Improvement method of fuzzy control theory membership function by using convolutional neural network

Abstract

The application provides an improved method for determining a membership function of fuzzy control theory by using a convolutional neural network, comprising the following steps: S1, calculation and determination of an input layer of the convolutional neural network; S2, building the convolutional neural network, performing pooling training on the input related power, obtaining convolution calculation, constructing a loss function J, and using a gradient descent method to optimize the convolution kernel weight until a convergence condition is reached, and finally outputting the membership function of the related power; S3, establishing a fuzzy rule and a fuzzy reasoning system, and converting the related membership function into optimized total power and fuel cell power by using a fuzzy logic control theory; and finally establishing an optimization function compared with a traditional offline algorithm, and verifying the feasibility of the convolutional neural network combined with the fuzzy logic control in solving dynamic problems. The technical scheme has the advantages of the traditional offline optimization algorithm and the ability of real-time decision making, and is used for calculating large-scale working condition data to improve efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of applying convolutional neural network to ship energy management strategy, in particular to an improved method for determining membership function of fuzzy control theory by using convolutional neural network. BACKGROUND

[0002] The optimization-based control strategy can be generally divided into two categories, one is global optimization control strategy, and the other is real-time optimization control strategy. The global optimization control strategy, also known as offline optimization control strategy, is a control strategy that calculates a given multivariable nonlinear function in a certain limited interval and uses optimal control theory to design a local optimal solution. However, this kind of optimization strategy calculation often depends on the working condition, and it is difficult to solve when facing high-dimensional variables and a large amount of calculation.

[0003] Dynamic programming (DP) and Pontryagin maximum principle (PMP) are common global optimization energy management strategy methods that can effectively solve the optimal control problem of nonlinear dynamic systems. However, in actual working condition calculation, there is no unified method to handle the problem, and different characteristics need to be referred to and unique skills are needed to complete the calculation when facing different problems. Fuzzy logic control, as a rule-based control strategy, can solve complex nonlinear time-varying problems, but the determination of membership function and rules often depends on the designer's experience and intuition, and the effect is not good. SUMMARY

[0004] Therefore, the purpose of the present application is to provide an improved method for determining membership function of fuzzy control theory by using convolutional neural network, which not only combines the advantages of traditional offline optimization algorithm, but also has the ability of real-time decision-making, and is used for calculating large-scale working condition data to improve efficiency.

[0005] To achieve the above purpose, the technical scheme adopted by the present application is as follows: the improved method for determining membership function of fuzzy control theory by using convolutional neural network comprises the following steps:

[0006] Step S1: calculation and determination of the input layer of the convolutional neural network;

[0007] Step S2: building the convolutional neural network, pooling training the input related power, and obtaining convolution calculation, constructing loss function J, optimizing the convolution kernel weight by using gradient descent method until the convergence condition is reached, and finally outputting the membership function of the related power;

[0008] Step S3: Establishing fuzzy rules and fuzzy inference system, using fuzzy logic control theory to convert relevant membership functions into optimized total power and fuel cell power; finally, establishing an optimization function compared with traditional offline algorithm to verify the feasibility of convolutional neural network combined with fuzzy logic control in solving dynamic problems.

[0009] In a preferred embodiment, the step S1 comprises the following steps:

[0010] Step S11: Given a fixed working condition and running state of a marine vessel, a hybrid fuel power system is constructed, and a relevant mathematical model is established;

[0011] Step S12: An offline optimization control strategy is used to calculate the target function, and the local optimal solution in a series of time is obtained, which is used as the input layer of the convolutional neural network.

[0012] In a preferred embodiment, the fixed working condition and running state of the marine vessel specifically includes working condition information of the vessel, navigation parameters of the vessel, and load power demand of the vessel.

[0013] In a preferred embodiment, a hybrid fuel cell power system of the vessel is established, which is composed of a fuel cell energy system and a lithium battery energy system, and a corresponding mathematical model is established according to the hybrid fuel cell system.

[0014] In a preferred embodiment, the output power of the fuel cell is corrected according to the dynamic change amount of the lithium battery SOC, so that the lithium battery is stabilized in the optimal state of charge region, and then the equivalent hydrogen energy minimum principle is used to convert the electric energy consumption of the lithium battery into the consumption amount of hydrogen, thereby establishing the total hydrogen consumption amount at each sampling time.

