A Multi-Electric Aircraft Hybrid Energy Energy Management Method Based on Stochastic Model Prediction

Abstract

The present invention discloses a multi-electric aircraft hybrid energy energy management method based on stochastic model prediction, which is used to solve the problems of limitations and poor applicability of common energy management strategies for hybrid energy systems. The hybrid energy system architecture consists of a three-stage generator, a lithium-ion battery, a supercapacitor, an AC / DC converter, and a bidirectional DC-DC converter. In the rolling optimization calculation, dynamic programming is combined with multi-step real-time difference, and the time step is weighted, ensuring that the output voltage of the generator fluctuates little and the output power is stable, and the SOCs of the lithium battery and the supercapacitor are both in the safe operating range, making the model have strong practicability.

Description

Technical Field

[0001] The present invention belongs to the field of energy management, and particularly relates to a method for an energy management system of a multi-electric aircraft hybrid energy based on stochastic model prediction. Background Art

[0002] With the continuous increase in the electrification level of aircraft, multi-electric aircraft adopt more and more power systems as their secondary energy systems. Therefore, multi-electric aircraft require a large-capacity power supply system. The 270V high-voltage DC system greatly increases the power supply capacity and has the characteristics of a light grid quality and easy implementation of uninterrupted power supply, and has been applied to military multi-electric aircraft F35 and F22. In order to achieve reliable 270V power supply, multi-electric aircraft usually adopt a hybrid energy power supply mode. The electric energy between different energy systems needs to be reasonably managed to ensure the safe and stable operation of multi-electric aircraft. The energy management strategies commonly used in the hybrid energy system of multi-electric aircraft include state machine control strategy, fuzzy logic control strategy, classical PI control strategy, and equivalent fuel consumption minimum strategy.

[0003] The literature "Wang T, Qi L, Chen W, et al. Application of energy management strategy based on state machine in fuel cell hybrid power system [C] / / 2017 IEEE Transportation Electrification Conference and Expo, Asia-Pacific (ITEC Asia-Pacific). IEEE, 2017." proposed an energy management strategy based on a state machine for a hybrid energy system combining air-cooled fuel cells and lithium batteries. This strategy is based on switch rule control and determines the reference output power of each power supply system according to the load power demand and the SOC of the lithium battery, which can meet the power demand of different loads and achieve dynamic distribution of energy. However, for different initial conditions of the system, the control effect of the state machine control strategy varies greatly and the adaptability is poor.

[0004] The literature "Xie C, Xu X, Bujlo P, et al. Fuel cell and lithium iron phosphate battery hybrid powertrain with an ultracapacitor bank using direct parallel structure[J]. Journal of Power Sources, 2015, 279: 487-494." proposed an energy management strategy based on fuzzy logic control for a hybrid energy system with fuel cells, lithium iron phosphate batteries, and supercapacitors in parallel, achieving the goal of stabilizing the DC bus voltage. However, the fuzzy logic control strategy has low control accuracy, poor dynamic quality, and lacks systematicness.

[0005] The literature "Motapon S N, Dessaint L A, Al-Haddad K. A Comparative Study of Energy Management Schemes for a Fuel-Cell Hybrid Emergency Power System of More-Electric Aircraft[J]. IEEE Transactions on Industrial Electronics, 2013, 61(3): 1320-1334." pointed out that the energy management strategy based on PI control can achieve online optimization to ensure that the fuel cell system stably provides the power required by the load. However, the classical PI control strategy will reduce the relative stability of the system, and its parameter tuning is difficult, not easy to adjust, and has low applicability.

[0006] The literature "Zhang G, Chen W, Jin Y, et al. Study on equivalent consumption minimization strategy for fuel cell hybrid tramway[C] / / Transportation Electrification Asia-pacific. IEEE, 2017." proposed an equivalent fuel consumption strategy applicable to fuel cell hybrid trams, which can ensure the effective distribution of the energy required by the load and the stability of the bus voltage. However, the value of the equivalent factor of the equivalent fuel consumption strategy needs to be adjusted according to the load conditions, otherwise the optimization effect will decline and the applicability is poor.

[0007] At the same time, the above strategies are all based on hybrid energy systems containing fuel cells, which have certain limitations and may not be suitable for energy management between generators, lithium batteries and supercapacitors. Summary of the invention

[0008] In view of the shortcomings of the above strategies, the present invention proposes a hybrid energy management method for more-electric aircraft based on random model prediction, which mainly solves the limitations and poor applicability of common control strategies, while achieving online optimization and better control accuracy. Specifically, the purpose of the present invention is to improve the following aspects:

[0009] 1. The existing control strategies have poor adaptability to changes in operating conditions and parameter adjustments.

