predicting temperature according to an embodiment of the present invention.

[0022] Fig.12 The present invention is a flowchart of a data preprocessing procedure for training a temperature prediction model according to an embodiment of the present invention.

[0023] Description of reference numerals:

[0024] 10: Energy storage system

[0025] 110: Power regulation system

[0026] 112: Computing Device

[0027] 114: Storage Media

[0028] 116: Sensing module

[0029] 120: Energy storage battery system

[0030] 122, 124: Electronic components

[0031] 140: Energy Management System

[0032] 142: Cloud Database

[0033] 150: Temperature prediction model

[0034] 152a~152n,152t: Model encoder

[0035] 154: Weight Model

[0036] 156: Model Decoder

[0037] 160: Municipal power grid

[0038] 162: Power module temperature sensor

[0039] 164: Water cooling system temperature sensor

[0040] 166: Power system numerical sensor

[0041] 170: Renewable Energy

[0042] 600: Temperature prediction device

[0043] 610: Processor

[0044] 620: Storage Media

[0045] 622: Program code

[0046] 810, 820: Training the Autoencoder

[0047] 812, 822: training encoder

[0048] 814, 824: training decoder

[0049] S310~S370, S710~S760, S910~S950, S1010~S1030, S1110~S1160, S1210~S1270: Steps

[0050] L1_1,…,L1_n: first layer

[0051] L2: Second layer

[0052] L_cat: concatenation layer

[0053] x1,…,xn,xt: sensor data

[0054] f1,…fn,ft: time series features

[0055] f1',…fn',ft': first relevant feature

[0056] f1”,…fn”,ft”: second relevant features

[0057] ft_pred: predicted temperature features

[0058] T_pred: predicted temperature

[0059] xt1, xt2: training sensor data

[0060] xt1_c, xt2_c: training compressed data

[0061] xt1', xt2': training reconstruction data DETAILED DESCRIPTION

[0062] The present disclosure is further described below in conjunction with the accompanying drawings and embodiments so that relevant persons in the technical field to which the present invention belongs can better understand the present invention and implement it accordingly, but the embodiments are not intended to limit the present invention.

[0063] Figure 1 FIG. 1 is a block diagram of an energy storage system (Energy Storage System) according to an embodiment of the present disclosure. Figure 1As shown, the energy storage system 10 is provided with a power control system (Power Control System) 110 , an energy storage battery system (Battery Energy Storage System) 120 and an energy management system (Energy Management System) 140 .

[0064] The power conditioning system 110 is coupled to the energy storage battery system 120 and the energy management system 140. The power conditioning system 110 receives the electric energy generated by the renewable energy (Renewable Energy) 170, and supplies power to the electric equipment in the neighboring area in real time. When the electric energy generated by the renewable energy 170 is greater than the power demand of the neighboring area, the power conditioning system 110 will store the excess electric energy in the energy storage battery system 120, so that when the electric energy generated by the renewable energy 170 is insufficient to supply the power demand of the neighboring area in real time, the energy storage battery system 120 can be used for power supply. When the battery capacity of the energy storage battery system 120 is full, the power conditioning system 110 will supply the electric energy generated by the renewable energy 170 to the municipal power grid 160.

[0065] The energy management system 140 includes a cloud database 142. The cloud database 142 is configured to store a relational database, a time series database, and a model database (not shown).

[0066] The energy management system 140 is a central control system that can grasp the status of all systems for monitoring and management. In one embodiment, the energy management system 140 can perform data screening, data pre-processing, deep learning model architecture training and use, diagnosis of system status, data curve trend analysis, abnormal reporting and model management. In addition, the energy management system 140 can also provide a graphical user interface for front-end status display.

[0067] In this way, the energy storage system 10 acts as a microgrid, which can independently supply power to a region, and can also operate in parallel with the city grid 160 .

[0068] In one embodiment, the energy storage battery system 120 is provided with an electronic component 122 for controlling charging and discharging power, adjusting heat dissipation and power value. For example, the electronic component 122 is a power module, a water cooling system component or a power system component. For the sake of simplicity of the specification, only the electronic component 122 is used to represent it.

[0069] In one embodiment, an electronic component 124 for controlling a power switch is disposed in the power regulation system 110. For example, the electronic component 124 is an insulated gate bipolar transistor (IGBT) for switching or regulating direct current, alternating current, and voltage.

[0070] In one embodiment, the electronic component 124 is also a component for controlling charging and discharging power, adjusting heat dissipation and power value, as described above and will not be repeated here.

[0071] Figure 2 FIG. 1 is a block diagram of a power regulation system according to an embodiment of the present disclosure. Figure 2 As shown, the power regulation system 110 includes a computing device 112 , a storage medium 114 , and a sensing module 116 . The storage medium 114 is coupled to the computing device 112 and the sensing module 116 .

[0072] In one embodiment, the sensing module 116 is configured to obtain the electronic components 122 and 124 (eg Figure 1 ) to generate a plurality of sensing data and each sensing data is associated with a type. The sensing data is, for example, data obtained from different types of electronic components 122.

