Speed stabilization control method and system for hybrid power range extender
By using an echo state network model and a discrete-time state-space prediction model, an optimal control signal sequence is generated, which solves the problems of torque pulsation and torsional vibration in the HEV power system during startup and load abrupt changes. It achieves active feedforward compensation for engine torque fluctuations, improves ride comfort, and reduces drivetrain impact.
Patent Information
- Application Number
- CN202511400039.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-28
- Publication Date
- 2025-11-18
AI Technical Summary
Hybrid electric vehicle (HEV) power systems experience torque pulsation and torsional vibration during startup and sudden load changes. Traditional vibration suppression methods are difficult to adapt to low-speed large-amplitude vibrations and changes in operating conditions, and model predictive control compensation accuracy decreases under time delay and disturbance conditions.
An echo state network model is used to learn the time-series dynamic characteristics of engine torque disturbance. By predicting the future disturbance sequence through multi-step look-ahead, and combining the discrete-time state-space prediction model and quadratic programming solution, the optimal control signal sequence is generated to achieve active feed-forward compensation for engine torque fluctuation.
It effectively suppresses the vibration of the HEV powertrain, improves ride comfort and reduces the impact load on the drivetrain, especially significantly reducing noise and vibration during startup.
Smart Images

Figure CN120963658A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of motor control technology, and in particular to a method and system for steady speed control of a hybrid range extender. Background Technology
[0002] The powertrain of a hybrid electric vehicle (HEV) consists of an internal combustion engine, two electric motors, and a planetary gear mechanism. During engine start-up and sudden load changes, the periodic combustion and complex transmission chain introduce significant torque pulsations and torsional vibrations, manifesting as multi-level fluctuations in output torque and speed. This reduces ride comfort and easily induces fatigue in transmission components. Traditional passive vibration suppression methods (such as spring-mass dampers and centrifugal pendulum vibration absorbers) rely on parameter tuning and have limited adaptability to low-speed, large-amplitude vibrations and changes in operating conditions. Active methods attempt to cancel out these vibrations by using the motor's output torque in the opposite phase; however, in situations with time delays in the actuator, very short sampling periods, and strong non-stationary disturbances, the controller struggles to robustly match phase and amplitude, often resulting in insufficient or over-compensation.
[0003] Model predictive control (MPC) is suitable for vehicle vibration suppression because it can explicitly handle multivariable constraints. However, MPC needs to complete finite-domain optimization within each control cycle. If the sampling period is in the millisecond range, there is a dead zone delay in the execution link, or future disturbances are unavailable, it is often difficult to provide a high-quality solution within the time limit. Especially when facing highly time-varying and random disturbances such as the start-up process, the accumulation of future disturbance errors leads to a decrease in compensation accuracy. While directly embedding complex recurrent neural networks into MPC for nonlinear modeling can improve the fitting, the training and online solution costs are high, and the interpretability and engineering feasibility are insufficient. Summary of the Invention
[0004] In view of the above technical problems, the present invention provides a method and system for steady speed control of a hybrid range extender to solve the problem that HEV power systems are prone to undercompensation or overcompensation during compensation.
[0005] Other features and advantages of this disclosure will become apparent from the following detailed description, or may be learned in part from practice of this disclosure.
[0006] According to one aspect of the present invention, a method for steady-speed control of a hybrid range extender is proposed, the method comprising: The system acquires disturbance signals during the operation of the HEV powertrain and trains an echo state network model to learn the time-series dynamic characteristics of the disturbance signals. The disturbance signals are the fluctuations in the crankshaft output torque of the internal combustion engine over time. The fluctuations are sampled to obtain a historical disturbance sequence composed of multiple past control cycles, with the sampling period being the discrete time step of the control cycle. The echo state network model adopts a three-layer structure of input layer, reservoir layer, and output layer. The input layer receives the historical disturbance sequence, and the reservoir layer consists of recurrent neural units with nonlinear activation functions and leakage integral characteristics to map the input to high-dimensional dynamic features. The output layer provides one-step or multi-step disturbance prediction values based on the reservoir state. During control operation, the current system state and current disturbance value of the HEV power system are acquired; the system state includes the relative angular displacement and relative angular velocity between the engine crankshaft and the planetary carrier, as well as the angular velocity of the second motor, and the disturbance value is the engine torque at the current moment; The current disturbance value and the historical disturbance sequence of a predetermined length are input into the trained echo state network model to predict the future disturbance sequence within multiple control cycles. During prediction, the echo state network model performs multi-step look-ahead prediction, covering the continuous moments required for both the control and prediction time domains, to obtain the future disturbance sequence arranged chronologically from the current moment. This multi-step look-ahead prediction establishes a time margin between control calculation and actual execution, and this time margin is not less than the sum of the computation time required for optimization and the lag time of the actuator. Based on the future disturbance sequence, a predictive control optimization model is constructed. During the construction, an optimization objective function is set with the goal of reducing the vibration of the HEV power system, and control input constraints and system output constraints are applied to the optimization model. The optimization model adopts a discrete-time state-space prediction model, which takes the system state described in the previous step as the initial value, combines the disturbance predictions at each future time, and gradually extrapolates along the prediction time domain to obtain the predicted system output at each future time. Within the current control cycle, the optimization model is solved to obtain the optimal control signal sequence covering the predetermined control prediction time domain. A real-time iterative quadratic programming solution strategy is adopted during the solution process. Under the given control time domain length and prediction time domain length configuration, the optimal motor torque command is calculated for multiple time steps forward from the current time. Based on the future disturbance sequence, the optimization result is fed forward to compensate for the engine torque fluctuations that will occur. The obtained control signal sequence satisfies all input and output constraints and achieves a trade-off between vibration suppression and control smoothness under the weight settings. The control signal corresponding to at least the next control cycle in the optimal control signal sequence is temporarily stored, and the control signal sequence is output to the actuator of the HEV power system when the next control cycle arrives, so as to implement the active vibration control of the power system in advance by taking advantage of the predicted future disturbances.
[0007] Furthermore, when training the echo state network model, the input weights and the weights within the reservoir are kept in a random sparse fixed connection, and only the weights of the output layer are trained, so as to minimize the error between the predicted perturbation and the actual perturbation on a given training set; wherein, the reservoir is a set of recursive nonlinear units with leakage integral characteristics, and its leakage coefficient is in the range of 0.2 to 0.5; the number of reservoir units is in the range of 200 to 1000; the input of the input layer is the historical perturbation sequence sampled according to the control cycle, and the output of the output layer is the point prediction value of the perturbation at the next time moment, and the future perturbation sequence with multiple steps of look-ahead is obtained through repeated iterations; The training data of the echo state network model covers the operating conditions under the engine starting scenario, including random variation samples of engine friction torque and starter speed. The weights of the output layer are solved by linear regression, and amplitude constraints or regularization terms are added to the regression objective to suppress overfitting.
[0008] Furthermore, the method also includes: establishing a probabilistic regression model for the time series of the perturbation signal by combining Gaussian process regression, which is used to adaptively correct and evaluate the reliability of the prediction results of the echo state network model; wherein: The time series of the disturbance signal is regarded as a random process that follows a joint Gaussian distribution, and a regression model is obtained by estimating the mean and covariance. The covariance is composed of the radial basis kernel of the time difference and the observation noise. The hyperparameter of the kernel function is obtained by maximizing the log-likelihood and using gradient descent or equivalent numerical methods. During the online control phase, the regression model is updated using the disturbance signal within the most recent time window, and the predicted mean is output as a correction for the deviation of the prediction of the echo state network model. The predicted variance is output as a measure of disturbance uncertainty for adaptive adjustment of control weights or constraint redundancy.
[0009] Furthermore, the HEV power system is composed of an internal combustion engine, a first motor, a second motor, and a planetary gear mechanism coupled together. Based on dynamic modeling, the stiffness and damping of the torsional damper are used as parameters, and a linear state-space representation is adopted with relative angular displacement and relative angular velocity as descriptors.
