A multi-modal fusion eVTOL energy consumption optimization system and dynamic modeling method

Abstract

The application discloses a multi-modal fusion eVTOL energy consumption optimization system and a dynamic modeling method, the system comprises an LSTM-PINN multi-modal space-time prediction module, an online learning module, an autonomous decision-making double-cycle DRL module and an energy-aware adaptive module; the long short-term memory network LSTM is used as the time sequence coding layer of the PINN, the online learning module comprises a mutation detection unit and an incremental learning unit, the autonomous decision-making double-cycle DRL module comprises an inner loop planning module and an outer loop strategy module; the multi-modal time sequence data stream is input into the LSTM-PINN multi-modal space-time prediction module, the LSTM time sequence coding layer decodes the multi-modal time sequence data stream to generate the initial condition loss item prediction value of the PINN loss function, the online learning module dynamically allocates the loss item weight of the PINN loss function, outputs the space-time prediction sequence, the autonomous decision-making double-cycle DRL module receives the space-time prediction sequence to generate the trajectory parameter, and the energy-aware adaptive module generates the control instruction sequence through the trajectory parameter.

Description

Technical Field

[0001] This invention relates to the field of advanced aircraft energy management technology, and in particular to a multimodal fusion eVTOL energy consumption optimization system and dynamic modeling method. Background Technology

[0002] Current eVTOL energy management technology faces three key challenges: First, traditional PID controllers rely on fixed parameter settings. During the transition from rotor to fixed-wing tilt, the lift source abruptly changes from rotor aerodynamics to the Bernoulli effect of the wing. The nonlinear characteristics of the power system cause a step jump in the energy consumption curve (measured data fluctuations can reach ±35%). Traditional control methods are prone to power overload or stall risks due to response delays. Second, existing single-mode prediction models are only designed for specific flight phases (such as hovering or level flight) and fail to establish a dynamic coupling mechanism for rotor modal aerodynamic equations, resulting in a high error rate in energy consumption prediction during mode switching. Third, while offline-trained machine learning models can achieve high accuracy in laboratory environments, they lack online adaptive capabilities when faced with sudden airflow disturbances and battery performance degradation in actual flight. Frequent factory recalibration is required, severely restricting operational efficiency. These systemic defects collectively indicate that current technologies cannot meet the real-time energy optimization requirements of eVTOL under complex operating conditions.

[0003] With the rapid development of urban air mobility (UAM), hybrid propulsion systems (rotor-fixed-wing hybrid power architecture) have become the mainstream design due to their ability to combine vertical takeoff and landing with high-speed cruise. However, the nonlinear energy consumption abrupt changes caused by multi-modal switching severely restrict range and safety. This technical solution deeply integrates physical model-driven and data-driven methods, and achieves dynamic energy consumption optimization of the aircraft under all operating conditions by constructing a multi-modal fusion system with autonomous decision-making capabilities. Summary of the Invention

[0004] The purpose of this invention is to address the problem of sudden energy consumption changes in current eVTOLs under complex operating conditions, and to provide a multimodal fusion eVTOL energy consumption optimization system.

[0005] Another aspect of the present invention provides a dynamic modeling method for a multimodal fusion-based eVTOL energy consumption optimization system.

[0006] The technical solution adopted to achieve the purpose of this invention is:

[0007] A multimodal fusion-based eVTOL energy optimization system is disclosed. The system includes an LSTM-PINN multimodal spatiotemporal prediction module, an online learning module, an autonomous decision-making dual-loop DRL module, and an energy-aware adaptive module. The LSTM-PINN multimodal spatiotemporal prediction module uses a Long Short-Term Memory (LSTM) network as the temporal encoding layer of a PINN, where the PINN is defined by a loss function... To optimize the online learning module, a mutation detection unit and an incremental learning unit are included. The autonomous decision-making dual-loop DRL module includes an inner loop planning module and an outer loop strategy module. The multimodal temporal data stream is input into the LSTM-PINN multimodal spatiotemporal prediction module, and the LSTM temporal coding layer decodes the multimodal temporal data stream to generate the PINN loss function. The predicted value of the initial conditional loss term, and simultaneously, the online learning module for the PINN loss function. The weights of the loss terms are dynamically allocated, and a spatiotemporal prediction sequence is output and fed to the autonomous decision-making dual-loop DRL module. The outer loop strategy module generates a global decision, and the inner loop planning module combines the global decision generated by the outer loop strategy module to generate temporally continuous trajectory parameters. The energy sensing adaptive module generates a control command sequence through the temporally continuous trajectory parameters generated by the autonomous decision-making dual-loop DRL module.

