Energy storage system performance prediction method and device based on combined feature selection

Abstract

The application provides a kind of energy storage system performance prediction method and device based on combination feature selection, the method comprises: obtaining the current performance parameter of each target subsystem in target battery system and the system category to which each target subsystem belongs;Determine the feature extraction strategy and target performance prediction model corresponding to each target subsystem based on the system category belonging;Based on the corresponding feature extraction strategy, the current target feature corresponding to each target subsystem is obtained by combination feature selection to current performance parameter;The corresponding current target feature is input into the corresponding target performance prediction model, to obtain the current performance prediction value of each target subsystem;According to the current performance prediction value of all target subsystems, determine the current performance prediction value of target battery system, by adaptively selecting the corresponding feature extraction strategy and the corresponding target performance prediction model for each target subsystem, real-time and effective performance prediction of target battery system is realized and the prediction accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of battery management technology, and in particular to a method and apparatus for predicting the performance of energy storage systems based on the selection of combined characteristics. Background Technology

[0002] Hydrogen fuel cells possess characteristics such as wide fuel adaptability, high energy conversion efficiency, all-solid-state operation, modular assembly, and zero pollution, making them promising mobile power sources for ships and vehicles. Hydrogen fuel cells are also being applied in distributed power generation, and distributed power stations are gradually becoming an important part of the energy supply due to their low cost and high maintainability.

[0003] The average system performance of a hydrogen fuel cell reflects its operational efficiency and is a key indicator of its lifespan. Predicting the average system performance is correlated with a series of variables, including the operating conditions, environmental factors, and control software version within the hydrogen fuel cell system. These variables are interconnected and mutually influential, making the evaluation of fuel cell performance and the relevant parameters for achieving optimal output performance extremely challenging. Current technologies typically rely on periodic manual monitoring to predict hydrogen fuel cell system performance; however, this method suffers from poor real-time performance and inconsistent accuracy. Summary of the Invention

[0004] This invention provides a method and apparatus for predicting the performance of energy storage systems based on the selection of combined features. This method addresses the shortcomings of existing methods that rely on periodic manual testing to predict the performance of hydrogen fuel cell systems, which suffer from poor real-time performance and difficulty in guaranteeing prediction accuracy. The invention aims to enhance the real-time performance and accuracy of hydrogen fuel cell system performance prediction.

[0005] This invention provides a method for predicting the performance of energy storage systems based on combined feature selection, comprising:

[0006] Obtain the current performance parameters of each target subsystem in the target battery system, as well as the system category to which each target subsystem belongs; the current performance parameters include multi-dimensional performance characteristics;

[0007] Based on the system category to which each target subsystem belongs, determine the feature extraction strategy and target performance prediction model corresponding to each target subsystem;

[0008] Based on the feature extraction strategy corresponding to each target subsystem, the multidimensional performance features are combined and selected to obtain the current target features corresponding to each target subsystem.

[0009] The current target features corresponding to each target subsystem are input into the target performance prediction model corresponding to each target subsystem to obtain the current performance prediction value of each target subsystem;

[0010] determining a current performance prediction value of the target battery system according to current performance prediction values of all the target subsystems;

[0011] The target performance prediction model corresponding to each target subsystem is obtained by training a pre-constructed deep learning model based on a decision tree algorithm, historical performance parameters of sample subsystems in a system category to which each target subsystem belongs at each historical time, and performance labels.

[0012] According to the energy storage system performance prediction method based on combined feature selection provided by the application, the method further comprises:

[0013] For each target subsystem, the following steps are performed:

[0014] Obtaining historical performance parameters of sample subsystems in a system category to which the current target subsystem belongs at each historical time and performance labels;

[0015] Based on a chi-square test algorithm, a first correlation coefficient between each performance feature in the historical performance parameters and the performance labels is calculated;

[0016] Based on a Spearman algorithm, a second correlation coefficient between each performance feature in the historical performance parameters and the performance labels is calculated;

[0017] According to the first correlation coefficient and the second correlation coefficient, a feature extraction strategy corresponding to the system category to which the current target subsystem belongs is determined;

[0018] A mapping relationship between the system category to which the current target subsystem belongs and the feature extraction strategy is established;

[0019] Based on the system category to which each target subsystem belongs, a feature extraction strategy corresponding to each target subsystem is determined, comprising:

[0020] According to the mapping relationship and the system category to which each target subsystem belongs, a feature extraction strategy corresponding to each target subsystem is obtained.

[0021] According to the energy storage system performance prediction method based on combined feature selection provided by the application, the determination of the feature extraction strategy corresponding to the system category to which the current target subsystem belongs according to the first correlation coefficient and the second correlation coefficient comprises:

[0022] The first correlation coefficient is compared with a first preset value;

[0023] According to the comparison result, at least one performance feature with the first correlation coefficient greater than the first preset value is determined in the historical performance parameters;

[0024] determining a performance feature in which the second correlation coefficient is greater than a second preset value;

[0025] determining a feature extraction strategy corresponding to a system category to which the current target subsystem belongs according to the performance feature in which the second correlation coefficient is greater than the second preset value.

[0026] According to the energy storage system performance prediction method based on combined feature selection provided by the application, the training steps of the target performance prediction model corresponding to each target subsystem include:

[0027] obtaining the historical performance parameters of the sample subsystem under the system category to which each target subsystem belongs and the performance label of the sample subsystem;

[0028] performing combined feature selection on the multi-dimensional performance features of the sample subsystem based on the feature extraction strategy corresponding to each sample subsystem to obtain the historical target features corresponding to the sample subsystem;

[0029] constructing a sample data set according to the historical target features and the performance label;

[0030] dividing the sample data set into a training set and a test set;

[0031] training the deep learning model according to the training set and the decision tree algorithm;

[0032] testing the model performance of the trained deep learning model according to the test set, and adjusting the model structure and / or model parameters of the trained deep learning model in the case that the trained deep learning model fails to pass the test;

[0033] continuing to train the adjusted deep learning model according to the training set and the decision tree algorithm until the trained deep learning model passes the test, thereby obtaining the target performance prediction model corresponding to each target subsystem.

[0034] According to the energy storage system performance prediction method based on combined feature selection provided by the application, before the method further includes:

[0035] determining whether there is a non-steady-state performance feature in the historical performance parameters of the sample subsystem at each historical time based on a steady-state processing algorithm;

[0036] In the case that there is a non-steady-state performance feature in the historical performance parameter of any historical moment, the historical performance parameter of the any historical moment and the performance label are deleted.

[0037] The application provides a performance prediction method of an energy storage system based on combined feature selection.

[0038] The current performance prediction values of all the target subsystems are weighted and added or averaged to obtain the current performance prediction value of the target battery system.

[0039] The application further provides an energy storage system performance prediction device based on combined feature selection.

[0040] The acquisition module is configured to acquire current performance parameters of each target subsystem in a target battery system and system categories to which each target subsystem belongs, and the current performance parameters include multi-dimensional performance features.

[0041] The first determination module is configured to determine, based on the system categories to which each target subsystem belongs, a feature extraction strategy and a target performance prediction model corresponding to each target subsystem.

[0042] The selection module is configured to perform combined feature selection on the multi-dimensional performance features based on the feature extraction strategy corresponding to each target subsystem to obtain current target features corresponding to each target subsystem.

[0043] The prediction module is configured to input the current target features corresponding to each target subsystem into the target performance prediction model corresponding to each target subsystem to obtain current performance prediction values of each target subsystem.

[0044] The second determination module is configured to determine, based on the current performance prediction values of all the target subsystems, a current performance prediction value of the target battery system.

[0045] The target performance prediction model corresponding to each target subsystem is obtained by training a pre-constructed deep learning model based on a decision tree algorithm, historical performance parameters of sample subsystems in each system category to which each target subsystem belongs at each historical moment, and performance labels.

[0046] The application further provides an electronic device including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, and the processor implements the performance prediction method of the energy storage system based on combined feature selection as described above when executing the program.

[0047] The application further provides a non-transitory computer-readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the energy storage system performance prediction method based on combined feature selection.

[0048] The application further provides a computer program product, which comprises a computer program, and the computer program is executed by a processor to implement the energy storage system performance prediction method based on combined feature selection.

