Multi-objective optimization method and system assisted by gradient boosting neural network, and device

Abstract

The present invention relates to the technical field of objective optimization. Disclosed are a multi-objective optimization method and system assisted by a gradient boosting neural network, and a device. The method comprises: acquiring sampling data of an energy control management system of a hybrid electric vehicle by means of experimental design; on the basis of a gradient boosting neural network algorithm, determining an approximation function of a multi-objective optimization problem of the energy control management system; and using a multi-objective optimization algorithm to solve for the approximation function to obtain an optimal solution. The present invention can solve the problems that function evaluation is time-consuming, and an optimal solution for a problem cannot be obtained, etc.

Description

Gradient-enhanced neural network-assisted multi-objective optimization method, system, and device

[0001] This application claims priority to the Chinese patent application filed with the China Patent Office on June 12, 2023, with application number 202310691691.3 and invention name “Gradient Enhanced Neural Network Assisted Multi-Objective Optimization Method, System and Device”, the entire contents of which are incorporated by reference into this application. Technical Field

[0002] The present invention relates to the field of target optimization technology, and in particular to a multi-target optimization method, system and device assisted by a gradient enhancement neural network. Background Art

[0003] Hybrid electric vehicles are powered by an internal combustion engine and an electric motor. In the design of hybrid electric vehicles, the energy control manager plays a key role in ensuring hybrid performance as a key element.

[0004] Typically, the energy control and management system for a hybrid electric vehicle (HEV), as shown in Figure 1, consists of an internal combustion engine, a power control system, a battery, and an electric motor. Before the HEV begins driving, the battery is charged from the power grid. During braking, the electric motor acts as a generator, recycling braking energy back into the battery. The energy control and management system also uses the battery state of charge (SOC(t)) at time t as input to determine its specific control strategies.

[0005] Given that vehicle speed and torque are controlled by the driver or external devices, they are difficult to quantify using specific parameters and models. The design of energy control and management systems requires simultaneous consideration of multiple optimization objectives, and this type of problem is a black-box optimization problem. While existing modal analysis methods, such as finite element analysis and experimental modal analysis, can closely approximate real-world physical models, they often result in complex models and time-consuming evaluation.

[0006] To address this issue, the industry has begun using machine learning techniques to fit complex physical models. For example, surrogate models such as the binomial approximation and support vector machine models are used to fit complex functions. While these methods address some of the shortcomings of modal analysis, they cannot meet the requirements of fitting complex multi-objective optimization functions and have certain limitations in fitting complex physical models.

[0007] Summary of the Invention

[0008] The purpose of the present invention is to provide a multi-objective optimization method, system and device assisted by a gradient enhancement neural network to solve problems such as time-consuming function evaluation and inability to obtain the optimal solution to the problem.

[0009] To achieve the above object, the present invention provides the following solutions:

[0010] In a first aspect, the present invention provides a multi-objective optimization method assisted by a gradient boosting neural network, comprising:

[0011] Obtain sampling data of the energy control management system of hybrid electric vehicles through experimental design;

[0012] Determine the approximate function of the multi-objective optimization problem of the energy control and management system based on the gradient boosting neural network algorithm;

[0013] The multi-objective optimization algorithm is used to solve the approximate function and obtain the optimal solution.

[0014] Optionally, the sampling data includes at least a maximum threshold range of the battery state of charge when the internal combustion engine is turned off, a minimum threshold range of the battery state of charge when the internal combustion engine is turned off, a minimum speed range for operating the internal combustion engine, a maximum speed range for operating the internal combustion engine, an internal combustion engine torque value range, and a speed range for turning off the internal combustion engine.

[0015] Optionally, the objectives in the multi-objective optimization problem are: battery voltage, operation conversion times, emissions, noise and battery state of charge.

[0016] Optionally, according to a gradient boosting neural network algorithm, an approximate function of a multi-objective optimization problem of an energy control and management system is determined, specifically including:

[0017] According to the sampling data and Kriging model, the gradient boosting neural network is trained to obtain the approximate function of the multi-objective optimization problem.

[0018] Optionally, a multi-objective optimization algorithm is used to solve the approximate function to obtain an optimal solution, specifically including:

[0019] The multi-objective optimization algorithm NSGA-II is used to solve the approximate function and obtain the optimal solution.

[0020] In a second aspect, the present invention provides a multi-objective optimization system assisted by a gradient boosting neural network, comprising:

[0021] A data acquisition module, used for acquiring sampling data of the energy control management system of the hybrid electric vehicle through experimental design;

[0022] An approximate function determination module is used to determine the approximate function of the multi-objective optimization problem of the energy control and management system based on the gradient boosting neural network algorithm;

[0023] The optimization solution module is used to solve the approximate function using a multi-objective optimization algorithm to obtain the optimal solution.

