A real-time energy management method based on CNN-BiLSTM prediction and PPO hot start model prediction control

Abstract

A real-time energy management method based on CNN-BiLSTM prediction and PPO hot-start model predictive control belongs to the field of aerospace power energy management technology. A model of the turboshaft engine, generator and motor controller, power battery, and power battery degradation of a hybrid-electric aircraft is established, and a whole-aircraft simulation model is built. A model predictive control framework is constructed, and the objective function and physical constraints of the energy management method are determined. The flight power demand is predicted using CNN-BiLSTM, and the optimal hot-start initial value is provided for the SQP solver of the MPC using an offline-trained PPO network, realizing real-time airborne energy management. This method improves the online solving speed of the MPC, meets the requirements of real-time airborne control, avoids the black-box constraints of reinforcement learning, and balances the requirements of airborne real-time performance, fuel economy, and battery life extension. It has been verified by HIL hardware-in-the-loop testing and has strong engineering deployability.

Description

Technical Field

[0001] This invention relates to the fields of energy management, intelligent control, and model predictive control technology for aviation hybrid power systems, specifically to a real-time energy management method for hybrid aircraft based on CNN-BiLSTM prediction and PPO hot-start MPC. Background Technology

[0002] Hybrid-electric aircraft employ a series electrical architecture consisting of a turboshaft engine and a generator coupled with a power battery. With its advantages of low fuel consumption, long range, and low emissions, it has become a key research area in the field of aviation propulsion. During flight, the aircraft's power demand fluctuates drastically, and the onboard embedded controller has limited computing resources, placing extremely high demands on the real-time performance, stability, and safety of energy management strategies.

[0003] Patent CN120116917A discloses a reinforcement learning-based model predictive control energy management method for plug-in hybrid electric vehicles (PHEVs). This method first uses a least squares support vector machine-background algorithm (LSSVM-BP) to predict the future speed of the PHEV in the time domain. Then, it calculates the vehicle's power demand (SOC) in the predicted time domain based on a vehicle dynamics model and defines a reference SOC curve for that domain. Finally, it uses a Q-learning reinforcement learning algorithm to solve the rolling optimization process of the model predictive control, obtaining a sequence of control variables (battery power) for each stage in the predicted time domain. The first element of the control variable sequence is then output to the PHEV's powertrain. However, this method directly outputs the control variable sequence for the MPC using Q-learning, which, as a black-box model, lacks explicit handling of system hard constraints. Ultimately, the control variables are not filtered by physical constraints, posing a risk of constraint violation.

[0004] The invention patent with publication number CN112561169A discloses a "hybrid MPC energy management method for parallel hybrid electric vehicles". This scheme employs piecewise affine technology to construct an engine fuel consumption model and a battery SOC dynamic model based on a battery equivalent circuit model to model the transmission system and electric motor of the parallel hybrid electric vehicle. It uses a BP neural network to predict the vehicle speed in the prediction time domain, thereby obtaining the vehicle's power demand. The HYSDEL compiler integrates the engine PWA model and the battery SAS model into a hybrid logic dynamics (MLD) framework. The energy management problem is transformed into a mixed integer linear quadratic programming problem, and the optimal power allocation and shift control sequence are obtained by solving it. However, this scheme uses piecewise affine linearization for the engine fuel consumption model and a switching affine model for the battery model, simplifying the original nonlinear model and reducing model fidelity.

[0005] Among the current mainstream energy management methods, nonlinear model predictive control (NMPC) can accurately handle multiple system constraints and achieve multi-objective optimization, making it suitable for the complex operating conditions of hybrid-electric aircraft. However, traditional NMPC requires iteratively solving the nonlinear optimization problem at each sampling time, resulting in numerous iterations and long computation time, which is difficult to meet the requirements of real-time sampling control on airborne systems. The commonly used RTI-SQP approximate solution method in engineering accelerates the solution by simplifying the nonlinear model and shifting the optimal solution from the previous time step as the initial value for iteration. However, it suffers from drawbacks such as decreased model fidelity, reduced solution accuracy, and energy allocation deviating from the global optimum.

[0006] Existing reinforcement learning-based intelligent energy management methods are black-box models, unable to explicitly handle various physical constraints of engines, generators, and batteries. Under untrained conditions, they are prone to control failures and breaches of system constraints, failing to meet aviation application standards in terms of safety. Furthermore, most existing hybrid power energy management technologies are designed for ground-based hybrid vehicles, whose powertrains involve mechanical coupling and shifting decisions. This is fundamentally different from the purely electric coupling and mechanically independent series architecture of hybrid aircraft, making direct application impossible. In addition, traditional energy management strategies focus solely on fuel economy, neglecting the long-term degradation costs of batteries, leading to excessively rapid battery lifespan decline and increased aircraft maintenance costs.

[0007] More importantly, existing technical solutions that integrate reinforcement learning and model predictive control only verify their effectiveness through offline software simulations, without conducting real-time hardware-in-the-loop testing, thus failing to demonstrate their engineering feasibility for airborne deployment. Given these shortcomings, there is an urgent need for a dedicated real-time energy management method for hybrid-electric aircraft that balances real-time performance, control accuracy, constraint safety, fuel economy, and battery life. Summary of the Invention

[0008] The purpose of this invention is to address the challenges of rapidly changing power demands during flight in hybrid-electric aircraft, limited onboard computing resources, and the difficulty of meeting real-time energy management requirements with existing nonlinear model predictive control (NMPC) online solution speeds. Current engineering practices often employ Real-Time Iterative Sequential Quadratic Programming (RTI-SQP) methods to approximate NMPC solutions; however, these methods typically rely on model simplification and local iterations, resulting in decreased solution accuracy, sensitivity to initial conditions, and limited energy allocation effectiveness. Therefore, this invention provides a real-time energy management method based on CNN-BiLSTM prediction and PPO hot-start model predictive control. This method improves the online solution speed of model predictive control while satisfying system constraints, and also considers fuel economy, battery state-of-charge stability, and battery life.

[0009] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0010] A real-time energy management method based on CNN-BiLSTM prediction and PPO hot-start model predictive control is applied to a hybrid-electric aircraft. The real-time energy management method includes the following steps:

[0011] Step 1, Model the hybrid-electric aircraft: Establish the turboshaft engine model, generator and motor controller model, power battery model, and power battery degradation model of the hybrid-electric aircraft respectively;

[0012] Step 2, Construct the Model Predictive Control (MPC) framework: Determine the objective function and physical constraints of the energy management method of model predictive control.

