LEO satellite life optimization method based on data prediction matching

Abstract

The invention discloses an LEO satellite life optimization method based on data prediction matching. Firstly, in order to measure the optimization effect of the life of the LEO satellite and the capability of ground business processing, the product of the average life of the LEO satellite and the energy of accumulated processing business is calculated by using an index, and the product is used for measuring the effects of different algorithms in the aspects of optimizing the life of the LEO satellite and the capability of processing the business. And secondly, predicting solar radiation energy in the LEO satellite orbit by using a convolutional neural network-long short-term memory network (CNN-LSTM) algorithm. Besides, the sub-satellite trajectory of the LEO satellite is obtained through STK software simulation, an LEO satellite service demand observation model is established according to the sub-satellite trajectory, and then the service demand energy of the LEO satellite is predicted by using an LSTM-RF algorithm. And finally, based on the Lyapunov theory, converting the energy management problem of the LEO satellite from a random optimization problem to a determined optimization problem of each time slot, and solving by using an improved adaptive model predictive control (AMPC) algorithm to obtain an energy management decision of each time slot of the LEO satellite. Simulation experiment results show that compared with an MPC algorithm, a Game Theory algorithm and a Greedy Algorithm algorithm, the effects of the method are improved by 10.79%, 25.22% and 58.16% respectively, the service life of an LEO satellite can be remarkably prolonged, the service processing capacity of the LEO satellite can be remarkably improved, and the operation cost of the LEO satellite can be remarkably reduced.

Description

Technical Field

[0001] The invention relates to the field of LEO satellite communications, and in particular to a method for optimizing the life of a LEO satellite based on data prediction and matching. Background Art

[0002] At present, the industry has reached a broad consensus on the vision of the sixth generation mobile communication system (6G), which can be summarized as "three alls and one strong": full connectivity, full coverage, all-round applications, and strong security. However, since the 6G network needs to achieve three-dimensional full coverage, the use of terahertz spectrum, and the application of AI functions, these applications have led to a substantial increase in the business volume of the 6G network, which has led to a particularly serious energy consumption problem for the 6G network. Therefore, in the early stage of the development of the 6G network system, it is necessary to adopt the energy-saving concept to build a green and endogenous 6G network.

[0003] Low Earth Orbit (LEO) satellite networks (LSNs) have become an important link in achieving full coverage of 6G with their advantages such as low latency, high transmission bandwidth and low power consumption. It is foreseeable that the amount of business that LSNs will need to undertake in the future will be huge, and LSNs satellites need to have powerful computing capabilities to handle these huge business volumes. Powerful computing power requires a lot of energy, which will lead to a sharp increase in the power demand of LSNs, which may quickly deplete the batteries of LSNs, thus affecting the energy-constrained LEO satellites in LSNs. On the other hand, in the absence of sunlight, LSNs must operate only on battery energy. The power consumption during these periods will place a heavy burden on satellite batteries and may shorten their service life. This will bring huge operating costs to LSNs. Once the LEO satellites in LSNs fail due to multiple deep battery charge and discharge cycles, they cannot be maintained. Satellite constellation operators can only launch new satellites to replace the failed satellites, which brings high economic costs. The total cost of manufacturing and launching a single Starlink satellite exceeds US$1 million. Even with the cost reductions, updating failed satellites for mega-constellations with tens of thousands of satellites is still a considerable expense.

[0004] In response to the challenge of improving the life of LEO satellites in LSNs, there are many papers that have studied the optimization of the energy supply system of LEO satellites. Some papers modeled and simulated the energy conversion of LEO satellite solar panels and compared the simulation results to determine the best choice of solar panels, thereby improving the solar radiation energy obtained by LEO satellites. Some papers proposed two efficient routing methods to reduce the depth of satellite discharge by optimizing traffic, thereby improving the battery life of LEO satellites and ensuring the continuous operation of LSNs. There are also papers that have developed a communication method that controls the transmission power and transmission gain of the satellite antenna according to the degradation state of the battery to extend the battery life.

