CN107728478B - Fuel cell oxygen excess coefficient neural network prediction control method - Google Patents
Fuel cell oxygen excess coefficient neural network prediction control method Download PDFInfo
- Publication number
- CN107728478B CN107728478B CN201710889277.8A CN201710889277A CN107728478B CN 107728478 B CN107728478 B CN 107728478B CN 201710889277 A CN201710889277 A CN 201710889277A CN 107728478 B CN107728478 B CN 107728478B
- Authority
- CN
- China
- Prior art keywords
- neural network
- fuel cell
- oxygen excess
- training sample
- learning
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
- G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
- G05B13/04—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators
- G05B13/042—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators in which a parameter or coefficient is automatically adjusted to optimise the performance
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Automation & Control Theory (AREA)
- Fuel Cell (AREA)
Abstract
A neural network prediction control method for an oxygen excess coefficient of a fuel cell belongs to the technical field of control. Aiming at the problem of oxygen excess coefficient control of the automobile fuel cell, the invention designs the controller by utilizing a neural network predictive control algorithm, so that the fuel cell system can obtain sufficient oxygen and ensure the optimal power. The method comprises the following steps: software selection, training sample design, neural network prediction model off-line learning, neural network prediction model learning and neural network prediction controller design. The invention extracts the system characteristics by directly learning the input and output data, effectively avoids complex mechanism modeling, can reduce the loss of the system characteristics in the online learning process, further improves the accuracy of multi-step prediction, and effectively processes the control problem of the nonlinear system with constraints.
Description
Technical Field
The invention belongs to the technical field of control.
Background
With the increasing environmental pollution and energy crisis, fuel cell vehicles are considered to be the ultimate form of vehicles because of their advantages of high energy conversion rate, zero emission, wide fuel (hydrogen) sources, and convenient fuel replenishment. The country also provides a great deal of preferential policies to support enterprises to research fuel cell automobiles, the fuel cell automobiles are listed in a three-longitudinal-three-transverse development framework of electric automobiles, and the breakthrough is integrally achieved through accumulation for more than ten years; as clearly proposed by "chinese manufacture 2025", about one thousand fuel cell vehicles were produced and operated in a demonstration manner by 2020. Both conventional and fuel cell vehicles involve air supply system control problems. Compared with the traditional automobile, the influence of the oxygen excess coefficient on the fuel cell automobile is more serious, and when the oxygen excess coefficient is too low, the service life of the battery can be shortened, and even the exchange membrane can be damaged. Therefore, how to design the controller to generate a reasonable voltage to drive the compressor and obtain sufficient air is one of the important issues in the research of the fuel cell of the automobile. The following problems are mainly involved in the oxygen excess coefficient control of the automobile fuel cell:
1. the fuel cell air supply system has a complex structure, mechanism modeling is difficult to perform, and even if a mechanism model is obtained, the model-based controller design is difficult to perform;
2. because the air transmission and the chemical reaction of the fuel cell require time, the inertia and the hysteresis characteristics of the system are serious, and a good effect is difficult to achieve by a common modeling mode;
3. the parameters of the fuel cell are time-varying due to problems such as changes in the environment, degradation of parts, and the like.
Disclosure of Invention
Aiming at the problem of oxygen excess coefficient control of the automobile fuel cell, the invention designs the controller by utilizing a neural network predictive control algorithm, so that the fuel cell system can obtain sufficient oxygen and ensure the optimal power. Since the fuel cell air supply system is complex, the equation of state is difficult to derive, and the parameters are time varying, it is difficult to design an efficient controller. The invention designs a neural network predictive control algorithm aiming at the system, can effectively process the problems of modeling and parameter time variation, and considers the constraint of the driving voltage of the compressor.