[0010] 2. Traditional optimization control strategies require the known aircraft operating conditions to obtain the global optimal solution and cannot achieve real-time control.

[0011] 3. General optimization control strategies require large amounts of calculations, slow calculation speeds, and high requirements for microprocessors.

[0012] 4. The systems currently being studied are usually fuel cell based, which has limitations and is not easily expandable to other systems.

[0013] In view of the shortcomings of the above strategies, the present invention proposes a hybrid energy management method for a more-electric aircraft based on random model prediction. The hybrid energy system adopts the prediction results based on random model predictive control according to the real-time power required by the dynamic load module and takes random disturbances into consideration, and distributes power to the three generators, lithium batteries and supercapacitors according to the energy management strategy.

[0014] The rolling optimization process of the hybrid energy management method adopts a method combining multi-step instant differential learning and dynamic programming;

[0015] The inter-component collaborative control module controls the power output of the three generators, the lithium battery and the supercapacitor according to the power distribution result;

[0016] Furthermore, the hybrid energy management method applied to a more-electric aircraft comprises the following steps:

[0017] S1 Establish random disturbance prediction model

[0018] By controlling the main generator (P mg1 , P mg2 ), auxiliary generator (P ag )、Lithium battery (P B ) and supercapacitor (P UC ) to achieve system power balance and ensure that each power supply device operates in the optimal state;

[0019] At sampling time k, the control variable is taken as:

[0020] u(k) = [P mg1 (k), P mg2 (k), P UC (k), P B (k), P ag (k)] T

[0021] where P mg1 (k) is the power of the main generator 1, P mg2 (k) is the power of the main generator 2, P UC (k) is the power of the supercapacitor module, P B (k) is the power of the lithium battery module, P ag (k) is the power of the auxiliary generator module.

[0022] Then:

[0023] Δu(k) = u(k) - u(k - 1) = [ΔP mg1 (k), ΔP mg2 (k), ΔP UC (k), ΔP B (k), ΔP ag (k)] T

[0024] where Δu(k) is the change in the control quantity between the current moment and the previous moment, ΔP mg1 (k) is the change in the power of the main generator 1, ΔP mg2 (k) is the change in the power of the main generator 2, ΔP UC (k) is the change in the power of the supercapacitor module, ΔP B (k) is the change in the power of the lithium battery module, ΔP ag (k) is the change in the power of the auxiliary generator module.

[0025] The state variable matrix is:

[0026] x(k) = u(k) = [P mg1 (k), P mg2 (k), P UC (k), P B (k), P ag (k), SOC UC (k), SOC B (k)] T where

[0027] SOC UC(k) is the state of charge of the supercapacitor, SOC B (k) is the state of charge of the lithium battery.

[0028] The output variables are:

[0029] y(k) = [P mg1 (k) + P mg2 (k) + P UC (k) + P B (k) + P ag (k), P mg1 (k), P mg2 (k), SOC UC (k), SOC B (k)] T

[0030] In the formula, SOC UC (k) and SOC B (k) are the state of charge of the supercapacitor and the lithium battery respectively. The relationship between the state of charge and power at adjacent sampling moments should satisfy:

[0031]

[0032] In the formula, SOC UC (k - 1) is the state of charge of the supercapacitor at the moment before the sampling moment, SOC B (k - 1) is the state of charge of the lithium battery at the moment before the sampling moment, E UC 、E B are the capacities of the supercapacitor and the lithium battery respectively; Δt is the sampling step,

[0033] The prediction model with random disturbance added is:

[0034]

[0035] In the formula, k is the current sampling moment;; Δu(k) is the change in the control quantity from the current moment to the previous moment, x(k + 1) is the state variable matrix at the next moment, y(k) is the output variable matrix at the current moment, and w(k) is the random disturbance generated by the outside world on the system at the k moment. A, B, and C are the state, input, and output matrices respectively

[0036]

[0037]

[0038]

[0039] S2 sets the constraint conditions

[0040] Based on the random model predictive control strategy, considering the characteristics of two energy storage devices, setting control and state constraints, and providing optimized control signals for each module of the hybrid energy system.

[0041] S21 Set the system output power constraint,

[0042] Assume the system has no losses, and the load power is the sum of the powers of two main generators (P mg1 、P mg2 ), auxiliary generator (P ag ), lithium battery (P B ), and supercapacitor (P UC ), satisfying:

[0043] P mg1 (k + i|k) + P mg2 (k + i|k) + P UC (k + i|k) + P B (k + i|k) + P ag (k + i|k) = P load

[0044] where x(k + i|k) is the predicted value of x at time k + i based on the current sampling time k; P load is the load power.