[0073] In one embodiment, the sensing module 116 includes a power module temperature sensor 162, a water cooling system temperature sensor 164, and a power system value sensor 166. The sensing data includes the power module temperature (e.g., chamber temperature), the water cooling system temperature (e.g., the water temperature of the water cooling system), and the power system value (e.g., the maximum drop value of the direct current, alternating current, voltage, or the three currents). In this embodiment, the sensing data is the chamber temperature, the water temperature of the water cooling system, the direct current, the alternating current, the voltage, and the maximum drop value of the three currents, so the sensing module 116 receives six types of sensing data. It is worth mentioning that the present disclosure does not limit the type of sensing data, and any data related to the power conditioning system 110 and the energy storage battery system 120 and can be collected and statistically analyzed to highlight the above system status is within the scope of the present disclosure.

[0074] In one embodiment, the sensing data x1 is associated with the first type, the sensing data x2 is associated with the second type, the sensing data xn is associated with the Nth type, and the sensing data xt is associated with the Tth type. In one embodiment, the sensing data xt is the power module temperature.

[0075] In one embodiment, the storage medium 114 is configured to store a plurality of sensing data and a temperature prediction model 150 .

[0076] In one embodiment, the computing device 112 is configured to perform an anomaly detection method by predicting the temperature. Figure 3 .

[0077] Figure 3 FIG. 1 is a flow chart of an abnormality detection method for an energy storage system according to an embodiment of the present disclosure. The abnormality detection method is applied to Figure 1 Energy storage system 10 and Figure 2 A power conditioning system 110 is provided.

[0078] In step S310 , a plurality of sensing data acquired from electronic components of an energy storage system is received.

[0079] In step S320 , a plurality of sensing data are respectively input into the temperature prediction model according to corresponding types.

[0080] In step S330 , each model encoder receives a plurality of sensing data of the same type and respectively calculates a plurality of time series features associated with the type from the plurality of sensing data of the same type.

[0081] In step S340 , a plurality of time series features of corresponding types are output to a weight model.

[0082] In step S350 , a predicted temperature feature is calculated from a plurality of time series features using a weight model and the predicted temperature feature is output to a model decoder.

[0083] In step S360 , the predicted temperature feature is reconstructed by a model decoder to generate a predicted temperature of the electronic component.

[0084] In step S370 , the error value between the predicted temperature and the current temperature is calculated to evaluate whether the operation of the electronic component is abnormal.

[0085] In step S310, the energy storage system 10 obtains sensing data of the electronic components 122 and 124 through the sensing module 116 of the power regulation system 110. Each sensing data is associated with a type. In one embodiment, the sensing data is associated with a time curve of a type. For example, the sensing data x1 is a time curve of the chamber temperature; the sensing data x2 is a time curve of the water temperature of the water cooling system; the sensing data xn is a time curve of the alternating current; and the power module temperature xt is a time curve of the temperature.

[0086] In step S320, the computing device 112 loads the temperature prediction model 150, wherein the temperature prediction model 150 includes a plurality of model encoders, a weight model, and a model decoder. For a better understanding of the operation of the temperature prediction model 150, please refer to Figure 4 .

[0087] Figure 4 FIG. 1 is a model architecture diagram of a temperature prediction model according to an embodiment of the present disclosure. Figure 4 As shown, the temperature prediction model 150 includes model encoders 152 a , 152 b , . . . , 152 n , 152 t , a reweighting module 154 and a model decoder 156 .

[0088] In one embodiment, the number of model encoders 152a to 152t corresponds to the number of the types. That is, the temperature prediction model 150 is provided with a corresponding model encoder for each type.

[0089] In one embodiment, the sensing data x1 associated with the first type is input to the model encoder 152a; the sensing data x2 associated with the second type is input to the model encoder 152b; and the sensing data xn associated with the Nth type is input to the model encoder 152n. For example, the sensing data processed by the model encoder 152a is a time curve of the chamber temperature; the sensing data processed by the model encoder 152b is a time curve of the water temperature of the water cooling system; the sensing data processed by the model encoder 152n is a time curve of the alternating current; and the sensing data processed by the model encoder 152t is a time curve of the power module temperature xt.

[0090] In one embodiment, the model encoders 152a, 152b, ..., 152n, 152t are encoders in a pre-trained autoencoder, such as Fourier transform, wavelet transform, time-frequency transform or statistical feature operation. The model encoders 152a, 152b, ..., 152n, 152t are used as feature acquisition models for time series data to extract time series features from sensor data. The method for training the model encoder will be Figure 8a , Figure 8b and Fig. 9 Provide explanation.

[0091] Please refer to Figure 3 In step S330 , the model encoders 152 a , 152 b , … , 152 n , 152 t process the sensing data of the respective corresponding types, and respectively calculate the time series features f1 , f2 , … , fn , ft of the respective corresponding types from the sensing data of the respective corresponding types.