[0010] Furthermore, the dynamic modeling and output definition of the HEV powertrain system satisfies the speed-torque coupling relationship of the planetary gear mechanism and explicitly considers the torsional damper; wherein: The speed of the second motor is given by a fixed linear combination of the speed of the carrier and the speed of the first motor according to the planetary transmission ratio, and the speed of the carrier and the pinion is given by a linear combination determined by the planetary transmission ratio. The ring gear torque is equal to the sum of the torque of the first motor converted by the planetary transmission ratio and the torque of the second motor. The frame torque is determined according to the proportional relationship between the planetary transmission ratio and the ring gear torque. The elastic potential energy of the torsional damper is proportional to the square of the relative angular displacement between the engine crankshaft and the carrier, the dissipated energy is proportional to the square of the relative angular velocity between the two, and the kinetic energy is the weighted sum of the moment of inertia and the square of the angular velocity of each rotating component. The engine torque is considered as an external disturbance acting on the HEV power system, and the motor torques of the first motor and the second motor are considered as constrained control inputs. Predictive control is calculated in the rolling time domain to achieve active cancellation of engine torque fluctuations.
[0011] Furthermore, establishing the time margin includes: The number of prediction steps covers both the control time domain and the prediction time domain, and the number of steps in the control time domain is no greater than the number of steps in the prediction time domain; The number of steps in the control time domain and the prediction time domain is configurable and matched with computing resources. The control time domain is set to N steps and the prediction time domain is set to be the control time domain plus 1 to N steps, so that the length of the future perturbation sequence is at least equal to the number of steps obtained in the prediction time domain. The time margin is defined as the sum of the time required for optimization calculation and the delay of the actuator. Through multi-step look-ahead, optimization and caching are ensured to be completed before the future control signal actually takes effect.
[0012] Furthermore, the optimization model employs a weighted quadratic objective function and uses a discrete-time state-space model for output stack prediction. The objective function includes two parts: an output deviation penalty and a control increment penalty. The output deviation penalty is used to suppress vibrations in the power system, minimizing the weighted deviation between the predicted angular velocity of the second motor and the zero-speed reference throughout the entire prediction time domain, while simultaneously suppressing torque oscillations at the output end. The control increment penalty is used to limit the rate and amplitude of change in the control quantities of the first and second motors, balancing vibration suppression effectiveness with energy consumption and execution smoothness. Wherein: The system output prediction is obtained by linear combination of the current state, the executed control quantity of the previous control cycle, the future control increment sequence, and the disturbance prediction sequence through a fixed coefficient matrix. The fixed coefficient matrix is constructed by recursively deriving the discretized system matrix and the control and disturbance transfer matrices along the time domain. By adjusting the output weights and control weights, a trade-off is achieved between vibration suppression effect, control smoothness, and execution constraints; The constraints of the optimization model specifically include: The input constraints are the upper and lower limits of the torque of the first motor and the second motor, given according to the manufacturer's specifications and thermomechanical capabilities, and applied time-by-time in the prediction time domain; The output constraint is the allowable operating range of the second motor's angular velocity and output torque; The control increment sequence is set to zero after exceeding the control time domain, so that the control at subsequent time points retains the final value to ensure that the problem is solvable and the solver is stable.
[0013] Furthermore, in implementing the timely delivery of the control signal sequence, the following steps are included: The optimal control signal sequence corresponding to each future control cycle is stored in a time-stamped ordered buffer queue. When the target time arrives, the signal is output from the buffer to the first motor and the second motor without waiting. When the actuator has a fixed or nearly fixed hysteresis characteristic, the buffer length is matched with the time margin so that the control signal sequence maintains a canceling relationship with the relative phase of the disturbance. Each control cycle repeats the rolling process of acquiring the state, predicting disturbances, solving optimization problems, and refreshing the cache to continuously achieve feedforward compensation for engine torque fluctuations.
[0014] Furthermore, the current system state is estimated using a discrete time-varying Kalman filter, wherein: The prediction step uses the discretized system matrix, the known control variables from the previous control cycle, and the disturbance estimate to update the state prediction values. The correction step uses the measured second motor angular velocity and output torque to correct the state estimate, and adaptively calculates the Kalman gain based on the covariance of process noise and measurement noise.
[0015] According to a second aspect of this disclosure, a hybrid range extender speed control system is provided, the system comprising: The disturbance modeling training module is used to acquire disturbance signals during the operation of the HEV powertrain system and train the echo state network model to learn the time-series dynamic characteristics of the disturbance signals. The disturbance signal is the fluctuation of the crankshaft output torque of the internal combustion engine over time. The fluctuation is sampled to obtain a historical disturbance sequence composed of multiple past control cycles, and the sampling period is the discrete time step of the control cycle. The echo state network model adopts a three-layer structure of input layer, reservoir layer, and output layer. The input layer is used to receive the historical disturbance sequence. The reservoir layer is a recursive neural unit with a nonlinear activation function and leakage integral characteristics to map the input to high-dimensional dynamic features. The output layer provides one-step or multi-step disturbance prediction values based on the reservoir state. The status acquisition module is used to acquire the current system status and the current disturbance value of the HEV power system at the control operation time; the system status includes the relative angular displacement and relative angular velocity between the engine crankshaft and the planetary carrier and the angular velocity of the second motor, and the disturbance value is the engine torque at the current time; The disturbance prediction module is used to input the current disturbance value and the historical disturbance sequence of a predetermined length into the trained echo state network model to predict the future disturbance sequence within multiple control cycles. During prediction, the echo state network model performs multi-step look-ahead prediction, covering the continuous moments required for both the control and prediction time domains, to obtain the future disturbance sequence arranged chronologically from the current moment. This multi-step look-ahead prediction establishes a time margin between control calculation and actual execution, and this time margin is not less than the sum of the computation time required for optimization and the lag time of the actuator. An optimization module is used to construct a predictive control optimization model based on the future disturbance sequence. During construction, an optimization objective function is set with the goal of reducing the vibration of the HEV power system, and control input constraints and system output constraints are applied to the optimization model. The optimization model adopts a discrete-time state-space prediction model, which takes the system state described in the previous step as the initial value, combines the disturbance predictions at each future time, and gradually extrapolates along the prediction time domain to obtain the predicted system output at each future time. The control solution module is used to solve the optimization model within the current control cycle to obtain the optimal control signal sequence covering the predetermined control prediction time domain. During the solution process, a real-time iterative quadratic programming strategy is employed. Given the control time domain length and prediction time domain length configuration, the optimal motor torque command is calculated for multiple time intervals from the current moment. Based on the future disturbance sequence, the optimization results are feedforward compensated for potential engine torque fluctuations. The obtained control signal sequence satisfies all input and output constraints and achieves a trade-off between vibration suppression and control smoothness under the set weights. The buffer execution module is used to temporarily store the control signal corresponding to at least the next control cycle in the optimal control signal sequence, and output the control signal sequence to the actuator of the HEV power system when the next control cycle arrives, so as to implement the active vibration control of the power system in advance by taking advantage of the predicted future disturbances.
[0016] The technical solution disclosed herein has the following beneficial effects: This disclosure utilizes an echo state network to learn the temporal evolution of engine torque disturbances, providing a multi-step look-ahead sequence of future disturbances. This constructs a "time margin" between control calculation and actual execution, transforming the hard real-time constraint of "optimization must be completed within the next cycle" into a scheduling problem of "pre-calculation and timely delivery." Combined with a state-space-based linear prediction model and quadratic programming solution, it can generate an optimal control sequence that phase-matches the upcoming disturbances, while satisfying constraints such as the upper and lower limits of motor torque and the safe output range. This keeps the second motor's angular velocity near zero, significantly reduces output torque fluctuations, effectively suppresses start-up jitter and noise, improves ride comfort, and reduces drivetrain impact load. In particular, the echo state network is trained only on the readout layer with fixed reservoir weights, making training efficient and easily generalizable. It can be combined with probabilistic regression for online correction of prediction uncertainties. The accompanying Kalman filter enables noisy state estimation, ensuring control stability even under sensor noise and parameter perturbations. Attached Figure Description
[0017] Figure 1 This is a flowchart of a hybrid range extender speed control method according to an embodiment of this specification; Figure 2 This is a structural block diagram of a hybrid range extender speed control system according to an embodiment of this specification. Detailed Implementation
[0018] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this disclosure more comprehensive and complete, and to fully convey the concept of the example embodiments to those skilled in the art. The described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a full understanding of embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure can be practiced with one or more of the specific details omitted, or other methods, components, apparatus, steps, etc., can be employed. In other instances, well-known technical solutions are not shown or described in detail to avoid obscuring various aspects of this disclosure.