[0008] In the above technical solution, the multimodal time-series data stream includes wind speed, turbulent kinetic energy, battery voltage, and rotor speed command streams.

[0009] In the above technical solution, the PINN loss function It consists of an initial condition loss term, a physical residual loss term, and a data-driven loss term.

[0010] In the above technical solution, the physical residual loss term includes the flight dynamics residual term. Rotor aerodynamic residuals and energy flow constraint residuals

[0011] In the above technical solution, the PINN loss function Defined as:

[0012]

[0013] In the formula, Let λ1, λ2, λ3, λ4, and λ5 be the weights of the PINN loss function. The mean square error of the initial condition loss term. This refers to the flight dynamics residual term within the physical residual loss term. This refers to the rotor aerodynamic residual term within the physical residual loss term. The energy flow constraint residual in the physical residual loss term. This is an error correction to compensate for the traditional flight dynamics model in the data-driven loss term.

[0014] In the above technical solution, the mean square error of the initial condition loss term The calculation formula is:

[0015]

[0016] In the formula, λ init Here, N represents the weighting coefficients, and N represents the number of spatial sampling points. This represents the initial conditional prediction value for the i-th spatial sampling point. Let x be the true physical field value at the initial moment of the i-th spatial sampling point. i Let i be the i-th sampling point.

[0017] In the above technical solution, the flight dynamics residual term in the physical residual loss term The calculation formula is:

[0018]

[0019] In the formula, N is the number of spatial sampling points, and m is the eVTOL mass. The location predicted by the network. It is the acceleration predicted by the network, x i For the i-th sampling point, t i For time i, F total The total external force consists of rotor thrust, gravity, and fluid resistance.

[0020] In the above technical solution, the rotor aerodynamic residual term in the physical residual loss term The calculation formula is:

[0021]

[0022] In the formula, N is the number of spatial sampling points, and t rotor For the actual rotor thrust, Q rotor C represents the actual rotor torque. T C is the thrust coefficient. Q Let ρ be the torque coefficient, Ω be the air density, Ω be the rotor angular velocity, and R be the blade radius.

[0023] In the above technical solution, the energy flow constraint residual in the physical residual loss term The calculation formula is:

[0024]

[0025] In the formula, P inputFor external energy input, T NN For rotor thrust, v z The power Q that does work on the vertical velocity rotor The rotor torque is from the aerodynamic residual term, ω is the rotational speed, and t is the rotor torque. i Let i be the time step i.

[0026] Another aspect of the present invention provides a dynamic modeling method for a multimodal fusion eVTOL energy consumption optimization system, comprising the following steps:

[0027] Step 1: Construct an LSTM-PINN multimodal spatiotemporal prediction module: An LSTM is connected in series at the front end of the PINN module as a spatiotemporal coding layer to extract the dynamic features of the multimodal time-series data stream and generate the predicted value of the initial conditional loss term of the PINN loss function. The PINN is optimized by a loss function consisting of an initial conditional loss term, a physical residual loss term, and a data-driven loss term. The physical residual loss term includes a flight dynamics residual term, a rotor aerodynamic residual term, and an energy flow constraint residual term. The weights of the initial conditional loss term, the flight dynamics residual term, the rotor aerodynamic residual term, the energy flow constraint residual term, and the data-driven loss term in the PINN loss function are λ1, λ2, λ3, λ4, and λ5, respectively. The LSTM-PINN multimodal spatiotemporal prediction module generates a spatiotemporal prediction sequence.

[0028] Step 2, Constructing an online learning module: An online learning module is constructed based on Bayesian mutation detection and residual attention mechanism (RBA). The online learning module includes a mutation detection unit and an incremental learning unit. The mutation detection unit analyzes the multimodal time-series data features of eVTOL in real time, identifies dynamic mutation points using a Bayesian probability model, and defines high-energy-consuming feature data blocks as the core input to trigger the incremental learning unit. When the mutation detection unit detects a mutation event, it triggers the RBA-based incremental learning unit. RBA assigns weights λ1, λ2, λ3, λ4, and λ5 to the loss terms of the dynamic PINN loss function.

