Unmanned aerial vehicle electric energy guarantee pre-scheduling method based on energy consumption prediction

Abstract

The invention relates to an unmanned aerial vehicle electric energy guarantee pre-scheduling method based on energy consumption prediction, and the method comprises the following steps: collecting historical task data and real-time states of an unmanned aerial vehicle, and carrying out the preprocessing of the historical task data and real-time states to generate a standardized feature set; according to the preprocessing features, constructing an LSTM prediction model, and integrating dynamic robust optimization to obtain prediction energy consumption and prediction uncertainty; generating an initial supply plan by using the predicted energy consumption and the predicted uncertainty, including charging station scheduling and time points; monitoring real-time data, and dynamically adjusting the supply plan by using the prediction uncertainty as a threshold value; estimating the overall performance, and generating a feedback data iterative optimization model. The historical task data and the real-time state of the unmanned aerial vehicle are utilized to construct the energy consumption prediction model, and the electric energy supply plan is generated in advance, so that energy exhaustion is prevented, and the task stability is ensured.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) technology, and particularly relates to a pre-scheduling method for ensuring the power supply of UAVs based on energy consumption prediction. Background Technology

[0002] In disaster response, drone swarm search and rescue operations require the deployment of multiple drones in earthquake-stricken areas for real-time monitoring and material delivery, ensuring continuous operation of the drone swarm in complex environments. Existing technology includes a Chinese patent application (CN119940869A) that discloses a drone swarm task optimization method based on master-slave game theory and steady-state matching. This method collects task information during drone deployment, clusters all users, and assigns a drone to each user cluster for task offloading. Drones engage in master-slave game theory with each user task, aiming to optimize their own energy consumption until a Nash equilibrium is reached. This method combines game theory and optimization strategies to allocate drone task assignments and manage energy consumption. While this method aims to optimize the energy consumption of individual drones and can achieve local optima, the ultimate goal for a drone swarm is to reduce the overall energy consumption of the entire swarm and achieve global optima. Summary of the Invention

[0003] This application addresses the energy depletion problem faced by unmanned aerial vehicles (UAVs) during long-duration missions by proposing a pre-scheduling method based on energy consumption prediction. This method utilizes historical mission data and real-time status of the UAV to construct an energy consumption prediction model, generating a power replenishment plan in advance to prevent energy depletion and ensure mission continuity and stability.

[0004] To achieve the above objectives, the present application discloses a pre-scheduling method for ensuring the power supply of unmanned aerial vehicles (UAVs) based on energy consumption prediction, which includes the following steps: S1 collects historical mission data and real-time status of UAVs, and preprocesses them to generate standardized feature sets; S2 constructs an LSTM prediction model based on preprocessed features, incorporates dynamic robust optimization, and obtains the predicted energy consumption. and forecast uncertainty ; S3 utilizes predicted energy consumption and forecast uncertainty Generate an initial supply plan, including charging station scheduling and timing. S4 monitors real-time data and uses predictive uncertainty. As a threshold, for the supply plan Make dynamic adjustments; S5 evaluates overall performance and generates feedback data to iteratively optimize the model.

[0005] Further, step S1 includes: Collect historical data ,in For task timestamps, For speed, For load, This represents actual energy consumption. Collect real-time data , This is the current battery level. For wind speed, For temperature; Calculate the wind resistance influence factor in air density, The drag coefficient, For the cross-sectional area of ​​the drone, For speed; Use an uncertainty-based imputation method to complete missing data. It is to replenish energy consumption; It has poor energy consumption; It is an uncertainty factor; It is the standard deviation of wind speed; It is the average wind speed; It is the optimal temperature; It is a thermodynamic constant; Constructing standardized feature vectors .

[0006] Further, step S2 includes: Training historical data using an LSTM network; Predicted energy consumption: It is a prediction of energy consumption. It is weight; It is a robustness factor. That is the maximum energy consumption; It is the total battery capacity. Actual energy consumption This is the current battery level. It is a factor affecting wind resistance; Calculate the uncertainty of prediction: , Indicates variance; Generate confidence intervals .