[0024] In a third aspect, the present invention provides an electronic device comprising a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute a multi-objective optimization method assisted by a gradient enhancement neural network according to the first aspect.

[0025] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0026] The present invention uses a gradient enhancement neural network algorithm to convert the multi-objective optimization problem of the energy control management system into an approximate function, and solves this approximate function through a multi-objective optimization algorithm, thereby solving problems such as time-consuming function evaluation and the inability to obtain the optimal solution to the problem.

[0027] Figures in the specification

[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0029] FIG1 is a schematic structural diagram of an energy control and management system for a hybrid electric vehicle provided by an embodiment of the present invention;

[0030] FIG2 is a schematic diagram of a flow chart of a multi-objective optimization method assisted by a gradient boosting neural network according to an embodiment of the present invention;

[0031] FIG3 is an overall block diagram of a multi-objective optimization method assisted by a gradient boosting neural network according to an embodiment of the present invention;

[0032] FIG4 is a flowchart of a multi-objective optimization algorithm according to an embodiment of the present invention. DETAILED DESCRIPTION

[0033] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0034] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0035] Example 1

[0036] Hybrid electric vehicle energy management systems present complex and expensive multi-objective optimization problems. Current methods for solving these problems are often very expensive, especially due to the time-consuming function evaluation, which severely challenges the use of conventional solvers. This paper proposes a multi-objective optimization method assisted by a gradient-boosted neural network to address issues such as time-consuming function evaluation and the inability to obtain optimal solutions, often associated with finite element analysis and other methods.

[0037] The method designed in the present invention only considers the control and management optimization of the internal combustion engine part.

[0038] The specific descriptions of the multi-objective optimization problems to be solved by the present invention are shown in Table 1 and Table 2.

[0039] Table 1 Specific definitions of some parameters in the energy control and management system

[0040] Hybrid electric vehicles need to be operated in accordance with regulations, the specific rules are as follows:

[0041] Rule 1: If the speed is lower than V off , then shut down the internal combustion engine.

[0042] Rule 2: If the battery state of charge is greater than SOC max , then shut down the internal combustion engine.

[0043] SOC max It is not the maximum value of the battery state of charge. It only indicates the battery state of charge level. Its state value indicates that the battery is no longer charging.

[0044] Rule 3: If the battery state of charge is less than SOC min , then turn off the internal combustion engine. This rule has a higher priority than rule 1.

[0045] Rule 4: If the internal combustion engine is on and the speed is less than V1, then perform action 1.

[0046] Rule 5: If the internal combustion engine is on and the speed is between V1 and V2, then perform action 2.

[0047] Rule 6: If the internal combustion engine is on and the speed is greater than V2, then perform action 3.

[0048] Operations 1, 2 and 3 actually correspond to the torque values ​​and engine speeds defined in Table 1.

[0049] Table 2 Specific definition of objectives in the multi-objective optimization problem of energy control management system

[0050] To solve the complex and expensive multi-objective optimization problems in energy control and management systems (also known as hybrid power systems), this paper proposes to use a neural network model to perform function fitting for multi-objective black-box optimization problems. Because gradient information is difficult to obtain, this paper uses a Kriging model to fit the problem function before using the neural network model to obtain the gradient information.

[0051] Kriging predicts the value of a function at a given point by calculating the weighted average of the function's sample points in its neighborhood. Mathematically, this method is closely related to regression analysis. Kriging interpolation is a linear, unbiased, optimal estimation method that not only fully considers the positional relationship between the observed and estimated points, but also the spatial relationship between them.

[0052] The Kriging model can be expressed in the following form: y(x)=F(β,x)+z(x)=f T (x)β+z(x).

[0053] Where β is the regression coefficient, f T (x) is a polynomial regression model that simulates the global estimation model of the design space, which is called deterministic drift; z(x) is a random distribution error that is used to simulate the approximate local deviation. Its statistical characteristics can be expressed as follows: E[z(x)] = 0, Var[z(x)] = δ 2 ,Cov[z(x i ), z(x j )]=σ 2 R(c,x i , x j ).

[0054] In the formula, R(c, x i , x j ) is the correlation coefficient function of any two points with parameter c, representing any two points x i and x j The correlation function is:

[0055] Where n is the number of design variables, x t i and x t j x i and x j The tth component of t is the Euclidean distance between two points.