[0029] Step 3, Design the Autonomous Decision-Making Dual-Loop DRL Module: The autonomous decision-making dual-loop DRL module includes an outer loop strategy module based on gating mechanism-near-end strategy optimization and an inner loop planning module based on Bezier parameterization and constrained strategy optimization. The outer loop strategy module receives the aerodynamic coupling tensor, the power margin assessment of the transition corridor, and the flight risk probability from the spatiotemporal prediction sequence to generate a global decision, and generates rotor mode commands and energy weights through a strategy gradient algorithm, which are then allocated to the inner loop planning module. The inner loop planning module receives the flow field pressure gradient and the transition corridor boundary from the spatiotemporal prediction sequence, and combines them with the global decision from the outer loop strategy module to generate time-domain continuous flight trajectory parameters.

[0030] Step 4: The energy perception and adaptive module uses the flight trajectory parameters output by the autonomous decision-making dual-cycle DRL module and, based on the model predictive control (MPC) framework, embeds the energy efficiency target into the dynamic optimization process. With minimizing global energy loss as the optimization objective, it prioritizes scheduling the high-efficiency rotor to bear the main thrust load. At the same time, it integrates turbulence disturbance compensation to perform anti-interference torque feedforward, generating a control command sequence that balances trajectory tracking accuracy and energy consumption.

[0031] Compared with the prior art, the beneficial effects of the present invention are:

[0032] 1. This invention accurately extracts the dynamic features of multimodal time-series data streams using Long Short-Term Memory (LSTM) networks and deeply couples them with the spatiotemporal conservation equations of Physical Information Neural Networks (PINN). By organically combining PINN and LSTM, it achieves modeling of the impact of historical states on current flow field prediction, breaking through the adaptability bottleneck of traditional control methods in cross-domain flight, and providing a real-time energy management solution for next-generation eVTOL intelligent aircraft.

[0033] 2. This invention innovatively introduces three-channel physical residual constraints (flight dynamics, rotor aerodynamics, and energy flow constraints), combining initial condition loss terms and data-driven loss terms to form a hybrid optimization objective, achieving high-precision, physically consistent dynamic modeling. This solves the sensitivity of traditional models to sudden environmental changes and energy efficiency deviations. Compared with traditional data-driven models, the prediction accuracy of key parameters such as thrust and torque is improved by more than 40%. The physical constraints based on blade element momentum theory (BEM) further solve the problem of high-order nonlinear modeling of rotor vortex fields. The hybrid loss function embedded in the physical constraints breaks through the dependence of pure deep learning on labeled data, systematically addressing the industry pain points of aerodynamic coupling inaccuracy and energy consumption prediction drift.

[0034] 3. This invention employs an inner and outer dual-loop reinforcement learning (DRL) module, based on a hierarchical collaborative mechanism of global task planning and local dynamic control, overcoming the response limitations of traditional single control loops. The outer loop strategy module achieves global strategy optimization for multiple objective constraints (energy consumption, safety, and timeliness) through reinforcement learning, while the inner loop planning module calculates dynamically feasible trajectories in real time using Bezier parameterization, significantly improving the energy efficiency and dynamic response speed of trajectory planning, and achieving an organic unity of long-term and short-term objectives in complex scenarios.

[0035] 4. This invention innovatively constructs a lightweight incremental online learning module. Through Bayesian mutation detection and dynamic residual weight allocation strategies, it achieves online millisecond-level tuning of model parameters. Combined with the real-time flow field-energy consumption prediction of the LSTM-PINN multimodal spatiotemporal prediction module, it generates safe and smooth flight trajectory parameters, ensuring that the control quantity always remains within the power and anti-stall boundaries. This enables safe, efficient, and comfortable autonomous tilt transition control under complex corridor constraints. Compared to traditional offline training schemes, the model's adaptive convergence speed is improved by 90% when encountering scenarios such as sudden turbulence and mechanical performance degradation. Furthermore, through the dual guarantees of physical consistency constraints and sparse updates, it avoids the catastrophic forgetting problem in incremental learning. Combined with the energy-sensing MPC control module, the system can dynamically adjust the thrust allocation weights based on the real-time rotor efficiency MAP and battery degradation characteristics, reducing energy loss throughout flight by 30%. This provides a technical foundation for the full lifecycle reliability and economy of eVTOL commercial operation.