[0007] Further, step S3 includes: Set the optimization objective to minimize the risk of energy depletion: , in, To replenish energy, It is a prediction of energy consumption; The formula for outputting the supply plan is as follows: It's a supply plan; It is a scheduling variable; It is the gravitational kinetic energy factor; It's about the quality of the drone; It is the efficiency coefficient; It is a high degree of change. It is about predicting uncertainty. It's the speed of the drone; Output the generated replenishment time series , For supply volume, This is the task timestamp.

[0008] Further, step S4 includes: Compare actual energy consumption With prediction Calculate the deviation ; Dynamic optimization generates an updated plan: in, It's an update plan; It's a deviation; It is the standard deviation of energy consumption. It is about predicting uncertainty; if This triggered a supply adjustment and update. for .

[0009] Further, step S5 includes: Calculate continuity index Total time It is a prediction of energy consumption. Current battery level; Generate a feedback score; the formula for the feedback score is as follows: It is a feedback score; It is continuity. That is the maximum energy consumption; Use feedback scores Update the parameter distribution of the Bayesian LSTM The feedback is then sent to step S2.

[0010] Further, in step S1, data cleaning involves multi-scale feature extraction using improved wavelet decomposition, and dynamically adjusting the threshold of the wavelet decomposition based on real-time environmental fluctuations in the UAV mission data, as shown in the following formula: in, This is the completed dataset; It is a wavelet decomposition function with dynamic threshold adjustment; It is historical mission data; It is real-time status data; It is a dynamic threshold that adaptively adjusts according to environmental fluctuations; It is the standard deviation of environmental parameters; It is the average value of the environmental parameters at the current moment; It is the change in energy consumption during task execution; This is the maximum energy consumption capacity of the drone; These are adjustment parameters, calibrated through regression analysis of historical data.

[0011] Furthermore, in step S1, the uncertainty filling formula is as follows: in, It is a Bayesian correction term. This is the energy consumption value after completion; It is the energy consumption value at the previous moment; It is the change in energy consumption; It is an uncertainty factor; uncertainty factor The calculation is as follows: Bayesian correction term The calculation formula is: Posterior distribution parameter update: It is the standard deviation of wind speed; It is the average wind speed; It is a smoothing parameter to avoid the denominator being zero; It is the current environmental and resource pressure index; It is an environmental carrying capacity parameter, which characterizes the maximum tolerance of the environment to fluctuations in energy consumption; This is the current ambient temperature; This is the optimal temperature for drone operation; It is a dynamic thermodynamic constant that adaptively adjusts according to mission type and environmental fluctuations; It involves adjusting parameters to control the impact of temperature deviation on uncertainty factors. It follows a posterior normal distribution; These are the prior distribution mean and variance; These are the mean and variance of the observed data; It is a confidence score for completing the energy consumption value; It is the maximum confidence level.

[0012] Furthermore, step S1 also includes: to cope with extreme weather or sudden disturbances, an environmental pressure triggering mechanism is introduced, which automatically switches to a high uncertainty mode when environmental parameters such as wind speed and temperature exceed preset thresholds, increasing the uncertainty. The weights are adjusted to tighten the variance of the Bayesian posterior distribution. Triggering condition formula: The adjusted factor is calculated as follows under the high uncertainty mode: It is the threshold of environmental fluctuation; It is the temperature deviation threshold; It is an uncertainty amplification factor; It is the adjusted factor under the high uncertainty mode.