[0049] The application provides an energy storage system performance prediction method and device based on combined feature selection, which comprises the following steps: acquiring current performance parameters of each target subsystem in a target battery system and system categories to which the target subsystems belong; acquiring the system categories to which the target subsystems belong, and determining feature extraction strategies corresponding to the target subsystems and target performance prediction models according to the acquired system categories; performing combined feature selection on the current performance parameters of the target subsystems according to the corresponding feature extraction strategies, and acquiring current target features corresponding to the target subsystems; inputting the current target features corresponding to the target subsystems into the corresponding target performance prediction models, taking output results as current performance prediction values of the target subsystems, integrating the current performance prediction values of the target subsystems to determine a current performance prediction value of the target battery system, and establishing a complete and automatic battery performance prediction process. The corresponding feature extraction strategies and the corresponding target performance prediction models are adaptively selected for each target subsystem, the target battery system is effectively and timely predicted, and the prediction accuracy is improved. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0051] Figure 1 is a flowchart of the energy storage system performance prediction method based on combined feature selection provided by the application;

[0052] Figure 2 is a structural schematic diagram of the target subsystem provided by the application;

[0053] Figure 3 is a correlation diagram between performance features provided by the application;

[0054] Figure 4 is a flowchart of determining the feature extraction strategy provided by the application;

[0055] Figure 5 is a flowchart of the target energy consumption prediction model training process provided by the present application;

[0056] Figure 6 is a flowchart of the model training process using the decision tree algorithm provided by the present application;

[0057] Figure 7 is a comparison diagram of the system performance prediction value and the actual system performance value provided by the present application;

[0058] Figure 8 is a flowchart of the target performance prediction model training process and prediction process corresponding to the target subsystem provided by the present application;

[0059] Figure 9 is a structural diagram of the energy storage system performance prediction device based on the combination feature selection provided by the present application;

[0060] Figure 10 is a structural diagram of the electronic device provided by the present application. DETAILED DESCRIPTION

[0061] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below with reference to the drawings in the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0062] It should be noted that the energy storage system performance prediction method and device based on the combination feature selection provided by the present embodiment can be applied to the performance prediction of various battery systems, such as hydrogen fuel cells, zinc-air batteries, etc. The present embodiment does not make specific limitations on this, and the following will be described taking the hydrogen fuel cell as an example, and other batteries can be adaptively replaced.

[0063] The hydrogen fuel cell is a power generation device that directly converts the chemical energy of hydrogen and oxygen into electric energy. The basic principle is that hydrogen is sent to the anode plate (negative electrode) of the fuel cell, and through the action of the catalyst (platinum), one electron in the hydrogen atom is separated. The hydrogen ion (proton) loses an electron and passes through the proton exchange membrane to reach the cathode plate (positive electrode) of the fuel cell. The electron cannot pass through the proton exchange membrane and can only reach the cathode plate of the fuel cell through an external circuit, thereby generating an electric current in the external circuit. After the hydrogen ion reaches the cathode plate, it combines with an oxygen atom and an electron to form water. The hydrogen fuel cell system mainly includes five parts: the cell stack, the hydrogen gas path, the air path, the cooling path, and the electrical path. The main function of the cell stack is to convert hydrogen and air into electric energy; the main function of the air path is to deliver air to the cell stack and discharge the water generated by the cell stack; the main function of the hydrogen gas path is to deliver hydrogen to the cell stack; the main function of the cooling path is to regulate the operating temperature of the cell stack; and the main function of the electrical path is to utilize the electric energy generated by the cell stack.

[0064] The hydrogen fuel cell has the characteristics of wide fuel adaptability, high energy conversion efficiency, full solid state, modular assembly, zero pollution, etc., and has broad application prospects as a mobile power source for ship power, vehicle power, space aviation, etc. The hydrogen fuel cell is applied to distributed power generation. Distributed power stations gradually become an important part of energy supply due to their low cost, high maintainability, etc. The system performance average of the hydrogen fuel cell reflects the operating efficiency of the cell and is a key indicator of cell life. The system performance average prediction is related to a series of variables such as operating conditions, environmental factors, and control software versions in the hydrogen fuel cell system. These variables are coupled and interact with each other, making it difficult to evaluate the related parameters of the fuel cell performance and obtain the best output performance. In the prior art, the hydrogen fuel storage system performance prediction based on combined feature selection is usually realized by artificial periodic detection, but this prediction method not only has poor real-time performance, but also has difficulty in ensuring detection accuracy.

[0065] To solve the above problems, the present application provides a hydrogen fuel storage system performance prediction method based on combined feature selection. By analyzing the dynamic changes between each variable and the current performance prediction value of the target cell system, an algorithm model is constructed with the current performance prediction value as the core, providing a more scientific and effective hydrogen fuel cell performance calculation and analysis method. While accurately and real-time calculating the current performance level, it can also analyze the factors affecting the performance of the fuel cell system. It can provide digital twin simulation means for performance analysis, design optimization and operation improvement of the hydrogen fuel cell system, and provide assistance for the design, manufacturing, operation and maintenance of the hydrogen fuel cell throughout its life cycle.

[0066] It should be noted that the execution subject of the method can be an electronic device, a component in the electronic device, an integrated circuit, or a chip. The electronic device can be a mobile electronic device or a non-mobile electronic device. Exemplarily, the mobile electronic device can be a mobile phone, a tablet computer, a notebook computer, a palm computer, a vehicle-mounted electronic device, a wearable device, an ultra-mobile personal computer (UMPC), a netbook, or a personal digital assistant (PDA), etc., and the non-mobile electronic device can be a server, a network attached storage (NAS), a personal computer (PC), a television (TV), a teller machine, or a self-service machine, etc., which are not limited in the present application.

[0067] The present application will be described below in combination with Figures 1-8 The present application provides a performance prediction method of an energy storage system based on combined feature selection.

[0068] Figure 1 FIG. 1 is a flowchart of the performance prediction method of the energy storage system based on combined feature selection provided by the present embodiment, as shown in the figure, the method comprises the following steps: Figure 1

[0069] In step 101, current performance parameters of each target subsystem in a current target battery system are obtained, and a system category to which each target subsystem belongs is obtained; the current performance parameters comprise multi-dimensional performance features.

[0070] The present embodiment is described below with the target battery system being a hydrogen fuel cell system. Each target subsystem is a same or different type of subsystem set in the current target battery system according to different positions, environments, and other factors.

[0071] Figure 2 FIG. 2 is a structural schematic diagram of the target subsystem provided by the present embodiment, as shown in the figure, the target subsystem can comprise a plurality of key elements, such as a stack, an electrical circuit, a cooling circuit, a hydrogen circuit, and an air circuit. Figure 2 The stack can comprise a battery stack composed of one or more hydrogen fuel cells; the electrical circuit can comprise a DCDC (Direct Current-Direct Current) and a motor; the cooling circuit can comprise a cooling fan, an electric heating device, a thermostat, and a cooling pump; the hydrogen circuit can comprise a hydrogen pump; and the air circuit can comprise an air compressor and a humidifier.

[0072] ​The performance parameters of each target subsystem are parameter data generated by each key element during current period operation of the battery, and can include set parameters and actual parameters. The set parameters can be parameter values set in advance according to rated operation of the target subsystem. The actual parameters can be actual loop parameters obtained according to feedback information in the loop of the target subsystem.

[0073] For example, the performance parameters of each target subsystem in the embodiment can include stack parameters, electrical loop parameters, cooling loop parameters, hydrogen loop parameters, air loop parameters, and environmental parameters.

[0074] The stack parameters can include control program version number parameters, system state parameters, system warning parameters, maximum and minimum voltage cell number parameters, performance tag parameters, stack feedback parameters, power limit flag parameters, low temperature flag parameters, maximum warning level parameters, hydrogen concentration parameters, emergency shutdown flag parameters, warning shutdown flag parameters, fuel cell temperature probe total number parameters, FCU (Fuel Cell Unit) heartbeat signal parameters, conductivity parameters, stack standard deviation parameters, and component maintenance frequency parameters. The maximum and minimum voltage cell number parameters can include maximum voltage cell number parameters and minimum voltage cell number parameters. The stack feedback parameters can include one or more controllable stack feedback data. The component maintenance frequency parameters can include one or more component maintenance frequency data.