[0036] 5. This invention constructs a high-efficiency energy consumption management and autonomous decision-making system for eVTOL multimodal flight through three major technological innovations: mechanism-data fusion modeling, hierarchical decision collaboration, and dynamic adaptive control. Compared with traditional methods, this system achieves a comprehensive breakthrough in dynamic response efficiency, multi-constraint collaboration capability, and environmental adaptability, providing disruptive technical support for the intelligent and commercial implementation of urban air transportation vehicles. Attached Figure Description

[0037] Figure 1 The diagram shows the topology of the eVTOL energy consumption optimization system architecture based on multimodal fusion of the present invention. Detailed Implementation

[0038] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0039] Example 1

[0040] A multimodal fusion-based eVTOL energy optimization system is disclosed. The system includes an LSTM-PINN multimodal spatiotemporal prediction module, an online learning module, an autonomous decision-making dual-loop DRL module, and an energy-aware adaptive module. The LSTM-PINN multimodal spatiotemporal prediction module uses a Long Short-Term Memory (LSTM) network as the temporal encoding layer of a PINN, where the PINN is defined by a loss function... To optimize the online learning module, a mutation detection unit and an incremental learning unit are included. The autonomous decision-making dual-loop DRL module includes an inner loop planning module and an outer loop strategy module. The multimodal temporal data stream is input into the LSTM-PINN multimodal spatiotemporal prediction module, and the LSTM temporal coding layer decodes the multimodal temporal data stream to generate the PINN loss function. The predicted value of the initial conditional loss term, and simultaneously, the online learning module for the PINN loss function. The weights of the loss terms are dynamically allocated, and a spatiotemporal prediction sequence is output and fed to the autonomous decision-making dual-loop DRL module. The outer loop strategy module generates a global decision, and the inner loop planning module combines the global decision generated by the outer loop strategy module to generate temporally continuous trajectory parameters. The energy sensing adaptive module generates a control command sequence through the temporally continuous trajectory parameters generated by the autonomous decision-making dual-loop DRL module.

[0041] An LSTM is cascaded in front of the PINN as a spatiotemporal coding layer to extract dynamic features from the multimodal temporal data stream, forming the LSTM-PINN multimodal spatiotemporal prediction module. This achieves deep coupling between the multimodal temporal data stream and the physical model through the construction of an LSTM-PINN fusion architecture. The loss function of the PINN... It consists of an initial condition loss term, a physical residual loss term, and a data-driven loss term.

[0042] Regarding the initial condition loss term: PINN uses a Long Short-Term Memory (LSTM) network as a spatiotemporal encoding layer to directly process the input of multimodal time-series data streams, including wind speed, turbulent kinetic energy, battery voltage, and rotor speed command streams, and outputs the hidden state h. t And decode to generate the initial conditional prediction value of PINN. Where D represents the decoding network. The true values ​​of the physical field at the known initial time are used. For monitoring, a mean squared error (MSE) loss term is constructed. :

[0043]

[0044] In the formula, λ init Here, N represents the weighting coefficients, and N represents the number of spatial sampling points. This represents the initial conditional prediction value for the i-th spatial sampling point. Let be the true value of the physical field at the initial moment of the i-th spatial sampling point.

[0045] Regarding the physical residual loss term: PINN's physical residual term is embedded in the flight dynamics residual term. Rotor aerodynamic residuals and energy flow constraint residual Apply constraints.

[0046] Flight dynamics residuals The flight trajectory predicted by the forced network conforms to the laws of physical conservation, and the flight dynamics residual term The formula is:

[0047]

[0048] In the formula, N is the number of spatial sampling points, and m is the eVTOL mass. The location predicted by the network. For the acceleration predicted by the network, x i For the i-th sampling point, t i For time i, F total The total external force consists of rotor thrust, gravity, and fluid resistance.