[0013] Furthermore, step S2 calculates the biomimetic robustness factor based on the dynamic adjustment mechanism of multi-agent cooperative behavior and in conjunction with the group energy allocation strategy. as follows: It is the first A drone in time The biomimetic robustness factor; It is the first A drone in time The wind drag function is estimated using real-time sensor data. It is based on the environmental adaptation weights of the attention mechanism, reflecting the first The drone's sensitivity to current wind resistance. It is the first The highest energy consumption in the history of drones It is the first The current battery level of the drone. It is the first The total battery capacity of the drone, It is a group synergy factor, reflecting the first A drone in time Influenced by the group's energy allocation strategy; Calculate environment-adaptive weights using a lightweight attention network. Dynamically assess wind resistance The importance of energy consumption prediction is calculated using the following formula: in, , , These are the queries, keys, and values ​​extracted from wind resistance features and environmental features through a small fully connected layer. For feature dimensions; Group synergy factor The calculation is as follows: in, For the number of drone swarms, For the first The remaining energy of the drone The average remaining energy of the population. To adjust the parameters; Using Bayesian LSTM, a posterior distribution is introduced into the model parameters to quantify the uncertainty of the prediction and provide confidence intervals for subsequent replenishment planning. The energy consumption prediction expression is as follows: It is the first A drone in time Predicted energy consumption It is the first The input feature sequence of the drone, The parameter distribution of the Bayesian LSTM follows a posterior distribution. , It is the uncertainty term in the prediction and follows a normal distribution; The confidence intervals output by the Bayesian LSTM provide a risk assessment basis for the generation of supply plans; By combining prediction error and uncertainty estimation through a composite loss function, energy consumption prediction and replenishment decisions can be optimized. in, Kullback-Leibler divergence is used to regularize the variational posterior distribution. With the true posterior The differences between them For balance parameters; In multi-drone collaborative tasks, dynamic replenishment scheduling is performed based on multi-agent reinforcement learning; the state space of the reinforcement learning includes the predicted energy consumption of each UAV. Remaining energy Given task priority, action space as a supply plan, and reward function: It is time Total group reward These are weighting coefficients that adjust the energy shortage penalty, supply cost, and task completion reward, respectively. It is the first A drone in time Supply costs, It is time Task completion metrics. Attached Figure Description

[0014] Figure 1 This is a flowchart of the UAV power supply pre-scheduling method based on energy consumption prediction provided in the embodiments of this application.

[0015] Figure 2 This is a framework diagram of a UAV power supply pre-scheduling system based on energy consumption prediction provided in an embodiment of this application. Detailed Implementation

[0016] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0017] refer to Figure 1 The method for pre-scheduling power supply for unmanned aerial vehicles (UAVs) based on energy consumption prediction disclosed in this application includes the following steps: S1 collects historical mission data and real-time status of UAVs, and preprocesses them to generate standardized feature sets; S2 constructs an LSTM prediction model based on preprocessed features, incorporates dynamic robust optimization, and obtains the predicted energy consumption. and forecast uncertainty ; S3 utilizes predicted energy consumption and forecast uncertainty Generate an initial supply plan, including charging station scheduling and timing. S4 monitors real-time data and uses predictive uncertainty. As a threshold, for the supply plan Make dynamic adjustments; S5 evaluates overall performance and generates feedback data to iteratively optimize the model.

[0018] Step S1: Data Acquisition and Preprocessing: Collect historical mission data and real-time status of the UAV, such as flight path, speed, and load, including current battery level, wind speed, and temperature, and perform preprocessing to generate a standardized feature set.

[0019] S11 Data Collection: Collect historical data ,in For task timestamps, For speed, For load, Actual energy consumption; real-time data , This is the current battery level. For wind speed, For temperature.

[0020] Calculate the preliminary wind resistance influence factor , in, air density, The drag coefficient, For the cross-sectional area of ​​the drone, air density; For speed.

[0021] S12 Data Cleaning and Completion: Introducing Uncertainty to Fill Missing Data It is to replenish energy consumption; It has poor energy consumption; It is an uncertainty factor; It is the standard deviation of wind speed; It is the average wind speed; It is the optimal temperature; It is a thermodynamic constant.

[0022] S13 Extracts standardized feature vectors This reduces the impact of noise and improves the accuracy of the prediction model.

[0023] In one embodiment, step S2 is based on preprocessing features. We construct an LSTM-based prediction model to predict future energy consumption curves.

[0024] Step S21 Model Initialization Training historical data using an LSTM network: ,in These are the model parameters.

[0025] S22 Predicted Energy Consumption: It is a predicted energy consumption, measured in joules; It is weight; It is the robustness factor, with a value between -1 and 1; That is the maximum energy consumption; This represents the total battery capacity. The above formula, combined with a sine function, simulates the periodic energy adjustments of organisms.

[0026] S23 Calculate prediction uncertainty: Prediction uncertainty , Generate confidence intervals . This step constructs a high-precision prediction model and outputs... and This provides a quantitative basis for pre-scheduling and ensures robustness in uncertain environments such as changes in wind speed during disasters.