[0075] The electrical loop parameters can include electrical loop current parameters, DCDC parameters, step-down DCDC parameters, and electrical loop set parameters. The electrical loop current parameters can include one or more electrical loop current feedback data. The DCDC parameters can include DCDC warning parameters, DCDC temperature parameters, DCDC heartbeat parameters, and DCDC state parameters. The step-down DCDC parameters can include step-down DCDC state parameters, step-down DCDC output current parameters, step-down DCDC output voltage parameters, step-down DCDC warning parameters, step-down DCDC temperature parameters, and step-down DCDC heartbeat parameters. The electrical loop set parameters can be parameter values set in advance for rated operation of one or more of the above-mentioned electrical loop current parameters, DCDC parameters, and step-down DCDC parameters.

[0076] The cold gas path parameters can include cooling path feedback parameters, water pump parameters, water pump feedback parameters, PTC (Pulse-Tube Cryocooler) parameters, PTC feedback parameters, cooling liquid parameters, and cooling path setting parameters. The cooling path feedback parameters can include one or more cooling path controllable feedback data and one or more cooling path non-controllable feedback data. The water pump parameters can include water pump warning parameters and water pump heartbeat parameters. The water pump feedback parameters can include one or more water pump feedback data. The PTC parameters can include PTC warning description parameters, PTC working state parameters, and PTC heartbeat parameters. The PTC feedback parameters can include one or more PTC feedback data. The cooling liquid parameters can include cooling liquid level parameters. The cooling path setting parameters can be parameter values pre-set for the rated operation of one or more of the above-mentioned cooling path feedback parameters, water pump parameters, water pump feedback parameters, PTC parameters, PTC feedback parameters, and cooling liquid parameters.

[0077] The hydrogen gas path parameters can include hydrogen circulation pump parameters, hydrogen circulation pump feedback parameters, hydrogen pump feedback parameters, hydrogen storage pressure parameters, hydrogen storage FCU heartbeat parameters, FCU hydrogen storage heartbeat parameters, hydrogen storage parameters, hydrogen gas path feedback parameters, and hydrogen gas path setting parameters. The hydrogen circulation pump parameters can include hydrogen circulation pump state parameters and hydrogen circulation pump heartbeat parameters. The hydrogen circulation pump feedback parameters can include one or more hydrogen circulation pump feedback data. The hydrogen pump feedback parameters can include one or more hydrogen pump feedback data. The hydrogen storage pressure parameters can include one or more hydrogen storage pressure value feedback data. The hydrogen storage parameters can include hydrogen storage maximum concentration parameters, hydrogen storage maximum concentration number parameters, hydrogen storage maximum temperature parameters, hydrogen storage state parameters, hydrogen storage warning level parameters, and hydrogen storage warning code parameters. The hydrogen gas path feedback parameters can include one or more hydrogen gas path feedback data. The hydrogen gas path setting parameters can be parameter values pre-set for the rated operation of one or more of the above-mentioned hydrogen circulation pump parameters, hydrogen circulation pump feedback parameters, hydrogen pump feedback parameters, hydrogen storage pressure parameters, hydrogen storage FCU heartbeat parameters, FCU hydrogen storage heartbeat parameters, hydrogen storage parameters, and hydrogen gas path feedback parameters.

[0078] The air path parameters can include air path feedback parameters, air volume adjustment times, air compressor parameters, air compressor feedback parameters, air compressor controller feedback parameters, and air path setting parameters. The air path feedback parameters can include one or more air path controllable feedback data and one or more air path uncontrollable feedback data. The air compressor parameters can include air compressor heartbeat parameters, air compressor software version parameters, and air compressor warning flag parameters. The air compressor feedback parameters can include one or more air compressor feedback data. The air compressor controller feedback parameters can include one or more air compressor controller feedback data. The air path setting parameters can be parameter values pre-set for the rated operation of one or more of the air path feedback parameters, the air volume adjustment times, the air compressor parameters, the air compressor feedback parameters, the air compressor controller feedback parameters, and the air path setting parameters.

[0079] The environmental parameters can include environmental feedback parameters, vehicle feedback parameters, vehicle setting parameters, vehicle parameters, cumulative mileage parameters, power bus parameters, and current limit parameters. The environmental feedback parameters can include one or more environmental feedback data. The vehicle feedback parameters can include one or more vehicle feedback data. The vehicle setting parameters can be parameter values pre-set for the rated operation of one or more of the vehicle feedback parameters. The vehicle parameters can include vehicle operating mode parameters and vehicle state of charge parameters. The power bus parameters can include power bus current parameters and power bus voltage parameters. The current limit parameters can include charge current limit parameters and discharge current limit parameters.

[0080] The current performance parameters of each target subsystem include data of the performance parameters of each target subsystem at the current time, or include data of the performance parameters of each target subsystem at the current time and at a plurality of historical times before the current time.

[0081] First, each target subsystem in the target battery system is scanned, and the current performance parameters of each target subsystem and the system category to which each target subsystem belongs are obtained based on the scanning results.

[0082] Optionally, determining the system category to which each target subsystem belongs can include first obtaining a subsystem number corresponding to each target subsystem, and then obtaining the system category to which each target subsystem belongs according to a pre-set mapping relationship between the subsystem number and the system category.

[0083] Step 102, based on the system category to which each target subsystem belongs, determining a feature extraction strategy corresponding to each target subsystem and a target performance prediction model;

[0084] Then, the feature extraction strategy corresponding to the target subsystem can be obtained according to the pre-set mapping relationship between each system category and the feature extraction strategy and the system category to which each target subsystem belongs. Similarly, the target performance prediction model corresponding to the target subsystem can be obtained according to the pre-set mapping relationship between each system category and the target performance prediction model and the system category to which each target subsystem belongs.

[0085] The target performance prediction model corresponding to each target subsystem is obtained by training a pre-constructed deep learning model based on a decision tree algorithm, historical performance parameters of sample subsystems in each system category at each historical time and performance labels.

[0086] Optionally, for the establishment of the target performance prediction model, a plurality of target performance prediction models can be established one by one for different system categories, and each system category and each performance prediction model can be stored in the form of a mapping relationship.

[0087] Before step 102 is performed, a corresponding performance prediction model needs to be established one by one for different system categories, and a one-to-one mapping relationship between different system categories and the corresponding performance prediction models needs to be established.

[0088] For the performance prediction model corresponding to each system category, the construction steps include:

[0089] First, the historical performance parameters and performance labels of sample subsystems in each historical time in the full life cycle under the system category are obtained.

[0090] After the historical performance parameters and performance labels of the sample subsystems are obtained, the historical performance parameters of the sample subsystems can be combined feature selected according to the feature extraction strategy corresponding to the sample subsystems to obtain the historical target features corresponding to the sample subsystems, then a sample data set is constructed according to the historical target features and the performance labels, and the sample data set is divided into a training set and a test set, and then the deep learning model is iteratively trained according to the training set and the decision tree algorithm until the trained deep learning model passes the test by using the test set, and the target performance prediction model corresponding to each target subsystem is obtained.

[0091] In step 103, the combined feature selection is performed on the multi-dimensional performance features based on the feature extraction strategy corresponding to each target subsystem to obtain the current target features corresponding to each target subsystem.

[0092] The features extracted by different feature extraction strategies are not the same.

[0093] Then, according to the feature extraction strategy corresponding to the target subsystem, the feature required to be selected for the current performance parameter is determined, so as to extract the corresponding current target feature from the current performance parameter, and input the current target feature into the target performance prediction model corresponding to the target subsystem to obtain the current performance prediction value of the target subsystem.

[0094] In step 104, the current target feature corresponding to each target subsystem is input into the target performance prediction model corresponding to each target subsystem to obtain the current performance prediction value of each target subsystem.

[0095] The current performance prediction value can be a prediction of the real-time performance value of the target subsystem at the current time by the target performance prediction model.

[0096] Optionally, the target feature is learned based on the target energy consumption prediction model, so as to predict and output the energy consumption prediction value of the target vehicle at the current time.

[0097] In step 105, the current performance prediction value of the target battery system is determined according to the current performance prediction values of all target subsystems.

[0098] Optionally, the current performance prediction value of the target battery system can be obtained by superimposing (weighted addition) the current performance prediction values of the target subsystems according to the actual scene, or can be obtained by inputting the current performance prediction values of the target subsystems into a pre-set scoring model for training, which is not limited in the embodiment.