[0049] Rotor aerodynamic residuals This approach can supplement empirical knowledge of rotor energy efficiency and dynamic response at the microscale. By applying dual-channel constraints to the lift coefficient and torque system through the blade element momentum theory (BEM) equations, it achieves joint modeling of the blade vortex field and wing aerodynamic loads. (Rotor aerodynamic residuals) The formula is:

[0050]

[0051] In the formula, N is the number of spatial sampling points, and T rotor For the actual rotor thrust, Q rotor C represents the actual rotor torque. T C is the thrust coefficient. Q Let ρ be the torque coefficient, Ω be the air density, Ω be the rotor angular velocity, and R be the blade radius.

[0052] Energy-constrained residuals Based on the real-time mapping of rotor torque parameters from BEM output and motor drive power, a power conservation residual term is constructed. This constraint forces the neural network to simultaneously satisfy aerodynamic lift requirements and electromechanical conversion efficiency during training, avoiding the pursuit of flow field accuracy at the expense of ineffective energy consumption caused by torque fluctuations. Energy flow constraint residual. The formula is:

[0053]

[0054] In the formula, P input For external energy input, T NN For rotor thrust, v z The power Q that does work on the vertical velocity rotor The rotor torque is from the aerodynamic residual term, ω is the rotational speed, and t is the rotor torque. i Let i be the time step i.

[0055] Regarding data-driven loss: Integrating multimodal data with physical equation constraints to achieve error correction of traditional flight dynamics models by compensating for losses with real flight data. By adjusting the weights of perception during the flight phase, we can enhance the learning of eddy current characteristics during hovering and improve the generalization of energy consumption prediction during cruise.

[0056] In summary, the loss function in the PINN hybrid loss function layer Defined as:

[0057]

[0058] In the formula, Let λ1, λ2, λ3, λ4, and λ5 be the weights of the PINN loss function. The mean square error of the initial condition loss term. This refers to the flight dynamics residual term within the physical residual loss term. The rotor aerodynamic residual term in the physical residual loss term and This refers to the energy flow constraint residual term within the physical residual loss term. This is an error correction to compensate for the traditional flight dynamics model in the data-driven loss term.

[0059] The design of the aforementioned LSTM-PINN multimodal spatiotemporal prediction module not only retains the physical consistency advantages of PINN but also introduces adaptability to dynamic changes in the actual flight environment through LSTM. This LSTM-PINN multimodal spatiotemporal prediction module, through the deep fusion of physical constraints, temporal characteristics, and data-driven approaches, provides eVTOL with high-precision and high-efficiency dynamic modeling capabilities, becoming a core enabling technology for multimodal energy consumption optimization.

[0060] The online learning module achieves dynamic model updates by integrating a mutation detection unit and a residual-based attention mechanism (RBA) incremental learning unit. First, a composite detection module is constructed to analyze the multimodal time-series data features of eVTOL in real time. A Bayesian mutation probability model is used to identify dynamic mutation points, and high-energy-consumption feature data blocks are defined as the core input to trigger the incremental learning unit, ensuring the model focuses on correcting energy consumption prediction biases. When a mutation event is detected, the RBA-based incremental learning unit is triggered. RBA allocates the weights λ1, λ2, λ3, λ4, and λ5 of the loss term in the dynamic PINN network loss function: dynamically adjusting the loss weights based on the residual size of the new data point. Points with larger residuals have higher weights. Incremental training guided by residual weights enables the model to simultaneously meet data fitting and physical consistency requirements when adapting to new operating conditions. Implementation requires combining lightweight residual calculation and dynamic weight allocation strategies to ensure the real-time performance and stability of the agent's online learning. RBA balances data fitting and physical constraints, and combines a sparse parameter update strategy to achieve rapid online model calibration, ensuring that the model adapts to new operating conditions during flight while avoiding catastrophic forgetting, and improving the model's robustness to time-varying environments and unknown disturbances.