[0027] In one embodiment, in step S3, using and Generate an initial supply plan, including charging station scheduling and timing.

[0028] Step S31: Set the optimization objective to minimize the risk of energy depletion. , in To replenish energy.

[0029] The S32 output supply plan, the formula is as follows: It's a supply plan; It is a scheduling variable; It is the gravitational kinetic energy factor; It's about the quality of the drone; It is the efficiency coefficient; It is highly variable.

[0030] Output the generated replenishment time series , This refers to the amount of supplies.

[0031] This step generates a supply plan in advance. To prevent energy depletion and ensure that drones are not interrupted in rescue operations due to power outages in disaster scenarios.

[0032] In one embodiment, step S4 monitors real-time data, using the data from the previous step. As a threshold pair Make dynamic adjustments.

[0033] S41 monitors in real time and compares actual energy consumption. and Calculate the deviation . S42 Dynamic Optimization: in, It's an update plan; It's a deviation; It is the standard deviation of energy consumption.

[0034] if This triggered a supply adjustment and update. for . This step involves real-time adjustments to the plan to ensure stability in dynamic environments with sudden wind speed changes. Achieve self-adaptation.

[0035] In one embodiment, step S5 evaluates the overall performance and generates feedback data to iteratively optimize the model.

[0036] S51 Calculate the continuity index: This represents the total time.

[0037] S52 generates feedback scores, calculated using the following formula: It is a feedback score; It is continuity.

[0038] use renew Feedback is sent to module 2.

[0039] This step evaluates the effectiveness of the plan and generates feedback. Iterate the system to improve long-term accuracy and stability.

[0040] In one embodiment, the data cleaning in step S1 involves multi-scale feature extraction using improved wavelet decomposition: the threshold of wavelet decomposition is dynamically adjusted based on real-time environmental fluctuations such as wind speed and temperature change rate in the UAV mission data, instead of using a fixed threshold.

[0041] The formula is as follows: in, This is the completed dataset; It is a wavelet decomposition function with dynamic threshold adjustment; It is historical mission data; It is real-time status data; It is a dynamic threshold that adaptively adjusts according to environmental fluctuations; It is the standard deviation of environmental parameters such as wind speed and temperature at the current moment; It is the average value of the environmental parameters at the current moment; It is the change in energy consumption during task execution; This is the maximum energy consumption capacity of the drone; These are adjustment parameters, calibrated through regression analysis of historical data.

[0042] Adaptive thresholds based on environmental fluctuations and changes in task energy consumption Wavelet decomposition can dynamically respond to environmental disturbances, resource fluctuations, and external pressures, and make corresponding adjustments accordingly.

[0043] In one embodiment, the uncertainty filling formula in step S1 is as follows: Among them, uncertainty factor The calculation is as follows: Bayesian correction term The calculation formula is: Posterior distribution parameter update: This is the energy consumption value after completion; It is the energy consumption value at the previous moment; It is the change in energy consumption; It is the improved uncertainty factor, ranging from 0 to 1; It is the standard deviation of wind speed; It is the average wind speed; It is a smoothing parameter to avoid the denominator being zero; It is the current environmental resource pressure index, calculated based on factors such as wind speed and temperature; It is an environmental carrying capacity parameter, which characterizes the maximum tolerance of the environment to fluctuations in energy consumption; This is the current ambient temperature; This is the optimal temperature for drone operation; It is a dynamic thermodynamic constant that adaptively adjusts according to mission type and environmental fluctuations; It involves adjusting parameters to control the impact of temperature deviation on uncertainty factors; It is a Bayesian correction term, which introduces probability distribution estimation; It follows a posterior normal distribution; These are the prior distribution mean and variance; These are the mean and variance of the observed data; It is a confidence score for completing the energy consumption value; It is the maximum confidence level, used for normalization.

[0044] The simulation term addresses the dynamic equilibrium of UAV energy consumption under environmental pressure, allowing uncertainty factors to better reflect the nonlinear impact of the environment on energy consumption. This is achieved through posterior distribution estimation and confidence correction terms. This provides a probability confidence interval for completing the energy consumption value, avoids the problem of uncertainty factors relying too much on static parameters, and improves the robustness of the model in extreme environments such as sudden strong winds or sudden temperature changes. and The parameters are adaptively updated based on the task type and real-time environmental fluctuations, enhancing the model's generalization ability.