The method comprises the following steps:
firstly, software selection: the controlled object of the control system and the simulation model of the controller are built through software Matlab/Simulink, the software version is Matlab R20012a, the solver selects code 3 respectively, the simulation step length is fixed step length, and the step length is selected to be 0.005 s;
II, training sample design: selecting input and output variables for neural network training, wherein the quantities influencing the oxygen excess coefficient mainly comprise the temperature, the pressure, the humidity, the load current and the current state of the oxygen excess coefficient of the cathode of the fuel cell; the input of the neural network is the current values of atmospheric pressure, load current, compressor driving voltage and oxygen surplus coefficient; gaussian noise is used as input to excite a fuel cell system to obtain a training sample set, the training sample set is selected, similar samples are removed, and therefore the minimum number of training samples is kept while the system characteristics are comprehensively extracted;
thirdly, offline learning of a neural network prediction model:
1) the neural network offline learning process: the neural network prediction model for a fuel cell is represented as:
wherein k represents the time of day, and,
and the predicted value is the oxygen excess coefficient at the k +1 th moment.
And
for the weights of the input layer to the hidden layer and the hidden layer to the output layer,
and
for the thresholds of the hidden and output layers, σ (x) is a logarithmic function, expressed as:
where e is a natural constant, and u (k) is the input to the neural network at time k, and is expressed as:
I(k)、Vc(k)、Patm(k) and
load current, compressor driving voltage, atmospheric pressure and actual values of oxygen excess coefficient at the kth moment are respectively;
fourthly, learning a neural network prediction model:
1) selecting a neural network online learning training sample:
the initial training sample set of online learning is the same as the training sample set of offline learning, and the initial weight and the initial threshold are obtained by offline learning; comparing the obtained new samples with samples in the old training sample set one by one, if the new samples are similar to the samples in the old training sample set, replacing the samples, and otherwise, directly adding the samples into the training sample set;
2) the neural network on-line learning and training process comprises the following steps:
the weight from the hidden layer to the output layer and the threshold value of the output layer are adjusted on line, and the weight from the hidden layer to the output layer and the threshold value of the output layer are adjusted by adopting an LM (Levenberg-Marquardt) algorithm;
designing a neural network predictive controller: iterating the neural network prediction model to obtain a prediction time domain NpInner prediction value, prediction time domain NpThe predicted values within are expressed as:
wherein
And
are respectively k +1, k +2 and k + NpPredicting a predicted value of the oxygen surplus coefficient at the moment;
wherein u (k), u (k +1) and u (k + N)p-1) are respectively the k, k +1, k + Np-neural network input at time 1;
wherein I (k +1), Vc(k+i)、Patm(k + i) and
load current, compressor drive voltage, atmospheric pressure and oxygen excess coefficient prediction values at the k + I-th time, respectively, due to I and PatmThe interference amount can be measured, and belongs to slow-changing parameters, so that:
I(k+i)＝I(k)，i＝1，2，…
Patm(k+i)＝Patm(k) (7)
if the control time domain is the same as the prediction time domain, the objective function J is represented as:
wherein min (-) represents the minimum value sought,
reference sequence for oxygen excess coefficient, in general
ΔU＝[ΔVc(k) ΔVc(k+1) … ΔVc(k+Np-1)]TIn incremental form of compressor drive voltage, Δ Vc(k+i)＝Vc(k+i)-Vc(k + i-1), i ═ 0, 1, 2, where V isc(k + i-1) is the compressor drive voltage at time k + i-1,uandywhich are weighting matrices for the input and output terms, respectively, the constraint of the compressor drive voltage is related to the load current, expressed as:
fmin(I)≤Vc≤fmax(I) (9)
wherein f ismin(I) And fmax(I) Respectively, a minimum and a maximum constraint function for the compressor drive voltage.
The invention mainly aims at the problem of oxygen excess coefficient control of the automobile fuel cell, and designs the controller by utilizing a neural network-based prediction control algorithm. The neural network-based predictive control algorithm is a perfect combination of the neural network and model predictive control, and a controller is designed through input and output data and characteristics obtained through data processing, so that complex mechanical modeling is effectively avoided. A prediction equation of the system can be obtained through the collected input and output data, then a cost function is constructed by using a model prediction control method, constraint conditions are considered, and optimized compressor driving voltage signals are obtained through optimization solution. The method can well solve the three problems, can effectively avoid complex system mechanism modeling based on a neural network predictive control algorithm, and simultaneously considers the constraint of the driving voltage of the compressor.