[0061] Design of an Autonomous Decision-Making Dual-Cycle DRL Module: eVTOL faces severe challenges during the transition between vertical takeoff and landing and cruise modes, particularly the aerodynamic coupling effect caused by rotor tilt, which leads to strong nonlinear coupling between lift / thrust vectors and airframe torque. Meanwhile, the power system must simultaneously combat sudden turbulence, meet power limits, and aerodynamic stability requirements during the transition. To ensure flight safety, international aviation standards define a transition corridor—a dynamically feasible region defined by the velocity-altitude envelope, whose boundaries are constrained by power margin and airframe stall resistance. However, existing control methods have significant limitations: traditional PID control struggles to handle time-varying nonlinearities during tilt, while single-layer DRL strategies are prone to safety violations under corridor constraints and lack synergistic optimization for energy efficiency. Therefore, an autonomous decision-making dual-cycle DRL module is proposed at the system's strategy planning layer. By combining the decision-making of the outer loop strategy layer and the trajectory planning of the inner loop planning module with the real-time predictive capabilities of PINN, safe, efficient, and comfortable multi-objective autonomous flight control within the transition corridor is achieved. This autonomous decision-making architecture employs a cascaded "policy-planning" mechanism. The outer loop policy module, based on a gated mechanism-near-end policy optimization (Gated-PPO), generates global decisions. It receives aerodynamic coupling tensors, power margin assessments of the transition corridor, and flight risk probabilities from the spatiotemporal prediction sequence. Through a policy gradient algorithm, it generates rotor mode commands and energy weights, which are then assigned to the inner loop planning module. The inner loop planning module, based on Bezier parameterization and constrained policy optimization methods, receives flow field pressure gradients and transition corridor boundaries from the spatiotemporal prediction sequence. Combined with the global decisions from the outer loop policy module, it generates time-continuous flight trajectory parameters. The control quantity is forced to always remain within the power and anti-stall boundaries, achieving safe, efficient, and comfortable autonomous tilt transition control under complex corridor constraints. This ensures flight within safe boundaries under power-limited conditions, avoiding entry into unstable regions.

[0062] The energy-sensing adaptive control module uses the flight trajectory parameters (position, velocity, attitude) generated by the autonomous decision-making dual-cycle DRL module. Based on the model predictive control (MPC) framework, it embeds the energy efficiency target into the dynamic optimization process: when solving the motor power distribution, rotor speed and tilt command, it uses the real-time battery remaining capacity and rotor efficiency MAP as coupling constraints. Through a quadratic programming algorithm, under the premise of satisfying the upper limit of the power system power and the dynamic boundary of the transition corridor, it takes minimizing global energy loss as the optimization objective, prioritizes scheduling the high-efficiency rotor to bear the main thrust load, and integrates turbulence disturbance compensation to feedforward the anti-interference torque. Finally, it generates a control command sequence that takes into account both trajectory tracking accuracy and energy consumption optimization, ensuring a coordinated balance between energy utilization and motion safety during flight.

[0063] The above description is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A multimodal fusion eVTOL energy consumption optimization system, characterized in that, The system includes an LSTM-PINN multimodal spatiotemporal prediction module, an online learning module, an autonomous decision-making dual-loop DRL module, and an energy-sensing adaptive module. The LSTM-PINN multimodal spatiotemporal prediction module uses a Long Short-Term Memory (LSTM) network as the temporal encoding layer of PINN, where PINN is defined by a loss function. Optimization is performed, including the PINN loss function. It includes initial condition loss terms, physical residual loss terms, and data-driven loss terms. The physical residual loss terms include flight dynamics residual terms. Rotor aerodynamic residuals and energy flow constraint residual The PINN loss function Defined as: ,in, , , , and The weights of the loss function, The mean square error of the initial condition loss term. This refers to the flight dynamics residual term within the physical residual loss term. This refers to the rotor aerodynamic residual term within the physical residual loss term. The energy flow constraint residual in the physical residual loss term. To compensate for errors in the traditional flight dynamics model in the data-driven loss term; the online learning module includes a mutation detection unit and an incremental learning unit, and the autonomous decision-making dual-loop DRL module includes an inner loop planning module and an outer loop strategy module; a multimodal time-series data stream containing command streams for wind speed, turbulent kinetic energy, battery voltage, and rotor speed is input to the LSTM-PINN multimodal spatiotemporal prediction module, and the LSTM temporal coding layer decodes the multimodal time-series data stream to generate the PINN loss function. The predicted value of the initial conditional loss term, and simultaneously, the online learning module for the PINN loss function. The weights of the loss terms are dynamically allocated, and a spatiotemporal prediction sequence is output and fed to the autonomous decision-making dual-loop DRL module. The outer loop strategy module generates a global decision, and the inner loop planning module combines the global decision generated by the outer loop strategy module to generate temporally continuous trajectory parameters. The energy sensing adaptive module generates a control command sequence through the temporally continuous trajectory parameters generated by the autonomous decision-making dual-loop DRL module.