[0099] The embodiment first obtains the current performance parameters of each target subsystem in the target battery system and the system category to which each target subsystem belongs, and obtains the system category to which each target subsystem belongs, determines the feature extraction strategy and the target performance prediction model corresponding to each target subsystem according to the obtained system category; then, according to the corresponding feature extraction strategy, the current performance parameters of each target subsystem are combined and the current target feature corresponding to each target subsystem is obtained; then, the current target feature corresponding to each target subsystem is input into the corresponding target performance prediction model, and the output result is taken as the current performance prediction value of each target subsystem, and then the current performance prediction values of each target subsystem are integrated to determine the current performance prediction value of the target battery system, thereby establishing a complete and automatic battery performance prediction process, adaptively selecting the corresponding feature extraction strategy and the corresponding target performance prediction model for each target subsystem, realizing real-time and effective performance prediction of the target battery system, and improving the prediction accuracy.

[0100] In some embodiments, the method further comprises:

[0101] For each target subsystem, the following steps are performed:

[0102] obtain historical performance parameters and performance labels of sample subsystems in each historical time under a system category to which the current target subsystem belongs; calculate first correlation coefficients between each performance feature in the historical performance parameters and the performance labels based on a chi-square test algorithm; calculate second correlation coefficients between each performance feature in the historical performance parameters and the performance labels based on a Spearman algorithm; determine a feature extraction strategy corresponding to the system category to which the current target subsystem belongs according to the first correlation coefficients and the second correlation coefficients; and establish a mapping relationship between the system category to which the current target subsystem belongs and the feature extraction strategy.

[0103] Based on the system category to which each target subsystem belongs, the feature extraction strategy corresponding to each target subsystem is determined, including: according to the mapping relationship and the system category to which each target subsystem belongs, the feature extraction strategy corresponding to each target subsystem is obtained.

[0104] First, according to the system category to which the current target subsystem belongs, the sample subsystem corresponding to the system category is obtained, and the sample subsystem is scanned, and the historical performance parameters and performance labels of the sample subsystem at each historical time in the whole life cycle are obtained according to the scanning result.

[0105] Optionally, the performance label can be obtained by pre-labeling according to the historical performance parameters of different sample subsystems at each historical time; in other words, there is a one-to-one mapping relationship between the historical performance parameters of the sample subsystem at each historical time and the performance label, and the performance label of the sample subsystem can be obtained according to the historical performance parameters of the sample subsystem at each historical time.

[0106] Optionally, the sample subsystem and its data can be classified and stored in the form of subsystem number according to the system category to which each target subsystem belongs in advance, so as to divide the original data of the battery system where the sample subsystem is located into sample subsystem data of multiple categories, and perform corresponding performance prediction model construction on the sample subsystem data of different categories, which can reduce the data amount of each performance prediction model construction, solve the problem of excessive data amount at the feature level, reduce time consumption, and increase the prediction effect.

[0107] Data classification is to merge data with certain common attributes or characteristics together, and distinguish data through its category attributes or characteristics. In other words, the same content, the same nature of information and the information required to be managed together, and the different and need to be separately managed information are distinguished, and then the relationship between each set is determined to form a well-organized classification system. The purpose of data classification is to assign new data objects to a correct category according to their attributes. Data classification emphasizes the division according to attributes and characteristics of different categories.

[0108] Then, feature selection is performed on each sample subsystem under each system category to determine the feature extraction strategy corresponding to each system category. The research of high-dimensional data is extremely challenging. Under the premise of ensuring the prediction accuracy of the learning algorithm, with the increase of feature dimension, the sample demand during training will increase exponentially. Feature selection can remove some irrelevant and redundant features, thereby achieving the effects of reducing dimension, reducing the number of features, reducing running time, and improving the running time of the algorithm, etc. Chi-square test algorithm and Spearman algorithm are two commonly used feature selection methods.

[0109] As shown in Figure 3 , taking a historical moment of a sample subsystem as an example, the correlations between performance characteristics are not the same, and the correlations between some performance characteristics are almost 0, which seriously affects the battery performance prediction.

[0110] Optionally, the first correlation coefficient between each performance characteristic and the performance label in the historical performance parameter can be calculated based on the chi-square test algorithm; wherein the first correlation coefficient can be a chi-square value calculated by the chi-square test algorithm.

[0111] In the chi-square test, X 2 value (i.e. chi-square value) describes the correlation between independent variables and dependent variables. Wherein X 2 is larger, indicating that the actual and expected gap is larger, and the independence between two variables is smaller, that is, more relevant; X 2 is smaller, indicating that the actual and expected gap is approximately, and the independence is larger, and the correlation is smaller, so X 2 value can be used for feature selection and other related work. The specific formula for calculating the chi-square value is as follows:

[0112]

[0113] Wherein, A is the actual frequency, T is the theoretical frequency, X 2 is the chi-square value.

[0114] Similarly, the second correlation coefficient between each performance feature in the historical performance parameter and the performance label can be calculated based on the Spearman algorithm; the second correlation coefficient can be a Spearman correlation coefficient calculated by the Spearman algorithm.

[0115] The Spearman correlation coefficient is a non-parametric index, and the calculation is based on the data ranks. When an external factor is the main feature affecting the mean of system performance, the value is close to 1, and vice versa. The calculation expression of the Spearman correlation coefficient is:

[0116]

[0117] wherein p is the Spearman correlation coefficient between two groups of vectors; n is the sample size; R and S are each performance feature and performance label; R i , S i are the i-th parameters of each performance feature and performance label vector, respectively; are the average ranks of each performance feature and performance label vector, respectively.

[0118] Then, according to the first correlation coefficient and the second correlation coefficient, the feature extraction strategy corresponding to the system category to which the current target subsystem belongs is determined;

[0119] Optionally, the feature extraction strategy corresponding to the system category to which the target subsystem belongs can be determined according to one or more performance features that are relatively more sensitive in the sample subsystem, i.e., one or more performance features with higher similarity, etc. For example, a performance feature set with first correlation coefficients and second correlation coefficients higher than a preset threshold can be selected, and the feature extraction strategy can be determined according to the performance feature set. Alternatively, the first correlation coefficients and the second correlation coefficients can be weighted and added to obtain a total correlation coefficient, and the feature extraction strategy can be determined according to a performance feature set with a total correlation coefficient higher than a preset threshold. The present embodiment does not make specific limitations in this regard.

[0120] Optionally, a mapping relationship between the system category to which the current target subsystem belongs and the feature extraction strategy can be established, which can be a mapping relationship directly existing between the system category to which the current target subsystem belongs and the feature extraction strategy. Alternatively, the mapping relationship can be established by setting an identifier for the system category to which the current target subsystem belongs and corresponding to the target feature extraction strategy. The present embodiment does not make specific limitations in this regard.

[0121] After the mapping relationship between the feature extraction strategy and the system category to which each target subsystem belongs is established, the target feature extraction strategy corresponding to the system category to which the target subsystem belongs can be obtained according to the mapping relationship and the system category to which the target subsystem belongs.

[0122] Correspondingly, when the feature extraction strategy corresponding to the system category to which the target subsystem belongs is acquired according to the mapping relationship, the feature extraction strategy can be directly acquired according to the mapping relationship between the two, or can be acquired according to the mapping relationship between the identifier pre-set for the system category to which the target subsystem belongs and the target feature extraction strategy. The embodiment is not limited in this regard.

[0123] In the embodiment, for each target subsystem, the sample subsystems under the system category to which the current target subsystem belongs and the historical performance parameters and performance labels of the sample subsystems at each historical moment are acquired. Then, based on the chi-square test algorithm and the Spearman algorithm, the first correlation coefficient and the second correlation coefficient between each performance feature in the historical performance parameters and the performance label are calculated, respectively. Then, according to the first correlation coefficient and the second correlation coefficient, the feature extraction strategy corresponding to the system category to which the current target subsystem belongs is determined, and the mapping relationship between the system category to which the current target subsystem belongs and the feature extraction strategy is established. Thus, the feature extraction strategy corresponding to each target subsystem can be acquired according to the mapping relationship and the system category to which each target subsystem belongs, and the feature extraction strategy applicable to the system category to which each target subsystem belongs can be acquired more flexibly, so as to help to realize real-time and effective performance prediction of the target battery system, reduce time consumption, and increase prediction effect.