[0045] S22 Set the charge and discharge power constraints:

[0046]

[0047] where x(k + i|k) is the predicted value of x at time k + i based on the current sampling time k; P mg_MAX is the maximum power of the main generator; P UC_MIN 、P UC_MAX are the minimum and maximum powers of the supercapacitor respectively; P B_MIN 、P B_MAX are the minimum and maximum powers of the lithium battery respectively; P ag_MAX is the auxiliary generator power.

[0048] S23 Set the state of charge constraints:

[0049]

[0050] SOC UC_MIN 、SOC UC_MAX are the minimum and maximum states of charge of the supercapacitor; SOC B_MIN 、SOC B_MAX are the minimum and maximum states of charge of the lithium battery. SOC UC (k + i|k)、SOC B(k+i|k) are the predicted output values of the state of charge of the supercapacitor and the lithium battery at the current sampling time k at time k+i, respectively.

[0051] S3 Rolling optimization

[0052] The optimization part in model predictive control is to perform rolling optimization within a finite time domain. At each sampling time, the future control actions that optimize the performance index function within the prediction time domain starting from that time are solved, and then advanced to the next sampling time, and the optimization time domain also advances forward.

[0053] Considering the safe and economic operation of the hybrid energy system, the system control objectives are mainly divided into two parts:

[0054] S31 During the operation of the system, while satisfying the load demand as much as possible at each sampling time, the power distribution balance of the system should be maintained to ensure the normal operation of each distributed unit.

[0055] S32 To protect the life of the generator, it is preferred to ensure that the output power of the generator remains constant.

[0056] The optimization model of the system uses the difference between the output value of the controlled object at future sampling points and the desired trajectory. Therefore, an optimization model that meets the control objectives, that is, the performance index function, is defined as:

[0057]

[0058] In the formula, k = 0, 1, 2...; Q is a positive definite weighting coefficient matrix of the prediction output error; P mean is the average power of the system, that is, the reference trajectory of the system; P mg1 (k+i|k), P mg2 (k+i|k) are the predicted output values of the power of two main generators at time k+i at the current sampling time k.

[0059] Currently, when obtaining the control actions in rolling optimization for the energy management control strategy based on stochastic model predictive control, dynamic programming is often used to obtain the exact solutions of the optimal value function and the optimal control variables in the prediction domain. Multi-step real-time difference learning is a special incremental learning method that uses the experience traversed by the system and the current incomplete state information to predict the actions of the system.

[0060] In the optimization calculation, a method combining multi-step immediate difference learning and dynamic programming is used to obtain the optimal control actions. In the early stage of the algorithm operation, the dynamic programming method is used to calculate the optimal control actions within the control domain and use them as ideal data samples.

[0061] First, within a certain period at the initial stage of the system operation, it is obtained by using the dynamic programming algorithm. The specific algorithm is as follows:

[0062] At the (N - 1)-th decision step,

[0063]

[0064] J * (x(N - 1)) is the optimal value function at the (N - 1)-th decision step and state x(N - 1). Pmean is the system average power, which is also the reference trajectory of the system; P mg1 (k + i|k), P mg2 (k + i|k) are the predicted output values of the powers of two main generators at time k + i at the current sampling time k.

[0065] At the k-th (k ∈ [0 ≤ k < N - 1)) decision step:

[0066]

[0067] Among them, J * (x(k)), J * (x(k + 1)) are the optimal value functions at the decision steps and states x(k) and x(k + 1) at the k-th and (k + 1)-th times respectively. P mean is the system average power, which is also the reference trajectory of the system; P mg1 (k + i|k), P mg2 (k + i|k) are the predicted output values of the powers of two main generators at time k + i at the current sampling time k. The problem can be solved by gradually iterating from back to front. The optimal control vector at the k-th decision step is as follows:

[0068] u * (k) = argmin u(k) J(x(k))

[0069] Among them, J * (x(k)) is the optimal value function at the decision step and state x(k + 1) at the k-th time.

[0070] After calculating the optimal predictive control sequence {u * (k), u * (k + 1), …, u * (N - 1)} at time k, let the control action u(k) at time k = u1 * (k).

[0071] In the control vector predicted by the multi-step immediate difference learning algorithm, the weight exponent λ ∈ (0, 1) is weighted for the past k time steps, that is, the weight change amount is:

[0072]

[0073] where ω is the weight, and Δω t is the change in weight within the control domain time interval t; u t is the control action at time t within the control domain predicted by the instant difference learning algorithm; α > 0 is the learning factor; λ is the weight exponent, and the learning error at time t is:

[0074]

[0075] where ω is the weight; e t is the learning error at time t; u k is the control action at the moment within the control domain predicted by the instant difference learning algorithm; λ is the weight exponent.