[0005] When optimizing the life of LEO satellites, the quality of service is lacking, which may affect the quality of service of satellite services. A dynamic routing algorithm for maximizing satellite battery life in low-orbit satellite networks is proposed in some literature. A multi-objective optimization problem with the goal of maximizing satellite battery life, minimizing end-to-end delay and packet loss rate is established. Satellite network routing is regarded as a Markov decision process, deep learning is combined with reinforcement learning, and a routing strategy is learned by using the powerful perception ability of deep learning and the decision-making ability of reinforcement learning to balance the use of inter-satellite batteries and reduce the service life of satellite batteries. A literature proposes to apply Q learning to the power allocation problem in satellite-to-ground communications of low-orbit satellites. Using this method, the life of LEO satellite batteries can be extended by sharing the workload of overloaded satellites with neighboring satellites with lower loads. A literature proposes an intelligent energy management strategy to control the risk of battery exhaustion by designing an online mission acquisition, processing and transmission integrated scheduling (MAPTIS) scheme, thereby further optimizing the life of LEO satellites while improving the mission service quality requirements.

[0006] Some literature considers the balance between service quality and satellite life, but the strategies need to rely on the assistance of other satellites in the satellite network, which will also increase the pressure of business demand on other satellites. The strategy proposed in another literature is limited to the online scheduling strategy and lacks the vision of global optimization. Summary of the invention

[0007] In view of the above background technology, the present invention provides a method for optimizing the life of LEO satellites based on data prediction and matching. The method is oriented to a single LEO satellite. Based on historical data, the method predicts the solar radiation energy situation and the distribution conditions of business needs during the operation time of the LEO satellite, and then plans the energy management of the LEO satellite from a global perspective, so that the LEO satellite can optimize the DOD of the LEO satellite battery under the condition of meeting the business needs, thereby extending the life of the LEO satellite. At the same time, the business potential of the satellite is fully tapped, and the energy is planned from a global perspective. Without relying on the assistance of other satellites, the life of the LEO satellite can be optimized while ensuring business services.

[0008] The technical solution for achieving the purpose of the present invention is:

[0009] The LEO satellite life optimization method based on data prediction matching includes:

[0010] (1) Building a network model:

[0011] Using the radial distance of the LEO satellite relative to the center of the earth, the LEO satellite orbit model is established, as shown in the following formula:

[0012] r(t)=(R+h)[1-ecosα(t)]

[0013] Where R is the radius of the Earth, h is the altitude of the satellite, e is the eccentricity of the satellite orbit, α(t) is the eccentric anomaly angle;

[0014] The orbital period of LEO satellite is as follows:

[0015]

[0016] Where G is the gravitational constant and M is the mass of the Earth;

[0017] (2) Constructing LEO satellite energy model:

[0018] Solar radiation energy is converted into electrical energy by the solar panels carried by LEO satellites, where the power of solar radiation energy is P r (t), the converted electric energy power is P g (t); The converted electric energy is distributed through the energy dispatcher; P d-b (t) is the energy power allocated to the payload module by the energy scheduler for use by various ground services; P d-B (t) is the power allocated to the battery module by the energy scheduler. When it takes a positive value, the battery is charged, and when it takes a negative value, the battery is discharged. The energy released by the discharge is used by the payload module; P b (t) Power requirements for LEO satellite processing of terrestrial services;

[0019] The moment when the satellite enters the sunshine area is taken as the origin, that is, the starting point of the solar radiation intensity analysis; the calculation starts from the position t of the satellite on the orbit, t = 0 represents the position where the satellite just enters the sunshine area, and the solar radiation intensity P received by the LEO satellite is r (t) The energy model is established as follows:

[0020]