[0013] Step 3: Offline training of CNN-BiLSTM predictor to predict the power demand of hybrid-electric aircraft: The CNN-BiLSTM network is trained offline using two sets of flight power demand data (i.e., the first set of flight data and the second set of flight data) collected through real flight missions. The bidirectional temporal dependency features of the flight power demand sequence are mined to obtain a CNN-BiLSTM power demand predictor that can predict the future continuous temporal power demand sequence of the aircraft.

[0014] Step 4, Offline training of the Proximal Policy Optimization (PPO) network: Using two sets of flight demand power data collected through real flight missions, input the hybrid-electric aircraft simulation model built in Step 1 to train the PPO network, learn the optimal power allocation prior policy of the hybrid-electric aircraft, and obtain a fully converged PPO policy network.

[0015] Step 5, Online PPO-MPC Energy Management: The proposed real-time energy management method based on CNN-BiLSTM prediction and PPO hot-start model predictive control is validated on a third set of flight data collected through real flight missions. The trained CNN-BiLSTM demand power predictor is used to predict the power demand of the hybrid aircraft in real time, and the trained PPO policy network is used to provide an iterative initial value sequence for the SQP solution of MPC, thereby accelerating the MPC solution speed and realizing real-time energy management of the hybrid aircraft.

[0016] The PPO policy network is only used to provide high-quality initial values ​​for SQP optimization iteration and accelerate algorithm convergence. It does not directly output whole-machine control commands. The final power distribution control quantity of the hybrid aircraft is obtained by solving the MPC optimization that satisfies all physical hard constraints.

[0017] The beneficial effects of this invention are:

[0018] (1) Significantly improved real-time performance: This invention utilizes an offline-trained PPO network to provide a warm-start iterative initial value sequence for the SQP solver of NMPC. The average computation time per control step is reduced from 0.3761s in traditional MPC to 0.0653s, improving computational efficiency by 82.64%, which meets the requirements. Real-time control requirements with a sampling period of 0.2s can be deployed on an embedded controller for hybrid-electric aircraft.

[0019] (2) Excellent fuel economy: In the third set of real flight data tests, the equivalent fuel consumption of the present invention is 80.973 kg, which is 99.46% of the global optimal value of dynamic programming (80.537 kg), 1.41% lower than the traditional MPC (82.13 kg) and 7.24% lower than the rule-based energy management method (87.294 kg).

[0020] (3) Balancing battery life: This invention considers fuel consumption cost, battery degradation cost and SOC maintenance penalty in the MPC objective function. The battery SOH degradation is the smallest among all real-time implementable strategies, second only to dynamic programming methods.

[0021] (4) Higher accuracy in demand power prediction: The proposed CNN-BiLSTM demand power predictor has an RMSE of 5.1517kW, a MAE of 2.7672kW and a MAPE of 2.76% under the condition of 20 steps in the prediction time domain. All of these are better than similar prediction models such as LSTM, CNN-LSTM, single-layer BiLSTM, double-layer BiLSTM, CNN-LSTM-Attention and CNN-BiLSTM-Attention.

[0022] (5) Balancing intelligent optimization and hard constraint guarantee: This invention uses the PPO network only as a provider of initial values ​​for hot start, and the final control decision is still generated by the MPC optimization solution that satisfies all hard constraints, thus avoiding the constraint violation problem that may be caused by the black box characteristics of pure reinforcement learning methods.

[0023] (6) Engineering feasibility has been verified: Hardware-in-the-loop test results based on Speedgoat real-time target machine show that HIL and offline simulation are almost completely consistent in SOC, SOH and power distribution, with very small relative deviation, proving that this method has the ability to be deployed in actual engineering.

[0024] Compared with the prior art, the advantages and technical effects of the present invention are as follows:

[0025] 1. This invention proposes a real-time energy management method for hybrid-electric aircraft based on CNN-BiLSTM prediction and PPO hot-start MPC. Unlike most existing MPC energy management methods for hybrid electric vehicles, this invention performs specialized modeling and energy management design for hybrid-electric aircraft, covering the turboshaft engine-generator system, power battery system, electric propulsion system, and battery degradation dynamics. It also uses an economical MPC framework to jointly optimize fuel consumption, battery health, and SOC adjustment, achieving 99.46% of the equivalent fuel consumption as globally optimal through dynamic programming.

[0026] 2. This invention innovatively uses the PPO agent as the initial value provider for the SQP solver of the NMPC, rather than directly providing control decisions. The offline-trained PPO network learns a near-globally optimal power allocation strategy from historical flight data, and its output serves as the initial iteration sequence for the SQP solver, enabling the SQP to converge to the optimal solution with only a few iterations. The average computation time per control step is reduced from 0.3761s in traditional MPC to 0.0653s, improving computational efficiency by 82.64%, which meets the requirements. Real-time control requirements with a sampling period of 0.2s can be implemented on an airborne embedded controller.

[0027] 3. Compared with RTI-SQP, this invention achieves real-time solution without simplifying the model or sacrificing solution accuracy. The traditional RTI-SQP method accelerates the solution by simplifying the model and shifting the previous optimal solution as the initial iteration sequence, inevitably at the cost of solution accuracy. This invention retains the complete nonlinear prediction model and economic objective function, and accelerates the solution only by providing higher quality hot-start initial values ​​through PPO. The equivalent fuel consumption is reduced by 1.41% compared to the traditional MPC and by 7.24% compared to rule-based energy management methods.

[0028] 4. This invention simultaneously considers fuel consumption cost, battery degradation cost, and SOC maintenance penalty in the objective function of MPC, achieving synergistic optimization of fuel economy and battery life. Simulation results show that the battery SOH degradation is the smallest among all real-time implementable strategies, second only to dynamic programming methods, demonstrating the ability of this invention to protect the long-term health of the battery.

[0029] 5. While ensuring the global learning advantage of the intelligent algorithm, this invention still relies on MPC to make the final control decision. It can explicitly handle multiple physical hard constraints such as engine power, engine speed, engine torque, generator power, generator torque, battery SOC, SOH, battery current and battery power, avoiding the constraint violation or performance degradation problems that may be caused by the black box characteristics of pure reinforcement learning methods, and taking into account both intelligent optimization and safety constraint protection.

