A fuel cell unmanned aerial vehicle energy management system, method and unmanned aerial vehicle

Abstract

The application provides a fuel cell unmanned aerial vehicle energy management system, method and unmanned aerial vehicle. The management system comprises a multi-source data acquisition module, an information fusion and state construction module, an integrated prediction and uncertainty quantification module, a risk-sensitive reinforcement learning decision module, a power distribution execution module and an online learning update module. The method of the application constructs a complete technical closed loop of "multi-source information fusion-integrated prediction and uncertainty quantification-risk-sensitive reinforcement learning decision-online evolutionary learning", realizes intelligent balance and unification between real-time performance, global economy and operation safety of energy management in a complex flight environment by constructing an enhanced state space that fuses multi-source real-time information, and combining an integrated prediction model with uncertainty quantification capability and a risk-sensitive deep reinforcement learning decision maker.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) energy management technology, specifically relating to a fuel cell UAV energy management system, method, and UAV. Background Technology

[0002] Fuel cell drones, due to their high energy density and long endurance potential, have become an important development direction for drone energy systems. The energy management strategy of hybrid power systems (typically composed of fuel cells and lithium batteries) directly affects the drone's endurance, reliability, and overall performance. Currently, technological development in this field mainly revolves around the following methods, but each has its own limitations:

[0003] (1) Rule-based and instantaneous optimization strategy: Although it has good real-time performance, it cannot adapt to complex working conditions, it is difficult to achieve global optimization, and it relies on precise parameter calibration.

[0004] (2) Global optimization-based strategies (such as dynamic programming): Although theoretically optimal, they are computationally complex, rely on accurate models and complete prior information, and cannot be applied online in real time.

[0005] (3) Machine learning-based strategies (such as reinforcement learning): These strategies do not require precise models, but generally suffer from low learning efficiency and high exploration risks. Although some studies have introduced predictive models (such as GRU) to improve efficiency, their predictions usually rely solely on historical load data and do not fully utilize multi-source information such as the environment, resulting in insufficient prediction accuracy and adaptability under complex flight conditions. More importantly, existing methods lack an evaluation mechanism for the reliability of the prediction results themselves, and fail to quantify this uncertainty in predictions into clear decision risk signals. Therefore, they cannot intelligently adjust decision-making strategies to prioritize system safety when predictions are unreliable, and there is a possibility of operational risks arising from blindly pursuing economic optimization.

[0006] In summary, existing fuel cell drone energy management technologies either struggle to balance real-time performance with global optimization, lack robustness to environmental uncertainties and model errors, or suffer from low learning efficiency and lack an intelligent mechanism capable of sensing and predicting uncertainty and making adaptive risk decisions. Therefore, there is an urgent need for a novel energy management method that can integrate multi-source real-time information, possess online quantification capabilities for predictive uncertainty, and dynamically adjust decision-making risk preferences based on this uncertainty. Summary of the Invention

[0007] To address the aforementioned technical problems, in a first aspect, the present invention provides a fuel cell unmanned aerial vehicle (UAV) energy management system, comprising:

[0008] The multi-source data acquisition module is used to acquire multi-source operational information of the UAV in real time;

[0009] The information fusion and state construction module is used to fuse multi-source information through an attention mechanism network to construct an enhanced state vector;

[0010] The integrated prediction and uncertainty quantification module is used to load and run the integrated prediction model after inputting the enhanced state vector sequence, and output the load power prediction mean and uncertainty metric Uncertainty.

[0011] The risk-sensitive reinforcement learning decision module is used to load and run a deep reinforcement learning decision engine based on the Dyna-Q network framework and output power allocation instructions.

[0012] The power distribution execution module is used to convert the power distribution command into a control signal, drive the fuel cell and lithium battery system to execute the output power distribution command, and complete the real-time power distribution between the fuel cell and the lithium battery.

[0013] The online learning update module manages the storage of state transition data and performs model fine-tuning on the integrated prediction and uncertainty quantification module and the risk-sensitive reinforcement learning decision module.

[0014] This invention aims to achieve an intelligent balance and unity between real-time performance, global economy, and operational safety in energy management under complex flight environments by constructing an enhanced state vector that integrates multi-source real-time information and combining an integrated prediction model with uncertainty quantification capabilities with a risk-sensitive reinforcement learning decision module.