[0124] In some embodiments, the determination of the feature extraction strategy corresponding to the system category to which the current target subsystem belongs according to the first correlation coefficient and the second correlation coefficient includes: comparing the first correlation coefficient with a first preset value; determining at least one performance feature in the historical performance parameters for which the first correlation coefficient is greater than the first preset value according to the comparison result; determining a performance feature in the at least one performance feature for which the second correlation coefficient is greater than a second preset value; and determining the feature extraction strategy corresponding to the system category to which the current target subsystem belongs according to the performance feature for which the second correlation coefficient is greater than the second preset value.

[0125] Figure 4 A flowchart for determining the feature extraction strategy is provided in the embodiment. As shown in FIG. 4, the determination of the feature extraction strategy corresponding to the system category to which the current target subsystem belongs includes the following steps: Figure 4

[0126] Step 401, data preparation, that is, to acquire the historical performance parameters and performance labels of the sample subsystems under the system category to which the current target subsystem belongs at each historical moment;

[0127] Step 402, after the data preparation is completed, the data can be preprocessed;

[0128] ​Step 403, after the calculation of the chi-square value is completed, each performance feature can be divided into a high interval and a low interval according to the first preset value and the calculated first correlation coefficient, and a division result is obtained;

[0129] Step 404, the performance features in the high interval are selected according to the division result to form a feature subset;

[0130] The first correlation coefficient of each performance feature in the historical performance parameter is compared with the first preset value, and at least one performance feature (i.e., a feature subset) with a first correlation coefficient greater than the first preset value is determined in the historical performance parameter according to a comparison result.

[0131] Step 405, after the calculation of the second correlation coefficient of the feature subset based on the Spearman algorithm is completed, the obtained second correlation coefficient is compared with a second preset value, and a comparison result is obtained;

[0132] Step 406, the features in the interval determined by the second preset value are removed, and the remaining performance features form a new feature subset; that is, in at least one performance feature, the performance feature with a second correlation coefficient less than or equal to the second preset value is removed, and the performance feature with a second correlation coefficient greater than the second preset value is determined to be retained.

[0133] Step 407, after the feature selection is completed, the set of performance features with a second correlation coefficient greater than the second preset value is retained as a feature extraction strategy corresponding to the system category to which the current target subsystem belongs, and the current step is ended.

[0134] The embodiment first compares the first correlation coefficient with the first preset value to obtain a comparison result, and determines at least one performance feature with a first correlation coefficient greater than the first preset value in the historical performance parameter according to the comparison result; then, in the at least one performance feature, a performance feature with a second correlation coefficient greater than the second preset value is determined, and a feature extraction strategy corresponding to the system category to which the current target subsystem belongs is determined according to the performance feature with the second correlation coefficient greater than the second preset value, so that the corresponding feature extraction strategy is adaptively selected for each target subsystem, which is helpful for real-time and effective performance prediction of the target battery system.

[0135] Figure 5 is a flowchart of a training process of a target energy consumption prediction model provided by the embodiment, as shown in Figure 5 In some embodiments, the training steps of the target performance prediction model corresponding to each target subsystem include:

[0136] Step 501, obtaining the historical performance parameters of sample subsystems under the system category to which each target subsystem belongs and the performance labels of the sample subsystems;

[0137] The historical performance parameters include performance parameters of the sample subsystem collected at one or more historical time points before the current time point.

[0138] It should be noted that for any sample subsystem at a historical time point, the historical performance parameters corresponding to the sample subsystem can include only the performance parameters at the historical time point, or include the performance parameters at the historical time point and the performance parameters at multiple historical time points before the historical time point.

[0139] In step 502, the multi-dimensional performance features of the sample subsystem are combined and selected based on the feature extraction strategy corresponding to each sample subsystem, to obtain historical target features corresponding to the sample subsystem.

[0140] According to the feature extraction strategy corresponding to the sample subsystem, the target features corresponding to the historical performance parameters can be combined and selected, so as to facilitate the training of the deep learning model.

[0141] Because the features in different sample subsystems are different, the selected features are also different. Therefore, for the historical performance parameters of the sample subsystem, an independent deep learning model is used for training, and the historical target features obtained by combining and selecting the historical performance parameters in the sample subsystem are selected as the input of the model, so as to realize effective differentiation of the sample subsystem and efficient learning and training of the deep learning model.

[0142] In step 503, a sample data set is constructed according to the historical target features and the performance labels.

[0143] The historical target features corresponding to each sample subsystem and the performance labels of the sample subsystem are stored in a one-to-one correspondence, and all sample subsystems under different system categories are formed into sample data sets corresponding to different system categories.

[0144] In step 504, the sample data set is divided into a training set and a test set.

[0145] In step 505, the deep learning model is trained according to the training set and the decision tree algorithm.

[0146] The LightGBM (Light Gradient Boosting Machine, decision tree) algorithm is used for model training. LightGBM is an integrated strong learning model based on distributed gradient boosting tree (Gradient Boosting Decision Tree, GBDT), which is applied to regression problems due to its advantages of fast speed, low internal consumption and high accuracy. LightGBM uses decision tree as the base learning algorithm, and its training formula can be expressed as:

[0147]

[0148] wherein H t is the t-th learner; and Θ is the set space of all learners.

[0149] Figure 6 is a flowchart of the model training process using the decision tree algorithm provided in this embodiment.

[0150] The LightGBM algorithm proposes a gradient-based one-side sampling algorithm (GOSS) for the problem of multiple samples, and an exclusive feature bundling algorithm (EFB) for the problem of multiple features. As shown in Figure 6 , the steps of model training using the decision tree algorithm are as follows:

[0151] First, start model training;

[0152] Next, select the large gradient training samples in the training set;

[0153] The LightGBM algorithm can use the obtained gradient to filter the training samples. The larger the gradient, the more under-learned it should be. If the large gradient training samples can be predicted correctly, the contribution to the gain will be greater, so it is hoped that the node splitting can accurately divide the large gradient training samples, and the small gradient training samples can have errors. Since so, when filtering the training samples, the large gradient training samples in the complete training set are retained.

[0154] Next, select the small gradient training samples in the training set;

[0155] Next, construct new small gradient training samples;

[0156] In the case of trying to keep the data distribution unchanged, sample the small gradient training samples in the training set.

[0157] Next, merge the samples;

[0158] Merge the complete large gradient training samples and the sampled small gradient training samples to obtain a new training set.

[0159] Next, train the weak learner;

[0160] Input the new training set into the model to train the weak learner.

[0161] Next, determine whether the iteration termination condition is reached;

[0162] If it is judged that the iteration termination condition is reached, the model parameters in the parameter space are input into the final prediction result of the deep learning model; otherwise, the step of selecting the large gradient training sample is returned to and the iterative training is continued.

[0163] Finally, the final prediction result is output.

[0164] In step 506, the model performance of the trained deep learning model is tested according to the test set, and in the case that the trained deep learning model fails the test, the model structure and / or model parameters of the trained deep learning model are adjusted.

[0165] In step 507, the adjusted deep learning model is continuously trained according to the training set and the decision tree algorithm until the trained deep learning model passes the test, and the target performance prediction model corresponding to each target subsystem is obtained.

[0166] After each training is completed, it is necessary to judge whether the deep learning model reaches the iteration termination condition, that is, the model performance of the trained deep learning model is tested by using the test set.

[0167] If the trained deep learning model fails the test, the model structure and / or model parameters of the trained deep learning model are adjusted, and then the adjusted deep learning model is trained according to the training set and the decision tree algorithm until the trained deep learning model passes the test, and the target performance prediction model corresponding to each target subsystem is obtained; if the trained deep learning model passes the test, the trained deep learning model is taken as the target performance prediction model.

[0168] Exemplarily, the model parameters can be a combination of one or more of the following parameters: num_leaves (number of leaf nodes), learning_rate (learning rate), feature_fraction (number of selected features), bagging_fraction (amount of data used), max_depth (maximum depth), min_child_weight (total weight of all samples in the minimum child node), and the like. Among them, num_leaves can be used to control the number of leaf nodes of each decision tree; learning_rate can be used to control the speed of updating the weight of each model in the gradient descent process; feature_fraction can be used to control how many features are selected when growing each decision tree; bagging_fraction can be used to control how much data the model will use in each iteration; max_depth can be used to determine the depth of the decision tree, i.e., the maximum depth reached by each leaf node; and min_child_weight can be the total weight of all samples in the minimum child node, which is not specifically limited in the embodiment.