[0019] Furthermore, the accompanying drawings are merely illustrative of this disclosure. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.
[0020] This invention provides a method for steady-speed control of a hybrid range extender. (Refer to...) Figure 1 The diagram shown is a flowchart illustrating a hybrid range extender speed control method according to an embodiment of the present invention. This method can be applied to electronic devices such as personal computers, servers, and controllers. The method can be executed by a device, which can be implemented by software and / or hardware. Specifically, the method may include the following steps S101-S106: In step S101, disturbance signals during the operation of the HEV powertrain are acquired, and the echo state network model is trained to learn the time-series dynamic characteristics of the disturbance signals. The disturbance signals are the fluctuations in the crankshaft output torque of the internal combustion engine over time. The fluctuations are sampled to obtain a historical disturbance sequence composed of multiple past control cycles, and the sampling period is the discrete time step of the control cycle. The echo state network model adopts a three-layer structure of input layer, reservoir layer, and output layer. The input layer is used to receive the historical disturbance sequence, and the reservoir layer is a recursive neural unit with a nonlinear activation function and leakage integral characteristics to map the input to high-dimensional dynamic features. The output layer provides one or more steps of disturbance prediction values based on the reservoir state.
[0021] The HEV power system consists of an internal combustion engine, a first motor, a second motor, and a planetary gear mechanism coupled together. Based on dynamic modeling, the stiffness and damping of the torsional damper are used as parameters, and a linear state-space representation is adopted with relative angular displacement and relative angular velocity as descriptors. The dynamic modeling and output definition of the HEV power system satisfy the speed and torque coupling relationship of the planetary gear mechanism and explicitly consider the torsional damper; wherein: the speed of the second motor is given by a fixed linear combination of the speed of the carrier and the speed of the first motor according to the planetary transmission ratio, and the speed of the carrier and the pinion is given by a linear combination determined by the planetary transmission ratio; the ring gear torque is equal to the sum of the torque of the first motor converted by the planetary transmission ratio and the torque of the second motor, and the carrier torque is determined according to the proportional relationship between the planetary transmission ratio and the ring gear torque; the elastic potential energy of the torsional damper is proportional to the square of the relative angular displacement between the engine crankshaft and the carrier, the dissipated energy is proportional to the square of the relative angular velocity between the two, and the kinetic energy is the weighted sum of the moment of inertia and the square of the angular velocity of each rotating component; the engine torque is regarded as an external disturbance acting on the HEV power system, and the motor torques of the first motor and the second motor are regarded as constrained control inputs, which are calculated by predictive control in the rolling time domain to achieve active cancellation of engine torque fluctuations. Furthermore, when training the echo state network model, the input weights and the weights within the reservoir are kept in a random sparse fixed connection, and only the weights of the output layer are trained, so as to minimize the error between the predicted disturbance and the actual disturbance on a given training set; wherein, the reservoir is a set of recursive nonlinear units with leakage integral characteristics, and its leakage coefficient is in the range of 0.2 to 0.5; the number of reservoir units is in the range of 200 to 1000; the input of the input layer is the historical disturbance sequence sampled according to the control cycle, and the output of the output layer is the point prediction value of the disturbance at the next moment, and the future disturbance sequence is obtained by repeated iterations; the training data of the echo state network model covers the working conditions under the engine starting scenario, including random variation samples of engine friction torque and starter speed; the weights of the output layer are solved by linear regression, and amplitude constraints or regularization terms are added to the regression objective to suppress overfitting.
[0022] In step S102, during the control operation, the current system state and the current disturbance value of the HEV power system are obtained; the system state includes the relative angular displacement and relative angular velocity between the engine crankshaft and the planetary carrier, and the angular velocity of the second motor; the disturbance value is the engine torque at the current moment.
[0023] As a supplement, the current system state is estimated using a discrete time-varying Kalman filter, wherein: the prediction step updates the state prediction value using the discretized system matrix, the known control quantity of the previous control cycle, and the disturbance estimate; the correction step corrects the state estimate using the measured second motor angular velocity and output torque, and adaptively calculates the Kalman gain based on the covariance of process noise and measurement noise.
[0024] As an explanation, step S101 defines the disturbance as the fluctuation of the crankshaft output torque of the internal combustion engine over time. This fluctuation is discretely sampled according to the control cycle, forming a historical disturbance sequence composed of multiple past control cycles. The sampling period is consistent with the control cycle (e.g., 1 ms). The collected historical disturbance sequence is used as time-series data input to the Echo State Network (ESN) model for learning and prediction. The goal is to extract the dynamic features of the time series without changing the physical source of the disturbance, and output one or more steps of forward disturbance prediction values. This provides external disturbance feedforward information for subsequent rolling time-domain predictive control, thereby achieving active cancellation of engine torque fluctuations. The ESN adopts a three-layer structure: input layer—reservoir—output layer. The input layer receives the aforementioned historical disturbance sequence; the reservoir is composed of recurrent neural units with nonlinear activation functions and has leakage integral characteristics, allowing the input to be mapped to a high-dimensional dynamic feature space; the output layer provides point predictions of the disturbance based on the reservoir state. The leakage integral update formula for the reservoir and the readout layer calculation formula are as follows: ; ; in Leakage coefficient, W and W represent the input weights and reservoir internal weights, respectively. During training, they remain randomly connected and fixed (generally randomly generated from a normal or uniform distribution). Only the output layer weights are trained. ;enter The output is the historical disturbance sequence obtained by sampling according to the control cycle. The predicted value of the point for the perturbation at the next time step is obtained through repeated iterations to achieve multi-step look-ahead prediction. The output layer weights are solved using linear regression, and amplitude constraints or regularization terms can be added to the regression objective to suppress overfitting. The reservoir leakage coefficient can be taken in the range of 0.2 to 0.5, and the reservoir size (number of cells) can be taken in the range of 200 to 1000. In actual implementation, a=0.3 and N=500 can be selected to balance prediction accuracy and computational cost. The training data in this embodiment covers the engine starting scenario, and random variations of engine friction torque and starter speed (e.g., normal perturbation with 5% variance) are introduced into the samples to enhance the model's generalization ability to typical starting transient fluctuations.
[0025] For the physical object of the HEV powertrain system, this embodiment adopts a coupled structure including an internal combustion engine, a first motor (M1), a second motor (M2), and a planetary gear mechanism. The torsional damper between the engine and the carrier is explicitly considered, and relative angular displacement and relative angular velocity are used as descriptive quantities in the modeling. Based on the geometric and dynamic relationships of the planetary gears, the rotational speed and torque satisfy a fixed linear coupling: the rotational speed of the second motor and the pinion (the planetary gear in the planetary gear set) can be given by a linear combination of the rotational speeds of the carrier and the first motor, specifically: ; The torque between the ring gear and the carrier satisfies: ; in This is the planetary gear ratio. The subscripts c, m1, m2, and p represent the carrier, the first motor, the second motor, and the pinion, respectively. , , , These represent the torques of the ring gear, the carrier, and the two motors, respectively. These relationships are used to ensure consistency of velocity and torque among the rotating components in the modeling. Furthermore, the elastic potential energy and dissipated energy of the torsional damper are characterized by relative quantities between the engine crankshaft and the carrier, and the kinetic energy of the entire system is a weighted sum of the moments of inertia and squares of the angular velocities of each rotating component. ; ; in For the corner, For rotational inertia, , For the stiffness and damping of the vibration damper, , The radii of the sun gear and planetary gears are... , Let the mass of a single planetary gear be given by its number. Based on this, construct the Lagrange function. With dissipation function The generalized force of the system is The Lagrange equation can be written as: ; Based on this, the dynamic equations are derived and expressed in a linear state space. Meanwhile, in the HEV structure, the engine is connected to the frame via a torsional damper, and the two motors are coupled to the output shaft (after reduction) via ring gears. The aforementioned speed / torque relationship and energy terms are consistent with the structure.