[0092] In one embodiment, the model encoder 152a processes the time curve of the chamber temperature and analyzes the time feature sequence f1 of the chamber temperature; the model encoder 152b processes the time curve of the water temperature of the water cooling system and analyzes the time feature sequence f2 of the temperature difference, and so on. Therefore, the multiple model encoders 152a, 152b, ..., 152n, 152t of the temperature prediction model 150 calculate the corresponding types of time series features f1, f2, ..., fn, ft.

[0093] In step S340 and step S350, the model encoder 152a, 152b,…, 152n outputs multiple time series features f1, f2,…, fn, ft of corresponding types to the weight model 154, and the weight model 154 calculates the predicted temperature feature ft_pred for the multiple time series features f1, f2,…, fn, ft and outputs the predicted temperature feature ft_pred to the model decoder 156.

[0094] In one embodiment, the weight model 154 is used to calculate the relationship between the received multiple time series features f1, f2, ..., fn, ft, so as to estimate the temperature feature associated with the time feature through the relationship between the time series features, so as to generate the predicted temperature feature ft_pred. Figure 5 The detailed description is as follows.

[0095] The weight model 154 outputs the predicted temperature feature ft_pred to the model decoder 156 , so that the model decoder 156 converts (reconstructs) the predicted temperature feature ft_pred into the predicted temperature.

[0096] In step S360 , the model decoder 156 reconstructs the predicted temperature feature ft_pred into the predicted temperature T_pred.

[0097] In this embodiment, the model decoder 156 is, for example, a component of a decoder of an auto encoder. The model decoder 156 is used to convert the predicted temperature feature ft_pred back into time domain data to generate the predicted temperature T_pred.

[0098] In one embodiment, the model encoder 152t calculates the time series feature ft based on the power module temperature xt and outputs it to the model decoder 156. The model decoder 156 reconstructs the predicted temperature feature ft_pred with reference to the time series feature ft. In this embodiment, the model decoder 156 directly references the time series feature ft associated with the power module temperature xt during the process of reconstructing the predicted temperature T_pred, so as to ensure that the reconstructed predicted temperature T_pred is highly correlated with the features of the power module temperature xt, thereby increasing the accuracy of the predicted temperature T_pred.

[0099] In step S370, the computing device 112 obtains the current temperature, calculates the error value between the predicted temperature T_pred and the current temperature, and evaluates whether the operation of the electronic component is abnormal based on the error value. In this embodiment, the energy storage system 10 mainly focuses on the temperature change of the electronic component; the current temperature refers to the currently detected temperature of the electronic component, such as the power module.

[0100] The predicted temperature T_pred may be regarded as the temperature of the electronic component after a period of time, and may be used as a temperature indicator of the electronic component under normal operation.

[0101] In one embodiment, when the error between the predicted temperature T_pred and the current temperature is greater than a threshold value, the computing device 112 accumulates an abnormal number of times. Within a detection cycle, when the abnormal number of times is greater than a tolerance value, the computing device 112 determines that the electronic component, such as a power module, is operating abnormally; otherwise, the computing device 112 determines that the electronic component, such as a power module, is operating normally.

[0102] In one embodiment, when the computing device 112 determines that the operation of the electronic component is abnormal, it will issue an abnormality warning to the monitoring system, such as Figure 1 Energy management system 140.

[0103] For further description of the weight model 154 in step S340 and step S350, please refer to Figure 5 .

[0104] Figure 5 FIG. 1 is a model architecture diagram of a weight model according to an embodiment of the present disclosure. Figure 5 As shown, the weight model 154 includes a plurality of first layers L1_1, L1_2, ..., L1_n, a second layer L2 and a concatenation layer (L_cat, Layerconcatenate).

[0105] In one embodiment, the first layer is a computational model, such as a model using convolution operations, linear operations, support vector machines (SVM) or support vector regression (SVR), kmeans operations, long short-term memory operations (LSTM), self attention operations, or a combination of the above operations.

[0106] Take the first layer L1_1 as an example. The time series feature f1 of the sensing data x1 is input to the computational model of the first layer L1_1. The computational model of the first layer L1_1 performs a model operation on the time series feature f1 to generate a first related feature f1' of the sensing data x1. Similarly, the computational model of the first layer L1_2 performs a model operation on the time series feature f2 to generate a first related feature f2' of the sensing data x2. The computational model of the first layer L1_n performs a model operation on the time series feature fn to generate a first related feature fn' of the sensing data xn.

[0107] In one embodiment, the first related features f1 ′, f2 , . . . , fn are input to the computational model of the second layer L2 .

[0108] In one embodiment, the second layer L2 is a computational model, such as a model using linear computation, attention computation, graph neural networks computation, or a combination of the above computations.