[0045] In one embodiment, step S1 further includes: to cope with extreme weather or sudden disturbances, an environmental pressure triggering mechanism is introduced, which automatically switches to a high uncertainty mode when environmental parameters such as wind speed and temperature exceed a preset threshold, thereby increasing the uncertainty. The weights are adjusted to tighten the variance of the Bayesian posterior distribution. This is to improve the conservatism of the completion results.

[0046] Triggering condition formula: Under high uncertainty mode: It is the threshold of environmental fluctuation; It is the temperature deviation threshold; It is an uncertainty amplification factor; It is the adjusted factor under the high uncertainty mode.

[0047] Through triggering mechanisms and dynamic adjustment strategies, the model automatically switches to conservative completion mode under extreme conditions. Combined with confidence quantification, this ensures that the completion results do not lead to overly optimistic or pessimistic energy consumption predictions. Dynamic parameter adjustment and adaptive threshold mechanisms make the model applicable to different task scenarios and environmental conditions. The Bayesian correction term provides a probability confidence interval for the completion results, improving the reliability and interpretability of the energy consumption prediction model.

[0048] In one embodiment, step S2 introduces a robust factor that dynamically adjusts based on multi-agent cooperative behavior, combined with a group energy allocation strategy, to simulate the energy reserves and environmental adaptation behavior of bird flocks during long-distance migration. Robust Factor The expression is as follows: It is the first A drone in time The biomimetic robustness factor has a value range of 1. . It is the first A drone in time The wind drag function is estimated using real-time sensor data. It is based on the environmental adaptation weights of the attention mechanism, reflecting the first The drone's sensitivity to current wind resistance. It is the first The highest energy consumption ever recorded for a drone. It is the first The current battery level of the drone. It is the first Total battery capacity of the drone. It is a group synergy factor, reflecting the first A drone in time Influenced by the group's energy allocation strategy.

[0049] A lightweight attention network is used to dynamically assess wind resistance. The importance of energy consumption forecasting. The calculation formula is: in, , , These are the queries, keys, and values ​​extracted from wind resistance and environmental features through a small fully connected layer. For feature dimensions.

[0050] Group synergy factor Simulating the group energy allocation behavior during bird foraging, computation is performed through information sharing in a multi-agent system: in, For the number of drone swarms, For the first The remaining energy of the drone The average remaining energy of the population. To adjust parameters, the stability of collaborative tasks is enhanced by encouraging drones with uneven energy distribution to adjust their own energy consumption predictions through "collective consensus."

[0051] In one embodiment, step S2 uses a Bayesian LSTM to introduce a posterior distribution into the model parameters, quantifying the uncertainty of the prediction and providing confidence intervals for subsequent replenishment planning. The energy consumption prediction expression is: It is the first A drone in time Predicted energy consumption. It is the first The input characteristic sequence of the drone includes historical energy consumption, flight speed, wind resistance, and mission payload. The parameter distribution of the Bayesian LSTM follows a posterior distribution. Approximations are derived through variational inference. It is the uncertainty term for prediction, which follows a normal distribution. ,in It is the variance estimate of the Bayesian LSTM output.

[0052] Confidence intervals (e.g., 95% confidence intervals) output by Bayesian LSTM This provides a risk assessment basis for generating replenishment plans. For example, when the lower limit of predicted energy consumption approaches the remaining energy threshold, a replenishment warning is triggered.

[0053] By combining prediction error and uncertainty estimation through a composite loss function, energy consumption prediction and replenishment decisions can be optimized simultaneously. in, Kullback-Leibler divergence is used to regularize the variational posterior distribution. With the true posterior The differences between them These are the balancing parameters.