[0030] 6. The engineering feasibility of this invention was verified on the Speedgoat real-time target machine hardware-in-the-loop (HIL) test platform. The SOC, SOH, and power distribution curves of the HIL experiment almost completely overlapped with the offline simulation results, with a relative deviation of only 0.008267% for terminal SOC and only 0.0000004% for terminal SOH, indicating that this method can be reliably executed in a real-time control environment and has the capability for practical engineering deployment. Attached Figure Description

[0031] Figure 1 A schematic diagram of the hybrid-powered aircraft used in the embodiment;

[0032] Figure 2 A flowchart of the real-time energy management method based on CNN-BiLSTM prediction and PPO hot-start model predictive control used in the embodiment;

[0033] Figure 3 The BSFC characteristic diagram of the turboshaft engine used in the embodiment at different speeds and power;

[0034] Figure 4 A MAP diagram showing the efficiency of the generator used in the example;

[0035] Figure 5 Efficiency MAP of the motor controller used in the embodiment;

[0036] Figure 6 This is the average reward convergence curve of the PPO network over 800 training epochs.

[0037] The specific explanations of the reference numerals in the attached diagram are as follows: 1. Flight controller; 2. Coordination controller; 3. Turboshaft engine; 4. Turboshaft engine controller; 5. Coupling; 6. Generator; 7. Motor controller; 8. Power battery; 9. Bidirectional DC / DC converter; 10. Battery management system; 11. Propeller; 12. Propeller motor; 13. Electronic speed controller. Detailed Implementation

[0038] The following embodiments, in conjunction with the accompanying drawings, will clearly and completely describe the technical solutions of the present invention. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] This invention proposes a real-time energy management method for hybrid-electric aircraft based on CNN-BiLSTM prediction and PPO hot-start model predictive control. A schematic diagram of the structure of this hybrid-electric aircraft is shown below. Figure 1As shown, it mainly includes flight controller 1, coordination controller 2, turboshaft engine 3, turboshaft engine controller 4, coupling 5, generator 6, motor controller 7, power battery 8, bidirectional DC / DC converter 9, battery management system 10, propeller 11, propeller motor 12, and electronic speed controller 13. Figure 1 As shown, the output shaft of the turboshaft engine 3 is connected to the rotating shaft of the generator 6 via a coupling 5. After the turboshaft engine 3 starts, it drives the generator 6 to rotate at high speed through the coupling 5. The three-phase AC power generated by the generator 6 is rectified by the motor controller 7 and output to the DC bus. The power battery 8 is connected to the same DC bus via a bidirectional DC / DC 9. The battery management system 10 can control the power battery 8 to switch between discharge mode and charging mode. The ESC 13 obtains DC power from the DC bus and inverts it into AC power, while driving the propeller motor 12 and controlling its speed. The propeller motor 12 drives the propeller 11 to rotate, providing power for the flight of the hybrid aircraft. The coordination controller is the control core of the hybrid aircraft. It works in conjunction with the flight control system 1 to coordinate and manage the turboshaft engine 3, generator 6, power battery 8 and other components in a unified manner according to the flight conditions and the power demand of the propeller motor 12, and sends the instructions of the real-time energy management method of the present invention to each component.

[0040] The overall flowchart of the real-time energy management method based on CNN-BiLSTM prediction and PPO hot-start model predictive control proposed in this invention is as follows: Figure 2 As shown, it includes two parts: an offline training phase and an online control phase.

[0041] The offline training phase includes two parallel training processes:

[0042] (1) Training of CNN-BiLSTM demand power predictor: The CNN-BiLSTM network was trained using the first and second sets of flight demand power data collected through real flight missions, so that it could accurately predict the demand power sequence of the next 20 steps based on the historical demand power of the previous 20 steps.

[0043] (2) PPO agent offline training: In the built Simulink simulation environment of hybrid aircraft, the PPO agent learns a near-global optimal energy allocation strategy through repeated interaction with the environment, and is trained for a total of 800 rounds.

[0044] Online control phase: at each sampling time ( =0.2s) Execute in a loop according to the following process:

[0045] (1) System status acquisition: Acquire the current status of the hybrid-electric aircraft system, including the current power demand. ,Battery ,Battery RC polarization voltage and the generator power at the previous moment. and battery power ;

[0046] (2) Demand power prediction: Input the historical demand power of the previous 20 steps into the trained CNN-BiLSTM predictor to predict the demand power sequence of the next 20 steps;

[0047] (3) Generation of initial value sequence for hot start iteration: The current system state parameters obtained in step (1) and the CNN-BiLSTM prediction sequence obtained in step (2) are used to construct a 14-dimensional observation vector (the composition is completely consistent with the training stage, including the 5 predicted power values ​​of the CNN-BiLSTM prediction sequence at steps 1, 5, 10, 15 and 20, as well as 9 system states and derived quantities such as current demand power, SOC, SOC tracking error, SOH, generator power, battery power, charge and discharge rate, power supply margin and generator power change rate). The vector is input into the trained PPO agent. The PPO agent rolls forward inference within Nc steps in the control time domain and outputs the generator power increment reference sequence. , serving as the initial value sequence for the hot-start iteration of the SQP solver in NMPC;

[0048] (4) NMPC optimization solution: As the initial iteration sequence of the SQP solver, it solves the constrained nonlinear optimization problem in the prediction time domain. Since the initial value is close to the optimal solution, the SQP solver only needs a small number of iterations to converge, which significantly reduces the online computation time.

[0049] (5) Control execution: The first control action of the obtained optimal control sequence. The physical systems applied to hybrid-electric aircraft;

[0050] (6) Rolling optimization: Update system status, time domain rolls to At time +1, repeat steps (1) to (5) until the flight mission ends.

[0051] Through the above design, this invention significantly reduces online computation time while preserving the complete nonlinear prediction model and solution accuracy, thus meeting the real-time requirements of the airborne controller for hybrid-electric aircraft. The specific implementation steps of the method are as follows:

[0052] Step 1: Modeling of the hybrid-electric aircraft, including: modeling of the turboshaft engine of the hybrid-electric aircraft, modeling of the generator and motor controller of the hybrid-electric aircraft, modeling of the power battery of the hybrid-electric aircraft, and modeling of the power battery degradation model of the hybrid-electric aircraft.

[0053] (1) Modeling of the turboshaft engine of the hybrid-electric aircraft: The turboshaft engine is the main energy source of the hybrid-electric aircraft. In this embodiment, the total fuel consumption of the hybrid-electric aircraft is obtained by integrating the instantaneous fuel rate over the flight time.