[0026] At the state-space level, for ease of control and observation definition, this embodiment selects the following state variables and output quantities: ; The control input and external disturbance are respectively: ; in (Engine torque) is explicitly defined as an external disturbance acting on the system. , (The torque of the two motors) serves as the constrained control input. Through Lagrange derivation and elimination using the geometric relationship of the planetary gears, a continuous-time linear time-invariant state-space model can be obtained: ; The matrix is: ; ; ; ; And the control input satisfies the amplitude constraint: ; In the above formula , , , For the planetary gear ratio The equivalent inertia coefficient is obtained by combining the system's rotational inertia with the mass / number of planetary gears; relative rotation angles are used in the modeling. With relative angular velocity Describes the generation of the damper torque. This state-output definition makes: on the one hand, the disturbance... It can serve as a learning / prediction target for ESN, on the other hand, The constraint can be explicitly addressed in subsequent rolling time-domain predictive control optimization, making the solution obtained... , Cancel within constraints This causes fluctuations in output torque.
[0027] Combining the above physical model and data-driven prediction, S101 of this embodiment is specifically executed as follows: First, data is collected under typical operating conditions such as engine starting. The time series is used to construct a historical disturbance sequence, the sampling period of which is equal to the control period; then, this historical sequence is used as the ESN input, employing a randomized and fixed... With leakage coefficient Reservoir recursion, training only the readout matrix To minimize prediction error and suppress overfitting by imposing magnitude or regularization constraints; again, predict the output in one step. Multi-step forward lookup with iterative rolling This is used as a future disturbance sequence for the predictive controller to solve for the constrained problem in the rolling time domain. This allows the torque applied by the two motors to actively counteract the torque applied by the motors while satisfying the constraints. Output torque and speed vibration caused by fluctuations.
[0028] In step S102, it is necessary to perform control operation at each time step. Two types of online information are obtained: the current system state and the current disturbance value. The system state is derived from the state variable definition of the controlled object. ; in This represents the relative angular displacement between the engine crankshaft and the planetary carrier. Let be the relative angular velocity between the two. The angular velocity of the second motor is given; the corresponding system output is defined as: ; The control input and disturbance are: ; Among them, as mentioned above, The engine torque is considered as an external disturbance acting on the system. The continuous-time linear time-invariant state space is represented by the aforementioned continuous-time linear time-invariant state space model; beyond this, the physical relationships between angular velocity and angular displacement are satisfied. , , In digital control implementation, the sampling period is used as the basis for... Discretizing (e.g., 1ms) yields: ; in and These represent process noise and measurement noise, respectively, used to describe the uncertainty of online measurements. To obtain the "current system state" in real time under noisy conditions, a discrete time-varying Kalman filter is embedded for estimation: ; ; ; ; ; Therefore, at time The current system state is taken as .in Output can be used directly Correction obtained, and The state equations and observations are jointly estimated by a filter; if necessary, the ring gear torque in the output is also included. This can be used as a measurement to improve estimation accuracy. This process ensures that even in the presence of noise and when components cannot be directly measured, a three-dimensional state estimate satisfying the definition of this step can still be provided within each control cycle.
[0029] The current disturbance value is the engine torque at the current moment. During control operation, data is obtained through available engine torque measurements (or equivalent calculated measurements). And use it as a disturbance input for state prediction and correction: First, in the prediction step, press Explicitly use the torque known at the previous moment. Secondly, the current output is incorporated into the correction step. After correction This allows the acquisition of the current system state. It is important to emphasize that the modeling of the controlled object explicitly defines engine torque as an external disturbance. Therefore, in this step, we directly use The disturbance value can be obtained instantaneously, meaning it is acquired only at the current moment. The acquisition of both the disturbance and the state is synchronized with the control cycle.
[0030] In step S103, the current disturbance value and the historical disturbance sequence of a predetermined length are input into the trained echo state network model to predict the future disturbance sequence within multiple control cycles. When making predictions, the echo state network model performs multi-step look-ahead prediction, covering the continuous moments required for both the control time domain and the prediction time domain, to obtain the future disturbance sequence arranged chronologically from the current moment. The multi-step look-ahead prediction is used to establish a time margin between control calculation and actual execution, and the time margin is not less than the sum of the calculation time required for optimization and the lag time of the actuator.
[0031] The establishment of the time margin includes: the prediction step count covering both the control time domain and the prediction time domain, with the number of steps in the control time domain not exceeding the number of steps in the prediction time domain; the number of steps in the control time domain and the prediction time domain are configurable and matched with computing resources, and the control time domain is set to N steps and the prediction time domain is set to be the control time domain plus 1 to N steps, such that the length of the future disturbance sequence is at least equal to the number of steps obtained in the prediction time domain; the time margin is defined as the sum of the time required for optimization calculation and the delay of the actuator, and multi-step look-ahead ensures that optimization and caching are completed before the future control signal actually takes effect.
[0032] For explanation purposes, at control time k, the current disturbance value and a predetermined-length historical disturbance sequence sampled at the control period are used as inputs and fed into a trained echo state network. During the prediction phase, the network maintains the randomly sparse fixed connections of the input weights and the internal reservoir weights, and only uses the trained readout weights for calculation; the reservoir is a recurrent non-linear unit with leaky integration characteristics, and the per-time update and readout respectively satisfy the leaky integration update formula of the reservoir and the readout layer calculation formula in step S101 above; where is the input sequence, is the reservoir state, is the network output, is the leakage coefficient. For the disturbance modeling of this embodiment, the input unit receives the engine torque waveforms (historical disturbance sequences) within a number of past control periods and the current disturbance value, and the output unit gives the disturbance point prediction for the next discrete step; by repeating this operation, a trajectory-tracking multi-step-ahead disturbance prediction sequence can be formed to cover the disturbance evolution within multiple future control periods.
[0033] To ensure that the prediction sequence is arranged in sequence from the current moment in time and covers the control time domain and the prediction time domain, a state-space-based rolling time domain representation is used to uniformly characterize the multi-step-ahead and the time domain length. Given the current estimated state , the prediction of the next state and output satisfies: ; where is the disturbance prediction given by the echo state network. Stacking the controls for the next steps, the outputs, and the disturbances in the rolling window in time, we get: ; ; ; where is the number of steps in the control time domain, is the number of steps in the prediction time domain and , , , , are block matrices discretized from . The future disturbance sequence is exactly the vector arranged in time sequence starting from in and covering up to ; its length is at least equal to the number of steps in the prediction time domain .
[0034] To establish a time margin between control calculation and actual execution, this embodiment defines the time margin as the time difference between the current calculation time and the time when the optimized calculation result is used for execution. Utilizing the future perturbation sequence provided by the echo state network, rolling optimization is performed in advance within this margin, and future control signals are buffered, only to be output to the actuator when the corresponding time arrives. The time margin can be set independently, but it must be no less than the sum of the optimization calculation time and the actuator lag, thus completing optimization and buffering before the future control signal actually takes effect. The network's multi-step look-ahead prediction must at least cover... Step, and ensure So that we can both prepare for the future The step-by-step control calculation provides sufficient disturbance feedforward and leaves enough margin on the time axis to compensate for calculation and lag.
[0035] In step S104, a predictive control optimization model is constructed based on the future disturbance sequence. During the construction, an optimization objective function is set with the goal of reducing the vibration of the HEV power system, and control input constraints and system output constraints are applied to the optimization model. The optimization model adopts a discrete-time state-space prediction model, which takes the system state described in the previous step as the initial value, combines the disturbance predictions at each future time, and gradually extrapolates along the prediction time domain to obtain the system output predictions at each future time.