[0076] Meanwhile,

[0077]

[0078] where u t+1 is the control action at time t + 1 within the control domain predicted by the instant difference learning algorithm; ω is the weight; e t is the learning error at time t; λ is the weight exponent.

[0079] S4 Feedback correction

[0080] Since the prediction model is not the true physical model of the controlled system, it is affected by factors such as system time-variation and non-linearity. There is an error between the system output in the prediction domain calculated by the prediction model and the actual system output. On this basis, the stochastic model predictive control method compares the stochastic model prediction based on the actual system output with the prediction output at each sampling moment, and then corrects the subsequent predictions to form a closed-loop optimization. It increases the stability of the algorithm and solves the problems of model mismatch and interference.

[0081] The feedback correction is based on the following formula:

[0082] y p (k + i) = y m (k + i|k) + e(k), i = 0, 1, …, N

[0083] where N P is the prediction domain; y m (k + i|k) and y p (k + i) are the system outputs at time k + i within the prediction domain of the prediction model and the output after feedback correction at time k respectively; e(k) is the error between the system output predicted by the prediction model at time k and the true output.

[0084] The error is:

[0085] e(k) = ym (k) - y(k)

[0086] Among them, y m (k) is the output of the system at time k predicted by the stochastic model.

[0087] The present invention mainly solves the problems of the limitations and poor applicability of common control strategies, and at the same time realizes online optimization and better control accuracy, and is applicable to the energy management of multi - electric aircraft. Its technical solution includes an energy management method based on stochastic model prediction. Description of the Drawings

[0088] By reading the following detailed description of the preferred embodiments, various other advantages and benefits will become clear to those of ordinary skill in the art. The drawings are only for the purpose of showing the preferred embodiments and are not considered to be a limitation of the present invention. Moreover, throughout the drawings, the same reference numerals are used to represent the same components. In the drawings:

[0089] Figure 1 is the block diagram of the hybrid - energy energy management system in the present invention.

[0090] Figure 2 is the control block diagram of the stochastic model predictive control.

[0091] Figure 3 is the flowchart of the method combining multi - step immediate difference learning and dynamic programming for the rolling optimization process. Detailed Embodiments

[0092] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Usually, the components of the embodiments of the present invention described and illustrated herein can be arranged and designed in various different configurations.

[0093] Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed present invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0094] Design of the energy management method based on stochastic model prediction:

[0095] First, establish model information based on the structure of the hybrid energy system. Then, comprehensively utilize historical information and model information to perform rolling optimization on the objective function through a method that combines multi-step immediate difference learning and dynamic programming to achieve a globally optimal control effect. Finally, compare the measured electrical signal with the predicted output, correct the output parameters, and the output parameters are the power distributions of the generator, lithium battery, and supercapacitor. Accordingly, the intelligent optimal energy distribution of the multi-electric aircraft under different operating conditions is realized.

[0096] A hybrid energy management method for a multi-electric aircraft based on stochastic model prediction. The hybrid energy system, considering random disturbances, adopts the prediction results based on stochastic model predictive control according to the real-time power required by the dynamic load module, and distributes the power among three generators, a lithium battery, and a supercapacitor according to the energy management strategy. The hybrid energy system is as Figure 1 shown.

[0097] The rolling optimization process of the hybrid energy management method uses a method that combines multi-step immediate difference learning and dynamic programming for data processing;

[0098] The component collaborative control module controls the power outputs of the three generators, the lithium battery, and the supercapacitor according to the power distribution results;

[0099] Energy management method based on stochastic model prediction:

[0100] S1 Establish a stochastic disturbance prediction model

[0101] By controlling the powers of the main generator (P mg1 , P mg2 ), the auxiliary generator (P ag ), the lithium battery (P B ), and the supercapacitor (P UC ), the power balance of the system is achieved, and each power supply device operates in an optimal state;

[0102] At the sampling time k, the control variable is taken as:

[0103] u(k) = [P mg1 (k), P mg2 (k), P UC (k), P B (k), P ag (k) T

[0104] where, P mg1 (k) is the power of the main generator 1, P mg2 (k) is the power of the main generator 2, P UC (k) is the power of the supercapacitor module, P B(k) is the power of the lithium battery module, P ag (k) is the power of the auxiliary generator module.