[0015] Furthermore, the integrated prediction model consists of multiple parallel, identically structured gated recurrent unit (GRU) neural network sub-models, and each neural network sub-model contains a conditional normalization layer with real-time environmental information as the conditional input.

[0016] Secondly, the present invention also provides an energy management method for a fuel cell unmanned aerial vehicle (UAV), employing a fuel cell UAV energy management system, comprising the following steps:

[0017] Step S1, Multi-source data acquisition, fusion, and enhanced state construction:

[0018] The multi-source data acquisition module collects multi-source operational data of the UAV in real time. The information fusion and state construction module fuses the multi-source operational data through an attention mechanism network to generate an enhanced state vector that comprehensively represents the system and the environmental situation.

[0019] Step S2, integrating prediction and uncertainty quantification:

[0020] The enhanced state vector sequence within the current and historical time windows is input into the integrated prediction model, which outputs the predicted mean of load power within a future set time domain and outputs the uncertainty metric of the current prediction.

[0021] Step S3, Risk-Sensitive Reinforcement Learning Decision:

[0022] The enhanced state vector, the mean load power prediction, and the uncertainty metric at the current moment are input together into a deep reinforcement learning decision-maker based on the Dyna-Q network framework. The decision-maker will guide the management strategy to become more conservative when the prediction uncertainty is high, based on its reward function, and output the optimal power allocation command for the fuel cell.

[0023] Step S4, Power Allocation Execution and Online Learning:

[0024] The power allocation execution module executes the optimal power allocation command to control the power output of the fuel cell and lithium battery to meet the load requirements;

[0025] Meanwhile, the online learning update module continuously collects state transition data generated by the actual operation of the UAV and stores it in the experience replay buffer. Then, it periodically samples data from the experience replay buffer to fine-tune the integrated prediction model and the deep reinforcement learning network in the decision maker online, so as to realize the continuous evolution of the energy management strategy.

[0026] This invention can intelligently balance the endurance economy and operational safety of drones, and improve their overall robustness and reliability in complex real-world environments.

[0027] Furthermore, the multi-source information mentioned in step S1 includes at least load power, battery state of charge (SoC), flight attitude information, environmental information, and current mission instructions;

[0028] The environmental information includes wind speed, wind direction, and temperature.

[0029] Furthermore, the method for constructing the enhanced state vector based on multi-source real-time information in step S1 is as follows:

[0030] The multi-source real-time information is input into an attention mechanism network to calculate the dynamic weight of each information source at the current time. Then, the normalized features of each information source are weighted and fused according to the dynamic weight to generate an enhanced state vector that comprehensively represents the real-time state of the system and the environment.

[0031] Furthermore, the attention mechanism network described in step S1 is a lightweight feedforward neural network, whose output is a weight vector normalized by the Softmax function.

[0032] Furthermore, the method for calculating the uncertainty metric Uncertainty mentioned in step S2 is as follows:

[0033] The integrated prediction model summarizes the prediction results of all network neural sub-models, outputs the predicted mean of load power, and then calculates the variance of the predicted values ​​output by each network neural sub-model.

[0034] Furthermore, the risk-sensitive reinforcement learning decision module described in step S4 is based on the Dyna-Q network framework, and its reward function... It is risk-sensitive, specifically in the form of:

[0035]

[0036] in, This represents the basic economic incentive item based on hydrogen consumption, battery SoC maintenance, and fuel cell durability design; λ is a risk sensitivity coefficient greater than zero.

[0037] The The expression is:

[0038]

[0039] in, Hydrogen consumption rate, This is a reference value for the battery's state of charge. For fuel cell current, , , It is a positive number.

[0040] Furthermore, the online fine-tuning in step S4 specifically includes: randomly sampling small batches of data from the stored state transition data, updating the parameters of the ensemble prediction model and the deep reinforcement learning network using a gradient descent algorithm at a learning rate lower than the initial training rate.

[0041] Thirdly, the present invention also provides a fuel cell unmanned aerial vehicle, including a flight platform, a hybrid power system, and a flight control computer, wherein the flight control computer is configured with the above-mentioned energy management system, or is configured to execute the above-mentioned methods.