[0015] First, define the reference value of the lithium battery SOC as

[0016]

[0017] Where SOC upper is the upper limit of the lithium battery state of charge; SOC lower is the lower limit of the lithium battery state of charge.

[0018] The output power P FC of the fuel cell is

[0019] P FC = P Load + k SOC P BA,ch

[0020] Where k SOC is the P FCAdjustment coefficient; P Load For the simulation of load power; P BA,ch For the standard charging power of lithium battery;

[0021] P = k * SOC FC Adjustment coefficient k SOC For

[0022]

[0023] Wherein μ is the proportional coefficient, used for fast balancing the charging and discharging state of lithium battery;

[0024] According to the dynamic programming method, in a continuous period of time, the average is divided into a series of sampling time, the optimal value function is established at each sampling time, and the optimal value function is solved, so as to obtain two groups of time t and load power P Load And fuel cell power P FC The related function group determines the output layer of convolutional neural network.

[0025] In a preferred embodiment, the convolutional neural network includes input layer, convolutional layer, pooling layer, and full connection layer.

[0026] The total power P Load And fuel cell power P FC The related function group at sampling time is input to the input layer;

[0027] The convolutional layer extracts features from the input load power P Load And fuel cell power P FC And converts the two groups of power into weighted power;

[0028] The pooling layer uses a combination of mixed operator symbols to pool the weighted power;

[0029] The full connection layer processes the results of the pooling layer to obtain the membership function f(P) of the related power;

[0030] In the training process of convolutional neural network, the loss function is constructed to predict the result data of training, and the gradient descent method is used to update the convolution kernel parameters until the loss function reaches the convergence condition, so as to output the final related power membership function f(P FC ) and f(P Load ).

[0031] In a preferred embodiment, fuzzy logic control theory is used to establish fuzzy reasoning system and fuzzy rules, and the input of the system is the load power P Loadand the comprehensive state of charge coefficient SOC, the output of the system is the hybrid fuel cell power P FC , the membership functions of the input and output are established; the input SOC has three membership sets, which are L, M and H respectively;

[0032] The input is the load power P Load and the output is the fuel cell power P FC has five membership sets, which are VL, L, M, H and VH respectively; wherein, VL, L, M, H and VH represent Very Low, Low, Medium, High and Very High respectively;

[0033] By adjusting the through current of the one-way DC / DC converter, the input comprehensive state of charge coefficient SOC is confirmed, the battery power distribution weight of the hybrid fuel system is confirmed, and the output load power P and the fuel cell power P

[0034] In a preferred embodiment, an optimization function is established between the power result obtained by the dynamic programming method and the power result calculated by the fuzzy logic control combined with the convolutional neural network training, and the optimization function is as shown below

[0035]

[0036] If the final calculation result of the optimization function does not exceed the cumulative error index, it can be explained that the calculation result of the fuzzy logic control strategy combined with the convolutional neural network training can be approximately replaced by the optimal real-time decision result of the ship working condition by the dynamic programming method.

[0037] Compared with the prior art, the present application has the following beneficial effects: the convolutional neural network is applied in the process of solving the membership function of the fuzzy logic control, the mathematical model of the hybrid fuel cell power system is established, the output power of the DC / DC converter is adjusted, the power of the related battery is obtained as the input layer, a multi-layer convolutional neural network is built for training and calculation, a loss function is constructed, and the membership function meeting the convergence condition is taken as the output layer to calculate the fuzzy logic control strategy. Since the past membership function relies on the experience and intuition of the designer, the introduction of the convolutional neural network can more efficiently calculate large-scale working condition data. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 It is a flowchart of the convolutional neural network combined with the fuzzy logic control of the preferred embodiment of the present application;

[0039] Figure 2 It is a ship navigation working condition diagram of the preferred embodiment of the present application;

[0040] Figure 3 A mathematical model diagram of a hybrid fuel cell power system as a preferred embodiment of the present application;

[0041] Figure 4 A flowchart of calculating a membership function by a convolutional neural network as an embodiment of the present application. DETAILED DESCRIPTION

[0042] The present application will be further described below in conjunction with the accompanying drawings and embodiments.

[0043] It should be noted that the following detailed description is illustrative only, and is intended to provide further description in order to provide a further understanding of the present application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0044] It should be noted that the terms used herein are only intended to describe specific embodiments and are not intended to limit the exemplary embodiments according to the present application; as used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it should be further understood that when the terms "comprise" and / or "include" are used in the specification, there is a presence of a feature, step, operation, device, component, and / or combinations thereof.