[0021] Among them, P max It is the radiation intensity of the direct sunlight point. Its physical meaning is the solar radiation energy received per second per unit area perpendicular to the sunlight outside the Earth's atmosphere. α is the proportion of the sunlight area in the orbital period, 0≤t≤αT s Indicates that the satellite is operating in the sunshine area, αT s <t≤T s Indicates that it is running in the earth's shadow area; σ represents the width of the controlled radiation intensity distribution, that is, the speed at which the radiation intensity decreases from the maximum value to the smaller value; N(t) is the noise effect of solar activity and cosmic rays on the solar radiation power;

[0022] Establish a discrete time system, divide an orbital period of the LEO satellite into K equal-length time slots, the length of the time slot is Δt; j represents any divided time slot, and the time slot sequence is

[0023] For the t j time slot, its solar radiation energy E r (t j ) is shown in the following formula:

[0024]

[0025] Then the solar radiation energy sequence received by the LEO satellite in each time slot is E r (t) = {E r (t1),E r (t2),···,E r (t j ),···,E r (t K )};

[0026] For the t j time slot, the energy converted by the solar panel is:

[0027] E g (t j )=Aη r-g E r (t j )

[0028] Where A is the area of ​​the solar panel, ηr-g is the energy conversion efficiency of the satellite solar panel; then the energy sequence converted by the solar panel in each time slot is E g (t) = {E g (t1),E g (t2),···,E g (t j ),···,E g (t K )};

[0029] The energy sequence allocated by the energy scheduler to the payload module in each time slot is E d-b (t) = {E d-b (t1),E d-b (t2)···E d-b (t j )···E d-b (t K )};

[0030] The energy sequence allocated to the battery charging is E d-B (t) = {E d-B (t1),E d-B (t2)···E d-B (t j )···E d-B (t K )};

[0031] The energy sequence allocated to the battery discharge is E B-d (t) = {E B-d (t1),E B-d (t2)···E B-d (t j )···E B-d (t K )}

[0032] The battery module can store the electric energy converted by the solar panel and release the energy for ground services when needed. The initial energy sequence of the battery module in each time slot is E B (t) = {E B (t1),E B (t2)···E B (t j )···E B (t K )};

[0033] The energy sequence required by terrestrial services in each time slot is E b (t) = {E b (t1),E b (t2)···E b (tj )···E b (t K )};

[0034] (3) Building a battery model:

[0035] The battery is used as the energy storage system, the energy storage model is the battery system model, and the main characterization parameter of the battery energy storage system is the energy state E of any time slot. B (t j ), and the charging threshold and discharge threshold They represent the maximum and minimum energy storage of the battery respectively; when the charge threshold is exceeded, the battery will enter the overcharge state, at which time the battery will no longer be charged; conversely, when the battery energy is lower than the discharge threshold, the battery will enter the overdischarge state, at which time the battery will no longer be discharged; the battery energy sequence satisfies the following relationship:

[0036]

[0037] Among them, η c-f is the battery charging efficiency, η f-c is the discharge efficiency of the battery;

[0038] Set the battery discharge depth DOD (t j ) is shown in the following formula:

[0039]

[0040] The battery life of the LEO satellite is defined as L, and the depth of discharge DOD is logarithmically related to the satellite life L, as shown in the following formula:

[0041] log 10 L+A·DOD max =B

[0042] Where A and B are constants that depend on the battery specifications, DOD max Indicates a complete charge and discharge, that is, the power is discharged from 100% to 0%. L is the benchmark life of the battery, which indicates the life consumed after a complete charge and discharge process;

[0043] is the average service life of LEO satellites, and the reduced average service life is ΔL, as follows:

[0044]

[0045] (4) Build a business model:

[0046] According to the different levels of activity of ground users, a day is divided into M time periods, and its time period sequence is expressed as in m It represents any divided time period;

[0047] Assuming that the business volume in each time period follows a Poisson distribution, then the τth m The business volume in this time period is

[0048]