[0054] With the output power of the turboshaft engine and the output shaft speed of the turboshaft engine As input, performance data covering the speed range of 10,000–21,000 rpm and power range of 30–250 kW were collected through bench tests. A two-dimensional lookup table model of brake ratio fuel consumption (BSFC) was constructed using cubic spline interpolation to obtain the instantaneous fuel consumption rate. The instantaneous fuel flow rate was modeled as a function of the turboshaft engine's output power and the output shaft speed of the turboshaft engine. In the formula, This refers to the output power of the turboshaft engine. The output shaft speed of the turboshaft engine is used; therefore, the fuel consumption throughout the flight is calculated integrally. In the formula, This refers to the total fuel consumption of a hybrid-electric aircraft during a single flight. Instantaneous fuel flow rate; The total flight time of the hybrid-electric aircraft during this flight is shown in the following figures. The BSFC characteristics of the turboshaft engine at different speeds and power outputs are also shown in the following figures. Figure 3 As shown, brake specific fuel consumption (BSFC) can be defined as: .

[0055] Considering that this invention studies the steady-state fuel consumption characteristics of a turboshaft engine, transient dynamic response and temperature change effects are ignored, and a quasi-steady-state model based on experimental data is adopted. Performance data of the turboshaft engine were collected through bench tests within the speed range of 10000–21000 rpm and output power range of 30–250 kW, and a two-dimensional lookup table model of the BSFC (Balanced Fiber Fuel Consumption) was constructed using cubic spline interpolation. The obtained BSFC characteristic diagrams of the turboshaft engine at different speeds and power are shown below. Figure 3 As shown. By Figure 3 It is evident that BSFC decreases significantly with increasing engine power, while engine speed has a relatively small impact on BSFC, with the lowest BSFC mainly occurring in the high-power region. Under low-power conditions, BSFC increases sharply across the entire speed range, resulting in poor fuel economy. The strong nonlinear characteristics of BSFC reveal a complex coupling relationship between turboshaft engine operating conditions and fuel consumption. This indicates that guiding the turboshaft engine to operate in the high-power range through a reasonable energy management strategy can significantly reduce overall fuel consumption. Therefore, this provides an optimization basis for incorporating fuel consumption costs into the MPC (Multi-Process Control) economic objective function and prioritizing the operation of the turboshaft engine-generator system in the high-efficiency operating range within the energy management strategy of this invention.

[0056] (2) Modeling of the generator and motor controller of the hybrid-electric aircraft: The generator shaft is connected to the output shaft of the turboshaft engine via a coupling, therefore the generator speed is equal to the turboshaft engine speed, i.e. In the formula, The generator's rotational speed is given by the input power and its rotational speed. The generator's output torque is calculated from the input power and rotational speed. In the formula, This represents the generator torque. Generator efficiency. It is a function of speed and torque: Generator efficiency Based on bench test data and bilinear interpolation, a system was constructed as follows: Figure 4 The diagram shows the efficiency MAP of the generator.

[0057] The three-phase AC power generated by the generator is rectified into DC power by the motor controller and then delivered to the DC bus. The efficiency of the motor controller is... It is also a function of speed and torque: Motor controller efficiency Similarly, based on bench test data and bilinear interpolation, a system was constructed as follows: Figure 5 The efficiency MAP diagram of the motor controller is shown.

[0058] The combined efficiency of the generator and motor controller is The generator output power to the bus is .

[0059] From the generator efficiency MAP ( Figure 4 ) and motor controller efficiency MAP ( Figure 5 As can be seen, both the generator and the motor controller exhibit superior energy conversion performance in the medium-to-high speed range, and their overall efficiency shows a clear condition-dependent characteristic. This indicates that the energy management strategy should prioritize maintaining the turboshaft engine-generator system's operating point within the high-efficiency range of medium-to-high speed and medium-to-high torque to improve the overall conversion efficiency from the turboshaft engine's mechanical energy to the DC bus's electrical energy. This provides a basis for optimizing the selection of the turboshaft engine-generator system's operating point in the MPC energy management strategy of this invention.

[0060] (3) Modeling of the power battery of the hybrid aircraft: In order to balance computational efficiency and reasonable representation of battery dynamics, this invention adopts a first-order RC (1-RC) equivalent circuit model to describe battery dynamics, and the model parameters are as follows: , , The values ​​were obtained by fitting publicly available lithium iron phosphate (LFP) battery cycle datasets. To prevent overcharging and over-discharging, the SOC of the battery in this embodiment was set between [0.4, 0.8]. The nominal capacity of the LFP single cell used in this embodiment is 1.1 Ah; the battery pack adopts a multi-series and multi-parallel topology. After equivalently merging the capacity contribution of the parallel branches into the battery model, the battery pack is modeled as an equivalent series structure, with an equivalent nominal capacity of 61.6 Ah. The equivalent power inside the battery is defined. for:

[0061]

[0062] In the formula, This indicates the battery's output / input power, where This indicates that the battery is discharging. This indicates that the battery is charging. For the battery's discharge efficiency, The charging efficiency of the battery.

[0063] Based on the 1-RC equivalent circuit model, the dynamic equation of the system is:

[0064]

[0065] In the formula, This is the battery open-circuit voltage; This refers to the terminal voltage of the power battery. It is the RC polarization voltage; The internal resistance of the battery is ohms. This is the internal resistance to polarization; Polarizing capacitor; This represents the battery current.

[0066] According to the terminal voltage expression Relationship with power Solving the system of equations simultaneously yields information about the battery current. The physically feasible solution to the quadratic equation is:

[0067]

[0068] In the formula, This refers to the terminal voltage of the power battery. This is the battery open-circuit voltage. RC polarization voltage Battery current, The internal resistance of the battery is in ohms. This refers to the internal input / output power of the battery.

[0069] Next, considering the capacity degradation caused by battery aging, the battery SOC is updated using the ampere-hour integration method:

[0070]

[0071] In the formula, This represents the initial usable capacity of the battery. For battery capacity loss, This is the time variable for integration.

[0072] (4) Modeling of the degradation model of the power battery of hybrid aircraft:

[0073] To balance the computability of energy management optimization with the interpretability of battery health evolution, this invention employs a semi-empirical degradation model based on ampere-hour (Ah) throughput to characterize the capacity decay of LFP batteries.

[0074] Define cumulative ampere-hour throughput For degenerate state variables, update them discretely as follows:

[0075]

[0076] In the formula, For discrete time steps, for The power battery current is always present. This represents the discrete sampling time.