[0042] The beneficial effects of this invention are as follows:

[0043] This invention constructs a complete technical closed loop of "multi-source information fusion - integrated prediction and uncertainty quantification - risk-sensitive reinforcement learning decision-making - online evolutionary learning," which has the following significant advantages compared with existing technologies:

[0044] 1. Unlike traditional single-data-source methods, this invention employs "attention-driven multi-source information fusion," which adaptively assigns different fusion weights to information such as load power, attitude, and wind speed based on real-time flight scenarios (e.g., cruise, hovering, and sudden gusts). This solves the problem of limited state perception dimensions and poor adaptability in complex flight environments, achieving accurate representation of the system and environmental situation. The constructed "enhanced state vector" can more sensitively and accurately capture the most critical information for energy management at the current moment, providing high-quality input for subsequent load prediction and decision-making, and significantly improving the model's environmental adaptability under varying operating conditions.

[0045] 2. This invention addresses the dynamic characteristics of unmanned aerial vehicles (UAVs) by combining ensemble learning with a GRU network featuring conditional normalization. The conditional normalization layer dynamically adjusts the distribution of features within the network based on real-time environmental information, enabling the prediction model to adapt to environmental changes such as weather conditions in real time, resulting in higher prediction accuracy. This invention not only outputs a more accurate mean prediction of future load power, but more importantly, it outputs a measure of prediction uncertainty (variance). This quantitative indicator reveals the reliability of the current prediction results, providing valuable "risk prediction" information for subsequent decision-making, allowing the system to "know what it doesn't know."

[0046] 3. This invention addresses the core contradiction of traditional strategies struggling to balance economy and safety in uncertain environments through a "risk-sensitive reinforcement learning reward function design," achieving risk-adaptive intelligent decision-making. When uncertainty is high: the reward value is significantly reduced, lowering the expected return of any "aggressive optimization" actions that could lead to state fluctuations. This automatically guides the decision-maker to adopt a conservative and safe power allocation strategy, prioritizing flight safety and avoiding energy system shocks or failures caused by blindly relying on unreliable predictions.

[0047] When uncertainty is low: the penalty term approaches zero, and the reward function is mainly composed of economic reward terms. The decision-maker focuses on refined optimization, pursuing the best fuel economy and battery life.

[0048] This mechanism is the first in the field of drone energy management to achieve dynamic, continuous, and closed-loop linkage between "prediction confidence" and "decision risk preference," fundamentally solving the problem of real-time trade-off between safety and economy.

[0049] 4. The Dyna-DQN framework employed in this invention combines model-based planning with model-free reinforcement learning, enabling "simulation and deduction" within a learning environment using limited real-world interaction data, significantly improving learning efficiency. By continuously collecting data during real-world flights and fine-tuning online with a small learning rate, the integrated predictive model and decision network of this invention can continuously adapt to battery aging, changes in airframe performance, and new environmental patterns. This gives the energy management strategy "self-evolution" capabilities, maintaining superior adaptability and robustness throughout the entire lifecycle of the UAV, truly achieving "becoming smarter with each flight." Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the core steps of the management method of this invention;

[0051] Figure 2 This is a schematic diagram of the overall architecture of the energy management system of the present invention;

[0052] Figure 3 This is a schematic diagram of the hardware structure of the energy management system of the present invention. Detailed Implementation

[0053] The following specific embodiments illustrate the implementation of the invention. Those skilled in the art can easily understand other advantages and effects of the invention from the content disclosed in this specification.

[0054] Exemplary embodiments of the invention are now described with reference to the accompanying drawings. However, the invention may be embodied in many different forms and is not limited to the embodiments described herein. These embodiments are provided to fully and completely disclose the invention and to fully convey its scope to those skilled in the art. The terminology used in the exemplary embodiments illustrated in the drawings is not intended to limit the invention.

[0055] Unless otherwise stated, the terms used herein (including technical terms) have their common meaning as understood by one of ordinary skill in the art. Furthermore, it is understood that terms defined in commonly used dictionaries should be understood to have a meaning consistent with the context of their relevant field, and not to be interpreted as having an idealized or overly formal meaning.