[0049] Where λ(τ m ) is the τth m The Poisson distribution parameter of a time period, k represents the number of requests for the service in that time period;

[0050] In one orbital cycle, the sub-satellite trajectory is represented as a set of regional sequences according to the traffic volume. Where o represents any one of the divided regions; then the τth m The nth time period i The business models for each region are as follows:

[0051]

[0052] in For nth o Economic factors of different regions are used to measure the impact of the total economic volume of different regions on business volume;

[0053] In the orbital period t j If it is in the τth time slot of the LEO satellite operation, m The nth time period u area, then the energy required for the service in this time slot is

[0054] E b (t j ) = η b-e ·S(τ m ,n o )

[0055] where η b-e It is the conversion coefficient from business volume to energy required for business. Its physical meaning is the proportion of energy required to process one unit of business volume.

[0056] (5) By controlling the decision-making and planning of the battery charge and discharge energy in each time slot, the lithium-ion battery in the LEO satellite operation can be kept in the DOD within the optimal use range; the battery energy state is updated as shown in the following formula:

[0057] E B (t j+1 )=E B (t j )+η c-f E d-B (t j)-η f-c E B-d (t j )

[0058] The battery status is limited by the battery capacity as follows:

[0059]

[0060] in The lower limit of battery capacity is the minimum power contained in the battery hardware design. If the power is lower than this, the battery will stop supplying power. is the maximum capacity of the battery;

[0061] In any time slot, the sum of the battery's charging energy and the battery's current energy cannot exceed the battery capacity limit, as shown in the following formula:

[0062]

[0063] In any time slot, the battery discharge energy cannot exceed the difference between the battery energy in the current time slot and the battery lower limit, as follows:

[0064]

[0065] Among them, η f-c The charging efficiency of the battery;

[0066] The energy in the system in any time slot satisfies the following relationship:

[0067] E g (t j )=E d-b (t j )+E d-B (t j )

[0068]

[0069]

[0070] in, Indicates in the future The energy demand of the business in a time slot is very large, and when the battery is charged to the maximum power it can reach in advance, it still cannot meet the energy demand of the business. If this difference exists for a period of time in the future, the business demand of certain time slots will be traded off from a global perspective to achieve the best global goal;

[0071] (6) Solar radiation energy prediction

[0072] A hybrid model (CNN-LSTM) of convolutional neural network (CNN) and long short-term memory network (LSTM) is used to predict solar radiation energy.

[0073] First, collect historical solar radiation data, including sunshine duration, radiation intensity, etc., and then remove missing values ​​and outliers to ensure data quality;

[0074] Secondly, the historical data of solar radiation energy is preprocessed, assuming x g (t) is the historical data set of solar radiation energy. The historical data set is discretized with a unit time length of Δt. The sequence of the discretized historical data set is The total number of time slots of historical data is J g ;

[0075] Then divide the historical data set into training sets according to a certain ratio and test set The training set serves as the input for model training, and the test set will be used to verify the accuracy of model predictions.

[0076] Then the training set is normalized as shown in the formula:

[0077]

[0078] where max(x g ) and min(x g ) are the maximum and minimum values ​​in the historical data respectively. The normalized data is imported into the model for training. After the model training is completed, it is verified with the test set. If the model training error is stable within the error limit L, the model parameters are saved;

[0079] The trained model is used to predict the solar radiation energy of the LEO satellite, and the data predicted by the model is denormalized to obtain the prediction result, as shown in the following formula:

[0080]

[0081] The energy sequence of the predicted solar radiation from LEO satellite is: In addition, after conversion according to formula (2-8), the solar energy sequence obtained by the LEO satellite is

[0082] In order to effectively evaluate the accuracy of the prediction model, the coefficient of determination (R 2 ) to evaluate the prediction accuracy, R 2 As shown below:

[0083]

[0084] in is the average value of solar energy;

[0085] (7) Business demand energy prediction: The business demand energy is predicted using the LSTM-RF algorithm. The historical data set of business demand energy is x b (t u )x b (t) = [x b (t1),···,x b (t u ),···,x b (t J )], and then preprocess and divide the training set according to a certain ratio and test set The data is normalized and then used as the input for model training and verification. If the model training is successful, the model is saved and the service demand energy of the LEO satellite is predicted. The data predicted by the model is then denormalized to obtain the predicted data.