[0077] Battery capacity loss was calculated using a power-law-Arrhenius temperature factor model. :

[0078]

[0079] In the formula, It is the exponential pre-degradation coefficient identified through the LFP dataset; It is the activation energy; This is the universal gas constant. The model reference thermodynamic temperature is T, and the battery's real-time operating thermodynamic temperature is T. For the first The cumulative ampere-hour throughput of the battery at any given time. It is a power-law exponent characterizing the nonlinear relationship between battery capacity loss and cumulative ampere-hour throughput.

[0080] As the battery ages, its usable capacity decreases proportionally with its state of health (SOH). The battery's state of health (SOH) and current usable capacity are defined as follows:

[0081]

[0082] Based on this, considering the capacity decay caused by battery aging, the discrete update form of battery SOC is as follows:

[0083]

[0084] In the formula, This represents the real-time available battery capacity after capacity loss correction.

[0085] To characterize the power capacity degradation caused by battery aging, this invention introduces an empirical model that correlates internal resistance with SOH evolution, and the ohmic internal resistance and polarization internal resistance are corrected as follows:

[0086]

[0087] In the formula, The reference internal resistance under healthy conditions, used as a function of SOC, is obtained through interpolation from the quasi-steady-state discharge curves and internal resistance measurements in the LFP dataset; the internal resistance growth coefficient... Identified by fitting battery cycle test data from the LFP dataset; reference internal resistance It was obtained by interpolation of quasi-static discharge data.

[0088] Step 2: Construct the Model Predictive Control (MPC) framework. Step 2 involves constructing the MPC framework and determining the objective function and physical constraints of the MPC energy management method.

[0089] At each sampling time A constrained nonlinear optimization problem is established with the objective function being the sum of fuel cost, battery degradation cost, and SOC maintenance penalty term at each time point Np in the prediction time domain. This optimization problem is solved in the prediction time domain to obtain the optimal control sequence. Only the first control action of this sequence is applied to the system, and the remaining control inputs are discarded. At the next sampling time, the system state is updated, and the above optimization process is repeated, thus forming a rolling time-domain optimization strategy to ensure real-time control performance and dynamic adaptability.

[0090] The predictive model for hybrid-electric aircraft can be expressed as:

[0091]

[0092] In the formula, For state variables, To control the quantity, For output quantity, This refers to the perturbation quantity. Specifically: the state variables of the prediction model. Control quantity Output Disturbance This represents exogenous power demand disturbances, i.e., the actual future propulsion power demand that is currently unknown. The generator's output power is updated through an incremental control strategy. .

[0093] To reduce fuel consumption, mitigate battery degradation, and maintain stable battery state of charge (SOC) in hybrid-electric aircraft, an economical MPC objective function is constructed. This function comprehensively considers fuel consumption costs, battery degradation costs, and SOC maintenance penalties to achieve a balance between fuel economy and battery life. The MPC objective function is expressed as follows:

[0094]

[0095] fuel costs Battery degradation cost Battery SOC maintenance penalty .

[0096] The main constraints during MPC solving include:

[0097]

[0098] In the formula, for Generator output power at all times for The output power of the turboshaft engine at all times. for Constantly monitor the rotational speeds of the engine and generator shafts. for The constant output torque of the turboshaft engine for At all times, the generator output torque, for The state of charge of the power battery at all times. for Always monitor the health status of the power battery. for The power battery current is always present. for The power battery output / input power at all times, This represents the discrete sampling time.

[0099] Step 3: Train the CNN-BiLSTM predictor offline to predict the power demand of the hybrid aircraft.

[0100] This invention uses power demand data collected from real-world flight missions of hybrid-powered aircraft to train a CNN-BiLSTM. Specifically, a first set of flight data and a second set of flight data are used as the training set.

[0101] The CNN-BiLSTM network consists of a one-dimensional convolutional layer, a bidirectional long short-term memory (BiLSTM) layer, and a fully connected layer connected in series.

[0102] (1) One-dimensional convolutional layer (Conv1D) for historical power sequences of length L Perform local feature extraction: The "same" padding method maintains the sequence length before and after convolution, and a batch normalization layer is added after the convolutional layer to accelerate network convergence and improve the model's generalization ability; where, This represents the local temporal feature sequence output by the convolutional layer; This represents the historical demand power sequence as input; The kernel weight matrix; The bias vector of the convolutional layer; To modify the activation function of the linear unit.

[0103] (2) The BiLSTM layer contains two sub-layers: a forward LSTM and a backward LSTM. These sub-layers process the feature sequences output by the convolutional layers in forward and reverse temporal order, respectively, thereby comprehensively capturing the temporal dependencies in both the forward and backward directions. At the final moment, the hidden states from both directions are concatenated into a structure with dimension [missing value]. The feature vectors capture the temporal dependencies of the sequence's history and future direction, followed by a Dropout layer to prevent overfitting.

[0104] Both its forward and backward units employ the standard LSTM gating mechanism, with the core calculations as follows:

[0105] Forgotten Gate:

[0106] Input Gate:

[0107] Candidate state:

[0108] Status Update:

[0109] Output gate:

[0110] Hidden state:

[0111] In the formula, Indicates the first The feature vectors input to the LSTM unit at each time step; , , These represent the outputs of the forget gate, input gate, and output gate, respectively. Indicates the state of candidate cells; and These represent the cell states at the current and previous moments, respectively. and These represent the hidden states at the current and previous time points, respectively. , , , These represent the input weight matrices corresponding to the forget gate, input gate, candidate cell state, and output gate, respectively. , , , These represent the cyclic weight matrices of the corresponding gate structures; , , , These represent the bias vectors of the corresponding gate structures; This represents the Sigmoid activation function; Represents the hyperbolic tangent activation function; This represents element-wise product.

[0112] Hide the positive state and reverse hidden state The final output is obtained by splicing. A Dropout layer is then added to prevent overfitting.

[0113] In the formula, Indicates the forward LSTM at the th The hidden state at each time step; Indicates the inverse LSTM at the 1st The hidden state at each time step; This indicates the bidirectional hidden state after splicing.

[0114] (3) The fully connected layer maps the BiLSTM output to the prediction time domain length, generating the prediction power sequence for the next Np steps: In the formula, This represents the power demand forecast sequence within the future forecast time domain; This represents the weight matrix of the fully connected layer; This represents the concatenated hidden state output by the BiLSTM at the last time step; This represents the bias term for the fully connected layer.