[0036] The optimization model employs a weighted quadratic objective function and uses a discrete-time state-space model for output stack prediction. The objective function includes two parts: an output deviation penalty and a control increment penalty. The output deviation penalty is used to suppress vibration of the power system, minimizing the weighted deviation between the predicted angular velocity of the second motor and the zero-speed reference throughout the entire prediction time domain, while simultaneously suppressing torque oscillation at the output end. The control increment penalty is used to limit the rate and amplitude of change in the control quantities of the first and second motors, balancing vibration suppression effectiveness with energy consumption and execution smoothness. The system output prediction is obtained by linearly combining the current state, the executed control quantity of the previous control cycle, the future control increment sequence, and the disturbance prediction sequence through a fixed coefficient matrix. The fixed coefficient matrix is constructed by recursively constructing the discretized system matrix and the control and disturbance transfer matrices along the time domain. By adjusting the output weight and control weight, a trade-off is achieved between vibration suppression effect, control smoothness, and execution constraints. The constraints of the optimization model specifically include: input constraints are the upper and lower limits of the torque of the first motor and the second motor, given according to the manufacturer's specifications and thermomechanical capabilities, and applied time-by-time in the prediction time domain; output constraints are the allowable operating range of the angular velocity and output torque of the second motor; the control increment sequence is set to zero after exceeding the control time domain, so that the control at subsequent time moments maintains the final value to ensure that the problem is solvable and the solver is stable.
[0037] Specifically, using the future disturbance sequence obtained in step S103 as feedforward information, a predictive control optimization model based on discrete-time state space is constructed; this model estimates the system state at the end of the previous control cycle. The initial value is used, and the forecast is combined with the perturbation prediction at each future time in the prediction time domain. Stepwise extrapolation output. Discrete prediction recursive writing: ; And it uses input incremental parameterization: ; in This refers to the control variables that were executed in the previous control cycle. It will be used in the future... Step control increment sequence and By stacking the step output and perturbation over time in a rolling window, a linear stacked prediction relationship can be obtained, as described above. , The coefficient matrix is derived from the discretized... Recursive construction in the time domain: , ,
[0038] ; Decision variables are The above expression fully expands the requirement for output stack prediction, obtained by linearly combining the current state, the control quantity from the previous moment, the future control increment sequence, and the disturbance prediction sequence through a fixed coefficient matrix, in matrix form.
[0039] The optimization model employs a weighted quadratic objective function, comprising output deviation penalty and control increment penalty, to suppress dynamic system vibration while balancing execution smoothness and energy consumption. The tracking cost function is written as follows: ; Where Q and R are diagonal weighted matrices. To suppress vibration, the output reference... The trajectory is set as the zero-speed reference for the second motor's angular velocity and suppresses torque oscillation at the output end, i.e., for Bias weighting with zero reference is applied across the entire prediction time domain to minimize Deviating from zero speed and suppressing The oscillation. By adjusting Q (e.g., increasing...) , The weights of the components and R (increasing the control increment weight) can achieve a trade-off between vibration suppression strength, control smoothness, and energy consumption. To ensure the problem is solvable and compatible with execution constraints, control increments exceeding the control time domain are set to zero, so that the control maintains its terminal value in subsequent time steps. ; Constraints are applied time-by-time in the prediction domain, including both input and output constraints. Input constraints are the upper and lower limits of the torque of the two motors, satisfying manufacturing specifications and thermomechanical capabilities. ; The output constraint is the angular velocity of the second motor. With output torque Permissible working scope: ; The above input and output boundaries can all be configured according to predetermined parameters and safety margins; in the model , pass get, This is generated by the aforementioned stack prediction. In this way, both types of constraints are satisfied during optimization: the torque limit of the first motor and the second motor, and the operating range of speed / torque.
[0040] The above objectives and constraints are rewritten in standard quadratic programming (QP) form for efficient solution. Utilizing... and take the reference stack We can obtain: ; in: ; , , , ; This QP balances minimizing output deviation with smoothing control increment. Time-by-time constraints with input / output Unified within a convex optimization framework; due to perturbation prediction As already provided, the solver directly obtains the constraint-satisfied values in the prediction time domain. Then through Generate future control sequences. The standard QP algorithm can be used for real-time iterative solution; combined with the aforementioned multi-step look-ahead prediction and incremental zeroing beyond the control time domain, while ensuring computational feasibility, a smooth and constrained control signal is provided to the actuator, thereby effectively suppressing [the problem] throughout the entire prediction time domain. , vibration In step S105, the optimization model is solved within the current control cycle to obtain the optimal control signal sequence covering the predetermined control prediction time domain. A real-time iterative quadratic programming solution strategy is adopted during the solution process. Under the given control time domain length and prediction time domain length configuration, the optimal motor torque command is calculated for multiple time intervals from the current time. Based on the future disturbance sequence, the optimization result is fed forward to compensate for the engine torque fluctuations that will occur. The obtained control signal sequence satisfies all input and output constraints and achieves a trade-off between vibration suppression and control smoothness under the weight setting.
[0041] In the current control period k, the discrete-time prediction model constructed in step S104 is combined with the future disturbance sequence to solve the optimization problem. The objective function is a weighted quadratic form, consisting of output deviation penalty and control increment penalty: ; And apply discrete state prediction and input / output constraints: ; ; ; Where Q and R are diagonal weighted matrices, corresponding to the output and control input weights, respectively. This represents the forward-looking prediction of the engine torque that will occur. To meet the solution requirements within the current control cycle, a real-time iterative strategy is adopted, updating the optimal solution at each sampling time. The problem is formulated within a finite time domain with constraints, and can be calculated by a standard solver updating the optimal solution at each step. To meet the solution requirements of real-time iterative quadratic programming, the above equation is equivalently rewritten as relating to the control increment stack. Standard QP: ; ; ; , ; in , , , From discretized Obtained by recursion in the time domain For reference stack, , This is a fixed matrix that maps the input magnitude and incremental constraints to linear inequalities time-by-time. The QP form described above explicitly maps the current state... Previous cycle control quantity Future control increment With perturbation prediction stack The output stack relationship obtained by linear combination is embedded in the cost and constraints, thus obtaining the optimal control signal sequence covering the control / prediction time domain in a single solution. Due to the use of a generalized representation, the standard QP algorithm can be directly called to meet the real-time calculation needs within the current control cycle.
[0042] The real-time iterative online solution process is as follows: at time k, the Kalman filter is used to obtain... As initial values, the future perturbation sequence Incorporation error With the right end item In this process, the current QP is formed and the solver is invoked to iteratively update the optimal solution; the approach of updating the optimal solution at each step and correcting it as the optimal solution changes is called real-time iteration. After the solution is completed, the following is obtained: With the corresponding This generates a sequence of torque commands for the two motors that extends "multiple times from the current time". .in, The multi-step look-ahead predictions provided by the echo state network directly participate in the cost and constraint calculation of QP, which is equivalent to explicitly introducing feed-forward compensation for the engine torque fluctuations that will occur during optimization. This mechanism is consistent with the controller structure that uses future disturbance predictions to leave time margin for optimization at subsequent times.
[0043] To ensure the feasibility and stability of the solution, the control increment is adjusted according to the control time domain. The process ensures that subsequent time-series control maintains the final value; input constraints are applied time-by-time in the prediction domain according to manufacturing and thermomechanical capabilities. Output constraints are applied within the allowable operating range. Due to QP for By tuning Q (emphasis) and (deviation penalty) and R (emphasis) Smoothing penalty), which can reduce vibration (reducing) The optimal balance is achieved between zero-speed deviation and suppression of output torque oscillation and control smoothing / energy consumption; the obtained control sequence is within the constraint set and satisfies the above trade-off.