[0086] If the model prediction error is stable within the error limit If the forecast result is within , the forecasted business demand energy sequence is as follows

[0087] (8) Life optimization decision: According to the Lyapunov optimization framework, the Lyapunov function is defined as follows:

[0088]

[0089] Where Φ(t)=E B (t); In addition, the Lyapunov drift plus penalty function in any time slot can be written as:

[0090]

[0091] In time slot t j The optimization problem P1 can be reformulated as P2:

[0092]

[0093] st

[0094] Where V is the system control parameter; the complex stochastic optimization problem P1 is transformed into a common optimization problem P2. P2 is constructed by adding the Lyapunov drift term on the basis of P1. The relationship between P2 and P1 can be expressed as follows:

[0095] P2=[VP1+ΔΦ(t j )|Φ(t j )]

[0096] When ΔΦ(t j ) is small enough, the result of P2 is the same as the result of solving P1; in this case, the objective function P2 is simplified to P1. Thus, P2 is reduced to P1; therefore, under the stability condition, P2 is mathematically equivalent to P1;

[0097] Due to the deviation between the predicted results and the actual results, and the dynamic nature of the battery state still exists, the improved AMPC algorithm is used to solve the problem. The specific process is as follows:

[0098] System modeling: Build a dynamic model of the system; this model is used to describe the relationship between the system state and the control input. In AMPC, the system model can be updated over time;

[0099] Initialization: Set the initial battery charge E B (t0), and read the service demand energy of each time slot Energy obtained by converting solar radiation energy

[0100] Model update: Introduce the function of adaptive improvement, automatically adjust the subsequent prediction results according to the error between the actual results and the predicted results, first define the solar radiation energy prediction error and business energy prediction error The two errors in any time slot satisfy the following relationship:

[0101]

[0102]

[0103] Corrections are divided into short-term corrections and long-term corrections. Short-term corrections are local optimizations within a smaller time window to ensure that the system behavior in the current time slot and the last few time slots meets expectations, as follows:

[0104]

[0105] Where u is the time slot scale of the short-term correction, λ corr is the short-term error weight, which is used to indicate the weight of error correction. The short-term correction of business demand energy is similar.

[0106] Long-term correction is to perform global optimization over a longer time window, accumulate statistics of errors over a period of time, and adjust subsequent results based on the accumulated errors to ensure the stability and performance of the system over a longer period of time;

[0107]

[0108] Where U is the time slot scale of the long-term correction, λ lcorr The same is true for the long-term error weights and the long-term correction of business demand energy.

[0109] Rolling optimization: Rolling optimization is an online optimization method that solves the optimization problem in each time slot, applies the decision of the first time slot, and then rolls to the next time slot and repeats the process. P2 is continuously solved. Finally, the optimal charging and discharging decision for each time slot is obtained.

[0110] Through the above technical solution, the present invention has the following advantages:

[0111] 1. Data prediction: The Convolutional Neural Networks-Long Short-Term Memory (CNN-LSTM) algorithm is used to predict the solar radiation energy received during the LEO satellite orbit period. In addition, the Long Short-Term Memory-Random Forest (LSTM-RF) algorithm is used to predict the business demand energy.

[0112] 2. Energy planning: The matching energy planning problem based on Lyapunov theory is solved using the adaptive model predictive control (AMPC) algorithm. This method effectively matches the energy supply of LEO satellites with business needs, optimizing the life of LEO satellites while ensuring business services.