[0115] Step 4: Train the Proximal Policy Optimization (PPO) network offline.

[0116] The PPO network adopts an Actor-Critic architecture: the policy network (Actor) takes a 14-dimensional normalized state vector as input, passes through two 128-node hidden layers (ReLU activation), and outputs the mean and standard deviation parameters of a Gaussian distribution, from which continuous actions are sampled; the value network (Critic) shares the same hidden layer structure and outputs a scalar estimate of the expected reward of the current state; the parameters of the two networks are optimized independently.

[0117] The 14-dimensional observation vector contains four types of information: power prediction components (predicted power values ​​at steps 1, 5, 10, 15, and 20 of the CNN-BiLSTM prediction sequence). Current system status (current power demand) Battery State of Charge (SOC) and SOC tracking error Battery health status (SOH) and power distribution status (generator power) Battery power ) and dynamic change characteristics (charge and discharge rate) Power supply margin Generator power change rate The action output, normalized to [-1, 1], is mapped to the generator power increment. .

[0118] PPO uses a trimming alternative to the objective function: In the formula, This represents the pruning alternative objective function of the PPO policy network; These are the learnable parameters for the policy network (Actor); This represents an empirical expectation for the sampling time step t; The probability ratio; For generalized advantage estimation; This means restricting the probability ratio to an interval. Inside; These are the trimming parameters.

[0119] Value network loss is calculated using mean squared error: In the formula, The value network loss function; These are the learnable parameters of the Critic value network. For parameters The value network for the current state Expected return estimate; Monte Carlo estimates of cumulative returns for discounts.

[0120] The overall training objective is: In the formula, This represents the overall training objective of the PPO network; The loss coefficient of the value function; This is the entropy regularization coefficient; Representation Strategy The entropy is used to encourage strategies to explore moderately and prevent premature convergence to local optima.

[0121] The reward function is designed as follows: J is the MPC objective function (which includes three terms: fuel consumption cost, battery degradation cost, and SOC maintenance penalty). To maintain a fixed positive bias reward and prevent the agent from degenerating into a low-response strategy during training, ensuring active power regulation exploration throughout, all training rounds are started with a fixed initial SOC of 0.6, for a total of 800 rounds. Figure 6 As shown, the reward curve enters a stable convergence plateau (5800-5900 range) after about 100 rounds, indicating that the policy has fully converged.

[0122] Step 5: Online PPO-MPC Energy Management. The proposed real-time energy management method based on CNN-BiLSTM prediction and PPO hot-start model predictive control was validated using a third set of flight data collected from real flight missions. A trained CNN-BiLSTM predictor was used to predict the power demand of the hybrid-electric aircraft in real time, and a trained PPO network was used to provide an iterative initial value sequence for the SQP solution of the MPC, accelerating the MPC solution speed and thus achieving real-time energy management of the hybrid-electric aircraft.

[0123] At each sampling time Perform online energy management as follows:

[0124] (1) Status acquisition: Obtain the current system status, including power demand. Battery SOC ( ), battery SOH( ) RC polarization voltage and the generator power at the previous moment. and battery power .

[0125] (2) Power prediction: Input the historical power demand of the previous 20 steps into the trained CNN-BiLSTM predictor to predict the power demand sequence of the next 20 steps. .

[0126] (3) The trained PPO network provides an iterative initial value sequence for the SQP solution of MPC: Construct a 14-dimensional observation vector from the current system state parameters and the CNN-BiLSTM prediction sequence, input it into the trained PPO network, and perform a rolling forward inference within Nc steps in the control time domain to generate a generator power increment reference sequence. This step only involves forward inference of the neural network, and the computational cost is relatively small.

[0127] (4) SQP solution: As the initial iteration sequence for the SQP solver, it solves constrained nonlinear optimization problems in the prediction time domain. It outperforms the initial iteration sequence provided by the PPO network, which is already close to the optimal solution. The SQP solver only requires a small number of iterations to converge, significantly reducing computation time. It is worth noting that... It serves only as the initial iteration sequence for SQP solution; the final control decision is still generated by MPC optimization that satisfies all constraints.

[0128] (5) Control execution: Apply the first control action of the MPC optimal control sequence to the physical system and update the generator output power command and the power battery output power command.

[0129] (6) Rolling optimization: Repeat steps (1) to (5) at the next sampling time to achieve real-time closed-loop control.

[0130] To verify the prediction accuracy, energy optimization effect, and engineering practicality of the real-time energy management method for hybrid-electric aircraft based on CNN-BiLSTM prediction and PPO hot-start MPC proposed in this invention, this invention relies on three sets of measured flight data collected from actual test flights of hybrid-electric aircraft to conduct multi-model power prediction simulation comparison experiments, multi-algorithm energy management benchmarking simulation experiments, and Speedgoat hardware-in-the-loop (HIL) semi-physical experiments in sequence. Among them, the first and second sets of measured demand power data from test flights were used to complete the offline training of the CNN-BiLSTM prediction network and the PPO policy network, and the third set of complete test flight data with a duration of 2145.2s was selected as the test condition for all experiments.

[0131] (1) Demand Power Prediction Experiment: Keeping the input backtracking window uniformly at 20 steps (corresponding to a duration of 4 seconds) and the prediction output length uniformly at the demand power of the next 20 steps, the CNN-BiLSTM model of this invention was compared horizontally with similar prediction models (LSTM, CNN-LSTM, single-layer BiLSTM, double-layer BiLSTM, CNN-LSTM-Attention, CNN-BiLSTM-Attention) on the same dataset. The experimental results show that the RMSE (5.1517kW), MAE (2.7672kW), and MAPE (2.76%) of the CNN-BiLSTM model proposed in this invention are all lower than those of other models. The optimal performance is achieved because the convolutional layer is responsible for local short-range feature extraction, and BiLSTM comprehensively models the temporal dependencies of the sequence through a bidirectional gating mechanism. The two complement each other to enhance the prediction ability of complex power sequences. At the same time, the prediction performance of CNN-LSTM-Attention and CNN-BiLSTM-Attention is weaker than that of the proposed CNN-BiLSTM model. This is because the backtracking window used for prediction is relatively short (20 steps, corresponding to 4 seconds), which limits the long-range contextual information that the attention layer can obtain. Therefore, the attention mechanism cannot play its expected advantage and instead introduces additional trainable parameters, increasing the risk of overfitting the model on a limited training set.