Compared with the prior art, the invention has the beneficial effects that:
1. most of traditional control algorithms are based on mechanism models, but in the actual fuel cell oxygen excess coefficient control process, the mechanism modeling is complex and has large workload, and in addition, the mechanism modeling also has difficulty in considering parameter time variation problems. The neural network learning algorithm extracts system characteristics directly by learning input and output data, and effectively avoids complex mechanism modeling. Secondly, the online learning method of the neural network can effectively process the parameter time-varying problem.
2. In the online learning process of the neural network, a training sample set which almost covers all working conditions is used as an initial training sample set, then a new sample obtained each time is compared with a sample in an old training sample set, if the new sample is similar to a certain old sample, the old sample is replaced, and otherwise, the new sample is added into the training sample set. The operation can not only solve the influence of time variation of parameters and reduce the calculation amount (reduced number of samples for online learning), but also reduce the loss of system characteristics in the online learning process so as to improve the accuracy of multi-step prediction.
3. The fuel cell oxygen excess coefficient controller system designed in the invention is a nonlinear system, and in consideration of the hard constraint of the compressor driving voltage, the traditional control algorithm cannot effectively process the constraint of the system, and the neural network predictive control algorithm can effectively process the control problem with the constraint of the nonlinear system.
Drawings
These and/or other aspects of the present invention will become apparent from the following further description of embodiments of the invention, as illustrated in the accompanying drawings. Wherein:
FIG. 1 is a block diagram of an oxygen surplus coefficient control block for an automotive fuel cell system implementing neural network predictive control in accordance with the present invention;
FIG. 2 is a verification of an off-line learning neural network prediction model for predicting a 1-step value, in which (a) is a comparison graph of an actual value of an oxygen surplus coefficient and a predicted value of a neural network, wherein a solid line is an actual oxygen surplus coefficient value, and a dotted line is a predicted oxygen surplus coefficient value of the neural network model; the graph (b) is an absolute error curve of the actual value and the predicted value of the neural network;
FIG. 3 is a verification of an off-line learning neural network prediction model for predicting 20-step values, in which (a) is a comparison graph of an actual value of an oxygen surplus coefficient and a neural network prediction value, in which a solid line is an actual oxygen surplus coefficient value and a dotted line is a neural network model prediction oxygen surplus coefficient value; the graph (b) is an absolute error curve of the actual value and the predicted value of the neural network;
FIG. 4 is a graph of load current change in units A when verifying controller effectiveness;
FIG. 5 is a graph of the change in oxygen excess coefficient under the control of the present invention, where the dashed line is the curve of the actual system output oxygen excess coefficient and the solid line is the desired curve of the oxygen excess coefficient.
Detailed Description
The research method is model prediction control based on a neural network, and comprises the following steps:
firstly, selecting proper input and output variables according to the internal structure of the system; secondly, designing proper training samples and test samples according to the dynamic characteristics of the system, selecting the training samples, and removing similar samples to ensure that the number of the training samples is minimum while comprehensively extracting the system characteristics; performing off-line training on the neural network by using the obtained training samples again to obtain an initial weight and a threshold of the neural network model; then, updating the weight value through online training, taking the training sample as an initial training sample of the online training, replacing the new sample when the obtained new sample is similar to the old sample data, and adding the new sample into a training set if the obtained new sample is not similar to the old sample data; and finally, considering the constraint of the driving voltage of the compressor, constructing a cost function by using a model predictive control algorithm, and solving an optimal problem corresponding to the cost function to obtain a control input to act on the system, thereby realizing the control of the system.