[0045] Embodiment One:

[0046] Figure 1 A flowchart of a method for applying a convolutional neural network to calculate an optimal power distribution for a ferry ship as an embodiment of the present application. In this embodiment, the following steps are included:

[0047] S1: Calculate and determine the input layer of the convolutional neural network.

[0048] S1.1: Given a fixed working condition and a running state of a marine ship, a hybrid fuel power system is constructed, and a relevant mathematical model is established.

[0049] S1.2: An offline optimization control strategy is used to split and calculate the objective function, and a local optimal solution is obtained within a series of time, and the local optimal solution is used as the input layer of the convolutional neural network.

[0050] S2: Build a convolutional neural network, pool and train the input related power, and obtain convolution calculation, construct a loss function J, and use gradient descent method to optimize the convolution kernel weight until the convergence condition is reached, and finally output the membership function of the related power.

[0051] S3: Establish fuzzy rules and fuzzy inference system, use fuzzy logic control theory to convert the relevant membership function into the optimized total power and fuel cell power. Finally, compared with the traditional off-line algorithm, the optimization function is established, so as to verify the feasibility of convolutional neural network combined with fuzzy logic control to solve dynamic problems.

[0052] According to the specific example implementation, the working condition information of the ship is given,

[0053] The navigation parameters of the ship,

[0054] The load power demand of the ship

[0055] According to the parameter information, a ship hybrid fuel cell power system is established, which is composed of two energy systems, specifically including:

[0056] One is the fuel cell energy system, because the output voltage of the fuel cell varies greatly and is lower than the bus voltage connected to it, so it needs to be connected in cascade with a unidirectional DC / DC converter to isolate the high-frequency impact of the load, and can better control the operating point, making the system more reliable and stable.

[0057] The other is the lithium battery energy system, because lithium batteries have the advantages of high efficiency and high power density, and can absorb the power grid impact generated when the fuel cell load suddenly changes, so as a secondary auxiliary power supply, it will make the system more stable, and in the case that the SOC of the lithium battery is high enough, the output power of the power system can be adjusted and distributed by adjusting the DC / DC converter connected in cascade with the fuel cell, so the working point is directly connected to improve the energy utilization rate of the system. Finally, according to the hybrid fuel cell system, a mathematical model is established,

[0058] In order to avoid complex calculation caused by multiple targets, the output power of the fuel cell is corrected according to the dynamic change of the lithium battery SOC, so that the lithium battery is stabilized in the optimal state of charge area, and then the equivalent hydrogen energy minimum principle is used to convert the lithium battery power consumption into hydrogen consumption, so that the total hydrogen consumption at each sampling time can be established,

[0059] First, define the reference value of lithium battery SOC as

[0060]

[0061] Where SOC upper is the upper limit of the lithium battery state of charge, %; SOC lower is the lower limit of the lithium battery state of charge, %.

[0062] The output power P FC of the fuel cell is

[0063]

[0064] where k SOC is the SOC state-based P Fc adjustment coefficient; P Load is the simulated load power, kW; P BA,；h is the standard charging power of the lithium battery, kW.

[0065] where k fC is the SOC state-based P SOC adjustment coefficient

[0066]

[0067] where μ is a proportional coefficient for quickly balancing the charging and discharging states of the lithium battery.

[0068] According to the dynamic programming method (DP algorithm), in a continuous period of time, the average is divided into a series of sampling times, the optimal value function is established at each sampling time, and the optimal value functions are solved to obtain two sets of correlation functions of the total power P Load and the fuel cell power P FC , and determine the output layer of the convolutional neural network.

[0069] The convolutional neural network comprises an input layer, a convolutional layer, a pooling layer, and a fully connected layer.

[0070] The total power P Load and the fuel cell power P FC at the sampling time are used as the input of the input layer.

[0071] The convolutional layer extracts features from the input load power P Load and the fuel cell power P FC , and converts the two sets of power into weighted power.

[0072] The pooling layer uses a combination of mixed operator symbols to pool the weighted power.

[0073] The fully connected layer processes the results of the pooling layer to obtain the membership function f(P) of the relevant power.