[0109] The computational model of the second layer L2 receives the first related features f1', f2', ..., fn', and respectively calculates the correlation between each related feature and the other related features to generate corresponding multiple second related features. For example, the second layer L2 calculates the correlation between the first related feature f1' and other first related features, namely, the first related features f2', ..., fn', to generate the second related feature f1"; the second layer L2 also calculates the correlation between the first related feature f2' and other first related features, namely, the first related features f1', f3', ..., fn', to generate the second related feature f2", and so on. The second related feature f1" represents the correlation between the first related feature f1' and the other first related features f2', ..., fn'. Similarly, the second related feature f2" represents the correlation between the first related feature f2' and the other first related features f1', f3', ..., fn'. The second related feature fn" represents the correlation between the first related feature fn' and the other first related features f1', ..., fn-1'.

[0110] In one embodiment, the cascade layer L_cat receives a plurality of second correlation features f1 ″, f2 ″, . . . , fn ″ from the second layer L2 . The cascade layer L_cat uses the second correlation features f1 ″, f2 ″, . . . , fn ″ to calculate the predicted temperature feature ft_pred.

[0111] In one embodiment, the concatenation layer L_cat is a computational model that uses a linear computational model to combine a plurality of second correlation features f1″, f2″, …, fn″ into one to obtain a predicted temperature feature ft_pred. The predicted temperature feature ft_pred is provided to the model encoder 156 ( Figure 4 ) to reconstruct the predicted temperature T_pred. The predicted temperature T_pred is as described above and will not be repeated here.

[0112] Figure 6 6 is a block diagram of a temperature prediction device according to an embodiment of the present disclosure. The temperature prediction device 600 includes a processor 610 and a storage medium 620. The processor 610 is coupled to the storage medium 620. The storage medium 620 stores a plurality of program codes 622.

[0113] In one embodiment, the processor 610 is, for example but not limited to, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a central processing unit (CPU), a system on chip (SoC), a field programmable gate array (FPGA), a network processor chip, or a combination of the above elements.

[0114] In one embodiment, the storage medium 620 is, for example but not limited to, a random access memory (RAM), a flash memory (Flash memory), a read only memory (ROM), a hard disk drive (HDD), a solid state drive (SSD), an optical storage device, or a combination of the above elements.

[0115] To further explain the various operations of the temperature prediction device 600, please refer to Figure 7 . Figure 7 The figure is a flow chart of a temperature prediction method according to an embodiment of the present disclosure.

[0116] In one embodiment, the temperature prediction device 600 is in communication or electrically connected to an energy storage system (eg Figure 1 The energy storage system 10 shown in FIG. Figure 6 The processor 610 is configured to receive multiple sensing data associated with multiple types of electronic components in the energy storage system. The processor 610 is configured to load multiple program codes 622 to perform multiple operations and calculate the predicted temperature T_pred by multiple sensing data associated with multiple types.

[0117] In step S710, the processor 610 receives a plurality of sensing data acquired from electronic components of the energy storage system. Each sensing data is associated with a type.

[0118] In step S720, the processor 610 inputs a plurality of sensing data into the temperature prediction model according to corresponding types, wherein the temperature prediction model includes a plurality of model encoders, a weight model and a model decoder.

[0119] In step S730 , each model encoder receives a plurality of sensing data of the same type and respectively calculates a plurality of time series features associated with the type from the plurality of sensing data of the same type.

[0120] In step S740 , each model encoder outputs a plurality of time series features of a corresponding type to a weight model.

[0121] In step S750 , a predicted temperature feature is calculated from a plurality of time series features using a weight model and the predicted temperature feature is output to a model decoder.

[0122] In step S760 , the predicted temperature feature is reconstructed by a model decoder to generate a predicted temperature of the electronic component.

[0123] Steps S710 to S760 are similar to Figure 3 Please refer to the description of steps S310 to S360 of Figure 4 Temperature prediction model 150 and Figure 5 The description of the weight model 154 is omitted here for the sake of brevity.

[0124] To further explain the training method of the plurality of model encoders 150a to 152t of the temperature prediction model 150, please refer to Figure 8a and Figure 8b .

[0125] Figure 8a and Figure 8b FIG. 1 is an architecture diagram of a training auto-encoder according to an embodiment of the present disclosure. The training auto-encoder may be composed of Figure 2 The computing device 112 or Figure 6 The temperature prediction device 600 is used to perform the training of the related computing model before performing abnormality detection or temperature prediction. To simplify the description, the computing device 112 is used as a representative for description.

[0126] like Figure 8a As shown, in one embodiment, a training autoencoder 810 includes a training encoder 812 and a training decoder 814 .

[0127] In one embodiment, the training auto-encoder 810 is used to train its own computational model for a type of training sensing data xt1. For example, the training sensing data xt1 is a time curve of the chamber temperature.

[0128] The training encoder 812 receives the training sensing data xt1 to calculate the training compressed data xt1_c and transmits it to the training decoder 814. The training decoder 814 uses the training compressed data xt1_c to generate the training reconstructed data xt1'. In this embodiment, the computing device 112 compresses and reconstructs the training sensing data xt1 to obtain the training reconstructed data xt1', so the training reconstructed data xt1' will be close to the training sensing data xt1. The computing device 112 compares the error between the training reconstructed data xt1' and the training sensing data xt1, and updates the weight parameters of the training encoder 812 through back propagation. The training auto-encoder 810 stops training after the error between the training reconstructed data xt1' and the training sensing data xt1 is less than a training threshold value. At this point, it means that the training auto-encoder 810 has been trained. For example, the training auto-encoder 810 for the time curve of the processing chamber temperature has been trained.