[0113] 3. Lifetime optimization: This paper proposes a new evaluation index, the Business Handling and Lifetime Optimization Index (BHLOI), which is used to evaluate the comprehensive performance of different energy management algorithms in meeting business needs and optimizing battery life. This index combines the cumulative business processing energy consumption of each algorithm and the optimized battery life to provide a standard for measuring the business processing capability of the algorithm and extending the battery life, which has important guiding significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0114] Figure 1 A network model diagram of an embodiment of the present invention;

[0115] Figure 2 A diagram of a LEO satellite energy system according to an embodiment of the present invention;

[0116] Figure 3 A schematic diagram of an energy storage model according to an embodiment of the present invention;

[0117] Figure 4 A schematic diagram of a business model of an embodiment of the present invention;

[0118] Figure 5 This is a diagram showing the prediction effect of different algorithms for LEO satellite solar energy according to an embodiment of the present invention;

[0119] Figure 6 This is a comparison chart of energy prediction effects of different algorithms for business demand according to an embodiment of the present invention;

[0120] Figure 7 A comparison diagram of LEO satellite life span changes under different algorithms in an embodiment of the present invention;

[0121] Figure 8 This is a comparison diagram of the energy of cumulatively processing services under different algorithms in the embodiment of the present invention;

[0122] Fig. 9 This is a comparison chart of indicators under different algorithms in the embodiment of the present invention;

[0123] Fig.10 This is a diagram showing the variation of the battery state with the Lyapunov control parameter V according to an embodiment of the present invention;

[0124] Fig.11 This is a comparison chart of average optimized lifespans of different battery capacities according to an embodiment of the present invention;

[0125] Fig.12 A specific life span variation diagram of a LEO satellite with different battery capacities according to an embodiment of the present invention;

[0126] Fig.13 This is a comparison chart of the energy consumed by LEO satellites with different battery capacities to process business needs in an embodiment of the present invention. DETAILED DESCRIPTION

[0127] The present invention is further described below in conjunction with the accompanying drawings and embodiments, but the contents of the present invention are not limited thereto.

[0128] Example:

[0129] The LEO satellite life optimization method based on data prediction matching includes:

[0130] 1. Building a system model

[0131] 1. Build a network model

[0132] System model such as Figure 1As shown in the figure, the LEO satellite alternates between the sunlit area and the shadow area in its orbit. The sunlit area refers to the area where the satellite can receive solar radiation energy, while the shadow area is the area where the LEO satellite cannot receive solar radiation energy. In a typical LEO satellite orbit, the time exposed to solar radiation accounts for about 66%-78% of each orbit.

[0133] In the sunlight area, LEO satellites collect energy and directly supply it to the subsystems, while storing excess energy in the energy storage system. In the shadow area, since they cannot receive solar radiation energy, the energy source of LEO satellites can only rely on battery storage. The batteries of LEO satellites are mainly lithium batteries.

[0134] During the operation of LEO satellites in the Earth's shadow orbit, the battery undergoes repeated charge and discharge cycles to provide sufficient energy. Different depths of discharge (DOD) will affect the life of lithium batteries. As an important part of the satellite energy storage system, the life and health of the battery determine the overall life of the satellite. Even in the sunshine area, although LEO satellites have sufficient power, unrestricted use will cause the battery life to age faster because it will increase the battery's DOD.

[0135] The energy planning strategy proposed in the present invention aims to optimize the charging and discharging strategy of the LEO satellite during operation, so that the DOD of the battery can be reduced to the greatest extent in both the sunlit area and the shadow area, thereby extending the battery life.

[0136] The present invention uses the radial distance of the LEO satellite relative to the center of the earth to establish the LEO satellite orbit model, such as:

[0137] r(t)=(R+h)[1-ecosα(t)]

[0138] Where R is the radius of the Earth, h is the altitude of the satellite, e is the eccentricity of the satellite orbit, and α(t) is the eccentric anomaly angle.