[0056] Example 1

[0057] This embodiment provides a fuel cell drone energy management system, including:

[0058] The multi-source data acquisition module is used to acquire multi-source operational information of the UAV in real time;

[0059] The information fusion and state construction module is used to fuse multi-source information through an attention mechanism network to construct an enhanced state vector;

[0060] The integrated prediction and uncertainty quantification module is used to load and run the integrated prediction model after inputting the enhanced state vector sequence, and output the load power prediction mean and uncertainty metric Uncertainty.

[0061] The risk-sensitive reinforcement learning decision module is used to load and run a deep reinforcement learning decision engine based on the Dyna-Q network framework and output power allocation instructions.

[0062] The power distribution execution module is used to convert the power distribution command into a control signal, drive the fuel cell and lithium battery system to execute the output power distribution command, and complete the real-time power distribution between the fuel cell and the lithium battery.

[0063] The online learning update module manages the storage of state transition data and performs model fine-tuning on the integrated prediction and uncertainty quantification module and the risk-sensitive reinforcement learning decision module.

[0064] The integrated prediction model consists of multiple parallel, identically structured gated recurrent unit (GRU) neural network sub-models, and each neural network sub-model contains a conditional normalization layer with real-time environmental information as the conditional input.

[0065] This invention aims to achieve an intelligent balance and unity between real-time performance, global economy, and operational safety in energy management under complex flight environments by constructing an enhanced state vector that integrates multi-source real-time information and combining an integrated prediction model with uncertainty quantification capabilities with a risk-sensitive reinforcement learning decision module.

[0066] Example 2

[0067] This embodiment provides an energy management method for fuel cell unmanned aerial vehicles, such as... Figure 1 As shown, it includes the following steps:

[0068] Step S1, Multi-source data acquisition, fusion, and enhanced state construction:

[0069] The multi-source data acquisition module collects multi-source operational data of the UAV in real time. The information fusion and state construction module fuses the multi-source operational data through an attention mechanism network to generate an enhanced state vector that comprehensively represents the system and the environmental situation.

[0070] Step S2, integrating prediction and uncertainty quantification:

[0071] The enhanced state vector sequence within the current and historical time windows is input into the integrated prediction model, which outputs the predicted mean of load power within a future set time domain and outputs the uncertainty metric of the current prediction.

[0072] Step S3, Risk-Sensitive Reinforcement Learning Decision:

[0073] The enhanced state vector, the mean load power prediction, and the uncertainty metric at the current moment are input together into a deep reinforcement learning decision-maker based on the Dyna-Q network framework. The decision-maker will guide the management strategy to become more conservative when the prediction uncertainty is high, based on its reward function, and output the optimal power allocation command for the fuel cell.

[0074] Step S4, Power Allocation Execution and Online Learning:

[0075] The power allocation execution module executes the optimal power allocation command to control the power output of the fuel cell and lithium battery to meet the load requirements;

[0076] Meanwhile, the online learning update module continuously collects state transition data generated by the actual operation of the UAV and stores it in the experience replay buffer. Then, it periodically samples data from the experience replay buffer to fine-tune the integrated prediction model and the deep reinforcement learning network in the decision maker online, so as to realize the continuous evolution of the energy management strategy.

[0077] To further illustrate the effects of this invention, this example uses a hexacopter UAV equipped with a hybrid power source of proton exchange membrane fuel cell and lithium battery as the implementation subject. Its core task is to perform a power line inspection for approximately 2 hours in a complex mountainous environment. This task encompasses various typical operating conditions such as cruising, hovering observation, headwind flight, and sudden maneuvers, placing extremely high demands on the adaptability and real-time performance of energy management strategies.

[0078] 1. System hardware and software configuration

[0079] (1) The NVIDIA Jetson Orin Nano embedded artificial intelligence module is used as the hardware core to provide computing power support for real-time inference and online learning of deep neural network models.

[0080] (2) Multi-source data acquisition system:

[0081] Power data: Employs a high-precision Hall current / voltage sensor to acquire load bus power at a frequency of 100Hz via the CAN bus. Fuel cell output current With voltage Lithium battery current With voltage It calculates the real-time battery state of charge (SoC) based on the ampere-hour integration method.

[0082] Flight status data: Based on an open-source flight controller (such as Pixhawk 6), the drone's pitch angle is acquired at a frequency of 50Hz via the MAVLink protocol. Roll angle Triaxial acceleration and flight altitude .