[0054] In one embodiment, step S2 treats the drone swarm as a multi-agent system, collaboratively optimizing the power supply plan. In this multi-drone collaborative task, dynamic replenishment scheduling is performed based on multi-agent reinforcement learning, jointly optimized with an energy consumption prediction model. The state space includes the predicted energy consumption of each drone. Remaining energy The task priority and action space are defined as a supply plan, including whether to supply, the amount of supply, and the supply time. The reward function is designed as follows: It is time The total group reward. These are weighting coefficients that adjust the energy shortage penalty, supply cost, and task completion reward, respectively. It is the first A drone in time The cost of resupply includes time and energy consumption. It is time Task completion metrics.

[0055] By using MARL, the uncertainty of energy consumption prediction, i.e. the Bayesian confidence interval, is directly embedded into the replenishment decision, forming a prediction-optimization closed-loop system, which improves the continuity and stability of the task.

[0056] The reward function incorporates group task completion rate to encourage collaborative behavior of drone swarms in energy allocation and supply scheduling, drawing on the "altruism" concept from biological group behavior.

[0057] Through group synergy factors Optimize with MARL, simulate bird flock energy allocation strategies, and improve multi-machine task collaboration. Dynamic environment attention weights. Adaptive adjustment to wind resistance enhances the model's adaptability to complex environments. Bayesian LSTM provides prediction confidence intervals, offering risk assessment for resupply decisions and addressing numerical estimation noise issues. A composite loss function and closed-loop optimization design achieve seamless integration between the prediction model and resupply scheduling.

[0058] In this embodiment, The system employs a multi-level dynamic adjustment and a Bayesian uncertainty-driven decision-making mechanism, and in practical deployments, quantitative calculations can be achieved through numerical integration, attention networks, and reinforcement learning algorithms, providing a more robust solution for UAV energy consumption prediction and power supply pre-scheduling. refer to Figure 2 The UAV power supply pre-scheduling system based on energy consumption prediction disclosed in this application includes the following modules: Standardized feature generation module: Collects historical mission data and real-time status of UAVs, and preprocesses them to generate standardized feature sets; LSTM prediction model module: Based on preprocessed features, an LSTM prediction model is constructed, incorporating dynamic robust optimization to obtain the predicted energy consumption. and forecast uncertainty ; Initial supply plan generation module: utilizing predicted energy consumption and forecast uncertainty Generate an initial supply plan, including charging station scheduling and timing. Supply plan dynamic adjustment module: monitors real-time data and uses prediction of uncertainty. As a threshold, for the supply plan Make dynamic adjustments; Iterative optimization module: Evaluates overall performance and generates feedback data for iterative optimization model.

[0059] To verify the effectiveness and robustness of this application, a series of experiments were designed to test the solution's mission continuity, energy consumption prediction accuracy, and dynamic adjustment capability of the supply plan in complex environments, focusing on drone swarm search and rescue missions in disaster response scenarios.

[0060] The experimental data comes from the following three parts: This dataset, collected from disaster response missions conducted by a drone operator between 2019 and 2022, contains 5,000 flight records covering scenarios such as earthquake monitoring and flood relief delivery. Each record includes fields such as timestamp, flight speed, payload weight, and actual energy consumption. The data was acquired through the drone's built-in sensors and mission logs. Some data is missing, with approximately 15% of energy consumption values ​​and 10% of environmental parameters missing.

[0061] Real-time data collection was performed using simulated disaster scenarios, including current battery level, wind speed, and ambient temperature. Data was sourced from drone sensors and ground-based weather stations, with a sampling frequency of 1 Hz.

[0062] Generated using meteorological datasets and wind resistance physical models published by the National Meteorological Administration, the model simulates extreme environments (such as strong winds and sudden temperature changes) in disaster scenarios to test its robustness under high uncertainty conditions.

[0063] The experiment used a high-performance computing server equipped with an NVIDIA RTX 3090 GPU, an Intel i9-10900K CPU, and 64GB of RAM. Python 3.8 was used as the development language, TensorFlow 2.5 was used to build LSTM and Bayesian LSTM models, MATLAB was used for wavelet decomposition and wind resistance simulation, and PyTorch was used for the attention mechanism and the implementation of Multi-Agent Reinforcement Learning (MARL). The Gazebo simulation environment was used to simulate multiple DJI Matrice 300 RTK drones performing disaster response missions, with a maximum flight time of approximately 55 minutes.

