CN-122022163-A - Multi-dimensional mode identification energy management method for hybrid propulsion system

CN122022163ACN 122022163 ACN122022163 ACN 122022163ACN-122022163-A

Abstract

The invention discloses a multi-dimensional mode identification energy management method of a hybrid propulsion system, which comprises the steps of constructing a flight-control-load multi-dimensional time sequence characteristic data space, identifying and softly classifying flight working conditions through multi-dimensional characteristics, making multi-mode fuzzy self-adaptive decisions based on working condition probability weights, conducting fuzzy system parameter off-line optimization based on a global optimal algorithm, extracting a DP optimal track under a full flight envelope, executing fuzzy system parameter self-adaptive calibration based on a dung beetle optimization algorithm, and obtaining a global reference track optimization vector capable of guiding power distribution in real time. According to the invention, a multidimensional time sequence characteristic data space is constructed, the real-time prediction working condition probability is identified by utilizing a depth network, and the global optimal parameter offline calibration is carried out by combining dynamic programming and a dung beetle optimizing algorithm, so that the power output instruction of the system can guide the power distribution in real time and approach to the global optimal track, and the self-adaptability and the fuel economy of the system under complex and changeable working conditions are improved.

Inventors

LIANG WEIHE
You Ruixian
WANG CHUNYAN
ZHAO WANZHONG
LUAN ZHONGKAI
ZHOU XIAOCHUAN
ZHANG ZIYU

Assignees

南京航空航天大学

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (5)