[0169] It can be understood that, after obtaining the target performance prediction model, in order to effectively evaluate the prediction effect, the following three commonly used evaluation indexes of regression models are studied in the embodiment:

[0170] Root Mean Squared Error (RMSE) is the square root of the square of the deviation between the predicted value and the true value and the number of observations. In actual measurement, the number of observations is always limited, and the true value can only be replaced by the most reliable value. The calculation method is as follows:

[0171]

[0172] Mean Absolute Error (MAE) is the average absolute error between the true value and the predicted value, and is usually used to measure the closeness of the prediction and the actual result. The calculation method is as follows:

[0173]

[0174] Coefficient of Determination (R 2 ) is used to measure the predicted value compared with only using the mean value, and its interval is usually between (0, 1). The calculation method is as follows:

[0175]

[0176] As shown in Table 1, R 2MAE and RMSE as evaluation indexes, the comparison between the prediction results of the decision tree algorithm (hereinafter referred to as LightGBM) used in this embodiment and the prediction results of the existing linear regression algorithm can be seen that the prediction errors of LightGBM are less than those of linear regression, indicating the superiority of the decision tree algorithm for training in this embodiment.

[0177] Table 1 Model performance prediction result summary table

[0178] Evaluation metrics LightGBM Linear regression [R 2 ]]> 0.9995 0.9644 MAE 0.0009 0.0082 RMSE 0.0012 0.0106

[0179] Figure 7 For the comparison between the system performance prediction value and the actual value of the system performance, as shown in Figure 7 , it can be found that the target performance prediction model of this embodiment can accurately and effectively predict the performance of the battery system, indicating the superiority of the target performance prediction model of this embodiment.

[0180] In this embodiment, first, sample subsystems under the system category to which each target subsystem belongs are obtained, and the historical performance parameters of the sample subsystems and the performance labels of the sample subsystems are obtained. Then, according to the feature extraction strategy corresponding to each sample subsystem, the historical target features of the sample subsystems are combined and feature selected to obtain the historical target features corresponding to the sample subsystems, and the historical target features and the performance labels are integrated to construct a sample data set. Then, the obtained sample data set is divided into a training set and a test set. According to the training set and a decision tree algorithm, a deep learning model is trained, and then the model performance of the trained deep learning model is tested according to the test set. If the trained deep learning model does not pass the test, the model structure and / or model parameters of the trained deep learning model are adjusted, and the adjusted deep learning model is continuously trained until the trained deep learning model passes the test. Finally, the target performance prediction model corresponding to each target subsystem is obtained, realizing the self-adaptive selection of the corresponding target performance prediction model for each target subsystem, which helps to realize the real-time and effective performance prediction of the target battery system while improving the prediction accuracy.

[0181] In some embodiments, before the combination feature selection of the multi-dimensional performance features of the sample subsystems according to the feature extraction strategy corresponding to each sample subsystem to obtain the historical target features corresponding to the sample subsystems, the method further comprises: determining whether there is non-steady-state performance feature in the historical performance parameters of the sample subsystems at each historical time based on a steady-state processing algorithm; in the case that it is determined that there is non-steady-state performance feature in the historical performance parameters at any historical time, the historical performance parameters at the any historical time and the performance label are deleted.

[0182] The steady-state processing algorithm is used to determine whether there is a non-steady-state performance characteristic in the historical performance parameters of the sample subsystem at each historical time. For the historical performance parameters of each historical time, if there is a non-steady-state performance characteristic in the historical performance parameters of the historical time, the historical performance parameters of the sample subsystem at the historical time and the performance label of the sample subsystem at the historical time need to be deleted; if there is no non-steady-state performance characteristic in the historical performance parameters of the historical time, the historical performance parameters of the sample subsystem at the historical time and the performance label of the sample subsystem at the historical time are retained.

[0183] The specific steps of processing the feature data by using the steady-state processing algorithm are as follows:

[0184] First step: filter and retain the data with continuous changes in the FCU to VCU heartbeat parameter;

[0185] Wherein, continuous change means that there is a change within adjacent 60 rows, if there is no change for more than 60 rows, only the first row of data is retained.

[0186] Second step: filter and retain the data with the highest warning level parameter being 1 or 0;

[0187] Third step: filter and retain the data with the electrical circuit setting parameter 1 (independent variable) being 192;

[0188] Fourth step: on the basis of the third step, filter and retain the data with more than 600 consecutive rows as a group;

[0189] Wherein, continuous means that no row in the group of data is filtered out.

[0190] Fifth step: filter and retain the data that simultaneously satisfies 74.5 < cooling path feedback parameter 3 (controllable) < 77.5 and 74.5 < cooling path feedback parameter 5 (controllable) < 77.5;

[0191] Sixth step: filter out the first 180 rows of data in each group.

[0192] After completing the steady-state processing, the historical performance parameters are converted into steady-state data. The battery steady state means the continuous stable operation state of the battery, and the steady-state performance evaluation index selected according to the specific rules is the premise of judging the performance of the hydrogen fuel cell system.

[0193] Figure 8 The flowchart of the process of training and prediction of the target performance prediction model corresponding to the target subsystem is shown. As shown in FIG. 1, taking the target battery system as a hydrogen fuel cell system as an example, the whole process steps of the target performance prediction model training and prediction are as follows: Figure 8

[0194] First, start model training;

[0195] ​Next, historical performance parameters in the hydrogen fuel cell system are collected;

[0196] Next, the historical performance parameters are data grouped to obtain historical performance parameters of a sample subsystem;

[0197] Next, the historical performance parameters are steady-state processed;

[0198] Next, feature selection is performed on the steady-state processed historical performance parameters to obtain historical target features;

[0199] It can be understood that after obtaining the historical target features, the historical target features and performance labels are integrated to construct a sample data set.

[0200] Next, the sample data set is split into a training set and a test set;

[0201] Next, the deep learning model is trained according to the training set and the LightGBM algorithm;

[0202] Next, the model performance of the trained deep learning model is model verified according to the test set to determine whether the trained deep learning model passes the test;

[0203] In the case where the trained deep learning model does not pass the test, the model structure and / or model parameters of the trained deep learning model are adjusted, and the adjusted deep learning model is trained according to the training set and the decision tree algorithm until the trained deep learning model passes the test.

[0204] Next, the target performance prediction model corresponding to each target subsystem is obtained, and the current target features corresponding to the target subsystem are input into the target performance prediction model corresponding to each target subsystem;

[0205] The current target features corresponding to the target subsystem are obtained by first obtaining the current performance parameters of each target subsystem in the target battery system and the system category to which each target subsystem belongs; the current performance parameters include multi-dimensional performance features, then based on the system category to which each target subsystem belongs, the feature extraction strategy and the target performance prediction model corresponding to each target subsystem are determined, and then based on the feature extraction strategy corresponding to each target subsystem, the combined feature selection is performed to obtain the current target features corresponding to each target subsystem.

[0206] Finally, the output result is taken as the current performance prediction value of the target subsystem.

[0207] The embodiment determines whether there is a non-steady-state performance characteristic in the historical performance parameters of the sample subsystem at each historical moment based on a steady-state processing algorithm, deletes the historical performance parameters and performance labels of any historical moment if it is determined that there is a non-steady-state performance characteristic in the historical performance parameters of any historical moment, and realizes the conversion of the historical performance parameters into steady-state data, which helps to improve the pool performance prediction accuracy.

[0208] In some embodiments, the determining of the current performance prediction value of the target battery system according to the current performance prediction values of all the target subsystems comprises: weightedly adding or averaging the current performance prediction values of all the target subsystems to obtain the current performance prediction value of the target battery system.

[0209] Optionally, the current performance prediction value of the target battery system can be obtained by weightedly adding or averaging the current performance prediction values of all the target subsystems. For example, in the embodiment, the current performance prediction value of each target subsystem is calculated according to the following formula to obtain the current performance prediction value of the target battery system:

[0210]

[0211] Wherein U1 is the current performance prediction value of the target subsystem; is the current performance prediction value of the target battery system.