In order to solve the technical problems, the invention is realized by adopting the following technical scheme:
the fuel cell oxygen excess coefficient control based on the neural network predictive control is realized by a software system. The software system is composed of Matlab/Simulink high-level simulation software.
Functionally, the present invention may include the following: a fuel cell system model, a fuel cell neural network prediction model, and a model predictive controller module. The function of each part is explained in detail as follows:
the fuel cell system model is mainly used for simulating a real controlled object, namely, the function of an air supply system of a real fuel cell can be accurately described, and training samples and test samples which can reflect the dynamic characteristics of the system can be provided for the learning of a fuel cell neural network prediction model;
the fuel cell neural network prediction model is mainly used for obtaining a prediction model which accords with the dynamic characteristics of a system through a training sample and providing a prediction model which accords with requirements for the design of a model prediction controller;
the model prediction controller module is mainly used for collecting various state information of the fuel cell air supply system model, then carrying out optimization operation, generating a control signal, namely a compressor driving voltage signal, and sending the signal to an actuating mechanism, namely a compressor, of the fuel cell air supply system.
The invention particularly relates to an oxygen surplus coefficient control method in the field of automobile proton exchange membrane fuel cells, and more particularly relates to a control method for learning structural characteristics of an air supply system of an automobile proton exchange membrane fuel cell by adopting a neural network so as to realize accurate control of an oxygen surplus coefficient based on model prediction control.
The present invention will be fully explained with reference to the accompanying drawings for illustrating technical contents, construction features, and objects of the invention in detail.
The control block diagram of the implementation of the oxygen excess coefficient control of the automobile fuel cell based on the neural network prediction control in the invention is shown in figure 1, wherein an NN prediction model (neural network prediction model) and an M are shown in the figureThe PC controller (model predictive controller) is written in the. The input to the controller is the desired oxygen surplus factor
Measurable disturbance variable load current I and atmospheric pressure PatmAnd predicting time domain N by neural network prediction modelpInner prediction value
The output of the controller is a compressor drive voltage Vc. The fuel cell system is built in Simulink, and the actual oxygen excess coefficient output by the fuel cell system
The measured interference amount and the compressor driving voltage can be used for carrying out online learning on the NN prediction model.
The control object of the present invention is that the controller controls the compressor driving voltage so that the oxygen excess coefficient of the fuel cell system is maintained at 2 and not lower than 1 during the disturbance change, according to the state of the fuel cell system.
The invention provides a set of devices based on the operation principle and the operation process. The construction and operation processes are as follows: 1. software selection
A controlled object of the control system and a simulation model of the controller are built through software Matlab/Simulink, the software version is Matlab R20012a, and the solver is selected to be ode 3. The simulation step size is a fixed step size, and the step size is selected to be 0.005 s.
2. Training sample design
Firstly, selecting input and output variables for neural network training. The output variable of the neural network is a predicted value of the controlled quantity, namely a predicted value of the oxygen surplus coefficient. The amount of influence on the oxygen excess coefficient is mainly the temperature, pressure, humidity, load current of the fuel cell cathode, and the current state of the oxygen excess coefficient, wherein the humidity and temperature are generally controlled to fixed values by a specific controller and thus may not be considered; the pressure at the cathode is determined primarily by the atmospheric pressure and the compressor drive voltage. Thus, the inputs to the neural network are the barometric pressure, the load current, the compressor drive voltage, and the current values of the oxygen surplus factor. And secondly, exciting the fuel cell system by adopting Gaussian noise as input to obtain a training sample set, selecting the training sample set, and removing similar samples to ensure that the number of training samples is kept minimum while comprehensively extracting system characteristics, wherein the number of the obtained training sample sets is 755 sets. And then, taking another group of Gaussian noises as an input excitation system to obtain a test sample set. And finally, carrying out normalization processing on the training sample set and the test sample set.