[0129] Similarly, Figure 8b As shown, in one embodiment, the training auto-encoder 820 includes a training encoder 822 and a training decoder 824 .

[0130] In one embodiment, the training autoencoder 820 trains its own computational model for another type of training sensing data xt2. For example, the training sensing data xt2 is a time curve of the water temperature of the water cooling system.

[0131] The training encoder 822 receives the training sensing data xt2 to calculate the training compressed data xt2_c and transmits it to the training decoder 824. The training decoder 824 uses the training compressed data xt2_c to generate the training reconstructed data xt2'. In this embodiment, the computing device 112 compresses and reconstructs the training sensing data xt2 to obtain the training reconstructed data xt2', so the training reconstructed data xt2' will be close to the training sensing data xt2. The computing device 112 compares the error between the training reconstructed data xt2' and the training sensing data xt2, and updates the weight parameters of the training encoder 822 through back propagation. The training auto-encoder 820 stops training after the error between the training reconstructed data xt2' and the training sensing data xt2 is less than a training threshold value. At this point, it means that the training auto-encoder 820 has been trained. For example, the training auto-encoder 820 for processing the time curve of the water temperature of the water cooling system has been trained.

[0132] By analogy, the computing device 112 will train a corresponding number of training auto-encoders based on the number of types that need to be processed.

[0133] It is worth mentioning that the training sensor data xt1 and xt2 are obtained from e.g. Figure 2 The sensing data obtained by the sensing module 116 or from Figure 1 The cloud database 142 obtains the sensing data. The training sensing data is, for example, sensing data obtained in real time or historical sensing data.

[0134] Fig. 9 FIG. 1 is a flowchart of a training model encoder according to an embodiment of the present disclosure. The training of the model encoder may be performed by Figure 2 The computing device 112 or Figure 6 The temperature prediction device 600 is used to perform the training of the related computing model before performing abnormality detection or temperature prediction. To simplify the description, the computing device 112 is used as a representative for description.

[0135] In step S910 , the computing device 112 trains a plurality of auto-encoders according to a plurality of training sensing data respectively.

[0136] In step S920, the computing device 112 determines whether the difference (error) between the training reconstruction data and the training sensing data input to the autoencoder is less than a training threshold value. If the difference is less than the training threshold value, step S930 is executed. If the difference is greater than or equal to the training threshold value, step S950 is executed.

[0137] In step S930 , the computing device 112 determines whether the auto-encoder training is completed.

[0138] In step S940 , the computing device 112 configures the training encoder as the model encoder.

[0139] In step S950, the computing device 112 uses the difference value to perform back propagation to update the weight parameters of the training encoder. After step S950, the computing device 112 returns to step S910 until the auto-encoder training is completed.

[0140] In step S910 to step S930 and step S950, as described above Figure 8a and Figure 8b The description will not be repeated here.

[0141] In step S940, the computing device 112 copies the model structure and weight parameters of the trained training encoder to one of the plurality of model encoders 152a to 152n in the temperature prediction model 150. For example, Figure 8a The training encoder 812 of the trained training autoencoder 810 is copied as Figure 4 Model encoder 152a. Figure 8b The training encoder 822 of the trained training autoencoder 820 is copied as Figure 4 Model encoder 152b.

[0142] As a result, each model encoder of the temperature prediction model 150 has a computational model that can accurately obtain the time series characteristics of the corresponding type of sensing data.

[0143] Fig.10 FIG. 1 is a flow chart of a training weight model according to an embodiment of the present disclosure. The training method of the weight model can be as follows: Figure 2 The computing device 112 or Figure 6 The temperature prediction device 600 is used to perform the above operation. To simplify the description, the computing device 112 is used as a representative for the following description.

[0144] In step S1010 , the computing device 112 uses a loss function to calculate a training error value.

[0145] In step S1020 , the computing device 112 determines the use of a plurality of training sensing data according to the training error value.

[0146] In step S1030 , the computing device 112 uses the adopted plurality of training sensing data to execute a back propagation algorithm to update the parameters of the weight model 154 .

[0147] In step S1010 , the computing device 112 uses the loss function to perform a gradient descent operation to update the parameters of the weight model 154 .

[0148] In one embodiment, the loss function may be one of formulas (1) to (4) or a combination thereof.

[0149]

[0150] Among them, loss feature,i is the error of the feature domain of the i-th training sensing data, f t is the time series feature generated by the model encoder, f t,super is the largest of all time series features. Formula (1) is used to calculate the degree of difference between two temperature features.

[0151] As described above, the time series features obtained from the training sensing data are, for example, related to the power system features, so what is calculated in formula (1) is the feature difference related to the voltage or current.

