A hybrid UAV energy management method based on deep learning and dynamic programming

Abstract

The application relates to a hybrid unmanned aerial vehicle energy management method based on deep learning and dynamic programming, and belongs to the technical field of prediction of energy management strategies. The method is based on a Bi-LSTM bidirectional memory network to accurately predict power demand of an aircraft in a future time domain, and a prediction module considering environmental information has high generalization and flexibility for model prediction, so that the prediction result is more comprehensive and accurate, and the parameter complexity of direct prediction is reduced; on the basis of deep learning prediction, a rolling optimization module is constructed based on a linear programming algorithm to feed back and correct the prediction result, and more accurate prediction results are provided; on the basis of minimum equivalent fuel consumption, an energy management optimization function is constructed, the overall fuel consumption is minimized and the power distribution of a power battery and an engine is optimized on the basis of meeting the predicted power, and the economy of flight is improved. The application is suitable for fields such as aerial reconnaissance, unmanned aerial vehicle agriculture and aerial traffic monitoring, reduces the overall fuel consumption of the aircraft, and improves the economy.

Description

Technical Field

[0001] The present invention relates to a hybrid unmanned aerial vehicle (UAV) energy management method based on deep learning and dynamic programming, and belongs to the technical field of prediction of energy management strategies. Background Art

[0002] To conserve resources, achieve energy conservation and emission reduction, and mitigate environmental pollution, hybrid power technology has become a key area of ​​powertrain research. The goal of energy management is to maximize system energy efficiency by achieving greater energy conservation and efficiency while meeting the driving requirements of the power equipment, based on the system characteristics of each equipment component and the equipment's operating conditions. However, current research on energy management strategies focuses primarily on energy management and control strategies for new energy vehicles, while research on energy management for aircraft is relatively insufficient. Currently, commonly used hybrid aircraft energy management strategies focus on rule-based energy management. However, aircraft operating environments are complex, with random instabilities, and flight paths rarely have the repeatability of car routes. Therefore, rule-based energy management struggles to comprehensively consider the actual flight environment and aircraft flight state parameters to arrive at an optimal energy management strategy. Deep learning-based energy management methods, with their high generalization and flexibility, offer a viable approach to addressing the shortcomings of rule-based energy management. Dynamic programming algorithms, with their excellent global optimization capabilities, offer a viable energy management strategy combining deep learning and dynamic programming. Summary of the Invention

[0003] Regarding rule-based energy management, it is difficult to comprehensively consider the actual flight environment and aircraft flight status parameters to obtain the optimal energy management strategy. The main purpose of this invention is to propose a hybrid UAV energy management method based on deep learning and dynamic programming. By obtaining the historical operating data of the aircraft, the energy management strategy in the future time domain can be predicted and optimized, thereby reducing the overall fuel consumption of the aircraft and improving its economy while ensuring the completion of the flight mission.

[0004] The object of the present invention is achieved through the following technical solutions:

[0005] The application discloses a hybrid unmanned aerial vehicle energy management method based on deep learning and dynamic programming, which is characterized in that: the method is based on a Bi-LSTM bidirectional memory network to accurately predict power demand of an aircraft in a future time domain, and a prediction module considering environmental information has high generalization and flexibility for model prediction, so that the prediction result is more comprehensive and accurate, and the parameter complexity of direct prediction is reduced; on the basis of deep learning prediction, a rolling optimization module is constructed based on a linear programming algorithm to feed back and correct the prediction result, so that a more accurate prediction result is provided; on the basis of minimum equivalent fuel consumption, an energy management optimization function is constructed to realize minimum overall fuel consumption and optimal power distribution of a power battery and an engine on the basis of meeting the predicted power, and flight economy is improved.

[0006] A hybrid unmanned aerial vehicle energy management method based on deep learning and dynamic programming comprises the following steps:

[0007] Step 1: constructing an aircraft dynamics model and collecting aircraft flight process parameters;

[0008] The aircraft operating parameter data are collected, and parameter data including but not limited to flight speed, flight altitude, wind speed, temperature, lithium battery SOC state and overall power demand of the aircraft are extracted and preprocessed.

[0009] Step 2: establishing a prediction module of future demand power of the aircraft; in the case that historical working condition data of the aircraft are known, the Bi-LSTM bidirectional memory network is used to accurately predict power demand of the aircraft in a short future time domain;

[0010] The Bi-LSTM predictor model is as follows:

[0011] Output=f BiLSTM (Input(t))

[0012]

[0013] Output(t)={P t+1 ,P t+2 ,···,P t+p}

[0014] Wherein f BiLSTM represents the predictor model of the Bi-LSTM, P represents power, H is flight altitude, t represents a current time, h represents a historical time, p represents a future time, v t is a current flight speed, and a t is a current flight acceleration.

[0015] Step 3: Establish an energy management prediction module that takes into account environmental information; select temperature, wind speed, and air pressure as environmental characteristic parameters to describe the flight environment of the aircraft. To improve the generalization ability of the predictor and reduce the complexity of the parameters, use the K-means clustering algorithm to classify the environment features, and establish an energy prediction management strategy based on step two that takes into account the environment;

[0016] Step 31: Set the temperature, wind speed, and air pressure of the environment where the aircraft is located as the characteristic parameters that affect the environmental characteristics, and set the judgment period for evaluating the environmental characteristics. Normalize the temperature, wind speed, and air pressure for each evaluation period.

[0017] Step 32: Divide the environmental characteristics into three categories:'mild', 'limit', and 'bad'. Use the K-means method to cluster the features, label the flight data based on the clustering results, and separate the labeled data into a training set and a test set. For a specific set of environmental parameters, use the KNN nearest neighbor algorithm to find the appropriate category.

[0018] Step 33: For a specific set of environmental parameters, calculate the Euclidean distance between the parameters and the training set parameters using the KNN nearest neighbor algorithm, and select the appropriate K value for classification. The constructed KNN classifier is:

[0019] {Input} KNN ={P1,P2,P3}

[0020] {Output} KNN ={M,L,B}=f KNN {Input} KNN

[0021] Where P1, P2, and P3 represent the characteristic parameters of temperature, wind speed, and air pressure, and M, L, and B represent the environmental characteristics, respectively, as mild, limit, and bad.

[0022] Step 34: Establish a Bi-LSTM power prediction model that takes into account environmental characteristics:

[0023] Output=f BiLSTM-environment (Input(t))

[0024]

[0025] Output(t)={P t+1 ,P t+2 ,···,P t+p}

[0026] Where f BiLSTM-environmentIt is a Bi-LSTM power predictor model that considers environmental characteristics. P represents power, H represents flight altitude, and Type e Environment t is the characteristic category of the environment, t is the current moment, h is the historical moment, p is the future moment, v t is the current flight speed, a t is the current flight acceleration.

[0027] Step 4: Based on the power predicted in step 3, with the goal of minimizing the equivalent fuel consumption of the aircraft, construct an aircraft energy management optimization objective function to achieve the optimal power distribution between the battery and the engine;

[0028] The optimization objective function is set as:

[0029]

[0030] Among them, E fe Represents the energy consumption of the engine system, E elec Represents the energy consumption of the power battery, ΔP ec represents the rate of change of engine system output power, χ and δ represent the weight coefficients of equivalent fuel consumption and engine system power output fluctuation, respectively. Represents the conversion factor from electricity to fuel consumption.

[0031] Step 5: Based on the power prediction in step 3 and the optimization objective function constructed in step 4, a rolling optimization module is established using a dynamic programming algorithm to achieve feedback correction of future time-domain power demand prediction and minimize overall fuel consumption, thereby improving flight economy.

[0032] The dynamic programming process for optimizing the objective function is as follows:

[0033] J N-k [SOC k , a k ]=min{L[SOC k ,a k ,u k ]+J N-(k+1) [SOC k+1 ,a k+1 ]}

[0034] Among them, SOC is the battery capacity, a is the engine speed, u is the control sequence, k is the control period, N is the total control period, J is the total control period, N-k [SOC k ,a k ] is based on SOC k and a k is the optimal performance functional of the Nk-segment process in the initial state, SOCk and a k determined by SOC0 and a0 and the previous k segments control, N-(k+1) [SOC k+1 ,a k+1 ] is the optimal performance functional of the rear N-(k+1) segment with the initial state of SOC k+1 and a k+1 , L[SOC k ,a k ,u k ] is the remaining cost of the k to k+1 segment control.

[0035] The prediction optimization process is obtained by dynamic programming to obtain the optimal control sequence of the engine, after the engine is executed, the new aircraft state is fed back to the power prediction module and the rolling optimization module, the power prediction and power distribution calculation are corrected according to the feedback information, and the new flight data is stored in the flight database, the generalization ability of the whole prediction management method is improved. Repeat the above prediction optimization process until the flight task is completed, so as to reduce the total fuel consumption of the aircraft and improve the economy.

[0036] Advantages

[0037] 1. The hybrid unmanned aerial vehicle energy management method based on deep learning and dynamic programming disclosed in the application can accurately predict power demand in the future time domain of the aircraft based on a Bi-LSTM bidirectional memory network, the prediction module considering environmental information has high generalization and flexibility for model prediction, so that the prediction result is more comprehensive and accurate, and the parameter complexity of direct prediction is reduced.

[0038] 2. The hybrid unmanned aerial vehicle energy management method based on deep learning and dynamic programming disclosed in the application can construct a rolling optimization module based on a linear programming algorithm on the basis of deep learning prediction, can correct the prediction result, and provide more accurate prediction results.

[0039] 3. The hybrid unmanned aerial vehicle energy management method based on deep learning and dynamic programming disclosed in the application can construct an energy management optimization function on the basis of equivalent fuel consumption, realize the minimum total fuel consumption and the optimal power distribution of the power battery and the engine on the basis of meeting the predicted power, and improve the economy of flight. DETAILED DESCRIPTION

[0040] Fig. 1 is a flowchart of the hybrid unmanned aerial vehicle energy management method based on deep learning and dynamic programming disclosed in the application; Figure 1 Fig. 1 is a flowchart of the hybrid unmanned aerial vehicle energy management method based on deep learning and dynamic programming disclosed in the application;

[0041] Fig. 1 is a flowchart of the hybrid unmanned aerial vehicle energy management method based on deep learning and dynamic programming disclosed in the application; Figure 2 Fig. 1 is a flowchart of the hybrid unmanned aerial vehicle energy management method based on deep learning and dynamic programming disclosed in the application;

[0042] attached Figure 3 A rolling optimization process in a hybrid unmanned aerial vehicle energy management method based on deep learning and dynamic programming is disclosed.

[0043] attached Figure 4 A prediction optimization process in a hybrid unmanned aerial vehicle energy management method based on deep learning and dynamic programming is disclosed. DETAILED DESCRIPTION

[0044] The technical solutions of the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only illustrative of the present application and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0045] The hybrid unmanned aerial vehicle energy management method based on deep learning and dynamic programming disclosed in the present embodiment has the following steps applied to a quadrotor unmanned aerial vehicle:

[0046] Step 1: Build an aircraft dynamics model and collect aircraft flight process parameters, including the following steps:

[0047] S11: Establish an aircraft dynamics model based on the principle of aircraft dynamics, and establish a quadrotor unmanned aerial vehicle dynamics model in matlab / simulink

[0048] S12: Collect flight state and flight environment information of the quadrotor unmanned aerial vehicle under actual flight task, including but not limited to flight speed, flight altitude, wind speed, temperature, lithium battery SOC state and overall power demand of the aircraft, wherein the overall demand power is calculated as:

[0049]

[0050] In the formula, P is the demand power, Cp is the rotor power coefficient, w is the rotor speed, and i is the rotor number.

[0051] The collected parameters are normalized:

[0052]

[0053] In the formula, X represents various types of original data, Y represents normalized data, and subscripts max and min represent maximum and minimum values, respectively.

[0054] Step 2: Build an aircraft power prediction information module based on Bi-LSTM, including the following steps:

[0055] S21:Figure 1 As shown, the Bi-LSTM is composed of two layers of LSTM network layers, one layer of which transmits information layer by layer forward in time from the past, and one layer of which transmits information layer by layer backward in time from the future.

[0056] The working principle of the Bi-LSTM prediction is as follows:

[0057] The hidden state expression of forward transmission is:

[0058] h t = sigmoid(W B1 x t +W h h t-1 +b h )

[0059] The hidden state expression of backward transmission is:

[0060] h t ' = sigmoid(W B '1x t +W h 'h t ' +1 +b' h )

[0061] The output y t expression is as follows:

[0062] y t = tanh(W B2 h t x t +W B '2h t '+b y )

[0063] In the formula, x t represents the input information at the current time, h t represents the hidden information at the current time in the forward transmission process, W B1 represents the weight matrix from the input layer to the hidden layer, W h represents the weight matrix from the hidden layer to the hidden layer, b h represents the bias from the input layer to the hidden layer, y t represents the output information at the current time, W B2 represents the weight matrix from the hidden layer to the input layer, and b y is the bias from the hidden layer to the output layer. The upper subscript'indicates the corresponding parameters in the backward transmission process, and the lower subscripts t, t-1, and t+1 represent the current time, the previous time, and the next time, respectively.

[0064] S23 selects a prediction time domain length of 10s, takes the normalized demand power, flight speed, flight acceleration and mission target flight height in step one as input, and uses Bi-LSTM to predict the future time domain demand power of the aircraft:

[0065] Output=f BiLSTM (Input(t))

[0066]

[0067] Output(t)={P t+1 ,P t+2 ,···,P t+p}

[0068] wherein f BiLSTM represents the predictor model of Bi-LSTM, P represents power, H is flight height, t represents the current time, h represents the historical time, p represents the future time, v t is the current flight speed, a t is the current flight acceleration, and the future demand power considering the working condition of the aircraft can be predicted by the Bi-LSTM power predictor.

[0069] Step 3: Construct an aircraft power prediction information module based on K-means environmental feature clustering, which specifically includes the following steps:

[0070] S31 takes the temperature, wind speed and air pressure of the environment where the aircraft is located as the feature parameters of the environmental features, sets the judgment period of the environmental features to 10s, and performs normalization calculation on the temperature, wind speed and air pressure every judgment period.

[0071] S32 uses K-means method to perform clustering calculation on the environmental parameters, divides the environmental features into three categories, namely'mild type', 'poor type' and'severe type', marks the flight data according to the clustering results, takes 80% of the marked data as the training set and 20% as the test set. KNN algorithm is used to learn the training set to find a suitable K value, and is used to test the classification effect of the test set.

[0072] S33 for a set of environmental features to be classified obtained by S31, the corresponding KNN classification model is:

[0073] {Input} KNN ={C1,C2,C3}

[0074] {Output} KNN ={M,L,B}=f KNN {Input} KNN

[0075] where C1, C2, C3 represent characteristic parameters of temperature, wind speed and air pressure, M, L, B represent environmental characteristics, respectively mild, limit, bad, f KNN {Input} KNN is the KNN classifier.

[0076] The power prediction process considering environmental characteristics is shown in FIG. 34. Figure 2 The power prediction process considering environmental characteristics is shown in FIG. 34.

[0077] The input and output of the Bi-LSTM power prediction model considering environmental characteristics are shown in FIG. 35.

[0078] Output = f BiLSTM-environment (Input(t))

[0079]

[0080] Output(t) = {P t+1 ,P t+2 ,···,P t+p}

[0081] where f BiLSTM-environment represents the Bi-LSTM power prediction model considering environmental characteristics, P represents power, H is flight altitude, Type e nvironment t represents the characteristic category of the environment, t represents the current time, h represents the historical time, p represents the future time, v t is the current flight speed, a t is the current flight acceleration.

[0082] Step 4: Construct the target optimization function with the least equivalent fuel consumption, which includes the following steps:

[0083] S41 The object of this embodiment is a hybrid aircraft, and the power sources are fuel engine and power lithium battery. The initial SOC of the power battery is not less than 60%, and the power of the power battery is only supplied by the engine during the task process, so the equivalent fuel consumption can be used to calculate the overall energy consumption of the aircraft. In addition, in order to better respond to the power change of the aircraft, the energy management optimization target of the hybrid aircraft is set as the constraint of the equivalent fuel consumption and the engine system output power change rate.

[0084] The aircraft energy management optimization objective function is constructed as follows:

[0085]

[0086] wherein E fe represents the energy consumption of the engine system, E elec represents the energy consumption of the power battery, ΔP ec represents the engine system output power change rate, χ and δ respectively represent the weight coefficients of equivalent fuel consumption and engine system power output fluctuation degree, represents the conversion coefficient of electric quantity into fuel consumption.

[0087] S42 sets the aircraft energy optimization state variable as the power battery SOC and the engine output power, and the control variable as the engine output power change rate, and the expression is as follows:

[0088] x = [SOC, P ec ]

[0089] u = [ΔP ec ] = f(a)

[0090] wherein x is the state variable, SOC is the lithium battery remaining capacity, P ec is the engine output power, u is the control variable, ΔP ec is the engine output power change rate, and a is the control object engine speed.

[0091] S43 takes into account the overall performance and constraints of the system, and the power battery, the engine system, and the drive motor need to meet the physical constraint expression as follows:

[0092]

[0093]

[0094]

[0095] wherein P ec is the engine output power, SOC is the lithium battery remaining capacity, P bat is the lithium battery power, I bat is the lithium battery current, P mot is the drive motor power, T mot is the drive motor torque, max corresponds to the allowed maximum value, and min corresponds to the allowed minimum value.

[0096] Step 5: On the basis of the predicted power in step 3 and the optimization function in step 4, a power prediction rolling optimization module and an energy management module based on the dynamic programming algorithm are constructed, which specifically include the following steps:

[0097] S51 establishes a rolling optimization process as shown in the attached Figure 3 As shown in the figure, the module mainly includes prediction, optimization, and feedback correction. The power prediction module is executed using Bi-LSTM, the control optimization module is executed using the dynamic programming algorithm, and finally the feedback correction between the prediction and actual output is realized through the optimization results.

[0098] S52 passes the predicted required power to the dynamic programming optimization module, and uses the dynamic programming algorithm to search for the optimal control sequence of the power battery SOC and engine power within the predicted time domain based on the objective function constructed in step 4.

[0099] S53 discretizes the battery SOC and engine speed according to a certain step size in the dynamic planning process.

[0100] S54 Dynamic programming optimization objective function (steps S42, S43):

[0101]

[0102] State variables and control quantities:

[0103] x=[SOC,P ec ]

[0104] u=[ΔP ec ]=f(a)

[0105] S55 dynamic programming recursion equation:

[0106] J N-k [SOC k , a k ]=min{L[SOC k ,a k ,u k ]+J N-(k+1) [SOC k+1 ,a k+1 ]}

[0107] Among them, SOC is the battery capacity, a is the engine speed, u is the control sequence, k is the control period, N is the total control period, J is the total control period, N-k [SOC k ,a k ] is based on SOC k and a k is the optimal performance functional of the Nk-segment process in the initial state, SOC k and a k Determined by SOC0, a0 and the first k segments of control, J N-(k+1) [SOC k+1 ,a k+1] is based on SOC k+1 and a k+1 is the optimal performance functional of the posterior N-(k+1) of the initial state, L[SOC k ,a k ,u k ] is the remaining cost of controlling the segment from k to k+1.

[0108] S56 as attached Figure 4 As shown, the optimal control sequence of the engine is obtained by recursive step-by-step dynamic programming, and the first step control sequence is sent to the control object engine speed.

[0109] After executing the first step, the S57 control object feeds back the new aircraft status to the power prediction module and the rolling optimization module, including flight speed, acceleration, power battery SOC, and required power.

[0110] S58 modifies the power prediction and power allocation calculation based on the feedback information; stores the new flight data in the flight database, and improves the generalization ability of the entire prediction management method.

[0111] S59 repeats the above steps until the flight mission is completed.

[0112] The entire S60 method flow is as follows Figure 1 As shown in the figure, the Bi-LSTM predictor that takes into account generalized environmental characteristics enables the aircraft to accurately predict the required power even in complex environmental conditions. Combined with the powerful optimization capability of dynamic programming, it ultimately achieves the goal of rolling optimization, significantly reduces aircraft fuel consumption, and has high algorithm stability.

[0113] The proposed method combines a bidirectional Bi-LSTM neural network with a dynamic programming algorithm. The Bi-LSTM power prediction algorithm, which takes environmental characteristics into account, is highly generalizable and flexible, accurately predicting the future power requirements of an aircraft by integrating multiple factors. The predictive optimization and energy management algorithm based on dynamic programming has powerful optimization capabilities. Compared with traditional rule-based energy management algorithms, dynamic optimization can significantly improve the energy distribution and utilization efficiency of hybrid power platforms. In summary, this method can bring many beneficial effects to energy management strategy research and engineering practice.

[0114] It should be understood that the above specific description is a further detailed explanation of the purpose, technical solutions and beneficial effects of the invention. The above is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A hybrid UAV energy management method based on deep learning and dynamic programming, characterized by: The following steps are included: Step 1: Build an aircraft dynamics model and collect aircraft flight process parameters; Step 2: Build a prediction module for the aircraft's future power requirements. Given the aircraft's historical operating data, use a Bi-LSTM bidirectional memory network to accurately predict the aircraft's power requirements in the short term. Step 3: Establish an energy management prediction module that takes environmental information into consideration; The implementation method of step 3 is: Step 31: Using the ambient temperature, wind speed, and air pressure of the aircraft as characteristic parameters affecting the environmental characteristics, setting a judgment cycle for judging the environmental characteristics, and performing normalization calculations on the temperature, wind speed, and air pressure for each judgment cycle; Step 32: Classify the environmental characteristics into three categories: "mild," "poor," and "severe." Use the K-means method to cluster the characteristics. Label the flight data based on the clustering results and split the labeled data into training and test sets. For a specific set of environmental parameters, use the KNN algorithm to find the appropriate category. Step 33: For a specific set of environmental parameters, calculate the Euclidean distance between it and each set of parameters in the training set according to the KNN proximity algorithm, select an appropriate K value to classify it, and the constructed KNN classifier is: {Input} KNN ={P1,P2,P3} {Output} KNN ={M,L,B}=f KNN {Input} KNN Among them, P1, P2, and P3 represent characteristic parameters of temperature, wind speed, and air pressure, and M, L, and B represent environmental characteristics, which are mild, poor, and severe respectively; Step 34: Establish a Bi-LSTM power prediction model that takes into account environmental characteristics: Output=f BiLSTM-environment (Input(t)) Output(t)={P t+1 ,P t+2 ,···,P t+p } where f BiLSTM-environment It is a Bi-LSTM power predictor model that considers environmental characteristics. P represents power, H represents flight altitude, and Type e Environment t is the characteristic category of the environment, t is the current moment, h is the historical moment, p is the future moment, v t is the current flight speed, a t is the current flight acceleration; Step 4: Based on step 3, with the goal of minimizing the equivalent fuel consumption of the aircraft, construct an aircraft energy management optimization objective function to achieve the optimal power distribution of the battery and engine; The implementation method of step 4 is: The optimization objective function is set as: Among them, E fe Represents the energy consumption of the engine system, E elec Represents the energy consumption of the power battery, ΔP ec represents the rate of change of engine system output power, χ and δ represent the weight coefficients of equivalent fuel consumption and engine system power output fluctuation, respectively. Represents the conversion factor from electricity to fuel consumption; Step 5: Based on the optimization objective function constructed in Steps 3 and 4, a rolling optimization module is established using a dynamic programming algorithm to achieve feedback correction of future time-domain power demand forecasts and minimize overall fuel consumption. The implementation method of step five is: The dynamic programming process for optimizing the objective function is as follows: J N-k [SOC k ,a k ]=min{L[SOC k ,a k ,u k ]+J N-(k+1) [SOC k+1 ,a k+1 ]} Among them, SOC is the battery capacity, a is the engine speed, u is the control sequence, k is the control period, N is the total control period, J is the total control period, N-k [SOC k ,a k ] is based on SOC k and a k is the optimal performance functional of the Nk-segment process in the initial state, SOC k and a k Determined by SOC0, a0 and the first k segments of control, J N-(k+1) [SOC k+1 ,a k+1 ] is based on SOC k+1 and a k+1 is the optimal performance functional of the posterior N-(k+1) of the initial state, L[SOC k ,a k ,u k ] is the remaining cost of controlling the segment from k to k+1.

2. The hybrid UAV energy management method based on deep learning and dynamic programming according to claim 1, characterized in that: The implementation method of step one is: Collect aircraft operating parameter data, extract parameter data including but not limited to flight speed, flight altitude, wind speed, temperature, lithium battery SOC status and aircraft overall power demand, and pre-process the parameter data.

3. The hybrid UAV energy management method based on deep learning and dynamic programming according to claim 2, characterized in that: The implementation method of step 2 is: The Bi-LSTM predictor model is as follows: Output=f BiLSTM (Input(t)) Output(t)={P t+1 ,P t+2 ,···,P t+p } where f BiLSTM represents the predictor model of Bi-LSTM, P represents power, H represents flight altitude, t represents current time, h represents historical time, p represents future time, v t is the current flight speed, a t is the current flight acceleration.

4. The hybrid UAV energy management method based on deep learning and dynamic programming according to claim 3, characterized in that: The predictive optimization process is to obtain the optimal engine control sequence through dynamic programming step by step recursion. After the engine is executed, the new aircraft state is fed back to the power prediction module and the rolling optimization module. The power prediction and power allocation calculation are corrected based on the feedback information. The new flight data is stored in the flight database to improve the generalization ability of the entire predictive management method. The above predictive optimization process is repeated until the flight mission is completed to reduce the overall fuel consumption of the aircraft.

Images