3. Neural network prediction model offline learning
1) Neural network offline learning process
And (3) performing offline learning on the model by adopting a MATLAB (matrix laboratory) self-contained BP neural network tool box. The neural network is designed into a three-layer network, the number of neurons of an input layer and an output layer is the same as the number of input and output of the neural network, the number of neurons is respectively 4 and 1, and the number of neurons of a hidden layer is set to be 10. The input layer and output layer transfer functions are linear functions and the hidden layer transfer function is a logarithmic function. The neural network predictive model of the fuel cell can be expressed as:
wherein k represents the time of day, and,
and the predicted value is the oxygen excess coefficient at the k +1 th moment.
And
for the weights of the input layer to the hidden layer and the hidden layer to the output layer,
and
for the thresholds of the hidden and output layers, σ (x) is a logarithmic function, which can be expressed as:
where e is a natural constant. u (k) is the input to the neural network at time k, and can be expressed as:
I(k)、Vc(k)、Patm(k) and
the load current, the compressor driving voltage, the atmospheric pressure, and the actual value of the oxygen excess coefficient at the k-th time are respectively.
After training, the obtained initial weight and initial threshold are as follows:
β2＝[-7.89]。
2) offline learning neural network model validation
And using the obtained test sample set for verifying the accuracy of the model. The offline learned neural network model is validated from two sets of experiments below.
FIG. 2 is the result of the model test of step 1 of prediction. Fig. 2(a) is a comparison between the actual value of the oxygen excess coefficient and the neural network prediction model prediction value (NN prediction value), and fig. 2(b) is an absolute error curve of these two values, and it can be seen from the graph that the prediction accuracy of the model is very high. Fig. 3 shows the result of the model verification for the prediction of 20 steps, and compared with the result of the prediction of 1 step, the model accuracy for 20 steps is reduced, but the dynamic characteristics of the fuel cell system can be fully embodied, so the model meets the design requirements of the controller.
4. Neural network prediction model learning
1) Neural network online learning training sample selection
The initial training sample set of the online learning is the same as the training sample set of the offline learning, and the initial weight and the initial threshold are obtained by the offline learning. In the online learning process, a newly obtained sample exists at each moment, dimension explosion can be caused if an old sample is not removed, and the calculation time is longer and longer; if the old sample is removed according to a common old sample removing mode, namely the first obtained sample is removed, the model characteristics are lost in the old sample removing process, and particularly for multi-step prediction, the prediction accuracy is greatly degraded. The following strategies are adopted in the invention: and comparing the obtained new samples with samples in the old training sample set one by one, if the new samples are similar to the samples in the old training sample set, replacing the samples, and otherwise, directly adding the samples into the training sample set. This approach also enables changes in system parameters to be identified while ensuring that the fuel cell system features are extracted as comprehensively as possible.
2) Neural network on-line learning and training process
For the BP neural network, the biggest disadvantage of implementing online learning is that training is slow and easily falls into local values. The main reason for this problem is that the hidden layer transfer function is non-linear, so the adjustment of the input layer to the hidden layer weights and hidden layer thresholds is more complicated. After off-line training, the neural network prediction model already extracts most characteristics of the system, and the influence of a parameter time-varying problem on the fuel cell system is limited, so in order to reduce on-line calculation time and accuracy, the method only adjusts the weight from the hidden layer to the output layer and the threshold value of the output layer on line. The adjustment of the weights of the hidden layer to the output layer and the threshold of the output layer uses the LM (Levenberg-Marquardt) algorithm.
5. Neural network predictive controller design
The controller in this patent aims to make the oxygen excess coefficient of the automotive fuel cell at 2 and not less than 1 in disturbance change while the compressor driving voltage satisfies the constraint conditions (avoiding compressor damage) at the constraint conditions when designing the neural network prediction controllerThe middle-jiao is embodied. The neural network predictive control algorithm adopted by the invention is formed by combining a neural network learning model and model predictive control. Compared with other control methods, the method has the main difference that the neural network realizes the prediction function, and directly iterates the neural network prediction model to obtain the prediction time domain NpThe predicted value of the interior. Predicting time domain NpThe predicted values within can be expressed as:
wherein
And
are respectively k +1, k +2 and k + NpAnd predicting the oxygen excess coefficient at the moment.