[0152]

[0153] Among them, loss reconst,i is the error between the temperature feature of the i-th training sensing data restored by the weight model 154 and the temperature feature of the training sensing data, T pred is the predicted temperature reconstructed by the model decoder 156, is the actual temperature of the training sensing data. Formula (2) is used to calculate the difference between the restored temperature feature and the actual temperature.

[0154]

[0155] Formula (3) is used to calculate the difference between the temperature characteristic curve after restoration and the actual temperature.

[0156] loss dis,i =|discriminator(T preD )-0|......Formula (4)

[0157] Wherein, discriminator is a discriminator used to judge the authenticity of the input data. Formula (4) is used to calculate whether the predicted temperature T_pred reconstructed from multiple time series features is different from the actual temperature of the input sensing data.

[0158] loss i =loss feature,i +loss reconst,i +loss cos,i +loss dis,i …Formula (5)

[0159] Wherein formula (5) is a linear combination of formula (1) to formula (4).

[0160] The operation device 112 performs back propagation through formula (6) to formula (9) to update the weights of the weight model 154.

[0161]

[0162]

[0163]

[0164]

[0165] Among them, loss batch is the training error set of the batch, μ is the error value loss i The average value, σ 2 is the error value loss i The standard deviation of θ is the parameter of the weight model 150, θ t+1 Update parameters for the batch.

[0166] The calculation result of the loss function is used to adjust and update the parameters of the weight model 154 so that the predicted temperature feature ft_pred subsequently calculated by the weight model 154 after reconstruction by the model decoder 156 will be as close as possible to the actual temperature of the training sensing data.

[0167] It is worth mentioning that the training sensing data in the training phase is a time curve, so the data processed in the training phase is also a corresponding characteristic curve.

[0168] In one embodiment, when the computing device 112 uses the loss function to perform back-propagation training, output values ​​whose values ​​are greater than a reasonable value are filtered out from among the output values ​​of the loss function.

[0169] Fig.11 The present invention is a flowchart of a method for performing anomaly detection by predicting temperature according to an embodiment of the present invention. Fig.11 The anomaly detection method can be Figure 2 The computing device 112 or Figure 6 The temperature prediction device 600 is used to perform the above operation. To simplify the description, the computing device 112 is used as a representative for the following description.

[0170] In step S1110 , the computing device 112 starts an abnormality detection procedure.

[0171] In step S1120 , the computing device 112 obtains real-time time-series sensing data.

[0172] In step S1130 , the computing device 112 performs data preprocessing on the real-time time-series sensing data to obtain preprocessed data.

[0173] In step S1140 , the computing device 112 slices the pre-processed data to obtain a plurality of slice time series data.

[0174] In step S1150 , the computing device 112 performs time sorting on the plurality of slice timing data and performs data normalization to obtain a plurality of normalized slice timing data.

[0175] In step S1160 , the computing device 112 uses the normalized slice timing data as the sensing data of step S310 and step S710 to execute the abnormality detection method and the temperature prediction method.

[0176] In step S1110 , the computing device 112 loads a program code for executing an anomaly detection program to perform anomaly detection using the predicted temperature.

[0177] In step S1120, the computing device 112 continuously obtains the real-time time series sensing data. In one embodiment, the real-time time series sensing data obtained by the computing device 112 is a continuous time curve. The real-time time series sensing data can be obtained from the cloud database 142 ( Figure 1 ) is obtained.

[0178] In step S1130 , data pre-processing includes data cleaning, data type adjustment, data integration or data transformation.

[0179] In step S1140 , since the real-time time-series sensing data obtained by the computing device 112 is a time curve, the pre-processed data is sliced ​​according to a time segment (eg, 5 minutes) to reduce the size of the curve data processed each time to one time segment.

[0180] In step S1150 , each slice time series data is sorted according to the time sequence, and the computing device 112 performs a data normalization operation on the sorted slice time series data so that all slice time series data are scaled to the same interval.

[0181] In step S1160, the data time series sensing data has been pre-processed, sliced ​​and normalized to remove the ambiguity between the original data, so these normalized sliced ​​time series data are suitable for subsequent feature acquisition. In this embodiment, the normalized sliced ​​time series data is used as the sensing data of the aforementioned steps S310 and S710, that is, the aforementioned sensing data is data that has been pre-processed, sliced ​​and normalized.

[0182] Fig.12 FIG. 1 is a flowchart of a data preprocessing procedure for training a temperature prediction model according to an embodiment of the present disclosure. Figure 2 The computing device 112 or Figure 6 The temperature prediction device 600 is used to perform the above operation. To simplify the description, the computing device 112 is used as a representative for the following description.

[0183] In step S1210 , the computing device 112 starts a training procedure for the temperature prediction model.

[0184] In step S1220 , the computing device 112 obtains historical time-series sensing data.

[0185] In step S1230 , the computing device 112 selects important historical data from the historical time-series sensing data.

[0186] In step S1240 , the computing device 112 performs data pre-processing on the important historical data to obtain pre-processed historical data.