[0139] The specific orbital period of LEO satellite is as follows:

[0140]

[0141] Where G is the gravitational constant and M is the mass of the Earth.

[0142] 2. Constructing LEO satellite energy model

[0143] like Figure 2 As shown in Figure 2, solar radiation energy is converted into electrical energy by the solar panels carried by LEO satellites. The power of solar radiation energy is P r (t), the converted electric energy power is P g(t). The converted electric energy is distributed by the energy dispatcher. d-b (t) is the energy power allocated to the payload module by the energy scheduler for use by various ground services. d-B (t) is the power allocated to the battery module by the energy scheduler. When it takes a positive value, it means the battery is charged, and when it takes a negative value, it means the battery is discharged. The energy released by the discharge is used by the payload module. b (t) is the power required by the LEO satellite to process ground services.

[0144] Gaussian distribution is usually used to fit the daily variation of solar radiation and is applicable to most areas. First, we start from the moment when the satellite enters the sunshine area as the origin, which is the starting point for the analysis of solar radiation intensity. The position of the satellite on the orbit t is calculated from this point, and t = 0 represents the position where the satellite just enters the sunshine area. The present invention is aimed at the solar radiation intensity P received by the LEO satellite r (t) The energy model is established as follows:

[0145]

[0146] Among them, P max It is the radiation intensity at the point where the sun shines directly, referring to the solar radiation energy constant. Its physical meaning is the solar radiation energy received per second per unit area outside the Earth's atmosphere, perpendicular to the sun's rays. α is the proportion of the sunlit area in the orbital period, 0≤t≤αT s Indicates that the satellite is operating in the sunshine area, αT s <t≤T s Indicates that it is operating in the Earth's shadow. σ represents the width of the radiation intensity distribution, that is, the speed at which the radiation intensity decreases from the maximum value to the smaller value. N(t) is the noise effect of solar activity and cosmic rays on the solar radiation power.

[0147] The present invention divides an orbital period of a LEO satellite into K equal-length time slots, where the length of the time slot is Δt. j represents any divided time slot, and the time slot sequence is

[0148] For the t j time slot, its solar radiation energy E r (t j ) is shown in the following formula:

[0149]

[0150] Then the solar radiation energy sequence received by the LEO satellite in each time slot is E r (t) = {E r (t1),E r (t2),···,E r (tj ),···,E r (t K )}.

[0151] For the t j time slot, the energy converted by the solar panel is:

[0152] E g (t j )=Aη r-g E r (t j )

[0153] Where A is the area of ​​the solar panel, η r-g The energy conversion efficiency of satellite solar panels.

[0154] Then the energy sequence converted by the solar panel in each time slot is E g (t) = {E g (t1),E g (t2),···,E g (t j ),···,E g (t K )}.

[0155] The energy sequence allocated by the energy scheduler to the payload module in each time slot is E d-b (t) = {E d-b (t1),E d-b (t2)···E d-b (t j )···E d-b (t K )},

[0156] The energy sequence allocated to the battery charging is E d-B (t) = {E d-B (t1),E d-B (t2)···E d-B (t j )···E d-B (t K )}.

[0157] The energy sequence allocated to the battery discharge is E B-d (t) = {E B-d (t1),E B-d (t2)···E B-d (t j )···E B-d (t K )}

[0158] The battery module can store the electric energy converted by the solar panel and release the energy for ground services when needed. The initial energy sequence of the battery module in each time slot is E B (t) = {E B (t1),E B (t2)···E B (t j )···E B (t K )}.

[0159] The energy sequence required by terrestrial services in each time slot is E b (t) = {E b (t1),E b (t2)···E b (t j )···E b (t K )}. 3. Build a battery model (battery life curve)

[0160] The battery of the present invention is used as an energy storage system. The energy storage model is a battery system model, and the schematic diagram is as follows Figure 3 As shown in the figure, the red arrow indicates the direction of the flow of electric energy. In any time slot, the charge and discharge of the battery is determined by the energy scheduler. The color of the battery status in the figure expresses that the battery has different degrees of influence on its life under different conditions. The right side is the relationship between DOD and battery life.