2. The eVTOL energy consumption optimization system as described in claim 1, characterized in that, The mean square error of the initial condition loss term The calculation formula is: In the formula, These are the weighting coefficients. N The number of spatial sampling points. For the first i Initial conditional prediction values ​​for each spatial sampling point For the first i The true physical field values ​​at the initial time of each spatial sampling point For the first i One sampling point.

3. The eVTOL energy consumption optimization system as described in claim 1, characterized in that, The flight dynamics residual term in the physical residual loss term The calculation formula is: In the formula, N The number of spatial sampling points. For eVTOL quality, The location predicted by the network. It is the acceleration predicted by the network. For the first i One sampling point, For the first i time, The total external force consists of rotor thrust, gravity, and fluid resistance.

4. The eVTOL energy consumption optimization system as described in claim 1, characterized in that, The rotor aerodynamic residual term in the physical residual loss term The calculation formula is: In the formula, N The number of spatial sampling points. For actual rotor thrust, This represents the actual rotor torque. For thrust coefficient, For torque system, air density, The rotor angular velocity, R Where is the blade radius.

5. The eVTOL energy consumption optimization system as described in claim 1, characterized in that, The energy flow constraint residual in the physical residual loss term The calculation formula is: In the formula, For external energy input, For rotor thrust, The power that does work for vertical velocity. The rotor torque is derived from the aerodynamic residual term. For rotational speed, For the first i time.

6. A dynamic modeling method for a multimodal fusion eVTOL energy consumption optimization system, characterized in that, Includes the following steps: Step 1: Construct an LSTM-PINN multimodal spatiotemporal prediction module: An LSTM is connected in series at the front end of the PINN module as a spatiotemporal encoding layer to extract dynamic features from a multimodal time-series data stream containing command streams for wind speed, turbulent kinetic energy, battery voltage, and rotor speed. This generates predicted values ​​for the initial condition loss term of the PINN loss function. The PINN is optimized by a loss function consisting of an initial condition loss term, a physical residual loss term, and a data-driven loss term. The physical residual loss term includes flight dynamics residuals, rotor aerodynamic residuals, and energy flow constraint residuals. The weights of these components in the PINN loss function are as follows: , , , and The LSTM-PINN multimodal spatiotemporal prediction module generates spatiotemporal prediction sequences; Step 2: Construct an online learning module: An online learning module is constructed based on Bayesian mutation detection and residual attention mechanism (RBA). This module includes a mutation detection unit and an incremental learning unit. The mutation detection unit analyzes the multimodal time-series data features of eVTOL in real time, identifies dynamic mutation points using a Bayesian probability model, and defines high-energy-consuming feature data blocks as the core input to trigger the incremental learning unit. When the mutation detection unit detects a mutation event, it triggers the RBA-based incremental learning unit. RBA assigns weights to the loss term of the dynamic PINN loss function. , , , and Distribute; Step 3, Design the Autonomous Decision-Making Dual-Loop DRL Module: The autonomous decision-making dual-loop DRL module includes an outer loop strategy module based on gating mechanism-near-end strategy optimization and an inner loop planning module based on Bezier parameterization and constrained strategy optimization. The outer loop strategy module receives the aerodynamic coupling tensor, the power margin assessment of the transition corridor, and the flight risk probability from the spatiotemporal prediction sequence to generate a global decision, and generates rotor mode commands and energy weights through a strategy gradient algorithm, which are then allocated to the inner loop planning module. The inner loop planning module receives the flow field pressure gradient and the transition corridor boundary from the spatiotemporal prediction sequence, and combines them with the global decision from the outer loop strategy module to generate time-domain continuous flight trajectory parameters. Step 4: The energy perception and adaptive module uses the flight trajectory parameters output by the autonomous decision-making dual-cycle DRL module and, based on the model predictive control (MPC) framework, embeds the energy efficiency target into the dynamic optimization process. With minimizing global energy loss as the optimization objective, it prioritizes scheduling the high-efficiency rotor to bear the main thrust load. At the same time, it integrates turbulence disturbance compensation to perform anti-interference torque feedforward, generating a control command sequence that balances trajectory tracking accuracy and energy consumption.

Images