[0083] Environmental data: Integrates a miniature ultrasonic anemometer and temperature / humidity sensor, and reads real-time wind speed at a frequency of 10Hz via UART serial port. Wind angle and ambient temperature .

[0084] Mission command data: Receive mission phase commands (such as "cruise", "hover", "inspection", "return") from the ground control station (GCS) via data transmission link and encode them as discrete state quantities.

[0085] (3) Actuator: A programmable bidirectional DC / DC converter is used as the power interface of the fuel cell. It supports CAN or Ethernet communication, receives power commands issued by the main controller, and has complete overcurrent, overvoltage and overheat protection functions.

[0086] (4) Software architecture: Based on Robot Operating System 2 (ROS 2 Humble), it achieves modularity and loose coupling.

[0087] High-level intelligent decision-making layer: The integrated GRU prediction model and risk-sensitive Dyna-DQN decision-maker of this invention are implemented using the PyTorch framework and encapsulated as an independent ROS 2 node (energy_manager).

[0088] The underlying real-time control layer, written in C++, implements functions such as data acquisition, state fusion, command issuance, and security monitoring, ensuring hard real-time performance with a 50-millisecond control cycle. Efficient and reliable data exchange between the two layers is achieved through ROS 2 topics and services.

[0089] After the system is powered on, the parameters (.pt file) of the integrated GRU prediction model and the weights (.pth file) of the Dyna-DQN decision network, which were pre-trained on the server using large-scale historical flight data, are loaded into the main controller memory to complete the algorithm initialization.

[0090] 2. Method Implementation Steps

[0091] like Figure 2 As shown, it includes the following steps:

[0092] Step S1: Multi-source information fusion and enhanced state construction

[0093] At the start of each 50-millisecond control cycle, the system synchronously acquires data from all sensors. Raw data preprocessing employs a first-order low-pass filter to smooth noise and utilizes offline statistical mean and variance for Z-score standardization. The preprocessed data forms the basic feature vector. .

[0094] Input this vector into a lightweight attention network (structure: input layer → fully connected layer (32 neurons, ReLU) → fully connected layer (6 neurons) → softmax output layer). The network outputs a 6-dimensional attention weight vector. Finally, an enhanced state vector is generated by element-wise weighted summation. This vector can dynamically reflect the importance of each information source under different operating conditions. For example, when flying against the wind, the weight of wind speed features automatically increases.

[0095] Step S2: Load forecasting and uncertainty quantification based on integrated GRU network

[0096] The system maintains a sliding window of length 20 to store the enhanced state vector sequence within the most recent second. This sequence is input into an ensemble prediction model consisting of five parallel and structurally identical improved GRU sub-models. Each sub-model is a two-layer GRU network with 128 hidden units per layer, and a Conditional Layer Normalization (CLN) module is inserted between layers. The scaling and translation parameters of the CLN module are determined by a small feedforward network based on real-time environmental characteristics. Dynamically generated.

[0097] Each GRU sub-model independently outputs a load power prediction sequence for the next 10 steps (i.e., 0.5 seconds). The ensemble model output is as follows:

[0098] 1) Predicted average load power: This will be used as the final load forecast value.

[0099] 2) Uncertainty measurement: The variance quantifies the consistency of predictions; the larger the variance, the higher the uncertainty.

[0100] Step S3: Dyna-DQN Reinforcement Learning Decision Based on Risk-Sensitive Rewards

[0101] Current enhancement status Load forecast average and uncertainty measurement Concatenate as input state for decision maker .

[0102] The decision-maker is based on the Dyna-DQN framework, and its deep Q-network adopts a dual-network structure (evaluation network). and target network Both are fully connected networks with two hidden layers (256 neurons each). The action space is the fuel cell output power, discretized into 11 levels: .

[0103] The reward function is designed to be risk-sensitive.

[0104]

[0105] in:

[0106] It is a basic economic incentive item.

[0107] The risk sensitivity coefficient is set to 0.4 in this embodiment.

[0108] To constrain the penalty value for violations, it is set to 5.0.

[0109] This design makes the total reward negatively correlated with prediction uncertainty, guiding the agent to adopt a conservative strategy when uncertainty is high. The decision-maker then... Calculate the Q-value for each action and select the optimal action using an ε-greedy strategy (online ε=0.02). (Target power level for fuel cells).