Wherein u (k), u (k +1) and u (k + N)p-1) are respectively the k, k +1, k + Np-neural network input at time 1.
Wherein I (k + I), Vc(k+i)、Patm(k + i) and
load current, compressor driving voltage, atmospheric pressure and oxygen excess coefficient predicted values at the k + i-th moment are respectively. Due to I and PatmThe interference amount can be measured, and belongs to slow-changing parameters, so that:
the control time domain is the same as the prediction time domain, and the objective function J can be expressed as:
wherein min (-) represents the minimum value sought,
reference sequence for oxygen excess coefficient, in general
ΔU＝[ΔVc(k) ΔVc(k+1) … ΔVc(k+Np-1)]TIn incremental form of compressor drive voltage, Δ Vc(k+i)＝Vc(k+i)-Vc(k + i-1), i ═ 0, 1, 2, …, where V isc(k + i-1) is the compressor drive voltage at time k + i-1.uAndywhich are the weighting matrices for the input and output items, respectively. The constraint on the compressor drive voltage is related to the load current and can be expressed as:
fmin(I)≤Vc≤fmax(I)， (9)
wherein f ismin(I) And fmax(I) Minimum and maximum constraint functions for the compressor drive voltage, respectively, are obtained by calibration. Predicting time domain N through massive debuggingpThe sum control time domain is set to 10, and the weighting matrix of the input term and the output term is set tou＝diag(1，1，1，1，1，1，1，1，1，1)，yBiag (5, 5, 5, 5, 5, 5, 5, 5, 5, 5). diag (·) represents a diagonal matrix with () intermediate numbers as diagonal numbers.
6. Experimental verification
In order to verify the control performance of the neural network predictive controller of the oxygen excess coefficient of the automobile fuel cell, the invention designs a group of experiments. Set atmospheric pressure PatmThe change of the load current I is a continuous step change as shown in fig. 4 at 1 atm. The change in the oxygen excess coefficient under this condition is shown in fig. 5. As can be seen from FIG. 5, when the load current suddenly changesIn time, the oxygen excess coefficient of the system output deviates from 2, but is larger than 1.5, and the oxygen excess coefficient is rapidly recovered to 2.
Therefore, the neural network prediction controller designed for the automobile fuel cell can well solve the problem of oxygen excess coefficient control.
Claims (1)
1. A fuel cell oxygen excess coefficient neural network prediction control method is characterized by comprising the following steps: the method comprises the following steps:
firstly, software selection: the controlled object of the control system and the simulation model of the controller are built through software Matlab/Simulink, the software version is Matlab R2012a, solvers are selected to be respectively ode3, the simulation step length is fixed step length, and the step length is selected to be 0.005 s;
II, training sample design: selecting input and output variables for neural network training, wherein the quantities influencing the oxygen excess coefficient mainly comprise the temperature, the pressure, the humidity, the load current and the current state of the oxygen excess coefficient of the cathode of the fuel cell; the input of the neural network is the current values of atmospheric pressure, load current, compressor driving voltage and oxygen surplus coefficient; adopting Gaussian noise as input to excite a fuel cell system to obtain a training sample set, selecting the training sample set, and removing similar samples to ensure that the number of training samples is kept minimum while comprehensively extracting system characteristics;
thirdly, offline learning of a neural network prediction model:
1) the neural network offline learning process: the neural network prediction model for a fuel cell is represented as:
wherein k represents the time of day, and,
the predicted value of the oxygen excess coefficient at the k +1 moment is obtained;
and
for the weights of the input layer to the hidden layer and the hidden layer to the output layer,
and
for the thresholds of the hidden and output layers, σ (x) is a logarithmic function, expressed as:
where e is a natural constant, and u (k) is the input to the neural network at time k, and is expressed as:
I(k)、Vc(k)、Patm(k) and