[0064] Experimental results show that in the 24-hour disaster response simulation, the mission continuity index of the UAV swarm reached 92.3%, higher than the target value of 90%. The confidence interval coverage of the improved Bayesian LSTM model was 93.5%. The dynamically adjusted resupply plan was triggered 18 times / 24 hours, with an average resupply delay of 3.2 minutes and an energy depletion event rate of 0.5%.

[0065] This application is compared with the following three benchmark methods: Traditional LSTM: lacks biomimetic robustness factor and uncertainty quantification.

[0066] Static wavelet decomposition (Static WD): without dynamic threshold adjustment and ecological dynamic uncertainty filling.

[0067] Rule-based replenishment scheduling: Replenishment is triggered based on a fixed power threshold, without the support of a predictive model.

[0068] Experimental results show that this application outperforms the benchmark method in terms of mission continuity, prediction accuracy, and resupply efficiency. In particular, it reduces the risk of energy depletion rate by more than 75%, verifying the effectiveness of the biomimetic robustness factor and Bayesian uncertainty quantification.

[0069] To further analyze the contribution of each innovative module, ablation experiments were conducted, and the following components were removed and the tests were repeated: No dynamic threshold adjustment: Wavelet decomposition uses a fixed threshold.

[0070] No group synergy factor: Remove In item.

[0071] No Bayesian uncertainty: Use traditional LSTM instead of Bayesian LSTM.

[0072] Removing any component leads to a performance degradation. Bayesian uncertainty contributes the most to prediction accuracy and energy depletion rate, while dynamic threshold adjustment and group synergy factor significantly improve task continuity.

[0073] Experimental results show that this application significantly improves the accuracy of UAV mission continuity and energy consumption prediction in disaster response scenarios, outperforming traditional methods and benchmark schemes. Ablation experiments further demonstrate the contributions of dynamic threshold adjustment, population synergy factor, and Bayesian uncertainty quantification to overall performance.

[0074] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A pre-scheduling method for ensuring power supply for unmanned aerial vehicles (UAVs) based on energy consumption prediction, characterized in that, Includes the following steps: S1 collects historical mission data and real-time status of UAVs, and preprocesses them to generate standardized feature sets; S2 constructs an LSTM prediction model based on preprocessing features, incorporates dynamic robust optimization, and obtains the predicted energy consumption and prediction uncertainty. S3 uses predicted energy consumption and predicted uncertainty to generate an initial replenishment plan, including charging station scheduling and timing. S4 monitors real-time data and uses forecast uncertainty as a threshold to dynamically adjust supply plans; S5 evaluates overall performance and generates feedback data to iteratively optimize the model.

2. The pre-scheduling method for UAV power supply based on energy consumption prediction according to claim 1, characterized in that, Step S1 includes: Collect historical data and collect real-time data; Calculate the wind resistance factor; Use an uncertain imputation method to complete the missing data; Construct standardized feature vectors.

3. The pre-scheduling method for UAV power supply based on energy consumption prediction according to claim 1, characterized in that, Step S2 includes: Training historical data using an LSTM network; Predict energy consumption; Calculate the uncertainty of forecasting; Generate confidence intervals.

4. The pre-scheduling method for UAV power supply based on energy consumption prediction according to claim 1, characterized in that, Step S3 includes: The optimization objective is set to minimize the risk of energy depletion; Output supply plan; Output the generated replenishment time series.

5. The pre-scheduling method for UAV power supply based on energy consumption prediction according to claim 1, characterized in that, Step S4 includes: Compare actual energy consumption with predictions and calculate the deviation; Dynamically optimize and generate updated plans; If the deviation exceeds the forecast uncertainty, a supply adjustment is triggered, and the plan is updated.

6. The pre-scheduling method for UAV power supply based on energy consumption prediction according to claim 1, characterized in that, Step S5 includes: Calculate continuity indicators; Generate feedback scores; use the feedback scores to update the parameter distribution of the Bayesian LSTM and feed them back to step S2.