[0132] (2) Energy Management Simulation Comparison Test: Three classic energy management algorithms—dynamic programming energy management method, traditional MPC energy management method, and rule-based energy management method—were selected as comparison schemes. Full-time domain simulation calculations were performed under the same 2145.2s flight test condition. The test results show that the method proposed in this invention, on the test set (the third set of data from a real flight mission of a hybrid-electric aircraft, with a flight duration of approximately 2145.2s), achieved a fuel consumption of 80.973kg, reaching 99.46% of the performance of the dynamic programming energy management method. (The fuel consumption of the dynamic programming energy management method is 80.537 kg), which is 1.41% lower than the traditional MPC energy management method (82.13 kg) and 7.24% lower than the rule-based energy management method (87.294 kg). The average computation time per control step is reduced from 0.3761 s in the traditional MPC energy management method to 0.0653 s, improving computational efficiency by 82.64%, meeting the requirements. =0.2s real-time control requirement; the battery SOH degradation is the smallest among all real-time implementable strategies, second only to dynamic programming energy management methods.

[0133] (3) Speedgoat Hardware-in-the-Loop (HIL) Semi-Physical Test: A semi-physical airborne power system test platform was built based on the Speedgoat real-time target aircraft. The control algorithm of this invention was written into the embedded controller. Closed-loop tests were carried out using the third set of measured flight data as the operating condition input to further verify the engineering feasibility of this method. The HIL test results were highly consistent with the offline simulation, proving that this method can be stably installed on airborne hardware and has the capability for practical engineering deployment.

[0134] This invention addresses the problems of rapidly changing power demand during flight in hybrid-electric aircraft, limited onboard computing resources, and the difficulty in meeting real-time control requirements with the online solution speed of existing nonlinear model predictive control. The invention solves the following technical problems:

[0135] 1. Existing nonlinear model predictive control (NMPC) methods require solving a constrained nonlinear optimization problem at each sampling time, resulting in numerous iterations and long computation times, which is difficult to meet the real-time requirements of airborne controllers for hybrid-electric aircraft. To address this problem, this invention utilizes an offline-trained proximal policy optimization (PPO) network to provide a high-quality, warm-start iterative initial value sequence for the NMPC's sequential quadratic programming (SQP) solver. Guided by this initial value, the SQP solver converges after a few iterations, reducing the average computation time per control step from 0.3761s in traditional MPC to 0.0653s, improving computational efficiency by 82.64%, thus meeting the requirements. Real-time control requirements with a sampling period of 0.2s.

[0136] 2. In existing engineering practices, the Real-Time Iterative Sequential Quadratic Programming (RTI-SQP) method is often used to approximate the solution of Non-Mean-Performance Computation (NMPC). This method accelerates the solution by simplifying the prediction model and using the optimal solution from the previous time step as the initial value sequence for the current time step. However, this method suffers from problems such as decreased solution accuracy, sensitivity to initial values, and deviation of energy allocation from the global optimum. To address these issues, this invention retains the complete nonlinear prediction model and the economical MPC objective function, and accelerates the solution by providing a hot-start iterative initial value sequence that is closer to the global optimum only through the PPO network. This significantly reduces the online computational burden without sacrificing model fidelity and constraint handling capabilities, achieving an equivalent fuel consumption of 99.46% of the global optimum of dynamic programming, which is 1.41% lower than traditional MPC.

[0137] 3. Existing energy management methods based on intelligent algorithms such as reinforcement learning are essentially black-box models, lacking the ability to explicitly handle the system's physical hard constraints such as engine power, engine speed, engine torque, generator power, generator torque, battery SOC, SOH, battery current, and battery power. Furthermore, in operating conditions where training data is not covered, there is a risk of control performance degradation or even exceeding the system's physical hard constraints. To address this issue, this invention uses the PPO network only as a hot-start initial value provider for the SQP solver of NMPC. The final control decision is still generated by the MPC optimization solution that satisfies all hard constraints, thus balancing the global learning advantages of intelligent algorithms with the safety constraint guarantee capabilities of MPC.

[0138] 4. Existing research on hybrid power system energy management mostly focuses on hybrid electric vehicles (HEVs), whose powertrain is characterized by the direct drive of wheels by an internal combustion engine via a mechanical transmission system. This involves mechanical coupling and shifting decision-making, making it difficult to directly transfer related modeling methods and control strategies to HEVs. Furthermore, there is a lack of research on MPC energy management methods for HEVs. To address this issue, this invention takes the series architecture of a HEV (turboshaft engine-generator system, power battery system, electric propulsion system, and coordinated control system) as its research object. Considering its purely electrical coupling and lack of a mechanical transmission chain, it models the turboshaft engine, generator and motor controller, power battery, and their degradation characteristics. Based on this, it constructs an MPC energy management framework for HEVs, filling a research gap in this field.

[0139] 5. Most existing research on energy management for hybrid electric vehicles focuses only on fuel economy and fails to explicitly consider the impact of battery degradation on long-term operating costs during energy management optimization. To address this issue, this invention incorporates fuel consumption costs, battery degradation costs, and battery SOC maintenance penalties into the objective function of MPC. This approach mitigates battery capacity decay while ensuring fuel economy, achieving the lowest SOC degradation among all real-time implementable strategies, second only to dynamic programming methods.

[0140] 6. Existing research methods integrating reinforcement learning and MPC are all verified only through software simulation, lacking hardware-in-the-loop (HIL) testing. Their real-time performance and engineering feasibility have not been fully proven. To address this issue, this invention verifies engineering feasibility on the Speedgoat real-time target machine hardware-in-the-loop testing platform. The HIL experimental results are highly consistent with offline simulations, demonstrating that this method has practical engineering deployment capabilities.