1. A method of multi-dimensional pattern recognition energy management for a hybrid propulsion system, comprising the steps of: step 1), constructing a flight-control-load multidimensional time sequence characteristic data space; the method comprises the steps of collecting multisource physical signals of flight pneumatics, driving control and a power system, extracting a change rate of a control lever reflecting driving maneuver intention and a power fluctuation characteristic reflecting environmental disturbance, packaging a multidimensional time sequence characteristic input vector containing historical time, and obtaining a power fluctuation characteristic of the control lever; Step 2), identifying and soft classifying the flight conditions through multidimensional features; Performing offline fuzzy C-means clustering to generate typical working condition semantic labels and membership samples, constructing an LSTM depth time sequence recognition network with a gating mechanism, and outputting a real-time predictive probability vector; Step 3), making a multi-mode fuzzy self-adaptive decision based on the working condition probability weight; establishing a sub-working condition Sugeno type fuzzy controller, formulating a multi-mode expert control rule base, and synthesizing a power output instruction based on probability weight; Step 4), performing off-line optimization on the parameters of the fuzzy system based on the global optimal algorithm; The method comprises the steps of obtaining a DP optimal track under a full flight envelope, executing a fuzzy system parameter self-adaptive calibration based on a dung beetle optimization algorithm, and obtaining a global reference track optimization vector capable of guiding power distribution in real time.
2. The hybrid propulsion system multidimensional pattern recognition energy management method of claim 1, wherein step 1) specifically comprises: 11 Collecting multisource physical signals of flight pneumatics, driving control and a power system; The original physical data of the flight-control-load is collected in real time through an airborne atmospheric data computer, a flight control system and a hybrid power controller, and the flight-control-load original physical data comprises a flight aerodynamic signal, namely a vacuum speed Vertical lifting rate Flying height Driving control signal, driver's joystick opening Power system status signal bus total power demand Rotational speed of gas generator of turboshaft engine And battery state of charge Performing time stamp calibration and synchronous alignment processing on the acquired multi-source data; 12 Extracting a change rate of a joystick reflecting driving maneuver intention and a power fluctuation feature reflecting environmental disturbance; Based on preset sliding time window length Carrying out sliding window interception on the data sequence synchronized in the step 11), and at each sampling moment Extracting dynamic characteristics of the data in the window; The dynamic feature extraction comprises extracting the motor intention feature of the driver, namely extracting the collected opening degree of the control lever of the driver The change rate of the control lever is obtained by operation : In which, in the process, For the opening of the joystick at the current moment, As the opening degree at the last sampling time, Extracting the total power requirement of the collected bus for the sampling period and the environmental disturbance characteristic Calculating to obtain power fluctuation standard deviation : In which, in the process, The total number of sampling points in the sliding window is% ), Is the inside of the window The power values at the individual historic moments in time, An arithmetic mean of the power demand within the current window; 13 Packaging a multi-dimensional timing feature input vector containing a historical time; the dynamic characteristics obtained by calculation And Time-step alignment and combination with original physical parameters to construct a history Multi-dimensional time sequence characteristic input vector of each time step 。
3. The hybrid propulsion system multidimensional pattern recognition energy management method of claim 2, wherein step 2) specifically comprises: 21 Performing offline fuzzy C-means clustering to generate typical working condition semantic labels and membership samples, wherein the method specifically comprises the following steps: 211 A normalized sample dataset is constructed; Establishing an off-line sample database containing full flight envelope, and extracting each time from the database According to the method described in step 12), and constructing a feature sample vector which is completely consistent with the output format of step 12) For sample vector Each component in the vector is normalized to obtain a normalized vector : In which, in the process, Is vector quantity The first of (3) A physical quantity; 212 Performing fuzzy C-means clustering iterations; Setting the clustering number of typical working conditions Fuzzy weighted index Constructing an objective function, minimizing all normalized samples To each cluster center Is a weighted distance of: , For the total number of samples in the off-line sample database, For the time-series index of the samples, , For the index of the cluster center, , Is the first The samples belong to Fuzzy membership degree of class working condition and meeting constraint conditions And is also provided with , Is the first Clustering center vectors of typical working conditions; iterative update membership And a cluster center Until convergence, finally obtaining a clustering center Is a vector with the input Vector of the same dimension ; The membership update formula is: , is the first The clustering center vector is used as a comparison standard, and the updating formula of the clustering center is as follows: ; 213 Calibrating the physical semantics of a typical working condition center; four cluster centers after analysis convergence The working condition calibration is carried out according to the physical meaning defined in the step 12) to form a standard library ; For the cluster center 1, calibrating to be in a ground idle mode and vacuum velocity component And power component A minimum vector; For the cluster center 2, calibrated as vertical take-off/landing/hover mode, vacuum velocity component Smaller but power component A maximum vector; for the cluster center 3, calibrated as steady-state cruise mode, vacuum velocity component Maximum, standard differential power of power fluctuation Very low vectors; for the cluster center 4, the power fluctuation standard deviation component is calibrated to be in a high dynamic/disturbance mode And a joystick rate of change component Component values are significantly higher than vectors at other centers; Matching and storing the membership vectors of each sample calculated by the fuzzy C-means algorithm with the corresponding time sequence feature vectors, and constructing a data-label mapping library; 22 The method for constructing the LSTM depth time sequence identification network with the gating mechanism specifically comprises the following steps: 221 Defining LSTM network topology and internal extraction logic; Constructing a depth time sequence recognition model composed of an input layer, an LSTM hidden layer containing a gating mechanism, a full connection layer and an output layer, wherein the model adopts a time sequence characteristic matrix output in the step 12) For input, the evolution trend of the physical parameters is subjected to recursive feature extraction through hidden layer internal forgetting, input and output gating logic, and then is subjected to spatial mapping and normalization processing, and a predictive probability vector of the current working condition is output through Softmax mapping Soft classification discrimination of the flight conditions is realized, and the soft classification discrimination is used for online simulation of the fuzzy membership distribution rule calculated offline in step 212); The input layer receives and aligns the feature sequence constructed in step 12) : In which, in the process, For the 8 multi-dimensional feature components determined in step 12), The LSTM hidden layer containing the gating mechanism passes through the forgetting gate Automatically