[0074] In the training process of the convolutional neural network, the loss function is used to predict the result data of the training, and the gradient descent method or the back propagation algorithm is used to train and update the convolution kernel parameters until the loss function reaches the convergence condition, so as to obtain the final membership functions f(P FC ) and f(P Load ) of the relevant power.

[0075] In this embodiment, the output membership function is fed back into the established hybrid fuel cell power system mathematical model to obtain the relevant power group And

[0076] The loss function J is established by L2 norm regularization:

[0077]

[0078] In the formula, N represents the number of time division, that is, the number of training samples, and a is the regularization strength. Deep learning back-propagates the weighting coefficient ω in each layer by gradient descent method i until the loss function J reaches the convergence condition. In order to avoid overfitting in the learning process, once the generalization ability of the learning model decreases, the learning process will stop or slow down, at which time a needs to be adjusted.

[0079] Finally, the fuzzy rule and fuzzy reasoning system are constructed to detect whether the calculation results of the relevant power output by the fuzzy logic control strategy meet the index that the relative cumulative error is within 5% compared to the results calculated by the dynamic programming method. If not, the convolutional neural network needs to be improved; if it meets the requirements, the convolutional neural network for calculating the fuzzy logic control membership function is obtained.

[0080] Example Two:

[0081] Figure 2 The working condition parameter table of the ship navigation is Figure 3 The schematic diagram of the hybrid fuel cell power system is

[0082] In this embodiment, according to the mathematical model of the hybrid fuel cell power system Figure 3 , it is assumed that the system is composed of three energy systems, specifically including:

[0083] One is the fuel cell energy system. Since the output voltage of the fuel cell varies greatly and is lower than the bus voltage connected thereto, it is necessary to connect a unidirectional DC / DC converter in series to isolate the high-frequency impact of the load and better control the operating point, making the system more reliable and stable.

[0084] One is the lithium battery energy system. Since the lithium battery has the advantages of high efficiency and high power density, and can timely absorb the power grid impact generated when the fuel cell load suddenly changes, it can make the system more stable as a secondary auxiliary power supply. In the case that the SOC of the lithium battery is high enough, the output power of the power system can be adjusted and distributed by adjusting the DC / DC converter connected in series with the fuel cell, so that the energy utilization rate of the system is improved.

[0085] The last one is super capacitor energy system, super capacitor has higher power density, much larger than general lithium ion battery, but its endurance is insufficient, through connecting with DC / DC converter, system efficiency can be improved and power energy can be distributed.

[0086] According to Figure 1 S1 in the formula (1)-(3), start to establish mathematical model of hybrid fuel cell power system, and according to equivalent hydrogen energy consumption minimum principle, avoid complex calculation amount caused by multiple targets, and then through dynamic programming method, a series of power groups of three energy systems in a period of time are obtained, as input layer of convolutional neural network.

[0087] According to S2, build convolutional neural network including input layer, convolutional layer, pooling layer, fully connected layer and output layer, the main function of convolutional layer is to identify and extract features of input layer, pooling layer is to reduce sampling processing of features extracted by convolutional layer, reduce complexity of convolution calculation and improve robustness of feature extraction, fully connected layer is a link of convolution unit, so that convolution network achieves complete and close learning and training process, and finally output related power membership function.

[0088] Establish L2 norm regularization loss function J, and through gradient descent method, update weight of convolution kernel and learn parameters, until loss function J reaches specific convergence condition, then output layer of convolutional neural network can be finally determined as membership function of related power.

[0089] According to S3, build fuzzy rule and fuzzy reasoning system, detect calculation results of related power output by fuzzy logic control strategy, whether the calculation results meet the index that relative cumulative error is within 5%, if not, improve convolutional neural network, if meet the requirement, obtain convolutional neural network for calculating fuzzy logic control membership function.

[0090] It should be noted that although the present application is specifically shown and introduced in combination with preferred embodiments, those skilled in the art should understand that various changes can be made to the present application in form and detail without departing from the spirit and scope of the present application defined in the appended claims, and all changes are within the protection scope of the present application.