[0110] Step S4: Power Allocation Execution and Online Learning

[0111] Power command The data is transmitted to the fuel cell DC / DC converter via the CAN bus. The output power of the lithium battery is determined by the power balance equation: .

[0112] Simultaneously, the online learning process is executed asynchronously. Transfer samples generated from actual interactions... An experience replay buffer with a capacity of 100,000 records is stored. The system initiates a fine-tuning learning process every 1000 control cycles (i.e., 50 seconds).

[0113] a) Randomly sample a batch of data (batch size=64) from the buffer.

[0114] b) Calculation The network's temporal difference loss, with a small learning rate (e.g. Perform one step of gradient descent to update its parameters.

[0115] c) Every 100 updates will Parameter soft update to Network (Update Coefficient) ).

[0116] d) Using newly acquired real load data, perform online fine-tuning of each sub-model in the integrated GRU prediction model with a small number of iterations.

[0117] This process enables continuous and incremental optimization of energy management strategies in real flight, giving the system long-term adaptability and robustness.

[0118] Example 3

[0119] Based on Example 2, this example provides a detailed explanation of how the four core steps (S1 to S4) of the energy management method described in this invention dynamically link together to form an intelligent closed loop of "perception-prediction-decision-learning" in a complete flight mission of a fuel cell UAV, which includes "climb-cruise-sudden maneuver-return".

[0120] (1) Phase 1: Climbing Phase (System Startup and Handling High Load)

[0121] a) S1 (Multi-source real-time information acquisition and fusion to enhance state construction): The system synchronously acquires data such as rapidly increasing load power, rapidly changing flight attitude, and initial wind speed. The attention network fuses this information, assigning higher weights to the current high power demand and attitude angle to generate an enhanced state vector representing "high-power climb".

[0122] b) S2 (Load Prediction and Uncertainty Quantification Based on Integrated GRU Network): This enhanced state sequence is input into the integrated prediction model. Due to the short historical sequence and drastic changes in load patterns at the beginning of the mission, the prediction results of each GRU sub-model show significant discrepancies, resulting in a moderately high level of uncertainty metric (prediction variance) in the output, indicating insufficient consistency in the model's predictions of the upcoming flight state.

[0123] c) S3 (Dyna-DQN Reinforcement Learning Decision Based on Risk-Sensitive Rewards): The decision-maker receives the augmented state with a moderately high uncertainty metric, based on the risk-sensitive reward function. At this point, the uncertainty penalty term As the load increases, the total reward value decreases relatively. This prompts the decision-maker to weigh the need to "satisfy the climbing power" against "preventing overly aggressive decisions under new operating conditions," outputting a robust fuel cell power command that ensures thrust while avoiding impact on the energy system.

[0124] d) S4 (Power Allocation Execution and Online Progressive Learning): The instruction is executed immediately. Simultaneously, the real data from this climb process is stored as the first important lesson in the experience replay buffer, initiating the online learning data accumulation for this task.

[0125] (2) Phase Two: Smooth Cruise Phase (Economic Optimization)

[0126] a) S1: The system continuously integrates data such as airspeed, stable attitude, and constant wind speed. At this point, environmental information and stable load become the main components of the state vector, and the attention network assigns high weights to these stable features.

[0127] b)S2: Based on long-term stable flight data, the GRU sub-models in the integrated prediction model show a high degree of consistency in their inferences about future load trends, and the output uncertainty measure (prediction variance) is very low, indicating high prediction reliability.

[0128] c)S3: Low uncertainty signals reduce the penalty term in the risk-sensitive reward function. The total reward is very small, and it mainly consists of economic incentives. Dominant. The decision-maker enters "economic optimization" mode, attempting to finely adjust the fuel cell power to operate within its high-efficiency range, and utilizes the battery for minor "peak shaving and valley filling" to minimize overall hydrogen consumption.

[0129] d)S4: The background learning thread starts and uses the current stable, high-quality data to fine-tune the parameters of the GRU sub-model and decision network in the integrated prediction model in batches, so that it can be more accurately adapted to today's cruise environment.