[0044] In step S106, the control signal corresponding to at least the next control cycle in the optimal control signal sequence is temporarily stored, and the control signal sequence is output to the actuator of the HEV power system when the next control cycle arrives, so as to implement the active vibration control of the power system in advance by utilizing the predicted future disturbances.
[0045] When implementing the timely delivery of the control signal sequence, the process includes: storing the segments of the optimal control signal sequence corresponding to each future control cycle in a time-stamped ordered buffer queue, and outputting them to the first and second motors without waiting when the target time arrives; when the actuator has a fixed or nearly fixed hysteresis characteristic, the buffer length is matched with the time margin, so that the relative phase of the control signal sequence and the arrival of the disturbance remains canceled; and repeating the rolling process of acquiring the state, predicting the disturbance, solving the optimization, and refreshing the buffer in each control cycle to continuously achieve feedforward compensation for engine torque fluctuations.
[0046] In step S106, the optimal control signal sequence obtained in step S105 is temporarily stored in chronological order, including at least the segment to be issued in the next control cycle. Upon reaching the corresponding moment, the motor torque command for that moment is output to the first and second motors without waiting, ensuring that the control is optimized and scheduled before actual execution. This implementation relies on the idea of pre-determining the optimal control for future moments and applying it to future states: instead of completing optimization only in the next control cycle, the future disturbance prediction given by the echo state network is used to pre-establish the optimization result for predictive control, ensuring that the optimal control signal is determined before the arrival of subsequent control cycles. Therefore, during the execution phase, only timely issuance and buffer queue reading are required. In other words, to facilitate real-time optimization, the control input is determined to be pre-applied to future states, and the optimal control signal required in the future is calculated by applying the predicted disturbance to predictive control in the future. Therefore, the optimal signal can be determined before the subsequent control cycle occurs. Accordingly, during project deployment, a time-stamped ordered cache queue is used to store the control sequence of multiple sampling points sequentially from the current moment, and each item is output from the cache to the execution mechanism when the target moment arrives, so as to achieve rolling execution with advance optimization and timely delivery. When the actuator has a fixed or nearly fixed lag / dead time, the buffer length should match the time margin to maintain phase cancellation of disturbances. As the set time delay increases, if the compensation and disturbance are not synchronized, control input phase reversal and compensation error amplification often occur. However, when the time margin is set reasonably, the output torque remains stable and the control input amplitude does not change significantly, indicating that the relative phase of the feedforward compensation and the disturbance maintains a canceling relationship. Based on this principle, in actual implementation, the buffer depth and the sending alignment point are set according to the estimated value of the total time delay (time margin = solution / communication / execution, etc.), so that the control segment of the i-th cycle in the buffer and the predicted disturbance of the i-th cycle are superimposed in phase at the execution level, thereby achieving active suppression of engine torque fluctuations.
[0047] The aforementioned buffer deployment mechanism operates continuously in a rolling manner at the control layer. Each control cycle first obtains the current system state through state observation and filtering, then uses an echo state network to generate a future disturbance sequence covering the control / prediction time domain. Based on this, a constrained quadratic programming problem is constructed and solved to obtain a new optimal control sequence. Subsequently, the buffer queue of undeployed future segments is refreshed, and only the one that should be effective at the current execution time is output. Because constraints such as amplitude limiting are applied hourly in the prediction time domain, the control commands for both motors satisfy the input and output boundaries at any time they are deployed. Simultaneously, by weighting the output deviation and control increment separately in the cost function, the optimal trade-off between vibration suppression and control smoothness is fully reflected in the buffer sequence. The rolling execution sequence can therefore be summarized as follows: obtain the state, predict the disturbance, solve the optimization, refresh the cache, issue on time, step into the next cycle, repeat in each cycle, so as to continuously achieve feedforward compensation and vibration suppression for engine torque fluctuations.
[0048] In one embodiment, the method further includes: establishing a probabilistic regression model for the time series of the perturbation signal by combining Gaussian process regression, used for adaptive correction and credibility assessment of the prediction results of the echo state network model; wherein: The time series of the disturbance signal is regarded as a random process that follows a joint Gaussian distribution, and a regression model is obtained by estimating the mean and covariance. The covariance is composed of the radial basis kernel of the time difference and the observation noise. The hyperparameter of the kernel function is obtained by maximizing the log-likelihood and using gradient descent or equivalent numerical methods. During the online control phase, the regression model is updated using the disturbance signal within the most recent time window, and the predicted mean is output as a correction for the deviation of the prediction of the echo state network model. The predicted variance is output as a measure of disturbance uncertainty for adaptive adjustment of control weights or constraint redundancy.
[0049] Specifically, a probabilistic regression model of the disturbance signal (the fluctuation of the internal combustion engine crankshaft torque over time) using Gaussian process regression is introduced to adaptively correct the look-ahead prediction of the echo state network and provide a reliability metric. This disturbance time series is treated as a stochastic process following a joint Gaussian distribution: for any input time... Given a time series input and its corresponding output When constructing a regression model The joint distribution is determined by the mean. With covariance Characterization, by estimating both, allows for the generation of a probabilistic model for a given time series. When considering measurement noise, the target regression can be derived from the covariance matrix. The calculation is performed, where the matrix elements are kernel functions. The constructed Gram matrix and observation noise sum ( (The Kronecker symbol is 1 when i=j); to adapt to the changes in interference statistics during operation, "adaptive estimation based on input data" is adopted, so that the model can be updated online with window data.
[0050] In the prediction phase, new inputs The regression mean and variance (corresponding to "the next control period and subsequent times") are given by the following formula: ; in It is a vector, whose components are given by each training point. Computational kernel function The above As a correction factor for the deviation of the echo state network output (i.e., the corrected perturbation point prediction), As a measure of disturbance uncertainty, during the online control phase, the disturbance samples collected within the most recent time window are updated. and Output Used for correction, output This is used to perform adaptive tuning in optimization, such as increasing the weight of output deviations or relaxing constraint redundancy, to maintain a robust trade-off between "vibration suppression strength - control smoothness / energy consumption - constraint feasibility".
[0051] The covariance is calculated as the sum of the radial basis function kernel of the time difference and the observation noise; the selected kernel function is: ; in For hyperparameters, maximize the log-likelihood. It can also be used in conjunction with gradient descent or equivalent numerical methods for estimation: ; In implementation, hyperparameters and noise accuracy Revaluation can be performed using a sliding window. and It adaptively updates over time, thereby continuously correcting the disturbance predictions of the echo state network online under conditions such as fluctuations in operating conditions, changes in the environment, or changes in combustion state, and using... This is a confidence threshold (e.g., tightening the trigger weights or relaxing the constraint redundancy).
[0052] In summary, the probabilistic regression and echo state network described above work together on the time axis. Each control cycle collects new perturbation samples and updates the probabilistic model, immediately outputting the next time step and several subsequent departure times. ,in This is superimposed on the aforementioned multi-step forward-looking perturbation trajectory to correct the prediction. The drive controller adaptively adjusts the weights or constraints redundancy within the prediction domain; then, it queues and buffers the optimized control sequence according to a predetermined time margin, and sends it out without waiting at the target time, so as to maintain the relative phase cancellation relationship with the arrival of disturbance and achieve robustness and real-time performance of active vibration control.