[0105] Then:

[0106] Δu(k) = u(k) - u(k - 1) = [ΔP mg1 (k), ΔP mg2 (k), ΔP UC (k), ΔP B (k), ΔP ag (k)] T

[0107] where Δu(k) is the change in the control variable from the current moment to the previous moment, ΔP mg1 (k) is the change in the power of the main generator 1, ΔP mg2 (k) is the change in the power of the main generator 2, ΔP UC (k) is the change in the power of the supercapacitor module, ΔP B (k) is the change in the power of the lithium battery module, ΔP ag (k) is the change in the power of the auxiliary generator module.

[0108] The state variable matrix is:

[0109] x(k) = u(k) = [P mg1 (k), P mg2 (k), P UC (k), P B (k), P ag (k), SOC UC (k), SOC B (k)] T where SOC UC (k) is the state of charge of the supercapacitor, SOC B (k) is the state of charge of the lithium battery.

[0110] The output variable is:

[0111] y(k) = [P mg1 (k) + P mg2 (k) + P UC (k) + P B (k) + P ag (k), P mg1 (k), P mg2 (k), SOC UC (k), SOC B (k)] T

[0112] In the formula, SOCUC (k) and SOC B (k) are the state of charge of the supercapacitor and the lithium battery respectively. The relationship between the state of charge and power at adjacent sampling times should satisfy:

[0113]

[0114] In the formula, SOC UC (k - 1) is the state of charge of the supercapacitor at the previous moment before the sampling time, and SOC B (k - 1) is the state of charge of the lithium battery at the previous moment before the sampling time. E UC and E B are the capacities of the supercapacitor and the lithium battery respectively; Δt is the sampling step,

[0115] The prediction model with random perturbation added is:

[0116]

[0117] In the formula, k is the current sampling time; Δu(k) is the change in the control quantity from the current moment to the previous moment, x(k + 1) is the state variable matrix at the next moment, y(k) is the output variable matrix at the current moment, and w(k) is the random perturbation generated by the outside world on the system at time k. A, B, and C are the state, input, and output matrices respectively

[0118]

[0119]

[0120]

[0121] S2 sets constraint conditions

[0122] Based on the stochastic model predictive control strategy, considering the characteristics of the two energy storage devices, control and state constraints are set to provide optimized control signals for each module of the hybrid energy system.

[0123] S21 sets the system output power constraint,

[0124] Assume that the system has no loss, and the load power is the sum of the powers of two main generators (P mg1 and P mg2 ), an auxiliary generator (P ag ), a lithium battery (P B ), and a supercapacitor (P UC ), and it satisfies:

[0125] P mg1 (k + i|k) + P mg2 (k + i|k) + P UC(k+i|k)+P B (k+i|k)+P ag (k+i|k)=P load

[0126] Where x(k+i|k) is the predicted value of x at the current sampling time k for the time k+i; P load is the load power.

[0127] S22 sets the charge and discharge power constraints:

[0128]

[0129] Where x(k+i|k) is the predicted value of x at the current sampling time k for the time k+i; P mg_MAX is the maximum power of the main generator; P UC_MIN , P UC_MAX are the minimum and maximum power of the supercapacitor respectively; P B_MIN , P B_MAX are the minimum and maximum power of lithium battery respectively; P ag_MAX Auxiliary generator power.

[0130] S23 sets the state of charge constraint:

[0131]

[0132] SOC UC_MIN , SOC UC_MAX SOC is the minimum and maximum state of charge of the supercapacitor. B_MIN , SOC B_MAX It is the minimum and maximum state of charge of lithium battery. UC (k+i|k), SOC B (k+i|k) are the predicted output values ​​of the state of charge of the supercapacitor and the lithium battery at the current sampling time k at the time k+i.

[0133] S3 scrolling optimization

[0134] The optimization part of model predictive control is a rolling optimization in a finite time domain. At each sampling moment, the future control action that optimizes the indicator function in the prediction time domain from that moment is solved, and then it is advanced to the next sampling moment, and the optimization time domain also moves forward.

[0135] Considering the safe and economical operation of the hybrid energy system, the system control objectives are mainly divided into two parts:

[0136] During the operation of the system, S31 should maintain the balance of system power distribution and ensure the normal operation of each distributed unit under the premise of meeting the load demand as much as possible at each sampling moment.

[0137] S32 To protect the lifespan of the generator, it is preferred to ensure that the output power of the generator remains constant.

[0138] The optimization model of the system uses the difference between the output value of the controlled object at future sampling points and the desired trajectory. Therefore, an optimization model that meets the control objective, that is, the objective function, is defined as:

[0139]

[0140] In the formula, k = 0, 1, 2…; Q is a positive definite weighted coefficient matrix of the prediction output error; P mean is the system average power, that is, the reference trajectory of the system; P mg1 (k + i|k), P mg2 (k + i|k) are the predicted output values of the powers of two main generators at the k + i moment at the current sampling moment k.