[0130] (3) Phase 3: Encountering sudden crosswinds and emergency maneuvers (safety first response)

[0131] a) S1: The sensor detects sudden changes in wind speed, drastic attitude adjustments, and sudden changes in load. The fusion network responds quickly, and the feature weights of wind speed and attitude angular velocity increase sharply in the enhanced state vector, thus depicting the "sudden disturbance" state in real time.

[0132] b)S2: Faced with an input that is very different from the historical smooth cruise mode, the prediction results of each GRU sub-model in the integrated prediction model show significant differences due to the lack of similar training experience. This causes the uncertainty measure of the output (prediction variance) to rise sharply to a high level, indicating that the system "realizes" that it is currently in a highly uncertain abnormal operating condition.

[0133] c) S3: The high-uncertainty signal is received by the decision-maker. According to the risk-sensitive reward function, the penalty term is then... The total reward value is significantly reduced as the system size increases dramatically. This means that any "optimization" action that could lead to state fluctuations has extremely low expected returns. As a result, the decision-maker ignores long-term optimization goals and switches to outputting a conservative and safe power command (such as maintaining the current power or making only minor adjustments) aimed at quickly stabilizing the system and ensuring power continuity.

[0134] d) S4: The complete state transition data for this critical emergency was marked as high-value experience and stored preferentially. This provides a core sample for subsequent online learning on how to respond to similar emergencies.

[0135] (4) Phase Four: Task Completion and Model Evolution

[0136] a) S1 / S2 / S3: As the drone returns and lands, the system resumes normal steady-state operation.

[0137] b) S4 (Offline Evolution): After the entire flight is completed, the ground station or cloud server gathers all the experience data accumulated during the mission (especially the rare data during the sudden maneuver phase) and performs a complete round of incremental training on the S2 integrated prediction model and the S3 decision network. After evolution, when encountering similar crosswind disturbances in the future, the S2 model's predictions will be more accurate, the prediction discrepancies among the sub-models will be smaller (i.e., the uncertainty metric will be lower), and the S3 decision-maker's response strategy will be better and more composed.

[0138] Example 4

[0139] This embodiment provides a fuel cell drone, including a flight platform, a hybrid power system, and a flight control computer. The flight control computer is configured with the energy management system described in Embodiment 1, or configured to execute the method described in Embodiment 2.

[0140] like Figure 3 As shown, the management system hardware includes a main controller, a fuel cell controller, a lithium battery controller, and a sensor array. The fuel cell controller, lithium battery controller, and sensor array are all electrically connected to the main controller. The fuel cell controller controls the output power of the fuel cell, and the lithium battery controller controls the output power of the lithium battery. The lithium battery and fuel cell provide power to the drone, controller, and sensor array.

[0141] In summary, this invention provides an energy management method for fuel cell unmanned aerial vehicles (UAVs) based on multi-source information fusion and uncertainty perception. This method constructs an enhanced state space that integrates real-time information from multiple sources, and combines an integrated prediction model with uncertainty quantification capabilities with a risk-sensitive deep reinforcement learning decision-maker to achieve an intelligent balance and unity between real-time performance, global economic efficiency, and operational safety in energy management under complex flight environments.

[0142] The examples above are merely illustrative of the invention and do not constitute a limitation on the scope of protection of the invention. Any design that is the same as or similar to the invention falls within the scope of protection of the invention.

Claims

1. A fuel cell unmanned aerial vehicle (UAV) energy management system, characterized in that, include: The multi-source data acquisition module is used to acquire multi-source operational information of the UAV in real time; The information fusion and state construction module is used to fuse multi-source information through an attention mechanism network to construct an enhanced state vector; The integrated prediction and uncertainty quantification module is used to load and run the integrated prediction model after inputting the enhanced state vector sequence, and output the load power prediction mean and uncertainty metric Uncertainty. The risk-sensitive reinforcement learning decision module is used to load and run a deep reinforcement learning decision engine based on the Dyna-Q network framework and output power allocation instructions. The power distribution execution module is used to convert the power distribution command into a control signal, drive the fuel cell and lithium battery system to execute the output power distribution command, and complete the real-time power distribution between the fuel cell and the lithium battery. The online learning update module manages the storage of state transition data and performs model fine-tuning on the integrated prediction and uncertainty quantification module and the risk-sensitive reinforcement learning decision module.