load current, compressor driving voltage, atmospheric pressure and actual values of oxygen excess coefficient at the kth moment are respectively;
fourthly, learning a neural network prediction model:
1) selecting a neural network online learning training sample:
the initial training sample set of online learning is the same as the training sample set of offline learning, and the initial weight and the initial threshold are obtained by offline learning; comparing the obtained new samples with samples in the old training sample set one by one, if the new samples are similar to the samples in the old training sample set, replacing the samples, and otherwise, directly adding the samples into the training sample set;
2) the neural network on-line learning and training process comprises the following steps:
the weight from the hidden layer to the output layer and the threshold value of the output layer are adjusted on line, and the weight from the hidden layer to the output layer and the threshold value of the output layer are adjusted by adopting an LM (Levenberg-Marquardt) algorithm;
designing a neural network predictive controller: iterating the neural network prediction model to obtain a prediction time domain NpInner prediction value, prediction time domain NpThe predicted values within are expressed as:
wherein
And
are respectively k +1, k +2 and k + NpPredicting a predicted value of the oxygen surplus coefficient at the moment;
wherein u (k), u (k +1) and u (k + N)p-1) are respectively the k, k +1, k + Np-neural network input at time 1;
wherein I (k + I), Vc(k+i)、Patm(k + i) and
load current, compressor driving voltage, atmospheric pressure and oxygen excess coefficient prediction values at the k + I-th time, respectively, due to I and PatmThe interference is measurable and belongs to slow-changing parameters, so that:
I(k+i)＝I(k)，i＝1，2，…
Patm(k+i)＝Patm(k) (7)
if the control time domain is the same as the prediction time domain, the objective function J is represented as:
wherein min (-) represents the minimum value sought,
is a reference sequence of oxygen excess coefficients,
ΔU＝[ΔVc(k) ΔVc(k+1) … ΔVc(k+Np-1)]Tin incremental form of compressor drive voltage, Δ Vc(k+i)＝Vc(k+i)-Vc(k + i-1), i ═ 0, 1, 2, …, where V isc(k + i-1) is the compressor drive voltage at time k + i-1,uandywhich are weighting matrices for the input and output terms, respectively, the constraint of the compressor drive voltage in relation to the load current is expressed as:
fmin(I)≤Vc≤fmax(I) (9)
wherein f ismin(I) And fmax(I) Respectively, a minimum and a maximum constraint function for the compressor drive voltage.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710889277.8A CN107728478B (en) | 2017-09-27 | 2017-09-27 | Fuel cell oxygen excess coefficient neural network prediction control method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710889277.8A CN107728478B (en) | 2017-09-27 | 2017-09-27 | Fuel cell oxygen excess coefficient neural network prediction control method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107728478A CN107728478A (en) | 2018-02-23 |
| CN107728478B true CN107728478B (en) | 2020-10-27 |
Family
ID=61207449
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710889277.8A Active CN107728478B (en) | 2017-09-27 | 2017-09-27 | Fuel cell oxygen excess coefficient neural network prediction control method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107728478B (en) |
Families Citing this family (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108681244B (en) * | 2018-05-11 | 2021-01-26 | 东南大学 | Fuel cell anode pressure dynamic matrix control method based on multi-model feedforward |
| CN109597362B (en) * | 2018-08-24 | 2021-07-09 | 吉林大学 | Design method of nonlinear controller of fuel cell air supply system |
| CN110335646B (en) * | 2019-07-12 | 2022-06-17 | 福州大学 | Vehicle fuel cell hydrogen peroxide ratio control method based on deep learning-prediction control |
| CN110909463B (en) * | 2019-11-14 | 2022-06-03 | 电子科技大学 | Active control and protection method and system for high-power millimeter wave gyrotron traveling wave tube |
| CN112597430B (en) * | 2020-12-22 | 2024-04-05 | 浙江中控技术股份有限公司 | Operation parameter optimization method for complex rectifying tower |