[0141] Currently, when obtaining the control action in the rolling optimization based on the energy management control strategy of stochastic model predictive control, dynamic programming is often used to obtain the exact solutions of the optimal value function and the optimal control variables in the prediction domain. Multi-step real-time difference learning is a special incremental learning method that uses the experience traversed by the system and the current incomplete state information to predict the actions of the system.

[0142] In the optimization calculation, a method combining multi-step immediate difference learning and dynamic programming is used to obtain the optimal control action. In the early stage of the algorithm operation, the dynamic programming method is used to calculate the optimal control action within the control domain and use it as an ideal data sample. The specific algorithm process is as Figure 3 .

[0143] First, within a certain period at the initial stage of the system operation, the dynamic programming algorithm is used to obtain. The specific algorithm is as follows:

[0144] At the (N - 1)-th decision step,

[0145]

[0146] J * (x(N - 1)) is the optimal value function at the (N - 1)-th decision step and the state x(N - 1). Pmean is the system average power, that is, the reference trajectory of the system; P mg1 (k + i|k), P mg2 (k + i|k) are the predicted output values of the powers of two main generators at the k + i moment at the current sampling moment k.

[0147] At the k-th (k ∈ [0 ≤ k < N - 1)) decision step:

[0148]

[0149] Among them, J * (x(k)) and J * (x(k + 1)) are the optimal value functions at the k-th moment and the k + 1-th moment for the decision step and states x(k) and x(k + 1) respectively. P mean is the system average power, that is, the reference trajectory of the system; P mg1 (k + i|k) and P mg2 (k + i|k) are the predicted output values of the power of two main generators at the k + i-th moment at the current sampling moment k. The problem can be solved by gradually iterating from the back to the front. The optimal control vector at the k-th decision step is as follows:

[0150] u * (k) = argmin u(k) J(x(k))

[0151] Among them, J * (x(k)) is the optimal value function at the k-th moment for the decision step and state x(k + 1).

[0152] After calculating the optimal predictive control sequence {u * (k), u * (k + 1), …, u * (N - 1)} at the k-th moment, let the control action at the k-th moment u(k) = u1 * (k).

[0153] In the control vector predicted by the multi-step immediate difference learning algorithm, the weight exponent λ ∈ (0, 1) is weighted for the past k time steps, that is, the weight change amount is:

[0154]

[0155] Among them, ω is the weight, Δω t is the weight change amount within the control domain time interval t; u t is the control action at the t-th moment within the control domain predicted by the immediate difference learning algorithm; α > 0 is the learning factor; λ is the weight exponent, then the learning error at the t-th moment is:

[0156]

[0157] Among them, ω is the weight; e t is the learning error at the t-th moment; u k is the control action at the moment within the control domain predicted by the immediate difference learning algorithm; λ is the weight exponent.

[0158] Meanwhile,

[0159]

[0160] Among them, u t+1 is the control action at time t+1 within the control domain predicted by the instant differential learning algorithm; ω is the weight; e t is the learning error at time t; λ is the weight exponent.

[0161] S4 Feedback correction

[0162] Since the prediction model is not the true physical model of the controlled system, it is affected by factors such as system time-variation and non-linearity. There is an error between the system output in the prediction domain calculated by the prediction model and the actual output of the system. On this basis, the stochastic model predictive control method compares the stochastic model prediction based on the actual output of the system with the predicted output at each sampling moment, and then corrects the subsequent predictions to form a closed-loop optimization. It increases the stability of the algorithm and solves the problems of model mismatch and interference.

[0163] The feedback correction is based on the following formula:

[0164] y p (k+i) = y m (k+i|k) + e(k), where i = 0, 1, …, N

[0165] Among them, N P is the prediction domain; y m (k+i|k) and y p (k+i) are respectively the system outputs at time k+i within the prediction domain of the prediction model and the output after feedback correction at time k; e(k) is the error between the system output predicted by the prediction model at time k and the true output.

[0166] The error is:

[0167] e(k) = y m (k) - y(k)

[0168] Among them, y m (k) is the output of the system at time k predicted by the stochastic model.

[0169] Beneficial effects

[0170] 1) The energy management strategy can achieve intelligent optimal allocation of energy among three generators, lithium batteries and supercapacitors when the peak load powers are different, with high control precision.