[0056] In the implementation of the gradient boosting machine, it is assumed that the prediction function after m iterations is Fm(x), and the corresponding loss function is L(y, Fm(x)). In order to achieve the goal of reducing the loss function as quickly as possible, the m+1th iteration sub-model function is constructed along the direction of the gradient descent of the loss function. The gradient descent direction at this time is:

[0057] Since the problem involved in the present invention is a multi-objective black-box optimization problem, it is impossible to obtain the problem gradient information. In order to solve this problem, the present invention establishes an approximate function of the black-box optimization problem by training the Kriging model, and then calculates the gradient of the problem by the above formula.

[0058] The loss function uses the commonly used L2, which calculates the square of the difference between the predicted value and the true value. The specific calculation formula is as follows:

[0059] Use the decision tree to construct the m+1th iteration sub-model function -gm(x) as the target variable to be predicted (in the case of the L2 loss function, the new target variable is the residual between the most recent prediction function and the original target variable). After obtaining the new decision tree model, the predicted value of all data is assumed to be h(x). Then after the m+1th iteration, the total prediction function should be: F m+1 (x) = F m (x)+β m+1 h(x).

[0060] Among them, β m+1 is the step size of each iteration, and the optimal step size of each iteration can be found by the least squares method.

[0061] According to the principle of gradient boosting neural network algorithm, the implementation route can be divided into the following steps:

[0062] 1. Select a training dataset, and all related models use the same dataset; give each data point in the dataset the same weight; select a basic model algorithm.

[0063] 2. Use the base model to predict the training data for the first time.

[0064] 3. Based on the first prediction error, each data point in the dataset is given a different weight; the greater the error of the data point, the higher the weight.

[0065] 4. Similarly, the model continuously trains the data set with adjusted weights, and each new model corrects the data that was misjudged by the previous model.

[0066] 5. The final model result is the weighted result of all models, so the effect is more ideal.

[0067] In each round of iteration, the negative gradient of the current model on all samples is first calculated, and then a new weak classifier is trained with this value as the target to fit and the weight of the weak classifier is calculated, and finally the model is updated.

[0068] Based on the above, this embodiment provides a gradient-enhanced neural network-assisted multi-objective optimization method, as shown in FIG2 , including the following steps.

[0069] Step 100: Acquire sampled data of the energy control and management system of the hybrid electric vehicle through experimental design. The sampled data includes at least a maximum threshold range of the battery state of charge when the internal combustion engine is turned off, a minimum threshold range of the battery state of charge when the internal combustion engine is turned off, a minimum speed range for operating the internal combustion engine, a maximum speed range for operating the internal combustion engine, a torque range of the internal combustion engine, and a speed range for turning off the internal combustion engine.

[0070] Step 200: Determine an approximate function for a multi-objective optimization problem of an energy control management system using a gradient boosting neural network algorithm. The objectives of the multi-objective optimization problem are: battery voltage, number of operation transitions, emissions, noise, and battery state of charge.

[0071] Step 300: Using a multi-objective optimization algorithm to solve the approximate function and obtain the optimal solution.

[0072] As a preferred embodiment, step 100 specifically includes:

[0073] The framework of the method provided in this embodiment is shown in FIG3 . Sample data is obtained through experimental design (DOE) and used for model training in S102 and S103 .

[0074] As a preferred implementation, step 200 specifically includes: training a gradient boosting neural network based on the sampled data and the Kriging model to obtain an approximate function of the multi-objective optimization problem.

[0075] In order to obtain the problem gradient information, the Kriging model is trained to establish an approximate analytical expression for the multi-objective optimization problem, which is used to calculate the gradient in the gradient enhanced neural network in S103. S102 is only a specific example of the overall algorithm process, which is not limited to the Kriging model. Other methods can also be used to obtain the problem gradient. That is, this application uses Kriging to first approximate the fitting function to obtain the gradient information, and then uses this gradient information to bring it into the gradient neural network to perform the final function fitting.

[0076] Based on the sample data in S101 and the Kriging model in S102, the GENN model is trained and an approximate function of the multi-objective optimization problem is obtained, which is used to evaluate the solution set in S104.

[0077] As a preferred implementation, step 300 specifically includes: using the multi-objective optimization algorithm NSGA-II to solve the approximate function to obtain the optimal solution.

[0078] The optimal solution is the optimal maximum battery state of charge threshold when shutting down the internal combustion engine, the minimum battery state of charge threshold when shutting down the internal combustion engine, the minimum speed of operating the internal combustion engine, the maximum speed of operating the internal combustion engine, the internal combustion engine torque value range, and the speed of shutting down the internal combustion engine.

[0079] This embodiment adopts the existing classic and widely used multi-objective optimization algorithm NSGA-II to process the multi-objective optimization problem fitted in step 300, wherein other multi-objective optimization algorithms are also suitable for embedding into the overall algorithm process.