[0053] Based on the same line of thought, such as Figure 2 The diagram shown is a structural block diagram of a hybrid range extender speed control system provided in an embodiment of the present invention. The system includes: The disturbance modeling training module 201 is used to acquire disturbance signals during the operation of the HEV powertrain system and train the echo state network model to learn the time-series dynamic characteristics of the disturbance signals. The disturbance signal is the fluctuation of the crankshaft output torque of the internal combustion engine over time. The fluctuation is sampled to obtain a historical disturbance sequence composed of multiple past control cycles, and the sampling period is the discrete time step of the control cycle. The echo state network model adopts a three-layer structure of input layer, reservoir layer and output layer. The input layer is used to receive the historical disturbance sequence. The reservoir layer is a recursive neural unit with a nonlinear activation function and leakage integral characteristics to map the input to high-dimensional dynamic features. The output layer provides one or more steps of disturbance prediction value based on the reservoir state. The status acquisition module 202 is used to acquire the current system status and the current disturbance value of the HEV power system at the control operation time; the system status includes the relative angular displacement and relative angular velocity between the engine crankshaft and the planetary carrier and the angular velocity of the second motor, and the disturbance value is the engine torque at the current time; The disturbance prediction module 203 is used to input the current disturbance value and the historical disturbance sequence of a predetermined length into the trained echo state network model to predict the future disturbance sequence within multiple control cycles. During prediction, the echo state network model performs multi-step look-ahead prediction, covering the continuous moments required for both the control and prediction time domains, to obtain the future disturbance sequence arranged chronologically from the current moment. This multi-step look-ahead prediction establishes a time margin between control calculation and actual execution, and the time margin is not less than the sum of the calculation time required for optimization and the lag time of the actuator. The optimization construction module 204 is used to construct a predictive control optimization model based on the future disturbance sequence. During construction, an optimization objective function is set with the goal of reducing the vibration of the HEV power system, and control input constraints and system output constraints are applied to the optimization model. The optimization model adopts a discrete-time state-space prediction model, which takes the system state described in the previous step as the initial value, combines the disturbance predictions at each future time, and gradually extrapolates along the prediction time domain to obtain the predicted system output at each future time. The control solution module 205 is used to solve the optimization model within the current control cycle to obtain the optimal control signal sequence covering the predetermined control prediction time domain. During the solution process, a real-time iterative quadratic programming strategy is employed. Given the control time domain length and prediction time domain length configuration, the optimal motor torque command is calculated for multiple time intervals from the current moment. Based on the future disturbance sequence, the optimization results are feedforward compensated for the upcoming engine torque fluctuations. The obtained control signal sequence satisfies all input and output constraints and achieves a trade-off between vibration suppression and control smoothness under the weight settings. The buffer execution module 206 is used to temporarily store the control signal corresponding to at least the next control cycle in the optimal control signal sequence, and output the control signal sequence to the actuator of the HEV power system when the next control cycle arrives, so as to implement the active vibration control of the power system in advance by utilizing the predicted future disturbances.
[0054] The specific details of the above system have been described in detail in the method section of the implementation plan. For any undisclosed details, please refer to the implementation plan of the method section, and therefore will not be repeated here.
[0055] This system uses an echo state network to learn the temporal evolution of engine torque disturbances and provides a multi-step look-ahead sequence for future disturbances. It constructs a "time margin" between control calculations and actual execution, transforming the hard real-time constraint of "optimization must be completed within the next cycle" into a scheduling problem of "pre-calculation and timely delivery." Combined with a state-space based linear prediction model and quadratic programming solution, it can generate an optimal control sequence that phase-matches the upcoming disturbances while satisfying constraints such as the upper and lower limits of motor torque and the safe output range. This keeps the second motor's angular velocity near zero, significantly reduces output torque fluctuations, effectively suppresses start-up jitter and noise, improves ride comfort, and reduces drivetrain impact load. In particular, the echo state network is trained only on the readout layer with fixed reservoir weights, making training efficient and easily generalizable. It can be combined with probabilistic regression to correct prediction uncertainties online. The accompanying Kalman filter enables noisy state estimation, ensuring control stability even under sensor noise and parameter perturbations.
[0056] The accompanying drawings are merely illustrative of the processes included in the methods according to exemplary embodiments of this disclosure and are not intended to be limiting. It is readily understood that the processes shown in the drawings do not indicate or limit the temporal order of these processes. Furthermore, it is readily understood that these processes may be executed synchronously or asynchronously, for example, in multiple modules.
[0057] It should be noted that although several modules or units of the system have been mentioned in the detailed description above, this division is not mandatory. In fact, according to exemplary embodiments of this disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.
[0058] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.
[0059] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.
Claims
1. A method for steady-speed control of a hybrid range extender, characterized in that, The method includes: The system acquires disturbance signals during the operation of the HEV powertrain and trains an echo state network model to learn the time-series dynamic characteristics of the disturbance signals. The disturbance signals are the fluctuations in the crankshaft output torque of the internal combustion engine over time. The fluctuations are sampled to obtain a historical disturbance sequence composed of multiple past control cycles, with the sampling period being the discrete time step of the control cycle. The echo state network model adopts a three-layer structure of input layer, reservoir layer, and output layer. The input layer receives the historical disturbance sequence, and the reservoir layer consists of recurrent neural units with nonlinear activation functions and leakage integral characteristics to map the input to high-dimensional dynamic features. The output layer provides one-step or multi-step disturbance prediction values based on the reservoir state. During control operation, the current system state and current disturbance value of the HEV power system are acquired; the system state includes the relative angular displacement and relative angular velocity between the engine crankshaft and the planetary carrier, as well as the angular velocity of the second motor, and the disturbance value is the engine torque at the current moment; The current disturbance value and the historical disturbance sequence of a predetermined length are input into the trained echo state network model to predict the future disturbance sequence within multiple control cycles. During prediction, the echo state network model performs multi-step look-ahead prediction, covering the continuous moments required for both the control and prediction time domains, to obtain the future disturbance sequence arranged chronologically from the current moment. This multi-step look-ahead prediction establishes a time margin between control calculation and actual execution, and this time margin is not less than the sum of the computation time required for optimization and the lag time of the actuator. Based on the future disturbance sequence, a predictive control optimization model is constructed. During the construction, an optimization objective function is set with the goal of reducing the vibration of the HEV power system, and control input constraints and system output constraints are applied to the optimization model. The optimization model adopts a discrete-time state-space prediction model, which takes the system state described in the previous step as the initial value, combines the disturbance predictions at each future time, and gradually extrapolates along the prediction time domain to obtain the predicted system output at each future time. Within the current control cycle, the optimization model is solved to obtain the optimal control signal sequence covering the predetermined control prediction time domain. A real-time iterative quadratic programming solution strategy is adopted during the solution process. Under the given control time domain length and prediction time domain length configuration, the optimal motor torque command is calculated for multiple time steps forward from the current time. Based on the future disturbance sequence, the optimization result is fed forward to compensate for the engine torque fluctuations that will occur. The obtained control signal sequence satisfies all input and output constraints and achieves a trade-off between vibration suppression and control smoothness under the weight settings. The control signal corresponding to at least the next control cycle in the optimal control signal sequence is temporarily stored, and the control signal sequence is output to the actuator of the HEV power system when the next control cycle arrives, so as to implement the active vibration control of the power system in advance by taking advantage of the predicted future disturbances.
2. The hybrid range extender speed control method according to claim 1, characterized in that, When training the echo state network model, the input weights and the weights within the reservoir are kept in a random, sparse, and fixed connection. Only the weights of the output layer are trained to minimize the error between the predicted and actual disturbances on a given training set. The reservoir is a set of recursive nonlinear units with leakage integral characteristics, and its leakage coefficient is in the range of 0.2 to 0.
5. The number of reservoir units is in the range of 200 to 1000. The input of the input layer is the historical disturbance sequence sampled according to the control cycle, and the output of the output layer is the point prediction value of the disturbance at the next time step. The future disturbance sequence is obtained through repeated iterations and multiple look-ahead steps. The training data of the echo state network model covers the operating conditions under the engine starting scenario, including random variation samples of engine friction torque and starter speed. The weights of the output layer are solved by linear regression, and amplitude constraints or regularization terms are added to the regression objective to suppress overfitting.
3. The hybrid range extender speed control method according to claim 1, characterized in that, The method further includes: establishing a probabilistic regression model for the time series of the perturbation signal by combining Gaussian process regression, which is used to adaptively correct and evaluate the reliability of the prediction results of the echo state network model; wherein: The time series of the disturbance signal is regarded as a random process that follows a joint Gaussian distribution, and a regression model is obtained by estimating the mean and covariance. The covariance is composed of the radial basis kernel of the time difference and the observation noise. The hyperparameter of the kernel function is obtained by maximizing the log-likelihood and using gradient descent or equivalent numerical methods. During the online control phase, the regression model is updated using the disturbance signal within the most recent time window, and the predicted mean is output as a correction for the deviation of the prediction of the echo state network model. The predicted variance is output as a measure of disturbance uncertainty for adaptive adjustment of control weights or constraint redundancy.