[0171] 2) The energy management strategy can keep the generator operating at the rated power all the time when it is put into use, without being disturbed by load changes, ensuring the maximum efficiency of the generator and extending the service life of the generator.

[0172] 3) The energy management strategy can use the supercapacitor to suppress the system power deficit, ensure that the lithium battery is in an ideal state of charge, and extend the service life of the lithium battery.

[0173] 4) The energy management method has a small amount of calculation, a fast calculation speed, reduces the requirements for the microprocessor, and reduces the cost.

[0174] As mentioned above, only the preferred specific embodiments of the present invention are described, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A multi-electric aircraft hybrid energy management method based on random model prediction, characterized in that, The hybrid energy system, according to the real-time power required by the dynamic load module, considering random disturbances, adopts the prediction results based on stochastic model predictive control, and distributes the power among three generators, a lithium battery, and a supercapacitor according to the energy management strategy; The rolling optimization process of the hybrid energy management method adopts a method combining multi-step immediate difference learning and dynamic programming; The component collaborative control module controls the power outputs of the three generators, the lithium battery, and the supercapacitor according to the power distribution results. The method for multi-electric aircraft hybrid energy management based on stochastic model prediction includes the following steps: S1 Establish a stochastic disturbance prediction model By controlling the power of the main generator (P mg1 , P mg2 ), auxiliary generator (P ag ), lithium battery (P B ) and supercapacitor (P UC ), the power balance of the system is achieved, ensuring that each power supply device operates in an optimal state; At the sampling time k, the control variable is taken as: u(k) = [P mg1 (k), P mg2 (k), P UC (k), P B (k), P ag (k)] T Among them, P mg1 (k) is the power of the main generator 1, P mg2 (k) is the power of the main generator 2, P UC (k) is the power of the supercapacitor module, P B (k) is the power of the lithium battery module, P ag (k) is the power of the auxiliary generator module; Then: Δu(k) = u(k) - u(k - 1) = [ΔP mg1 (k), ΔP mg2 (k), ΔP UC (k), ΔP B (k), ΔP ag (k)] T where, Δu(k) is the change in the control quantity from the current moment to the previous moment, ΔP mg1 (k) is the change in the power of the main generator 1, ΔP mg2 (k) is the change in the power of the main generator 2, ΔP UC (k) is the change in the power of the supercapacitor module, ΔP B (k) is the change in the power of the lithium battery module, ΔP ag (k) is the change in the power of the auxiliary generator module; The state variable matrix is: x(k) = u(k) = [P mg1 (k), P mg2 (k), P UC (k), P B (k), P ag (k), SOC UC (k), SOC B (k)] T Among them, SOC UC (k) is the state of charge of the supercapacitor, and SOC B (k) is the state of charge of the lithium battery; the output variables are: y(k) = [P mg1 (k) + P mg2 (k) + P UC (k) + P B (k) + P ag (k), P mg1 (k), P mg2 (k), SOC UC (k), SOC B (k)] T where SOC UC (k) and SOC B (k) are the state of charge of the supercapacitor and the lithium battery respectively, and the relationship between the state of charge and the power at adjacent sampling times should satisfy: Where SOC UC (k - 1) is the state of charge at the moment before the sampling moment of the supercapacitor, and SOC B (k - 1) is the state of charge at the moment before the sampling moment of the lithium battery, E UC 、E B are the capacities of the supercapacitor and the lithium battery respectively; △t is the sampling step size, The prediction model with added random disturbances is: In the formula, k is the current sampling time; Δu(k) is the change in the control quantity from the current moment to the previous moment, x(k + 1) is the state variable matrix at the next moment, y(k) is the output variable matrix at the current moment, w(k) is the random disturbance generated by the outside world on the system at time k, and A, B, and C are the state, input, and output matrices respectively. S2 Set constraint conditions Based on the stochastic model predictive control strategy, considering the characteristics of the two energy storage devices, control and state constraints are set to provide optimized control signals for each module of the hybrid energy system; S3 Rolling optimization The optimization part in model predictive control is to perform rolling optimization within a finite time domain; at each sampling time, the future control actions that optimize the index function within the prediction time domain starting from this time are solved, and then advanced to the next sampling time, and the optimization time domain will also advance forward; S4 Feedback correction The feedback correction is based on the following formula: y p (k + i)= y m (k + i|k)+ e(k), i = 0, 1, ···, N where, N P is the prediction domain; y m (k + i|k) and y p (k + i) are the system outputs at the (k + i)-th moment within the prediction domain of the prediction model and the output after feedback correction at the k-th moment, respectively; e(k) is the error between the system output predicted by the prediction model at the k-th moment and the true output; The error is: e(k) = y m (k) - y(k) where y m (k) is the output of the system at time k predicted by the random model.