7. The pre-scheduling method for UAV power supply based on energy consumption prediction according to claim 1, characterized in that, Step S1: Data cleaning: Multi-scale feature extraction is performed using improved wavelet decomposition, and the threshold of wavelet decomposition is dynamically adjusted according to the real-time environmental fluctuations of the UAV mission data, as shown in the following formula: in, This is the completed dataset; It is a wavelet decomposition function with dynamic threshold adjustment; It is historical mission data; It is real-time status data; It is a dynamic threshold that adaptively adjusts according to environmental fluctuations; It is the standard deviation of environmental parameters; It is the average value of the environmental parameters at the current moment; It is the change in energy consumption during task execution; This is the maximum energy consumption capacity of the drone; These are adjustment parameters, calibrated through regression analysis of historical data.

8. The pre-scheduling method for UAV power supply based on energy consumption prediction according to claim 1, characterized in that, In step S1, the uncertainty filling formula is as follows: in, It is a Bayesian correction term. This is the energy consumption value after completion; It is the energy consumption value at the previous moment; It is the change in energy consumption; It is an uncertainty factor; uncertainty factor The calculation is as follows: Bayesian correction term The calculation formula is: Posterior distribution parameter update: It is the standard deviation of wind speed; It is the average wind speed; It is a smoothing parameter to avoid the denominator being zero; It is the current environmental and resource pressure index; It is an environmental carrying capacity parameter, which characterizes the maximum tolerance of the environment to fluctuations in energy consumption; This is the current ambient temperature; This is the optimal temperature for drone operation; It is a dynamic thermodynamic constant that adaptively adjusts according to mission type and environmental fluctuations; It involves adjusting parameters to control the impact of temperature deviation on uncertainty factors. It follows a posterior normal distribution; These are the prior distribution mean and variance; These are the mean and variance of the observed data; It is a confidence score for completing the energy consumption value; It is the maximum confidence level.

9. The pre-scheduling method for UAV power supply based on energy consumption prediction according to claim 1, characterized in that, Step S1 also includes: to cope with sudden disturbances, when environmental parameters exceed a preset threshold, automatically switching to a high uncertainty mode to increase the uncertainty. The weights are adjusted to tighten the variance of the Bayesian posterior distribution. ; The adjusted factor is calculated as follows under the high uncertainty mode: It is the threshold of environmental fluctuation; It is the temperature deviation threshold; It is an uncertainty amplification factor; It is the adjusted factor under the high uncertainty mode.

10. The pre-scheduling method for UAV power supply based on energy consumption prediction according to claim 1, characterized in that, Step S2 calculates the biomimetic robustness factor based on the dynamic adjustment mechanism of multi-agent cooperative behavior and combined with the group energy allocation strategy. as follows: It is the first A drone in time The biomimetic robustness factor; It is the first A drone in time The wind drag function is estimated using real-time sensor data. It is based on the environmental adaptation weights of the attention mechanism, reflecting the first The drone's sensitivity to current wind resistance. It is the first The highest energy consumption in the history of drones It is the first The current battery level of the drone. It is the first The total battery capacity of the drone It is a group synergy factor, reflecting the first A drone in time Influenced by the group's energy allocation strategy; Calculate environment-adaptive weights using a lightweight attention network. Dynamically assess wind resistance The importance of energy consumption prediction is calculated using the following formula: in, , , These are the queries, keys, and values ​​extracted from wind resistance features and environmental features through a small fully connected layer. For feature dimensions; By using Bayesian LSTM, a posterior distribution is introduced into the model parameters to quantify the uncertainty of the prediction and provide confidence intervals for the generation of subsequent supply plans. The confidence intervals output by the Bayesian LSTM provide a risk assessment basis for the generation of supply plans; By combining prediction error and uncertainty estimation through a composite loss function, energy consumption prediction and replenishment decisions can be optimized. in, Kullback-Leibler divergence is used to regularize the variational posterior distribution. With the true posterior The differences between them For balance parameters; In multi-drone collaborative tasks, dynamic replenishment scheduling is performed based on multi-agent reinforcement learning; the state space of the reinforcement learning includes the predicted energy consumption of each UAV. Remaining energy Given task priority, action space as a supply plan, and reward function: It is time Total group reward These are weighting coefficients that adjust the energy shortage penalty, supply cost, and task completion reward, respectively. It is the first A drone in time Supply costs, It is time Task completion metrics.

Images