eliminating random fluctuation irrelevant to working condition judgment in the data of the step 12), wherein a calculation formula is as follows: In which, in the process, The function is activated for Sigmoid, To-be-trained weight matrix for forgetting gate, For the hidden state vector output by the last sampling instant LSTM cell, Is that The feature vectors that are input into the network at the moment, Bias vector to be trained for forgetting gate through input gate Identifying the feature most contributing to the identification at the current time and solidifying it to the cell state In the method, hidden state vectors containing long-term and short-term dynamics semantics are finally output In which, in the process, Is that The full connection layer receives the feature vector output by the LSTM layer at the last moment and maps the feature vector to the number of typical working conditions defined in the step 21) Equal output space: In which, in the process, As a matrix of weights, the weight matrix, The output layer adopts a Softmax activation function to output a score vector of the full connection layer Conversion to normalized probability distribution In which, in the process, For the current time belongs to Probability of class typical working conditions; 222 Training a long-term memory neural network; training LSTM network by using the off-line data-label mapping library constructed in step 21), and training the multi-dimensional time sequence feature matrix output in step 12) As a characteristic input to the network, the fuzzy membership vector at the corresponding moment generated by the fuzzy C-means algorithm in step 212) is used The nonlinear mapping logic from the time sequence physical signal to the working condition classification probability is learned and solidified by the network through minimizing the deviation between the prediction result and the true value; Constructing a loss function for measuring network identification accuracy In which, in the process, For the loss function value in the network training process, The number of sample batches selected for a single random gradient descent is determined by using a back-propagation algorithm over time based on a loss function Calculating weight matrix of each gating unit in network Offset vector Carrying out iterative correction on the parameters by using an optimizer until the loss function value is converged below a preset error threshold or reaches the maximum iterative times; 23 Outputting a real-time predictive probability vector; after training, solidifying and packaging the converged network weight parameters into an online working condition prediction module, so that the online working condition prediction module directly according to the real-time sequence characteristics input in the step 12) in a real-time operation stage Fast deducing corresponding working condition predictive probability vector 。
4. The hybrid propulsion system multidimensional pattern recognition energy management method of claim 3, wherein step 3) specifically comprises: 31 The method comprises setting input parameters of a fuzzy system, defining semantic set and membership function, setting output parameters of the fuzzy system, defuzzifying and outputting parameters, Setting the input variable of the fuzzy system as the total power demand of the bus And battery state of charge ; Setting the output variable of the fuzzy system as the first Output power of turboshaft engine under fuzzy rule table Suggested values; mapping the input physical quantity to fuzzy set formed by five grades, namely extremely low VL, low L, medium M, high H and extremely high VH, and selecting a Gaussian membership function And (3) performing feature conversion: In which, in the process, In order to input the amount of the input, And Respectively reflecting the function center mean value and standard deviation of the working condition characteristics; 32 A multi-mode expert control rule base is formulated; respectively establishing four sets of logic fuzzy rules aiming at four typical flight conditions calibrated in the step 213), and performing deblurring calculation on the reasoning result to obtain the accurate output power of the turboshaft engine ; 33 Synthesizing a power output instruction based on the probability weight; The system acquires the current sampling time in real time Calculating the final self-adaptive power instruction issued to the executing end of the turboshaft engine : In which, in the process, For the probability vector output by the LSTM identification network in step 222) The first of (3) A component representing the current time belonging to the first Probability confidence of class condition.
5. The hybrid propulsion system multidimensional pattern recognition energy management method of claim 4, wherein step 4) specifically comprises: 41 Extracting the DP optimal track under the full flight envelope; For the offline sample database containing the full flight mission profile in the step 21), performing reverse recursion optimization in the time domain by using a dynamic programming algorithm, searching a path with the lowest accumulated fuel consumption in a discretized state grid, and obtaining specific required power at each sampling time And battery state Is assigned a true value; 411 Setting a system state variable; Selecting a state of charge of a battery As a system state variable: In which, in the process, Is the current first The ratio of the remaining capacity of the battery in stages, And Is at present The open-circuit voltage and the internal resistance of the battery, For the rated capacity of the power battery pack, Net output power of the battery; 412 Setting a system control variable; selecting a recommended output power command for a turboshaft engine As a system control variable, the value range is strictly limited in physical constraint ; 413 Constructing a cumulative cost function; targeting total fuel consumption minimization within a full mission profile, constructing an objective function In which, in the process, For the instantaneous fuel consumption rate of a turboshaft engine at the current power output, A penalty factor is maintained for the amount of power, Is the reference battery state of charge; 414 Performing reverse recursion and forward backtracking of a dynamic programming algorithm; after performing reverse recursion and forward backtracking, the system output is calculated by And corresponding to the optimal control amount The formed offline optimal decision track is taken as a reference track to be used as a follow-up step 42) to guide a dung beetle optimization algorithm to learn the parameters of the fuzzy controller; 42 Executing self-adaptive calibration of parameters of a fuzzy system based on a dung beetle optimization algorithm; 421 A multidimensional decision vector to be optimized is designed; Design optimization variables Comprises fuzzy system input variables 、 Center point mean value of each fuzzy grade Gaussian membership function Sum width of Fuzzy system output variable Output constant term of (2) Confidence weight coefficient of each IF-THEN rule in fuzzy rule base Classifying the parameters according to the working conditions Concatenating into decision vectors As individual positions in the dung beetle population search space; 422 Building a global optimum guide-based comprehensive fitness function ; Comprehensive fitness function In which, in the process, For the DP optimal trajectory generated in step 41), For the total number of off-line training sample points corresponding to the working condition, For the safe out-of-limit number of times, The penalty weight coefficient; 423 Performing the iterative evolution of the dung beetle search behavior parameters; global exploration and local development mechanism of dung beetle optimization algorithm is utilized to optimize vectors in multidimensional variable space Performing rolling correction and iterative evolution to enable the output characteristic of the fuzzy controller to continuously approach the DP global optimal track, and outputting an optimal vector after the fitness function converges And solidifying and packaging the power distribution vector into an online working condition self-adaptive control library, so as to obtain a global reference track optimization vector capable of guiding the power distribution in real time.