2. The multi-electric aircraft hybrid energy energy management method based on model prediction according to claim 1, wherein The specific constraint conditions set in S2 are: S21 Set the system output power constraint, Assume that the system has no losses, and the load power is the sum of the powers of two main generators (P mg1 , P mg2 ), auxiliary generator (P ag ), lithium battery (P B ), and supercapacitor (P UC ), satisfying: P mg1 (k + i|k) + P mg2 (k + i|k) + P UC (k + i|k) + P B (k + i|k) + P ag (k + i|k) = P load where x(k+i|k) is the predicted value of x at time k+i predicted at the current sampling time k; P load is the load power; S22 Set the charge and discharge power constraint: where, x(k+i|k) is the predicted value of x at time k+i predicted at the current sampling time k; P mg_MAX is the maximum power of the main generator; P UC_MIN , P UC_MAX are the minimum and maximum powers of the supercapacitor respectively; P B_MIN , P B_MAX are the minimum and maximum powers of the lithium battery respectively; P ag_MAX is the power of the auxiliary generator; S23 Set the state of charge constraint: SOC UC_MIN and SOC UC_MAX are the minimum and maximum state of charge of the supercapacitor; SOC B_MIN and SOC B _ MAX are the minimum and maximum state of charge of the lithium battery; SOC UC (k + i|k) and SOC B (k + i|k) are the predicted output values of the state of charge of the supercapacitor and the lithium battery at the current sampling time k at time k + i, respectively.

3. A model predictive-based hybrid energy energy management method for a multi-electric aircraft according to claim 1, characterized in that, The specific rolling optimization in S3 is: The system control objective is divided into two parts: S31 During the operation of the system, under the premise of satisfying the load demand as much as possible at each sampling time, the power distribution balance of the system should be maintained to ensure the normal operation of each distributed unit; S32 To protect the life of the generator, give priority to ensuring that the output power of the generator remains unchanged; The optimization model of the system uses the difference between the output value of the controlled object at future sampling points and the desired trajectory; define the optimization model that meets the control objective, that is, the objective function is: where \(k = 0, 1, 2,\cdots\); \(Q\) is a positive definite weighting coefficient matrix of the predicted output error; \(P\) mean is the system average power, that is, the reference trajectory of the system; \(P\) mg1 (k + i|k), \(P\) mg2 (k + i|k) are the predicted output values of the powers of two main generators at the current sampling time \(k\) at the time \(k + i\); First, within a certain period at the initial stage of the system operation, the dynamic programming algorithm is used to obtain; The specific algorithm is as follows: At the (N - 1)-th decision step, J * (x(N - 1)) is the optimal value function at the (N - 1)-th decision step and state x(N - 1); P mean is the system average power, which is also the reference trajectory of the system; P mg1 (k + i|k), P mg2 (k + i|k) are the predicted output values of the powers of two main generators at the current sampling time k at time k + i; At the k-th (k ∈ [0 ≤ k < N - 1)) decision step: Among them, J * (x(k)) and J * (x(k + 1)) are the optimal value functions at the k-th moment and the (k + 1)-th moment for the decision step and states x(k) and x(k + 1) respectively; P mean is the system average power, which is also the reference trajectory of the system; P mg1 (k + i|k) and P mg2 (k + i|k) are the predicted output values of the powers of two main generators at the (k + i)-th moment at the current sampling moment k; The problem can be solved by gradually iterating from the back to the front; The optimal control vector at the k-th decision step is as follows: u * (k) = argmin u(k) J(x(k)) Among them, J * (x(k)) is the optimal value function at the k-th decision step and state x(k + 1); After calculating the optimal predictive control sequence {u * (k), u * (k + 1), …, u * (N - 1)} at time k, let the control action u(k) at time k be u1 * (k); Among the control vectors predicted by the multi-step immediate difference learning algorithm, the weight exponent λ ∈ (0, 1) is weighted for the past k time steps, that is, the weight change amount is: where ω is the weight, and Δω t is the change in weight within the control domain time interval t; u t is the control action at time t within the control domain predicted by the immediate difference learning algorithm; α > 0 is the learning factor; λ is the weight exponent, then the learning error at time t is: where ω is the weight; e t is the learning error at time t; u k is the control action at the moment within the control domain predicted by the instant difference learning algorithm; λ is the weight exponent; Meanwhile, Among them, u t+1 is the control action at the (t + 1)-th moment within the control domain predicted by the instant differential learning algorithm; ω is the weight; e t is the learning error at time t; λ is the weight exponent.

Images