4. The hybrid range extender speed control method according to claim 1, characterized in that, The HEV powertrain consists of an internal combustion engine, a first motor, a second motor, and a planetary gear mechanism coupled together. Based on dynamic modeling, the stiffness and damping of the torsional damper are used as parameters, and a linear state-space representation is adopted with relative angular displacement and relative angular velocity as descriptors.
5. The hybrid range extender speed control method according to claim 4, characterized in that, The dynamic modeling and output definition of the HEV powertrain system satisfy the speed and torque coupling relationship of the planetary gear mechanism and explicitly consider the torsional damper; wherein: The speed of the second motor is given by a fixed linear combination of the speed of the carrier and the speed of the first motor according to the planetary transmission ratio, and the speed of the carrier and the pinion is given by a linear combination determined by the planetary transmission ratio. The ring gear torque is equal to the sum of the torque of the first motor converted by the planetary transmission ratio and the torque of the second motor. The frame torque is determined according to the proportional relationship between the planetary transmission ratio and the ring gear torque. The elastic potential energy of the torsional damper is proportional to the square of the relative angular displacement between the engine crankshaft and the carrier, the dissipated energy is proportional to the square of the relative angular velocity between the two, and the kinetic energy is the weighted sum of the moment of inertia and the square of the angular velocity of each rotating component. The engine torque is considered as an external disturbance acting on the HEV power system, and the motor torques of the first motor and the second motor are considered as constrained control inputs. Predictive control is calculated in the rolling time domain to achieve active cancellation of engine torque fluctuations.
6. The hybrid range extender speed control method according to claim 1, characterized in that, When establishing the time margin, the following are included: The number of prediction steps covers both the control time domain and the prediction time domain, and the number of steps in the control time domain is no greater than the number of steps in the prediction time domain; The number of steps in the control time domain and the prediction time domain is configurable and matched with computing resources. The control time domain is set to N steps and the prediction time domain is set to be the control time domain plus 1 to N steps, so that the length of the future perturbation sequence is at least equal to the number of steps obtained in the prediction time domain. The time margin is defined as the sum of the time required for optimization calculation and the delay of the actuator. Through multi-step look-ahead, optimization and caching are ensured to be completed before the future control signal actually takes effect.
7. The hybrid range extender speed control method according to claim 1, characterized in that, The optimization model employs a weighted quadratic objective function and uses a discrete-time state-space model for output stack prediction. The objective function includes two parts: an output deviation penalty and a control increment penalty. The output deviation penalty is used to suppress vibration of the power system, minimizing the weighted deviation between the predicted angular velocity of the second motor and the zero-speed reference throughout the entire prediction time domain, while simultaneously suppressing output torque oscillation. The control increment penalty is used to limit the rate and amplitude of change in the control inputs of the first and second motors, balancing vibration suppression effectiveness with energy consumption and execution smoothness. Wherein: The system output prediction is obtained by linear combination of the current state, the executed control quantity of the previous control cycle, the future control increment sequence, and the disturbance prediction sequence through a fixed coefficient matrix. The fixed coefficient matrix is constructed by recursively deriving the discretized system matrix and the control and disturbance transfer matrices along the time domain. By adjusting the output weights and control weights, a trade-off is achieved between vibration suppression effect, control smoothness, and execution constraints; The constraints of the optimization model specifically include: The input constraints are the upper and lower limits of the torque of the first motor and the second motor, given according to the manufacturer's specifications and thermomechanical capabilities, and applied time-by-time in the prediction time domain; The output constraint is the allowable operating range of the second motor's angular velocity and output torque; The control increment sequence is set to zero after exceeding the control time domain, so that the control at subsequent time points retains the final value to ensure that the problem is solvable and the solver is stable.
8. The hybrid range extender speed control method according to claim 1, characterized in that, When implementing the timely delivery of the control signal sequence, the following are included: The optimal control signal sequence corresponding to each future control cycle is stored in a time-stamped ordered buffer queue. When the target time arrives, the signal is output from the buffer to the first motor and the second motor without waiting. When the actuator has a fixed or nearly fixed hysteresis characteristic, the buffer length is matched with the time margin so that the control signal sequence maintains a canceling relationship with the relative phase of the disturbance. Each control cycle repeats the rolling process of acquiring the state, predicting disturbances, solving optimization problems, and refreshing the cache to continuously achieve feedforward compensation for engine torque fluctuations.
9. The hybrid range extender speed control method according to claim 1, characterized in that, The current system state is estimated using a discrete time-varying Kalman filter, where: The prediction step uses the discretized system matrix, the known control variables from the previous control cycle, and the disturbance estimate to update the state prediction values. The correction step uses the measured second motor angular velocity and output torque to correct the state estimate, and adaptively calculates the Kalman gain based on the covariance of process noise and measurement noise.
10. A hybrid range extender speed control system, the system comprising: The disturbance modeling training module is used to acquire disturbance signals during the operation of the HEV powertrain system and train the echo state network model to learn the time-series dynamic characteristics of the disturbance signals. The disturbance signal is the fluctuation of the crankshaft output torque of the internal combustion engine over time. The fluctuation is sampled to obtain a historical disturbance sequence composed of multiple past control cycles, and the sampling period is the discrete time step of the control cycle. The echo state network model adopts a three-layer structure of input layer, reservoir layer and output layer. The input layer is used to receive the historical disturbance sequence, and the reservoir layer is a recursive neural unit with a nonlinear activation function and leakage integral characteristics to map the input to high-dimensional dynamic features. The output layer provides one or more steps of disturbance prediction based on the reservoir state; The status acquisition module is used to acquire the current system status and the current disturbance value of the HEV power system at the control operation time; the system status includes the relative angular displacement and relative angular velocity between the engine crankshaft and the planetary carrier and the angular velocity of the second motor, and the disturbance value is the engine torque at the current time; The disturbance prediction module is used to input the current disturbance value and the historical disturbance sequence of a predetermined length into the trained echo state network model to predict the future disturbance sequence within multiple control cycles. During prediction, the echo state network model performs multi-step look-ahead prediction, covering the continuous moments required for both the control and prediction time domains, to obtain the future disturbance sequence arranged chronologically from the current moment. This multi-step look-ahead prediction establishes a time margin between control calculation and actual execution, and this time margin is not less than the sum of the computation time required for optimization and the lag time of the actuator. An optimization module is used to construct a predictive control optimization model based on the future disturbance sequence. During construction, an optimization objective function is set with the goal of reducing the vibration of the HEV power system, and control input constraints and system output constraints are applied to the optimization model. The optimization model adopts a discrete-time state-space prediction model, which takes the system state described in the previous step as the initial value, combines the disturbance predictions at each future time, and gradually extrapolates along the prediction time domain to obtain the predicted system output at each future time. The control solution module is used to solve the optimization model within the current control cycle to obtain the optimal control signal sequence covering the predetermined control prediction time domain. During the solution process, a real-time iterative quadratic programming strategy is employed. Given the control time domain length and prediction time domain length configuration, the optimal motor torque command is calculated for multiple time intervals from the current moment. Based on the future disturbance sequence, the optimization results are feedforward compensated for potential engine torque fluctuations. The obtained control signal sequence satisfies all input and output constraints and achieves a trade-off between vibration suppression and control smoothness under the set weights. The buffer execution module is used to temporarily store the control signal corresponding to at least the next control cycle in the optimal control signal sequence, and output the control signal sequence to the actuator of the HEV power system when the next control cycle arrives, so as to implement the active vibration control of the power system in advance by taking advantage of the predicted future disturbances.
Citation Information
Cited By
Shock absorber performance optimization control method based on model fusion
CN121300099A