[0141] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A real-time energy management method based on CNN-BiLSTM prediction and PPO hot-start model predictive control, characterized in that, Includes the following steps: S1. Build a nonlinear simulation model of the hybrid-electric aircraft: Build models of the turboshaft engine, generator and motor controller, power battery, and power battery degradation to adapt to the actual flight conditions of the hybrid-electric aircraft; retain the inherent nonlinear characteristics of each component and build a simulation model of the hybrid-electric aircraft. S2. Construct a multi-objective, fully constrained nonlinear MPC optimization framework: Build a model predictive control MPC energy management framework, construct a multi-objective optimization objective function that simultaneously integrates fuel consumption cost, battery degradation cost, and SOC maintenance penalty term, match the full-dimensional physical constraints of the hybrid aircraft power system, and form a nonlinear rolling time domain optimization solution system with complete hard constraints; S3. Offline training of CNN-BiLSTM flight demand power predictor: Using multiple sets of flight demand power data collected from real flight missions, the CNN-BiLSTM network is trained offline to obtain a CNN-BiLSTM demand power predictor that can predict the future continuous time domain demand power sequence of the aircraft. S4. Offline training of PPO hot-start policy network: Based on the hybrid-electric aircraft simulation model built in step S1, combined with real flight condition data, the PPO reinforcement learning agent is trained in multiple rounds of offline environment interaction to learn the optimal power allocation prior policy of the hybrid-electric aircraft and obtain a fully converged PPO policy network. S5, PPO hot-start assisted SQP MPC online real-time energy management: In the online control phase, the real-time operating status of the hybrid aircraft system is periodically collected at a fixed sampling period; the historical demand power sequence is input into the trained CNN-BiLSTM demand power predictor to predict the future flight demand power sequence in real time; the real-time system status and the predicted power sequence are fused to construct a normalized multi-dimensional observation vector, which is input into the converged PPO policy network to infer and generate a continuous generator power increment reference sequence; this power increment reference sequence is used as the initial value for the hot-start iteration of the SQP algorithm in the MPC solver to solve the MPC nonlinear optimization problem with full physical constraints, and the optimal power allocation control command is executed through rolling time-domain optimization to achieve real-time energy management of the hybrid aircraft.

2. The real-time energy management method based on CNN-BiLSTM prediction and PPO hot-start model predictive control according to claim 1, characterized in that, In step S1, the turboshaft engine model is a quasi-steady-state fuel consumption model. Based on bench test data, a two-dimensional lookup table model of BSFC is constructed using cubic spline interpolation. The generator and motor controller models utilize the rigid coupling characteristics of the generator and turboshaft engine speeds, and combine bench test data and bilinear interpolation to construct efficiency MAP diagrams for the generator and motor controller respectively. The comprehensive energy conversion efficiency of the electromechanical system is obtained by multiplying the two MAP diagrams. The power battery model is a first-order RC equivalent circuit dynamic model. The real-time operating current of the battery is solved by simultaneously solving the power balance equation and the circuit dynamic equation, and the battery SOC is iteratively updated using the ampere-hour integral method. The power battery degradation model is a semi-empirical degradation model, using the battery's cumulative ampere-hour throughput as the degradation state variable, and matching the internal resistance aging correction model to characterize the battery's full life cycle degradation characteristics.

3. The real-time energy management method based on CNN-BiLSTM prediction and PPO hot-start model predictive control according to claim 2, characterized in that, include: Formula for calculating total fuel consumption for flight: In the formula, This refers to the total fuel consumption of a hybrid-electric aircraft during a single flight. Instantaneous fuel flow rate; This represents the total flight time of the hybrid-electric aircraft during this flight. Battery cumulative ampere-hour throughput For degenerate state variables, update them discretely as follows: In the formula, For discrete time steps, for The power battery current is always present. Represents discrete sampling time; Battery capacity loss was calculated using a power-law-Arrhenius temperature factor model. : In the formula, It is the exponential pre-degradation coefficient identified through the LFP dataset; It is the activation energy; This is the universal gas constant. The model reference thermodynamic temperature is T, and the battery's real-time operating thermodynamic temperature is T. The cumulative ampere-hour throughput of the battery at time k. It is a power-law exponent characterizing the nonlinear relationship between battery capacity loss and cumulative ampere-hour throughput.

4. The real-time energy management method based on CNN-BiLSTM prediction and PPO hot-start model predictive control according to claim 1, characterized in that, In step S2, the multi-objective optimization objective function weightedly integrates fuel consumption cost, battery degradation cost, and SOC deviation penalty term to achieve multi-objective coordinated optimization of fuel economy, battery life durability, and SOC steady-state maintenance. The expression of the multi-objective optimization objective function is as follows: The objective function of MPC is expressed as: fuel costs Battery degradation cost Battery SOC maintenance penalty ; The physical constraints include turboshaft engine speed constraints, turboshaft engine output power constraints, generator torque constraints, power battery SOC upper and lower limits constraints, battery charging and discharging current constraints, battery power constraints, and generator power increment variation constraints.

5. A real-time energy management method based on CNN-BiLSTM prediction and PPO hot-start model predictive control according to claim 1, characterized in that, In step S3, the CNN-BiLSTM demand power predictor is composed of a one-dimensional convolutional layer, a batch normalization layer, a bidirectional LSTM layer, a Dropout layer, and a fully connected layer connected in series. The one-dimensional convolutional layer extracts local short-term features of historical demand power, and the bidirectional LSTM layer mines the long-term and short-term dependencies of the power sequence from both forward and reverse time series. After offline training and convergence with real flight mission data, it stably outputs the flight demand power prediction sequence for the future continuous time domain.

6. The real-time energy management method based on CNN-BiLSTM prediction and PPO hot-start model predictive control according to claim 1, characterized in that, In step S4, the PPO policy network adopts an Actor-Critic dual-network independent architecture. The Actor policy network outputs continuous generator power increment actions, and the Critic value network evaluates the expected return of the current system state. The PPO policy network uses a custom reward function with a fixed positive bias, and the reward function formula is: In the formula, J is the objective function of MPC. To fix the positive bias reward and prevent the agent from falling into the degenerate behavior of low-response policies during training, the network iterative update is completed by using the MPC multi-objective optimization cost function as the reward benchmark and combining PPO pruning to replace the objective function. The network is trained through multiple rounds of offline environment interaction until the policy fully converges, and a power allocation prior policy close to the global optimum is obtained.

7. A real-time energy management method based on CNN-BiLSTM prediction and PPO hot-start model predictive control according to claim 6, characterized in that, The input to the PPO strategy network is a 14-dimensional normalized multidimensional observation vector, which includes: power components of key nodes at steps 1, 5, 10, 15, and 20 of the CNN-BiLSTM prediction sequence, real-time demand power, SOC, SOC tracking error, SOH, generator power, battery power, charge / discharge rate, power supply margin, generator power change rate, and system state components.

8. A real-time energy management method based on CNN-BiLSTM prediction and PPO hot-start model predictive control according to claim 1, characterized in that, In step S5, the generator power increment reference sequence output by the PPO strategy network provides the SQP algorithm with an iterative hot start initial value close to the optimal solution, thereby reducing the number of iteration steps of the SQP algorithm.

Images