| CN113222387B (en) * | 2021-04-30 | 2023-11-24 | 北京理工新源信息科技有限公司 | Multi-objective scheduling and collaborative optimization method for hydrogen fuel vehicle |
| CN113644301B (en) * | 2021-06-03 | 2022-07-01 | 吉林大学 | Vehicle-mounted fuel cell cathode flow and pressure control method |
| CN113594513B (en) * | 2021-08-02 | 2023-05-09 | 重庆交通职业学院 | Proton exchange membrane fuel cell model optimization control method |
| CN115762653B (en) * | 2022-11-11 | 2023-08-11 | 哈尔滨工程大学 | Fuel combustion mechanism optimization method based on evolutionary algorithm and deep learning |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| NZ512063A (en) * | 1999-09-24 | 2002-12-20 | Kevin H Slade | Method for signal and material processing using self-organizing field effects |
| CN102663219B (en) * | 2011-12-21 | 2015-03-04 | 北京理工大学 | Fuel cell output prediction method and system based on mixing model |
| CN104951836A (en) * | 2014-03-25 | 2015-09-30 | 上海市玻森数据科技有限公司 | Posting predication system based on nerual network technique |
| US9678664B2 (en) * | 2015-04-10 | 2017-06-13 | Google Inc. | Neural network for keyboard input decoding |
-
2017
- 2017-09-27 CN CN201710889277.8A patent/CN107728478B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN107728478A (en) | 2018-02-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107728478B (en) | Fuel cell oxygen excess coefficient neural network prediction control method | |
| CN112733462B (en) | Ultra-short-term wind power plant power prediction method combining meteorological factors | |
| CN111339712A (en) | Method for predicting residual life of proton exchange membrane fuel cell | |
| CN109242212A (en) | A kind of wind-powered electricity generation prediction technique based on change Mode Decomposition and length memory network | |
| CN112949894B (en) | Output water BOD prediction method based on simplified long-short-term memory neural network | |
| CN112287605B (en) | Power flow checking method based on graph convolution network acceleration | |
| CN111832839B (en) | Energy consumption prediction method based on sufficient incremental learning | |
| CN111723982A (en) | Medium-and-long-term power load combined prediction method based on gray-Markov chain | |
| CN115718263B (en) | Attention-based lithium ion battery calendar aging prediction model and method | |
| CN115456287A (en) | Long-and-short-term memory network-based multi-element load prediction method for comprehensive energy system | |
| CN115759415A (en) | Power consumption demand prediction method based on LSTM-SVR | |
| Hongwei et al. | An interpretable data-driven method for degradation prediction of proton exchange membrane fuel cells based on temporal fusion transformer and covariates | |
| CN110880044A (en) | Markov chain-based load prediction method | |
| CN110956304A (en) | Distributed photovoltaic power generation capacity short-term prediction method based on GA-RBM | |
| Xu et al. | Short-term electricity consumption forecasting method for residential users based on cluster classification and backpropagation neural network | |
| CN117634737A (en) | Photovoltaic power generation prediction model construction method, photovoltaic energy supply regulation and control method and device | |
| CN117312835A (en) | Fuel cell performance prediction method based on mixed probability data driving model | |
| CN117077327A (en) | Bearing life prediction method and system based on digital twin | |
| CN114234392B (en) | Air conditioner load fine prediction method based on improved PSO-LSTM | |
| CN111061708A (en) | Electric energy prediction and restoration method based on LSTM neural network | |
| CN114384427B (en) | Quantum neural network-based lithium ion battery capacity nonlinear degradation prediction method | |
| CN115660195A (en) | Carbon emission prediction method and system for industrial park multi-energy complementary system | |
| CN116432812A (en) | New energy power prediction method for optimizing LSTM (least squares) by using Zun sea squirt algorithm | |
| CN115716469A (en) | Output power distribution control method of hybrid power system | |
| CN113610665B (en) | Wind power generation power prediction method based on multi-delay output echo state network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |