A tilt-rotor unmanned aerial vehicle energy management optimization method based on deep learning

Abstract

The application discloses a tilt-rotor unmanned aerial vehicle energy management optimization method based on deep learning, and belongs to the field of aircraft power technology. The method comprises the following steps: constructing a power prediction neural network model based on collected unmanned aerial vehicle flight condition parameters, and predicting power demand under different length time domains; obtaining optimal power demand by using a dynamic programming algorithm according to different condition data, constructing a prediction time domain selection strategy based on optimal delta P approximation based on the optimal power demand, selecting a DNN neural network to establish a delta P prediction model, obtaining the optimal prediction time domain of the neural network and the power prediction result in the optimal prediction time domain based on the delta P prediction model, combining the constructed unmanned aerial vehicle system cost function, and adopting a dynamic programming rolling optimization hybrid power system energy distribution strategy until the optimal energy distribution strategy of the hybrid power system is obtained, so that the minimum equivalent fuel consumption of the system is realized, and the flight economy and the endurance time are improved.

Description

Technical Field

[0001] This invention belongs to the field of aircraft propulsion technology and relates to an energy management optimization method for tilt-rotor unmanned aerial vehicles based on deep learning. Background Technology

[0002] Traditional drones are mainly divided into two types: multi-rotor and fixed-wing. Multi-rotor drones can achieve vertical takeoff and landing and have low requirements for terrain conditions, but their flight speed is generally low and their maneuverability is poor. Fixed-wing drones can achieve higher cruising speeds, but their takeoff and landing are constrained by terrain and require a sufficiently long runway, thus limiting their range of use. Tiltrotor drones combine the advantages of multi-rotor vertical takeoff and landing and fixed-wing high-speed cruising, making them a drone type with great potential.

[0003] In traditional deep learning-based research on energy management strategies for hybrid power systems, the prediction time domain length is often a fixed value. Different prediction time domain lengths have a significant impact on the accuracy of the prediction results, which in turn affects the overall energy management effect and computational cost of the system. The optimal prediction time domain length is usually obtained offline. By training under different operating conditions, the energy consumption under different prediction time domain lengths is obtained, and the minimum energy consumption is selected to determine the optimal prediction time domain length for the corresponding operating condition. However, it is impossible to adjust in real time according to changes in operating conditions. Summary of the Invention

[0004] To address the aforementioned technical problems in this field, the main objective of this invention is to provide a deep learning-based energy management optimization method for tiltrotor unmanned aerial vehicles (UAVs). This method constructs a power prediction neural network model based on collected UAV flight parameters to predict power demands in different time domains. Based on data from different flight conditions, a dynamic programming algorithm is used to obtain the optimal power demand. A prediction time domain selection strategy based on optimal ΔP approximation is then constructed based on the optimal power demand. This yields the optimal prediction time domain of the neural network and the power prediction results within the optimal prediction time domain. Combined with the constructed UAV system cost function, a dynamic programming rolling optimization strategy for the hybrid power system's energy allocation is employed until the optimal energy allocation strategy for the hybrid power system is obtained. This achieves the minimum equivalent fuel consumption of the system, improving flight economy and endurance.

[0005] The objective of this invention is achieved through the following technical solution.

[0006] This invention discloses a deep learning-based energy management optimization method for tilt-rotor unmanned aerial vehicles (UAVs), comprising the following steps:

[0007] Step 1: Extract operating parameters from historical flight data of tiltrotor UAVs, including power demand, flight speed, flight acceleration, flight altitude, and state of charge (SOC) of the power battery. Normalize the collected flight data operating parameters and construct a dataset for training the power prediction neural network model in Step 2.

[0008] Step 1.1: Extract flight status data, including power demand, flight speed, flight acceleration, flight altitude, and battery state of charge. The power demand is mapped as follows:

[0009]

[0010] In the formula, P is the required power, Cp is the rotor power coefficient, w is the rotor speed, and i is the rotor number.

[0011] Step 1.2: Normalize the flight data parameters and construct a dataset for training the power prediction neural network model built in Step 2.

[0012] Step 2: By analyzing the operational parameters of the tiltrotor aircraft flight data from Step 1, a power prediction neural network model is constructed for predicting the required power during flight. The inputs to the power prediction neural network model include required power, flight speed, flight acceleration, and flight altitude, and the output includes the required power for different prediction time domains.

[0013] To more accurately predict power demand, the power prediction neural network model is preferably constructed based on a Bi-LSTM bidirectional network.

[0014] Step 2.1: Preferably, a power demand prediction module for the UAV is constructed using a Bi-LSTM bidirectional network. The Bi-LSTM consists of two LSTM network layers: one layer transmits information forward layer by layer from past historical moments, and the other layer transmits information backward layer by layer from future moments.

[0015] Step 2.2: The Bi-LSTM prediction network is as follows:

[0016] Forward pass-through of hidden state expression:

[0017] h t =sigmoid(W B1 x t +W h h t-1 +b h )

[0018] Backpropagation of hidden state expression:

[0019] h t= sigmoid(W B '1x t +W h 'h t ' +1 +b' h )

[0020] Output y t The expression is as follows:

[0021] y t =tanh(W B2 h t x t +W B '2h t '+b y )

[0022] In the formula, x t h represents the input information at the current moment. t W represents the hidden information at the current moment during the forward propagation process. B1 W represents the weight matrix from the input layer to the hidden layer. h Let b represent the weight matrix from hidden layer to hidden layer. h y represents the deviation from the input layer to the hidden layer. t W represents the output information at the current moment. B2 Let b represent the weight matrix from the hidden layer to the input layer. y It represents the deviation from the hidden layer to the output layer. The superscript ' indicates the corresponding parameter representation during the backpropagation process, and the subscripts t, t-1, and t+1 represent the current time, the previous time, and the next time, respectively.

[0023] Step 2.3: Select the prediction time domain length, and use the normalized power demand, flight speed, flight acceleration, and flight altitude from Step 1 as inputs to predict the future time domain power demand of the aircraft using Bi-LSTM:

[0024] Output = f BiLSTM (Input(t))

[0025]

[0026] Output(t) = {P} t+1 ,P t+2 ,…,P t+p}

[0027] Where f BiLSTM This represents the predictor model of Bi-LSTM, where P represents power, H is flight altitude, t represents the current time, h represents a historical time, p represents a future time, and v t Given the current flight speed, a tGiven the current flight acceleration, the Bi-LSTM power predictor can predict the power demand for UAV flight conditions in different time domains in the future.

[0028] Step 3: Based on the UAV flight data under different operating conditions obtained in Step 1, use dynamic programming algorithm to calculate the optimal power demand of the UAV at each moment under different operating conditions, calculate the optimal power change ΔP between each moment and the next moment, and establish a one-to-one correspondence between the flight state data of the tiltrotor aircraft at each moment and the optimal power change ΔP to establish a prediction database with strong correlation between the UAV flight state and the optimal power change ΔP.

[0029] Step 3.1: Calculate the optimal P trajectory under the task conditions, and then calculate the optimal ΔP for each step.

[0030] Step 3.2: Based on Step 3.1, extract the UAV flight status data for each step and establish a database corresponding to the flight status information and the optimal ΔP as follows:

[0031]

[0032]

[0033] (X1,Y1),…,(X k ,Y k ),(X k+1 ,Y k+1 ),…,(X k+n ,Y k+n )

[0034] Where H is the flight altitude, v is the flight speed, a is the flight acceleration, P is the required power, Y represents the optimal ΔP, X is the flight state data corresponding to the optimal ΔP, and the subscript indicates the time period.

[0035] Step 4: Based on the database of strong correlation between UAV flight status data and optimal power change ΔP constructed in Step 3, a deep neural network is constructed to predict the optimal power change ΔP, and the optimal ΔP value under the current flight status is found through the deep neural network.

[0036] In order to map the relationship between UAV flight status data and the optimal power change ΔP and to achieve efficient prediction calculation, as a preferred method, the deep neural network used to predict the optimal power change ΔP in step 4 is constructed based on a DNN neural network.

[0037] Step 4.1: Preferably, a DNN neural network is selected to establish a ΔP prediction model. The inputs to the prediction model are the UAV's flight speed, flight acceleration, power demand, and flight altitude. The prediction model is established as follows:

[0038] Y(t) = fDNN (X(t))

[0039] X(t)=H k ,v k ,a k ,P k t = k

[0040] Y(k)=ΔP k

[0041] Where H is the flight altitude, v is the flight speed, a is the flight acceleration, P is the required power, and the subscript k indicates the time period.

[0042] Step 4.2: Divide the ΔP prediction database into a training set and a test set, train the DNN model, and use the test set to verify the model's prediction results until the trained DNN model meets the preset accuracy requirements.

[0043] Step 5: Based on the power demand in different time domains predicted in Step 2, obtain the ΔP trajectory in different time domains. Compare the ΔP value in Step 1 with the optimal ΔP value predicted in Step 4 in different time domains, and take the prediction time domain that is closest to the optimal ΔP value as the optimal prediction time domain.

[0044] Step 5.1: Calculate the first step ΔP of the P trajectory for different time domain lengths.

[0045] Step 5.2: Compare the difference between the first step ΔP and the optimal ΔP in different time domains:

[0046] min|ΔP-ΔP * |

[0047] Where ΔP is the optimal ΔP predicted by the DNN, ΔP * The first step ΔP represents the prediction of the power trajectory by Bi-LSTM under different time domain lengths. The time domain with the smallest difference between the two is selected as the optimal prediction time domain.

[0048] Step 6: Based on the Hamiltonian optimization function, construct an equivalent fuel consumption function model for the UAV that considers engine fuel consumption and equivalent fuel consumption of electric energy. On this basis, consider the impact of frequent engine power jumps on engine performance and lifespan, and construct the overall cost function of the hybrid system. Combined with the rolling optimization process in Step 7, minimize the cost function of the UAV system, and achieve minimum fuel consumption and stable engine operation of the UAV system.

[0049] Step 6.1: The SOC change rate of the power battery is:

[0050]

[0051] Where I is the battery bus output current and Q is the battery capacity.

[0052] Ignoring battery internal resistance, the battery power is calculated as: P b =IV b Where I is the battery output current and Vb is the battery open-circuit voltage, the rate of change of SOC is written as:

[0053]

[0054] The equivalent fuel consumption of the powertrain system can be expressed using the Hamiltonian function as follows:

[0055]

[0056] Among them The fuel consumption rate of the internal combustion engine, λ, is a costate variable that satisfies the dynamic equation in the optimization principle. In this method, it is equivalent to the weight of battery power consumption in the total fuel consumption, H. f This represents the total equivalent fuel consumption rate.

[0057] Step 6.2: Considering the impact of frequent engine start-stop or frequent speed fluctuations on engine life, and taking into account the power change ΔPe, combined with the equivalent fuel consumption rate mentioned above, the overall cost function of the UAV system is:

[0058]

[0059] Where λ and β are weighting coefficients.

[0060] Step 7: Based on the optimal predicted power of the UAV system within the optimal time domain obtained from Steps 2 to 5, and combined with the system cost function constructed in Step 6, the optimal power allocation of the power system is obtained using a dynamic programming algorithm, that is, the optimal control sequence of the engine throttle is obtained. The first step of the control sequence is to control the engine. After the first step control command is executed, the new flight status data is expanded into the flight condition database, and the power demand prediction is updated. The above process is repeated to achieve rolling optimization until the optimal energy allocation strategy of the hybrid power system is obtained, so as to achieve the minimum equivalent fuel consumption of the system and improve flight economy and endurance.

[0061] Step 7.1: For the constructed system cost function, select the dynamic programming algorithm to establish a rolling optimization process, and select the state variables and control variables as follows:

[0062] x = [SOC, P] e ]

[0063] u=[ΔP e ]=f(a)

[0064] Where x is the state variable, u is the control variable, SOC is the state of charge of the power battery, ΔPe is the change in engine power, and a is the engine throttle of the controlled object.

[0065] The dynamic programming recurrence relation is:

[0066] 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 ]}

[0067] Where SOC is the battery charge, a is the engine speed, u represents the control sequence, k represents the control period, N represents the total control period, and J N-k [SOC k ,a k [Based on SOC] k and a k The optimal performance functional for the subsequent Nk segments of 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 [Based on SOC] k+1 and a k+1 L[SOC] is the optimal performance functional for the latter part N-(k+1) of the initial state. k ,a k ,u k [This represents the remaining cost of controlling the segment from k to k+1.]

[0068] Step 7.2: The optimal engine control sequence is obtained by dynamic programming through step-by-step recursion. The first step control sequence is sent to the engine throttle of the controlled object. After the controlled object executes the first step instruction, it feeds back the new UAV state to the optimal time domain selection module and the power prediction module to obtain the optimal prediction time domain length. The predicted power demand within the optimal prediction time domain is then used for dynamic programming. Dynamic programming is used to continuously optimize the hybrid power system energy allocation strategy until the optimal energy allocation strategy of the hybrid power system is obtained, so as to achieve the minimum equivalent fuel consumption of the UAV system and improve flight economy and endurance.

[0069] Beneficial effects:

[0070] 1. This invention discloses a deep learning-based energy management optimization method for tiltrotor unmanned aerial vehicles (UAVs). Based on acquired flight data of the UAV under different operating conditions, a dynamic programming algorithm is used to calculate the optimal power demand of the UAV at each moment under different operating conditions, and the optimal power change ΔP between each moment and the next moment is calculated. A one-to-one correspondence is established between the flight state data of the tiltrotor UAV at each moment and the optimal power change ΔP, establishing a prediction database with a strong correlation between the UAV flight state and the optimal power change ΔP. A deep neural network is constructed to predict the optimal power change ΔP. The optimal ΔP value under the current flight state is found through the deep neural network. Compared with the ΔP obtained from the power demand predicted by the power predictor at different time domain lengths, the optimal power demand in the optimal prediction time domain is obtained. The proposed ΔP approximation method can select the optimal prediction time domain under different operating conditions in real time, and perform future power demand prediction of the UAV within the optimal time domain, which can improve the accuracy of power prediction and the overall optimization effect of the UAV system.

[0071] 2. This invention discloses a deep learning-based energy management optimization method for tilt-rotor unmanned aerial vehicles (UAVs). It constructs an equivalent fuel consumption function model for the UAV based on the Hamiltonian optimization function, considering engine fuel consumption and equivalent fuel consumption from electrical energy. Furthermore, it considers the impact of frequent engine power fluctuations on engine performance and lifespan, incorporating the weighted engine power change as part of the cost function to construct the overall cost function of the UAV system. Combined with a rolling optimization process using dynamic programming, it obtains the optimal power allocation between the engine and the power battery, minimizing the cost function of the UAV system. This minimizes fuel consumption and ensures stable engine operation, improving flight economy and endurance.

[0072] 3. This invention discloses a deep learning-based energy management optimization method for tiltrotor unmanned aerial vehicles (UAVs). Based on the optimal power demand within the optimal time domain obtained from a power predictor and a constructed system cost function, a dynamic programming algorithm is used to obtain the optimal power allocation for the propulsion system and the optimal engine throttle control sequence. The first step of the control sequence is to control the engine. After the first step control command is executed, the optimal power demand prediction is updated based on new flight status data. This process is repeated cyclically to achieve rolling optimization. The rolling optimization process enables timely feedback and correction of the predicted power demand and real-time optimal power allocation for the propulsion system, effectively improving the UAV's endurance.

[0073] 4. This invention discloses a deep learning-based energy management optimization method for tiltrotor unmanned aerial vehicles (UAVs). By analyzing the flight data and operating parameters of the tiltrotor aircraft, a power prediction neural network model is constructed for predicting power demand under different flight conditions. In constructing the power prediction neural network model, altitude is used to characterize the impact of the environment on the power of the UAV under different flight conditions, making the power prediction neural network model constructed in this invention applicable to predicting the flight power of UAVs at different altitudes. Attached Figure Description

[0074] Figure 1 This invention discloses a method flow for optimizing energy management of tilt-rotor unmanned aerial vehicles based on deep learning.

[0075] Figure 2 This invention discloses an optimal time-domain selection strategy for an energy management optimization method for tilt-rotor unmanned aerial vehicles based on deep learning.

[0076] Figure 3 This invention discloses a rolling optimization process for an energy management optimization method for tilt-rotor unmanned aerial vehicles based on deep learning. Detailed implementation method:

[0077] like Figure 1 As shown in the figure, this embodiment discloses a deep learning-based energy management optimization method for tilt-rotor unmanned aerial vehicles (UAVs), and the specific implementation steps are as follows:

[0078] Extract operating condition information from historical flight data of tiltrotor UAVs, including power demand, flight speed, flight acceleration, flight altitude, and state of charge (SOC) of the power battery.

[0079] Demand power is mapped as follows:

[0080]

[0081] In the formula, P is the required power, Cp is the rotor power coefficient, w is the rotor speed, and i is the rotor number.

[0082] The extracted parameters are normalized:

[0083]

[0084] In the formula, X represents the extracted source data, Y represents the normalized data, and max and min represent the maximum and minimum values, respectively.

[0085] A power demand prediction module for unmanned aerial vehicles (UAVs) is constructed using a Bi-LSTM bidirectional network. The Bi-LSTM consists of two LSTM network layers: one layer propagates information forward from past historical moments, and the other layer propagates information backward from future moments. Its prediction principle is as follows:

[0086] Forward pass-through of hidden state expression:

[0087] h t =sigmoid(W B1 x t +W h h t-1 +b h)

[0088] Backpropagation of hidden state expression:

[0089] h t = sigmoid(W B '1x t +W h 'h t ' +1 +b' h )

[0090] Output y t The expression is as follows:

[0091] y t =tanh(W B2 h t x t +W B '2h t '+b y )

[0092] In the formula, x t h represents the input information at the current moment. t W represents the hidden information at the current moment during the forward propagation process. B1 W represents the weight matrix from the input layer to the hidden layer. h Let b represent the weight matrix from hidden layer to hidden layer. h y represents the deviation from the input layer to the hidden layer. t W represents the output information at the current moment. B2 Let b represent the weight matrix from the hidden layer to the input layer. y It represents the deviation from the hidden layer to the output layer. The superscript ' indicates the corresponding parameter representation during the backpropagation process, and the subscripts t, t-1, and t+1 represent the current time, the previous time, and the next time, respectively.

[0093] Choose the prediction time domain length, such as [5s, 10s, 15s]. Use the normalized power demand, flight speed, flight acceleration, and flight altitude from the collected data as input, and use the established power predictor to predict the future power demand of the aircraft.

[0094] Output = f BiLSTM (Input(t))

[0095]

[0096] Output(t) = {P} t+1 ,P t+2 ,…,P t+p}

[0097] Where fBiLSTM This represents the power predictor model of Bi-LSTM, where P represents power, H is the flight altitude, t represents the current time, h represents the historical time, p represents the future time, and v t Given the current flight speed, a t Given the current flight acceleration, the Bi-LSTM power predictor can predict the power demand of the aircraft under different time domains in the future, taking into account the flight conditions.

[0098] The process for selecting the optimal prediction time domain for power demand is shown in the attached figure. Figure 2 As shown, the process is as follows:

[0099] Establish a ΔP prediction database; calculate the optimal power demand trajectory for the task conditions using a dynamic programming algorithm, and calculate the optimal ΔP for each step.

[0100] Extract the UAV flight status data for each step and establish a database corresponding to the flight status information and the optimal ΔP as follows:

[0101]

[0102]

[0103] (X1,Y1),…,(X k ,Y k ),(X k+1 ,Y k+1 ),…,(X k+n ,Y k+n )

[0104] Where H is the flight altitude, v is the flight speed, a is the flight acceleration, P is the required power, Y represents the optimal ΔP, X is the flight state data corresponding to the optimal ΔP, and the subscript indicates the time period.

[0105] Based on a database of UAV flight status data and optimal ΔP, a deep neural network is established to predict the optimal ΔP and find the optimal ΔP value under the current flight status.

[0106] A ΔP prediction model is established using a DNN neural network. The inputs to the prediction model are the UAV's flight speed, flight acceleration, power demand, and flight altitude. The prediction model is as follows:

[0107] Y(t) = f DNN (X(t))

[0108] X(t)=H k ,v k ,a k ,P k t = k

[0109] Y(k)=ΔP k

[0110] Where H is the flight altitude, v is the flight speed, a is the flight acceleration, P is the required power, and the subscript k indicates the time period.

[0111] The ΔP prediction database is divided into a training set and a test set. The DNN model is trained until the accuracy requirement is met, and the model prediction results are verified using the test set.

[0112] Compare the difference between the first step ΔP of the power trajectory predicted by the power predictor and the optimal ΔP predicted by the ΔP prediction network under different prediction time domains:

[0113] min|ΔP-ΔP * |

[0114] Where ΔP is the optimal ΔP predicted by the DNN, ΔP * The first step ΔP represents the prediction of the power trajectory by Bi-LSTM under different time domain lengths. The time domain with the smallest difference between the two is selected as the optimal prediction time domain.

[0115] The state of charge (SOC) change rate of the power battery is:

[0116]

[0117] Where I is the battery bus output current and Q is the battery capacity.

[0118] Ignoring battery internal resistance, the battery power is calculated as: P b =IV b Where I is the battery output current and Vb is the battery open-circuit voltage, the rate of change of SOC can be written as:

[0119]

[0120] The equivalent fuel consumption of the powertrain system can be expressed using the Hamiltonian function as follows:

[0121]

[0122] Among them The fuel consumption rate of the internal combustion engine, λ, is a costate variable that satisfies the dynamic equation in the optimization principle. In this method, it can be simply regarded as the weight of battery power consumption in the total fuel consumption, H. f This represents the total equivalent fuel consumption rate.

[0123] Considering the impact of frequent engine start-stop or frequent speed fluctuations on engine life, and taking into account the power change ΔPe, combined with the equivalent fuel consumption rate mentioned above, the system cost function is designed as follows:

[0124]

[0125] Where λ and β are weighting coefficients.

[0126] The dynamic programming algorithm is chosen to establish the rolling optimization process, and the state variables and control variables are selected as follows:

[0127] x = [SOC, P] e ]

[0128] u=[ΔP e ]=f(a)

[0129] Where x is the state variable, u is the control variable, SOC is the state of charge of the power battery, ΔPe is the change in engine power, and a is the engine throttle of the controlled object.

[0130] The dynamic programming recurrence relation is:

[0131] 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 ]}

[0132] Where SOC is the battery charge, a is the engine speed, u represents the control sequence, k represents the control period, N represents the total control period, and J N-k [SOC k ,a k [Based on SOC] k and a k The optimal performance functional for the subsequent Nk segments of the initial state, SOC k and a k Determined by SOC0, a0, and the first k segments of control.

[0133] J N-(k+1) [SOC k+1 ,a k+1 [Based on SOC] k+1 and a k+1 L[SOC] is the optimal performance functional for the latter part N-(k+1) of the initial state. k ,a k ,u k [This represents the remaining cost of controlling the segment from k to k+1.]

[0134] From the appendix Figure 1 and attached Figure 3As shown, dynamic programming iteratively obtains the optimal control sequence for the engine. The first step control sequence is sent to the engine throttle of the controlled object. After the controlled object executes the first step instruction, it feeds back the new aircraft state to the optimal time domain selection module and the power prediction module to obtain the optimal prediction time domain length. The predicted power demand within the optimal prediction time domain is then used for dynamic programming to perform rolling optimization, thereby achieving the optimal power allocation between the engine and the battery and the minimum system cost.

[0135] This invention combines deep learning and dynamic programming algorithms. The proposed optimal time-domain selection method based on ΔP approximation can improve the overall optimization effect and power prediction accuracy of the system. Based on the Hamiltonian function, a cost function considering the change in engine power is proposed. Combined with the dynamic programming rolling optimization algorithm, the power allocation of the system can be optimized and the cost minimized. In summary, this invention can bring many beneficial effects to the research and engineering practice of energy management strategies for tiltrotor UAVs.

[0136] It should be understood that the above-described specific details are a further detailed explanation of the purpose, technical solution and beneficial effects of the invention. The above description 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 within the scope of protection of the present invention.

Claims

1. A deep learning-based energy management optimization method for tilt-rotor unmanned aerial vehicles (UAVs), characterized in that: Includes the following steps, Step 1: Extract operating parameters from historical flight data from the tiltrotor UAV flight database. The flight data operating parameters include power demand, flight speed, flight acceleration, flight altitude, and state of charge (SOC) of the power battery. Normalize the collected flight data operating parameters and construct a dataset for training the power prediction neural network model in the subsequent Step 2. Step 2: By analyzing the operational parameters of the tiltrotor aircraft flight data in Step 1, a power prediction neural network model for predicting the required power during flight is constructed; the input of the power prediction neural network model includes required power, flight speed, flight acceleration and flight altitude, and the output includes required power corresponding to different prediction time domains; Step 3: Based on the UAV flight data under different operating conditions obtained in Step 1, use dynamic programming algorithm to calculate the optimal power demand of the UAV at each moment under different operating conditions, calculate the optimal power change ΔP between each moment and the next moment, and establish a prediction database with strong correlation between UAV flight status and optimal power change ΔP by mapping the flight status data of the tiltrotor aircraft at each moment to the optimal power change ΔP. Step 4: Based on the database of strong correlation between UAV flight status data and optimal power change ΔP constructed in Step 3, a deep neural network is constructed to predict the optimal power change ΔP, and the optimal ΔP value under the current flight status is found through the deep neural network. Step 5: Based on the power demand in different time domains predicted in Step 2, the ΔP trajectory in different time domains is obtained. The ΔP value in Step 1 is compared with the optimal ΔP value predicted in Step 4 in different time domains. The prediction time domain that is closest to the optimal ΔP value is taken as the optimal prediction time domain. Step 6: Based on the Hamiltonian optimization function, construct an equivalent fuel consumption function model for the UAV that considers engine fuel consumption and equivalent fuel consumption of electric energy. On this basis, consider the impact of frequent engine power jumps on engine performance and lifespan, and construct the overall cost function of the hybrid system. Combined with the rolling optimization process in Step 7, minimize the cost function of the UAV system, and achieve minimum fuel consumption and stable engine operation of the UAV system. Step 7: Based on the optimal predicted power of the UAV system within the optimal time domain obtained from Steps 2 to 5, and combined with the system cost function constructed in Step 6, the optimal power allocation of the power system is obtained using a dynamic programming algorithm, that is, the optimal control sequence of the engine throttle is obtained. The first step of the control sequence is to control the engine. After the first step control command is executed, the new flight status data is expanded into the flight condition database, and the power demand prediction is updated. The above process is repeated to achieve rolling optimization until the optimal energy allocation strategy of the hybrid power system is obtained, so as to achieve the minimum equivalent fuel consumption of the system and improve flight economy and endurance.

2. The energy management optimization method for tilt-rotor UAVs based on deep learning as described in claim 1, characterized in that: Step 1 is implemented as follows: Step 1.1: Extract flight status data, including power demand, flight speed, flight acceleration, flight altitude, and battery state of charge. The power demand is mapped as follows: In the formula, P is the required power, Cp is the rotor power coefficient, w is the rotor speed, and i is the rotor number; Step 1.2: Normalize the flight data parameters and construct a dataset for training the power prediction neural network model built in Step 2.

3. The energy management optimization method for tilt-rotor UAVs based on deep learning as described in claim 2, characterized in that: The power prediction neural network model is constructed based on a Bi-LSTM bidirectional network. Step 2 is implemented as follows: Step 2.1: Construct a UAV power demand prediction module using a Bi-LSTM bidirectional network. The Bi-LSTM consists of two LSTM network layers: one layer transmits information forward layer by layer from past historical moments, and the other layer transmits information backward layer by layer from future moments. Step 2.2: The Bi-LSTM prediction network is as follows: Forward pass-through of hidden state expression: h t =sigmoid(W B1 x t +W h h t-1 +b h ) Backpropagation of hidden state expression: h t '=sigmoid(W B '1x t +W h 'h t ' +1 +b' h ) Output y t The expression is as follows: y t =tanh(W B2 h t x t +W B '2h t '+b y ) In the formula, x t h represents the input information at the current moment. t W represents the hidden information at the current moment during the forward propagation process. B1 W represents the weight matrix from the input layer to the hidden layer. h Let b represent the weight matrix from hidden layer to hidden layer. h y represents the deviation from the input layer to the hidden layer. t W represents the output information at the current moment. B2 Let b represent the weight matrix from the hidden layer to the input layer. y It represents the deviation from the hidden layer to the output layer. The superscript ' indicates the corresponding parameter representation during the back propagation process, and the subscripts t, t-1, and t+1 represent the current time, the previous time, and the next time, respectively. Step 2.3: Select the prediction time domain length, and use the normalized power demand, flight speed, flight acceleration, and flight altitude from Step 1 as inputs to predict the future time domain power demand of the aircraft using Bi-LSTM: Output=f BiLSTM (Input(t)) Output(t)={P t+1 ,P t+2 ,…,P t+p } Where f BiLSTM This represents the predictor model of Bi-LSTM, where P represents power, H is the flight altitude, t represents the current time, h represents a historical time, p represents a future time, and v t Given the current flight speed, a t Given the current flight acceleration, the power demand of the UAV is predicted in different time domains considering the flight conditions of the UAV, using a Bi-LSTM power predictor.

4. The energy management optimization method for tilt-rotor UAVs based on deep learning as described in claim 3, characterized in that: Step 3 is implemented as follows: Step 3.1: Calculate the optimal P trajectory under the task conditions, and then calculate the optimal ΔP for each step. Step 3.2: Based on Step 3.1, extract the UAV flight status data for each step and establish a database corresponding to the flight status information and the optimal ΔP as follows: (X1,Y1),…,(X k ,AND k ),(X k+1 ,AND k+1 ),…,(X k+n ,AND k+n ) Where H is the flight altitude, v is the flight speed, a is the flight acceleration, P is the required power, Y represents the optimal ΔP, X is the flight state data corresponding to the optimal ΔP, and the subscript indicates the time period.

5. The energy management optimization method for tilt-rotor UAVs based on deep learning as described in claim 4, characterized in that: The deep neural network used in step 4 to predict the optimal power change ΔP is constructed based on a DNN neural network. The implementation method of step 4 is as follows: Step 4.1: Select a DNN neural network to build a ΔP prediction model. The input to the prediction model is the UAV's flight speed, flight acceleration, power demand, and flight altitude. The prediction model is built as follows: Y(t)=f DNN (X(t)) X(t)=H k ,v k ,a k ,P k t=k Y(k)=△P k Where H is the flight altitude, v is the flight speed, a is the flight acceleration, P is the required power, and the subscript k indicates the time period; Step 4.2: Divide the ΔP prediction database into a training set and a test set, train the DNN model, and use the test set to verify the model's prediction results until the trained DNN model meets the preset accuracy requirements.

6. The energy management optimization method for tilt-rotor UAVs based on deep learning as described in claim 5, characterized in that: Step 5 is implemented as follows: Step 5.1: Calculate the first step ΔP of the P trajectory for different time domain lengths. Step 5.2: Compare the difference between the first step ΔP and the optimal ΔP in different time domains: min|ΔP-ΔP * | Where ΔP is the optimal ΔP predicted by the DNN, ΔP * The first step ΔP represents the prediction of the power trajectory by Bi-LSTM under different time domain lengths. The time domain with the smallest difference between the two is selected as the optimal prediction time domain.

7. The energy management optimization method for tilt-rotor UAVs based on deep learning as described in claim 6, characterized in that: Step 6 is implemented as follows: Step 6.1: The SOC change rate of the power battery is: Where I is the battery bus output current and Q is the battery capacity. Ignoring battery internal resistance, the battery power is calculated as: P b =IV b Where I is the battery output current and Vb is the battery open-circuit voltage, the rate of change of SOC is written as: The equivalent fuel consumption of the powertrain system can be expressed using the Hamiltonian function as follows: Among them The fuel consumption rate of the internal combustion engine, λ, is a costate variable that satisfies the dynamic equation in the optimization principle. In this method, it is equivalent to the weight of battery power consumption in the total fuel consumption, H. f Total equivalent fuel consumption rate; Step 6.2: Considering the impact of frequent engine start-stop or frequent speed fluctuations on engine life, and taking into account the power change ΔPe, combined with the equivalent fuel consumption rate mentioned above, the overall cost function of the UAV system is: Where λ and β are weighting coefficients.

8. The energy management optimization method for tilt-rotor UAVs based on deep learning as described in claim 7, characterized in that: Step 7 is implemented as follows: Step 7.1: For the constructed system cost function, select the dynamic programming algorithm to establish a rolling optimization process, and select the state variables and control variables as follows: x=[SOC,P e ] u=[ΔP e ]=f(a) Where x is the state variable, u is the control variable, SOC is the state of charge of the power battery, ΔPe is the change in engine power, and a is the engine throttle of the controlled object. The dynamic programming recurrence relation is: 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 ]} Where SOC is the battery charge, a is the engine speed, u represents the control sequence, k represents the control period, N represents the total control period, and J N-k [SOC k ,a k [Based on SOC] k and a k The optimal performance functional for the subsequent Nk segments of 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 [Based on SOC] k+1 and a k+1 L[SOC] is the optimal performance functional for the latter part N-(k+1) of the initial state. k ,a k ,u k [This refers to the remaining cost of controlling the segment from k to k+1;] Step 7.2: The optimal engine control sequence is obtained by dynamic programming through step-by-step recursion. The first step control sequence is sent to the engine throttle of the controlled object. After the controlled object executes the first step instruction, it feeds back the new UAV state to the optimal time domain selection module and the power prediction module to obtain the optimal prediction time domain length. The predicted power demand within the optimal prediction time domain is then used for dynamic programming. Dynamic programming is used to continuously optimize the hybrid power system energy allocation strategy until the optimal energy allocation strategy of the hybrid power system is obtained, so as to achieve the minimum equivalent fuel consumption of the UAV system and improve flight economy and endurance.

Images