[0212] The embodiment obtains the current performance prediction value of the target battery system by weightedly adding or averaging the current performance prediction values of all the target subsystems, which helps to realize real-time and effective performance prediction of the target battery system and improve the prediction accuracy.

[0213] The energy storage system performance prediction device based on combined feature selection provided by the present application is described below. The energy storage system performance prediction device based on combined feature selection described below can be referred to in correspondence with the energy storage system performance prediction method based on combined feature selection described above.

[0214] As shown in Figure 9 FIG. 1 is a structural schematic diagram of the energy storage system performance prediction device based on combined feature selection provided by the present application. The device comprises:

[0215] The acquisition module 901 is configured to acquire the current performance parameters of each target subsystem in the target battery system and the system categories to which each target subsystem belongs. The current performance parameters comprise multi-dimensional performance characteristics.

[0216] The first determination module 902 is configured to determine the feature extraction strategy and the target performance prediction model corresponding to each target subsystem based on the system categories to which each target subsystem belongs.

[0217] The selection module 903 is configured to perform combined feature selection on the multi-dimensional performance features based on the feature extraction strategies corresponding to the target subsystems, to obtain current target features corresponding to the target subsystems.

[0218] The prediction module 904 is configured to input the current target features corresponding to the target subsystems into target performance prediction models corresponding to the target subsystems, to obtain current performance prediction values of the target subsystems.

[0219] The second determination module 905 is configured to determine a current performance prediction value of the target battery system according to the current performance prediction values of all the target subsystems.

[0220] The target performance prediction model corresponding to each target subsystem is obtained by training a pre-constructed deep learning model based on a decision tree algorithm, historical performance parameters and performance labels of sample subsystems in the system category to which the target subsystem belongs at each historical moment.

[0221] The energy storage system performance prediction device based on combined feature selection provided in the embodiment is configured to first acquire current performance parameters of each target subsystem in a target battery system and a system category to which each target subsystem belongs; acquire the system category to which each target subsystem belongs, and determine a feature extraction strategy and a target performance prediction model corresponding to each target subsystem according to the acquired system category; then perform combined feature selection on the current performance parameters of each target subsystem according to the corresponding feature extraction strategy, to acquire current target features corresponding to each target subsystem; then input the current target features corresponding to each target subsystem into the corresponding target performance prediction model, and take the output result as a current performance prediction value of each target subsystem; and then integrate the current performance prediction values of each target subsystem to determine a current performance prediction value of the target battery system, to establish a complete and fully automated battery performance prediction process. The corresponding feature extraction strategy and the corresponding target performance prediction model are adaptively selected for each target subsystem, to achieve real-time and effective performance prediction of the target battery system and improve the prediction accuracy.

[0222] In some embodiments, the first determining module 902 is specifically configured to perform the following steps for each target subsystem: obtaining historical performance parameters of sample subsystems in each historical time and performance labels of the sample subsystems in a system category to which a current target subsystem belongs; calculating first correlation coefficients between each performance feature in the historical performance parameters and the performance labels based on a chi-square test algorithm; calculating second correlation coefficients between each performance feature in the historical performance parameters and the performance labels based on a Spearman algorithm; determining a feature extraction strategy corresponding to the system category to which the current target subsystem belongs according to the first correlation coefficients and the second correlation coefficients; establishing a mapping relationship between the system category to which the current target subsystem belongs and the feature extraction strategy; and determining the feature extraction strategy corresponding to each target subsystem based on the system category to which each target subsystem belongs, including: obtaining the feature extraction strategy corresponding to each target subsystem according to the mapping relationship and the system category to which each target subsystem belongs.

[0223] In some embodiments, the first determining module 902 is further configured to: compare the first correlation coefficients with a first preset value; determine at least one performance feature in the historical performance parameters in which the first correlation coefficients are greater than the first preset value according to a comparison result; determine a performance feature in the at least one performance feature in which the second correlation coefficients are greater than a second preset value; and determine the feature extraction strategy corresponding to the system category to which the current target subsystem belongs according to the performance feature in which the second correlation coefficients are greater than the second preset value.

[0224] In some embodiments, the energy storage system performance prediction device based on combined feature selection further comprises a training module, which is specifically configured to: obtain the historical performance parameters of sample subsystems in each system category to which each target subsystem belongs and performance labels of the sample subsystems; perform combined feature selection on the multi-dimensional performance features of the sample subsystems to obtain historical target features corresponding to the sample subsystems based on the feature extraction strategy corresponding to each sample subsystem; construct a sample data set according to the historical target features and the performance labels; divide the sample data set into a training set and a test set; train the deep learning model according to the training set and the decision tree algorithm; test the model performance of the trained deep learning model according to the test set, and adjust the model structure and / or model parameters of the trained deep learning model in a case where the trained deep learning model fails the test; continue to train the adjusted deep learning model according to the training set and the decision tree algorithm until the trained deep learning model passes the test, and obtain a target performance prediction model corresponding to each target subsystem.

[0225] In some embodiments, the training module is further configured to determine, based on the steady-state processing algorithm, whether there is a non-steady-state performance feature in the historical performance parameter of each historical time point of the sample subsystem;

[0226] In a case where it is determined that there is a non-steady-state performance feature in the historical performance parameter of any historical time point, the historical performance parameter of the any historical time point and the performance label are deleted.

[0227] In some embodiments, the second determination module 905 is specifically configured to: add or average all the current performance prediction values of the target subsystems to obtain the current performance prediction value of the target battery system.

[0228] Figure 10 An example of an entity structure diagram of an electronic device is shown in Figure 10 As shown, the electronic device can include a processor 1001, a communications interface 1002, a memory 1003, and a communications bus 1004, wherein the processor 1001, the communications interface 1002, and the memory 1003 complete mutual communication through the communications bus 1004. The processor 1001 can invoke a logical instruction in the memory 1003 to execute a performance prediction method of an energy storage system based on combined feature selection, which includes: obtaining current performance parameters of each target subsystem in a target battery system, and system categories to which each target subsystem belongs; the current performance parameters include multi-dimensional performance features; determining a feature extraction strategy and a target performance prediction model corresponding to each target subsystem based on the system categories to which each target subsystem belongs; performing combined feature selection on the multi-dimensional performance features based on the feature extraction strategy corresponding to each target subsystem to obtain current target features corresponding to each target subsystem; inputting the current target features corresponding to each target subsystem into the target performance prediction model corresponding to each target subsystem to obtain current performance prediction values of each target subsystem; and determining a current performance prediction value of the target battery system according to the current performance prediction values of all the target subsystems; wherein the target performance prediction model corresponding to each target subsystem is obtained by training a pre-constructed deep learning model based on a decision tree algorithm, historical performance parameters of sample subsystems in each historical time point under the system categories to which each target subsystem belongs, and performance labels.

[0229] The electronic device provided by the embodiments of the present application is used to execute the above-mentioned method embodiments, and the specific process and detailed content are referred to the above-mentioned embodiments, which will not be repeated here.

[0230] Further, the logic instructions in the memory 1003 described above can be implemented in the form of software functional units and sold or used as independent products, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0231] In another aspect, the present application also provides a computer program product, which comprises a computer program, the computer program can be stored on a non-transitory computer readable storage medium, and the computer program can be executed by a processor to enable a computer to execute the energy storage system performance prediction method based on combined feature selection provided by the above-mentioned methods. The method comprises: obtaining current performance parameters of each target subsystem in a target battery system, and a system category to which each target subsystem belongs; the current performance parameters comprise multi-dimensional performance features; determining a feature extraction strategy and a target performance prediction model corresponding to each target subsystem based on the system category to which each target subsystem belongs; performing combined feature selection on the multi-dimensional performance features based on the feature extraction strategy corresponding to each target subsystem, to obtain current target features corresponding to each target subsystem; inputting the current target features corresponding to each target subsystem into the target performance prediction model corresponding to each target subsystem, to obtain current performance prediction values of each target subsystem; and determining a current performance prediction value of the target battery system according to the current performance prediction values of all target subsystems; wherein the target performance prediction model corresponding to each target subsystem is obtained by training a pre-constructed deep learning model based on a decision tree algorithm, historical performance parameters of sample subsystems in the system category to which each target subsystem belongs at each historical time, and performance labels.

[0232] The electronic device provided by the embodiments of the present application is used to execute the above-mentioned method embodiments, and the specific process and detailed content are referred to the above-mentioned embodiments, which will not be described here.