Claims

1. An improved method for determining the membership function of fuzzy control theory using convolutional neural networks, characterized in that, Includes the following steps: Step S1: Calculate and determine the input layer of the convolutional neural network; Step S2: Build a convolutional neural network, perform pooling training on the relevant power of the input, obtain convolution calculation, construct the loss function J, and use gradient descent to optimize the convolution kernel weights until the convergence condition is met. Finally, output the membership function of the relevant power. Step S3: Establish fuzzy rules and a fuzzy inference system, and use fuzzy logic control theory to convert the relevant membership functions into optimized total power and fuel cell power; finally, establish an optimization function to compare with traditional offline algorithms, thereby verifying the feasibility of using convolutional neural networks combined with fuzzy logic control to solve dynamic problems; The convolutional neural network consists of four parts: an input layer, a convolutional layer, a pooling layer, and a fully connected layer. The total power and fuel cell power The relevant function set at the sampling time is used as the input to the input layer; The load power of the convolutional layer on the input and fuel cell power Feature extraction is performed, and the two sets of power are converted into weighted power. The pooling layer uses a combination of mixed operators to pool weighted power; The fully connected layer processes the pooling results from the pooling layer to obtain the membership function of the relevant power. ; During the training process of a convolutional neural network, a loss function is constructed to predict the training results data, and gradient descent is used to update the convolutional kernel parameters until the loss function reaches the convergence condition, thereby outputting the final relevance power membership function. and ; Using fuzzy logic control theory, a fuzzy inference system and fuzzy rules are established. The system inputs are the load power of the integrated fuel cell. Combined with the overall charge-state coefficient (SOC), the system output is the hybrid fuel cell power. Establish membership functions for input and output variables; the input variable SOC has three membership sets, namely L, M and H; Input load power and output fuel cell power There are five membership sets, namely VL, L, M, H and VH; where VL, L, M, H and VH represent Very Low, Low, Medium, High and Very High, respectively. By adjusting the current through the unidirectional DC / DC converter and inputting the overall state of charge (SOC), the battery power allocation weight of the hybrid fuel system is determined, and the load power is obtained from the output. and fuel cell power .

2. The improved method for determining the membership function of fuzzy control theory using a convolutional neural network according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Given a fixed operating condition and navigation state of a marine vessel, construct a hybrid fuel power system and establish a relevant mathematical model; Step S12: Use an offline optimization control strategy to decompose and calculate the objective function to obtain local optimal solutions over a series of time intervals, and use the local optimal solutions as the input layer of the convolutional neural network.

3. The improved method for determining the membership function of fuzzy control theory using a convolutional neural network according to claim 2, characterized in that, The fixed operating conditions and navigation status of the ship at sea specifically include the ship's operating condition information, the ship's navigation parameters, and the ship's load power requirements.

4. The improved method for determining the membership function of fuzzy control theory using a convolutional neural network according to claim 2, characterized in that, A hybrid fuel cell power system for ships is established, which consists of a fuel cell energy system and a lithium battery energy system. A corresponding mathematical model is then developed based on the hybrid fuel cell system.

5. The improved method for determining the membership function of fuzzy control theory using a convolutional neural network according to claim 1, characterized in that, The output power of the fuel cell is corrected based on the dynamic change of the lithium battery's SOC, so that the lithium battery is stabilized in the optimal state of charge region. Then, the electrical energy consumption of the lithium battery is converted into the hydrogen consumption through the principle of minimum equivalent hydrogen energy, and the total hydrogen consumption at each sampling time is established. First, define the reference value for the SOC of a lithium battery as follows: in This is the upper limit set for the nuclear power state of lithium batteries; This is the set limit for the nuclear power state of lithium batteries; fuel cell output power for in For SOC state based Adjustment coefficient; To simulate load power; Standard charging power for lithium batteries; The formula is based on the SOC state. Adjustment coefficient for Where μ is a proportionality coefficient used to quickly balance the charge and discharge states of the lithium battery; According to dynamic programming, a continuous time interval is divided into a series of sampling times. An optimal value function is established for each sampling time, and these optimal value functions are solved to obtain two sets of time t and load power. and fuel cell power The relevant function set is used to determine the output layer of the convolutional neural network.

6. The improved method for determining the membership function of fuzzy control theory using a convolutional neural network according to claim 1, characterized in that, An optimization function is established between the power result obtained by dynamic programming and the power result calculated by combining convolutional neural network training with fuzzy logic control. The optimization function is shown below. If the final calculation result of the optimization function does not exceed the cumulative error index, it means that the calculation result of the convolutional neural network training combined with the fuzzy logic control strategy can approximately replace the optimal real-time decision result of the dynamic programming method for the ship's operating conditions.

Images