Description

Multi-dimensional mode identification energy management method for hybrid propulsion system Technical Field The invention belongs to the technical field of energy management of hybrid propulsion systems, and particularly relates to a multi-dimensional mode identification energy management method of a hybrid propulsion system. Background With the rapid development of the hybrid propulsion system technology, a series hybrid propulsion system formed by a vortex shaft engine unit and a lithium power battery is adopted, so that the energy density and the power density can be effectively balanced, and the series hybrid propulsion system becomes a main flow power scheme with a long voyage. However, the hybrid propulsion system faces extremely complex working conditions in actual operation, and covers various modes such as ground idling, vertical take-off/landing/hovering, steady-state cruising, high dynamic maneuvering and the like, and in the switching process of the modes, the power requirement of the system has the characteristics of rapid transient change and large amplitude and is deeply influenced by the operation intention of a driver and the disturbance of the atmospheric environment. At present, although energy management strategies of hybrid propulsion systems have become research hotspots, the prior art still faces significant challenges when dealing with complex and varied flight conditions. The prior art has single recognition dimension on working conditions, and is difficult to accurately capture the coupling relation between driving intention and environmental disturbance. The conventional energy management method is mostly based on a fixed threshold value or a simple logic rule for mode switching, and the change rate of the opening degree of a control lever of a driver and the fluctuation characteristic of load power are not fully considered. Because the multidimensional time sequence association among the flight parameters, the operation behaviors and the external loads is ignored, when the system enters a high maneuver or complex meteorological working condition, the control logic cannot be adjusted in time, and the energy distribution is not matched, so that the flight safety is affected. For example, in the chinese patent application No. CN202410432818.4, entitled "a multi-mode hydrogen energy unmanned aerial vehicle energy management method based on online identification", a fuel cell stack model including unknown parameters and measurement noise is built, a system power model is built, real-time output characteristics of the stack are extracted, the flight mode is identified as cruising and non-cruising, online energy management strategies are designed for the two flight modes based on the results of parameter identification and feature extraction, although energy management strategies in different modes can be formulated for the unknown parameters and measurement noise, the identified and classified modes are not accurate enough, and influence of the opening of a driver joystick and the load power fluctuation characteristics on the energy management strategies is not considered. Disclosure of Invention Aiming at the defects of the prior art, the invention aims to provide a multi-dimensional mode identification energy management method of a hybrid propulsion system, which is characterized in that a multi-dimensional time sequence characteristic data space is constructed, the real-time prediction working condition probability is identified by utilizing a depth network, and the global optimal parameter offline calibration is carried out by combining dynamic programming and a dung beetle optimization algorithm, so that the problems that the working condition identification dimension is single, the coupling relation between driving intention and environment disturbance is difficult to capture and the mode classification is inaccurate in the existing energy management strategy are solved, the power output instruction of the system is ensured to guide the power distribution in real time and approach to the global optimal track, and the self-adaptability and the fuel economy of the system under complex and changeable working conditions are improved. In order to solve the technical problems, the invention provides the following technical scheme: a method of multi-dimensional pattern recognition energy management for a hybrid propulsion system, comprising the steps of: step 1), constructing a flight-control-load multidimensional time sequence characteristic data space; the method comprises the steps of collecting multisource physical signals of flight pneumatics, driving control and a power system, extracting a change rate of a control rod reflecting driving maneuver intention and power fluctuation characteristics reflecting environmental disturbance, and packaging a multidimensional time sequence characteristic input vector containing historical time. Step 2), identifying and soft classifying the fli