[0187] In step S1250 , the computing device 112 slices the pre-processed historical data to obtain a plurality of sliced ​​historical time series data.

[0188] In step S1260 , the computing device 112 performs time sorting on the plurality of slice historical time series data and performs data normalization to obtain a plurality of normalized historical slice time series data.

[0189] In step S1270, the computing device 112 uses the normalized historical slice time series data as Fig. 9 and Fig.10 The training sensor data is used to train the temperature prediction model.

[0190] In step S1210 , the computing device 112 loads a program code for executing a training procedure of a temperature prediction model to execute the training procedure.

[0191] In step S1220, the computing device 112 obtains the historical time-series sensing data. For example, the historical time-series sensing data is the sensing data continuously obtained in the past year. In one embodiment, the historical time-series sensing data obtained by the computing device 112 is a continuous time curve. The historical time-series sensing data can be obtained from the cloud database 142 ( Figure 1 ) was obtained.

[0192] In step S1240 , the computing device 112 extracts temperature-related sensing data from a plurality of historical time-series sensing data.

[0193] The description of step S1250 to step S1260 is similar to the description of the aforementioned step S1130 to step S1150 and will not be repeated here.

[0194] In step S1270, the historical time series sensing data has been pre-processed, sliced ​​and normalized to remove the differences between the original historical data. Therefore, these normalized historical sliced ​​time series data are suitable for subsequent feature acquisition and training. In this embodiment, the normalized historical sliced ​​time series data is used as the aforementioned Fig. 9 and Fig.10 The training sensing data, that is, the aforementioned training sensing data is data that has been pre-processed, sliced ​​and normalized.

[0195] In one embodiment, the computing device 112 randomly batch samples the normalized historical slice time series data of a unit time (ie, a time segment) to obtain a plurality of random batches of data, and performs data enhancement with the plurality of random batches of data to improve the training effect of the temperature prediction model.

[0196] In one embodiment, the computing device 112 and the temperature prediction device 600 are, for example, devices installed in the cloud. The present disclosure does not limit the devices executing the aforementioned method processes to be local devices or cloud devices.

[0197] In summary, the abnormality detection method, power regulation system and temperature prediction device for energy storage systems proposed in the present disclosure use multiple model encoders, weight models and model decoders of the temperature prediction model to accurately predict the temperature, so as to use the predicted temperature to compare with the current temperature, thereby evaluating whether the operating status of the energy storage system is normal. In this way, maintenance personnel can prepare for equipment failure or power outage problems in advance, and take countermeasures before the system is damaged and shut down, so as to solve the problem of overall system failure caused by damage to some electronic components of the equipment and achieve the technical effect of maintaining the stability of the overall system and uninterrupted power supply.

[0198] The above description is only a specific example of the present disclosure, and does not limit the claims of the present disclosure. Therefore, all equivalent changes made using the contents of the present disclosure are equally included in the scope of the present disclosure and are hereby stated.

Claims

1. A method for detecting anomalies in an energy storage system, comprising: (a) receiving a plurality of sensing data acquired from an electronic component of an energy storage system, wherein each of the sensing data is associated with a type; (b) inputting the plurality of sensing data into a temperature prediction model according to the type of each of the sensing data, wherein the temperature prediction model comprises a plurality of model encoders, a weight model and a model decoder; (c) receiving the plurality of sensing data of the same type through each of the model encoders and respectively calculating a plurality of time series features associated with the type from each of the plurality of sensing data of the type; (d) outputting the multiple time series features of each type to the weight model; (e) calculating a predicted temperature feature for the multiple time series features using the weight model and outputting the predicted temperature feature to a model decoder; (f) reconstructing the predicted temperature characteristic by the model decoder to generate a predicted temperature of the electronic component; and (g) evaluating whether the operation of the electronic component is abnormal by calculating an error value between the predicted temperature and a current temperature.

2. The anomaly detection method according to claim 1, wherein step (e) further comprises: Receiving the multiple time series features respectively through the multiple first layers of the weight model and calculating a first related feature through each of the first layers; Receiving the first correlation feature from each of the first layers through a second layer of the weight model and calculating the relationship between the first correlation feature of each of the first layers and a plurality of the first correlation features of the other plurality of first layers to generate a second correlation feature; and A plurality of the second related features from the second layer are received through a cascade layer of the weight model and the predicted temperature feature is calculated.

3. The anomaly detection method of claim 1 , wherein some of the plurality of time series features received by the model decoder are associated with a power module temperature, and step (f) further comprises: The model decoder calculates the predicted temperature feature based on the multiple time series features and the time series feature associated with the power module temperature, so that the model decoder reconstructs the predicted temperature feature with reference to the time series feature of the power module temperature.

4. The anomaly detection method according to claim 1, wherein step (g) further comprises: When the error value is greater than a threshold value, an abnormal number is accumulated; as well as When the number of abnormalities within a detection cycle is greater than a tolerance value, the operation of the electronic component is determined to be abnormal.