[0080] According to the general process of multi-objective optimization algorithm, the NSGA-II implementation process is shown in Figure 4.

[0081] S201: Initialize the population according to the constraints of the optimization problem and the upper and lower bounds of the problem. Specifically, randomly initialize and generate N individuals that meet the constraints, where N is set to 100.

[0082] S202: Offspring Generation. Specifically, parent individuals are first selected using the tournament selection method. This method is a local competitive selection method that randomly selects k individuals from the population for comparison. For the multi-objective optimization problem to be minimized, individuals with the smallest fitness function values ​​are selected to enter the parent population. This is repeated N times until the parent population meets the required size. Polynomial crossover and mutation operations are then performed to generate new offspring.

[0083] S203: Environment Selection. Specifically, the generated offspring and parent generations are merged (200 individuals), and 100 outstanding individuals are selected from them. The selection method is non-dominated sorting. Then, the clustering distance of individuals in each non-Pareto-dominated layer is calculated, and individuals with large clustering distances are preferentially retained.

[0084] S204: Determine whether a termination condition is met. Specifically, the termination condition is the number of population cycle iterations.

[0085] S205: Output an approximately optimal solution set.

[0086] Example 2

[0087] In order to execute the method corresponding to the above embodiment 1 and achieve the corresponding functions and technical effects, a multi-objective optimization system assisted by a gradient enhancement neural network is provided below.

[0088] This embodiment provides a multi-objective optimization system assisted by a gradient boosting neural network, including:

[0089] The data acquisition module is used to obtain sampling data of the energy control management system of the hybrid electric vehicle through experimental design.

[0090] The approximate function determination module is used to determine the approximate function of the multi-objective optimization problem of the energy control and management system based on the gradient enhancement neural network algorithm.

[0091] The optimization solution module is used to solve the approximate function using a multi-objective optimization algorithm to obtain the optimal solution.

[0092] Example 3

[0093] An embodiment of the present invention provides an electronic device including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the multi-objective optimization method assisted by a gradient enhancement neural network of embodiment one.

[0094] Optionally, the above-mentioned electronic device may be a server.

[0095] In addition, an embodiment of the present invention further provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the multi-objective optimization method assisted by a gradient enhancement neural network of embodiment one.

[0096] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0097] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A multi-objective optimization method assisted by a gradient enhanced neural network, characterized in that: include: Obtaining sampling data of the energy control management system of hybrid electric vehicles through experimental design; Determine the approximate function of the multi-objective optimization problem of the energy control management system based on the gradient boosting neural network algorithm; The multi-objective optimization algorithm is used to solve the approximate function and obtain the optimal solution.

2. A multi-objective optimization method assisted by a gradient enhancement neural network according to claim 1, characterized in that: The sampling data includes at least a maximum threshold range of the battery state of charge when the internal combustion engine is turned off, a minimum threshold range of the battery state of charge when the internal combustion engine is turned off, a minimum speed range for operating the internal combustion engine, a maximum speed range for operating the internal combustion engine, an internal combustion engine torque value range, and a speed range for turning off the internal combustion engine.

3. A multi-objective optimization method assisted by a gradient enhancement neural network according to claim 1, characterized in that: The objectives in the multi-objective optimization problem are: battery voltage, operation switching times, emissions, noise, and battery state of charge.

4. The multi-objective optimization method assisted by a gradient enhancement neural network according to claim 1, characterized in that: According to the gradient enhancement neural network algorithm, the approximate function of the multi-objective optimization problem of the energy control management system is determined, including: According to the sampling data and Kriging model, the gradient enhancement neural network is trained to obtain the approximate function of the multi-objective optimization problem.

5. The multi-objective optimization method assisted by a gradient enhancement neural network according to claim 1, characterized in that: The multi-objective optimization algorithm is used to solve the approximate function and obtain the optimal solution, including: The multi-objective optimization algorithm NSGA-II is used to solve the approximate function and obtain the optimal solution.

6. A multi-objective optimization system assisted by a gradient-enhanced neural network, characterized in that: include: A data acquisition module, used for acquiring sampling data of an energy control management system of a hybrid electric vehicle through experimental design; An approximate function determination module, used to determine the approximate function of the multi-objective optimization problem of the energy control management system according to the gradient enhancement neural network algorithm; The optimization solution module is used to solve the approximate function using a multi-objective optimization algorithm to obtain the optimal solution.

7. An electronic device, characterized in that: It comprises a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform a gradient enhanced neural network assisted multi-objective optimization method according to any one of claims 1 to 5.

Images