2. The fuel cell unmanned aerial vehicle energy management system according to claim 1, characterized in that, The integrated prediction model consists of multiple parallel, identically structured gated recurrent unit (GRU) neural network sub-models, and each neural network sub-model contains a conditional normalization layer with real-time environmental information as the conditional input.

3. A method for energy management of a fuel cell unmanned aerial vehicle (UAV), employing the fuel cell UAV energy management system described in claim 1 or 2, characterized in that, Includes the following steps: Step S1, Multi-source data acquisition, fusion, and enhanced state construction: The multi-source data acquisition module collects multi-source operational data of the UAV in real time. The information fusion and state construction module fuses the multi-source operational data through an attention mechanism network to generate an enhanced state vector that comprehensively represents the system and the environmental situation. Step S2, integrating prediction and uncertainty quantification: The enhanced state vector sequence within the current and historical time windows is input into the integrated prediction model, which outputs the predicted mean of load power within a future set time domain and outputs the uncertainty metric of the current prediction. Step S3, Risk-Sensitive Reinforcement Learning Decision: The enhanced state vector, the mean load power prediction, and the uncertainty metric at the current moment are input together into a deep reinforcement learning decision-maker based on the Dyna-Q network framework. The decision-maker will guide the management strategy to become more conservative when the prediction uncertainty is high, based on its reward function, and output the optimal power allocation command for the fuel cell. Step S4, Power Allocation Execution and Online Learning: The power allocation execution module executes the optimal power allocation command to control the power output of the fuel cell and lithium battery to meet the load requirements; Meanwhile, the online learning update module continuously collects state transition data generated by the actual operation of the UAV and stores it in the experience replay buffer. Then, it periodically samples data from the experience replay buffer to fine-tune the integrated prediction model and the deep reinforcement learning network in the decision maker online, so as to realize the continuous evolution of the energy management strategy.

4. The energy management method for a fuel cell unmanned aerial vehicle according to claim 3, characterized in that, The multi-source information mentioned in step S1 includes at least load power, battery state of charge (SoC), flight attitude information, environmental information, and current mission instructions; The environmental information includes wind speed, wind direction, and temperature.

5. The energy management method for a fuel cell unmanned aerial vehicle according to claim 3, characterized in that, The method for constructing the enhanced state vector based on multi-source real-time information in step S1 is as follows: The multi-source real-time information is input into an attention mechanism network to calculate the dynamic weight of each information source at the current time. Then, the normalized features of each information source are weighted and fused according to the dynamic weight to generate an enhanced state vector that comprehensively represents the real-time state of the system and the environment.

6. The energy management method for a fuel cell unmanned aerial vehicle according to claim 3, characterized in that, The attention mechanism network described in step S1 is a lightweight feedforward neural network, whose output is a weight vector normalized by the Softmax function.

7. The energy management method for a fuel cell unmanned aerial vehicle according to claim 3, characterized in that, The method for calculating the uncertainty metric Uncertainty mentioned in step S2 is as follows: The integrated prediction model summarizes the prediction results of all network neural sub-models, outputs the predicted mean of load power, and then calculates the variance of the predicted values ​​output by each network neural sub-model.

8. The energy management method for a fuel cell unmanned aerial vehicle according to claim 3, characterized in that, The risk-sensitive reinforcement learning decision module described in step S4 is based on the Dyna-Q network framework, and its reward function is... It is risk-sensitive, specifically in the form of: ； in, This represents the basic economic incentive item based on hydrogen consumption, battery SoC maintenance, and fuel cell durability design; λ is a risk sensitivity coefficient greater than zero. The The expression is: ； in, Hydrogen consumption rate, This is a reference value for the battery's state of charge. For fuel cell current, , , It is a positive number.

9. The energy management method for a fuel cell unmanned aerial vehicle according to claim 3, characterized in that, The online fine-tuning in step S4 specifically includes: randomly sampling small batches of data from the stored state transition data, updating the parameters of the ensemble prediction model and the deep reinforcement learning network using a gradient descent algorithm at a learning rate lower than the initial training rate.

10. A fuel cell unmanned aerial vehicle (UAV), comprising a flight platform, a hybrid power system, and a flight control computer, characterized in that, The flight control computer is configured with the energy management system as described in claim 1 or 2, or is configured to perform the method as described in any one of claims 3-9.

Images