5. The anomaly detection method as claimed in claim 1, wherein before executing step (b), the temperature prediction model is trained, comprising: training a plurality of autoencoders according to a plurality of training sensing data respectively, wherein each of the autoencoders comprises a training encoder and a training decoder and each of the autoencoders is associated with one of the types respectively; When a difference between a training reconstruction data of each of the auto-encoders and each of the input training sensing data is less than a training threshold value, determining that the training of each of the auto-encoders is completed; as well as The multiple training encoders are set as the multiple model encoders.

6. The anomaly detection method as claimed in claim 5, wherein before executing step (d), the weight model is trained, comprising: Use a loss function to calculate a training error value; Determining whether to adopt the plurality of training sensing data according to the training error value; as well as A back propagation algorithm is performed using the adopted plurality of training sensing data to update parameters of the weight model. 7 . The abnormality detection method as claimed in claim 1 , wherein the type of each of the sensing data comprises a power module temperature, a water cooling system temperature, and a power system value. 8 . The abnormality detection method as claimed in claim 1 , wherein each of the sensing data is a time curve associated with the type.

9. A power regulation system comprising: a sensing module configured to obtain a plurality of sensing data of an electronic component, wherein each of the sensing data is respectively associated with a type; A storage medium coupled to the sensing module and configured to store a temperature prediction model, wherein the temperature prediction model includes: A plurality of model encoders configured to receive the plurality of sensing data, wherein each of the sensing data is respectively associated with a type; a weight model coupled to the plurality of model encoders; and a model decoder coupled to the weight model; and A computing device, coupled to the sensing module and the storage medium, is configured to use the temperature prediction model to perform a plurality of operations, including: receiving the plurality of sensing data of the same type through each of the model encoders and respectively calculating a plurality of time series features associated with the type from each of the plurality of sensing data of the type; Outputting the plurality of time series features of each type to the weight model; Calculate a predicted temperature feature from the multiple time series features through the weight model and output the predicted temperature feature to a model decoder; Reconstructing the predicted temperature characteristic by the model decoder to generate a predicted temperature of the electronic component; and Whether the operation of the electronic component is abnormal is evaluated by comparing the error value between the predicted temperature and a current temperature.

10. The power regulation system of claim 9, wherein the computing device is configured to: Using a plurality of first layers of the weight model to respectively receive the plurality of time series features and calculating a first correlation feature by each of the first layers; Using a second layer of the weight model to receive the first correlation features from each of the first layers and respectively calculate the relationship between the first correlation features of each of the first layers and a plurality of the first correlation features of the other first layers to generate a plurality of second correlation features; and A cascade layer using the weight model receives a plurality of the second correlation features from the second layer and calculates the predicted temperature feature.

11. A power regulation system as described in claim 9, wherein some of the multiple time series features received by the model decoder are associated with a power module temperature, and the computing device is configured to use the model decoder to calculate the predicted temperature feature for the multiple time series features and the time series feature associated with the power module temperature, so that the model decoder reconstructs the predicted temperature feature with reference to the time series feature of the power module temperature.

12. The power regulation system of claim 9, wherein the computing device is configured to: When it is determined that the error value is greater than a threshold value, accumulating an abnormal number of times; and When the number of abnormalities in a detection cycle is greater than a tolerance value, the operation of the electronic component is determined to be abnormal.

13. The power regulation system of claim 9, wherein the computing device is configured to: training a plurality of autoencoders according to a plurality of training sensing data respectively, wherein each of the autoencoders comprises a training encoder and a training decoder and each of the autoencoders is associated with one of the types respectively; When a difference between a training reconstruction data of each auto-encoder and each input training sensing data is less than a training threshold value, determining that the training of each auto-encoder is completed; and The multiple training encoders are set as the multiple model encoders.

14. The power regulation system of claim 13, wherein the computing device is configured to: Use a loss function to calculate a training error value; Determining whether to adopt the plurality of training sensing data according to the training error value; and A back propagation algorithm is performed using the adopted plurality of training sensing data to update parameters of the weight model. 15 . The power regulation system as claimed in claim 9 , wherein the type of each of the sensing data comprises a power module temperature, a water cooling system temperature, and a power system value.

16. A temperature prediction device, comprising: a storage medium configured to store a plurality of program codes; as well as A processor, coupled to the storage medium and configured to load the plurality of program codes to perform the following operations: Receiving a plurality of sensing data acquired from an electronic component of an energy storage system, wherein each of the sensing data is associated with a type; Inputting the plurality of sensing data into a temperature prediction model according to the type of each sensing data, wherein the temperature prediction model includes a plurality of model encoders, a weight model and a model decoder; receiving the plurality of sensing data of the same type through each of the model encoders and respectively calculating a plurality of time series features associated with the type from each of the plurality of sensing data of the type; Outputting the plurality of time series features of each type to the weight model; Calculating a predicted temperature feature from the multiple time series features using the weight model and outputting the predicted temperature feature to a model decoder; and The predicted temperature characteristic is reconstructed by the model decoder to generate a predicted temperature of the electronic component.

Images