[0233] In yet another aspect, the present application also provides a non-transitory computer-readable storage medium having stored thereon a computer program, which, when executed by a processor, implements a method for predicting performance of an energy storage system based on combined feature selection, the method comprising: obtaining current performance parameters of each target subsystem in a target battery system, and a system category to which each target subsystem belongs; the current performance parameters comprising multi-dimensional performance features; determining a feature extraction strategy and a target performance prediction model corresponding to each target subsystem based on the system category to which each target subsystem belongs; performing combined feature selection on the multi-dimensional performance features based on the feature extraction strategy corresponding to each target subsystem to obtain current target features corresponding to each target subsystem; inputting the current target features corresponding to each target subsystem into the target performance prediction model corresponding to each target subsystem to obtain a current performance prediction value of each target subsystem; and determining a current performance prediction value of the target battery system according to the current performance prediction values of all target subsystems; wherein the target performance prediction model corresponding to each target subsystem is obtained by training a pre-constructed deep learning model based on a decision tree algorithm, historical performance parameters of sample subsystems in the system category to which each target subsystem belongs at each historical time, and performance labels.

[0234] The electronic device provided by the embodiments of the present application is used to execute the above-mentioned method embodiments, and the specific process and detailed content are referred to the above-mentioned embodiments, which will not be described here.

[0235] The device embodiments described above are only schematic, wherein the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment. Those skilled in the art can understand and implement without creative labor.

[0236] From the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be realized by means of software plus necessary general hardware platforms, and of course can also be realized by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part that contributes to the prior art can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0237] It should be pointed out finally that the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit the same; and although the present application has been described in detail with reference to the foregoing embodiments, it should be appreciated by those skilled in the art that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features thereof can be replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for performance prediction of energy storage system based on combined feature selection, characterized in that, The method comprises the following steps: obtaining current performance parameters of each target subsystem in a target battery system, and a system category to which each target subsystem belongs; the current performance parameters comprise multi-dimensional performance characteristics; the target subsystems comprise a stack, an electrical circuit, a cooling circuit, a hydrogen circuit and an air circuit, wherein the stack comprises a battery stack composed of one or more hydrogen fuel cells, the electrical circuit comprises a direct current converter and a motor, the cooling circuit comprises a cooling fan, an electric heating device, a thermostat and a cooling pump, the hydrogen circuit comprises a hydrogen pump, and the air circuit comprises an air compressor and a humidifier; determining a feature extraction strategy and a target performance prediction model corresponding to each target subsystem based on the system category to which each target subsystem belongs; performing combined feature selection on the multi-dimensional performance characteristics based on the feature extraction strategy corresponding to each target subsystem to obtain current target features corresponding to each target subsystem; inputting the current target features corresponding to each target subsystem into the target performance prediction model corresponding to each target subsystem to obtain current performance prediction values of each target subsystem; determining a current performance prediction value of the target battery system according to the current performance prediction values of all target subsystems; wherein the target performance prediction model corresponding to each target subsystem is obtained by training a pre-constructed deep learning model based on a decision tree algorithm, historical performance parameters of sample subsystems in the system category to which each target subsystem belongs at each historical time, and performance labels.

2. The method of claim 1, wherein, The method further comprises the following steps: for each target subsystem, the following steps are performed: obtaining historical performance parameters of sample subsystems in the system category to which the current target subsystem belongs at each historical time, and performance labels; calculating a first correlation coefficient between each performance characteristic in the historical performance parameters and the performance labels based on a chi-square test algorithm; calculating a second correlation coefficient between each performance characteristic in the historical performance parameters and the performance labels based on a Spearman algorithm; determining a feature extraction strategy corresponding to the system category to which the current target subsystem belongs according to the first correlation coefficient and the second correlation coefficient; establishing a mapping relationship between the system category to which the current target subsystem belongs and the feature extraction strategy; determining the feature extraction strategy corresponding to each target subsystem based on the system category to which each target subsystem belongs, comprises: obtaining the feature extraction strategy corresponding to each target subsystem according to the mapping relationship and the system category to which each target subsystem belongs.

3. The method of claim 2, wherein, The determination of the feature extraction strategy corresponding to the system category to which the current target subsystem belongs according to the first correlation coefficient and the second correlation coefficient comprises: comparing the first correlation coefficient with a first preset value; determining at least one performance characteristic in the historical performance parameters for which the first correlation coefficient is greater than the first preset value according to the comparison result; determining a performance characteristic in the at least one performance characteristic for which the second correlation coefficient is greater than a second preset value; determining the feature extraction strategy corresponding to the system category to which the current target subsystem belongs according to the performance characteristic for which the second correlation coefficient is greater than the second preset value.

4. The method of predicting performance of an energy storage system based on combined feature selection according to any one of claims 1-3, characterized in that, The training step of the target performance prediction model corresponding to each target subsystem comprises: obtaining the historical performance parameters of sample subsystems under the system category to which each target subsystem belongs and the performance labels of the sample subsystems; based on the feature extraction strategy corresponding to each sample subsystem, the multi-dimensional performance features of the sample subsystems are combined feature selected to obtain the historical target features corresponding to the sample subsystems; according to the historical target features and the performance labels, a sample data set is constructed; the sample data set is divided into a training set and a test set; according to the training set and the decision tree algorithm, the deep learning model is trained; according to the test set, the model performance of the trained deep learning model is tested, and in the case that the trained deep learning model fails to pass the test, the model structure and / or model parameters of the trained deep learning model are adjusted; continue to train the adjusted deep learning model according to the training set and the decision tree algorithm until the trained deep learning model passes the test, and obtain the target performance prediction model corresponding to each target subsystem.

5. The method of performance prediction of energy storage system based on combined feature selection according to claim 4, characterized in that, Before the combination feature selection of the multi-dimensional performance features of the sample subsystems based on the feature extraction strategy corresponding to each sample subsystem, the method further comprises: based on a steady-state processing algorithm, it is determined whether there is non-steady-state performance feature in the historical performance parameters of the sample subsystem at each historical time; in the case that there is non-steady-state performance feature in the historical performance parameters at any historical time, the historical performance parameters at the historical time and the performance labels are deleted.

6. The method of predicting performance of an energy storage system based on combined feature selection according to any one of claims 1-3, characterized in that, The determination of the current performance prediction value of the target battery system according to the current performance prediction values of all target subsystems comprises: the current performance prediction values of all target subsystems are weighted and added or averaged to obtain the current performance prediction value of the target battery system.

7. An energy storage system performance prediction device based on combined feature selection, characterized by, comprises: an acquisition module for acquiring the current performance parameters of each target subsystem in a target battery system and the system category to which each target subsystem belongs; the current performance parameters comprise multi-dimensional performance features; the target subsystems include stacks, electrical circuits, cooling circuits, hydrogen circuits and air circuits, wherein the stacks include battery stacks composed of one or more hydrogen fuel cells, the electrical circuits include direct current converters and motors, the cooling circuits include cooling fans, electric heating devices, thermostats and cooling pumps, the hydrogen circuits include hydrogen pumps, and the air circuits include air compressors and humidifiers; a first determination module for determining the feature extraction strategy and the target performance prediction model corresponding to each target subsystem based on the system category to which each target subsystem belongs; a selection module for combining feature selection of the multi-dimensional performance features based on the feature extraction strategy corresponding to each target subsystem to obtain the current target features corresponding to each target subsystem; The prediction module is configured to input the current target feature corresponding to each target subsystem into a target performance prediction model corresponding to each target subsystem to obtain a current performance prediction value of each target subsystem. The second determination module is configured to determine a current performance prediction value of the target battery system according to the current performance prediction values of all the target subsystems. The target performance prediction model corresponding to each target subsystem is obtained by training a pre-constructed deep learning model based on a decision tree algorithm, historical performance parameters of sample subsystems in a system category to which each target subsystem belongs at each historical moment, and performance labels.

8. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the energy storage system performance prediction method based on combined feature selection according to any one of claims 1 to 6 when executing the program. 9.A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program implements the energy storage system performance prediction method based on combined feature selection according to any one of claims 1 to 6 when executed by the processor.

10. A computer program product comprising a computer program, characterized in that, The computer program implements the energy storage system performance prediction method based on combined feature selection according to any one of claims 1 to 6 when executed by the processor.

Images