[0161] The main characterization parameter of the battery energy storage system is the energy state E in any time slot. B (t j ), and the charging threshold and discharge threshold Respectively represent the maximum energy storage and minimum energy storage of the battery. When the charging threshold is exceeded, the battery will enter the overcharge state, and the battery will no longer be charged. Conversely, when the battery energy is lower than the discharge threshold, the battery will enter the overdischarge state, and the battery will no longer be discharged. The battery energy sequence of the present invention satisfies the following relationship:

[0162]

[0163] Among them, η c-f is the battery charging efficiency, η f-c is the discharge efficiency of the battery.

[0164] Set the battery discharge depth DOD (t j ) is shown in the following formula:

[0165]

[0166] In addition, the present invention defines the battery life of the LEO satellite as L. The depth of discharge DOD is logarithmically related to the satellite life L, as shown in the following formula:

[0167] log 10 L+A·DOD max =B

[0168] Where A and B are constants that depend on the battery specifications, DOD max It indicates a complete charge and discharge, that is, the power is discharged from 100% to 0%. L is the benchmark life of the battery, which indicates the life consumed after a complete charge and discharge process.

[0169] is the average service life of the LEO satellite of the present invention, then the average service life reduced by the present invention is ΔL, specifically as follows:

[0170]

[0171] 4. Build a business model

[0172] Satellite Tool Kit (STK) is a space simulation and analysis tool developed by AGI (Analytical Graphics), which is mainly used in the fields of satellite orbit analysis, space visualization and space target defense. The business model of the present invention will conduct simulation experiments based on STK software to obtain the specific sunshine area time and earth shadow area time in the operation of LEO satellites, as well as the coverage time of each service area in the LEO satellite orbit.

[0173] The business volume of LEO satellites is mainly determined by two aspects: one is the difference in time dimension, and the other is the difference between regions. Through the simulation of STK software, the business needs of LEO satellites at different times and in different regions can be accurately simulated and analyzed, thereby providing data support for energy planning and optimization strategies.

[0174] For the service time model, the present invention divides a day into M time periods according to the different activity levels of ground users, and its time period sequence is expressed as in m Indicates an arbitrary time period.

[0175] Assuming that the business volume in each time period follows a Poisson distribution, then the τth m The business volume in this time period is

[0176]

[0177] Where λ(τ m ) is the τth m The Poisson distribution parameter of a time period, k represents the number of service requests in this time period.

[0178] In one orbital cycle, the sub-satellite trajectory is represented as a set of regional sequences according to the traffic volume. Where o represents any one of the divided regions.

[0179] Then the τth m The nth time period i The business models for each region are as follows:

[0180]

[0181] in For nth o The economic factors of different regions are used to measure the impact of the total economic volume of different regions on business volume.

[0182] In the orbital period t j If it is in the τth time slot of the LEO satellite operation, m The nth time period u area, then the energy required for the service in this time slot is

[0183] E b (t j ) = η b-e ·S(τ m ,n o ) (2-8)

[0184] where η b-e It is the conversion coefficient from business volume to energy required for business. Its physical meaning is the proportion of energy required to process one unit of business volume.

[0185] Business model diagram Figure 4 As shown, Figure 4 It shows the energy demand in different time periods and economic totals. The horizontal axis represents the 24-hour time period of a day, the vertical axis represents the level of economic totals in different regions, and the vertical axis is energy demand (kWh). As can be seen from the figure, energy demand presents a clear bimodal structure, reaching peaks at specific times during the day and night. This reflects that people have a higher demand for energy during active time periods during the day and night, which usually coincides with the peak periods of work and life activities. At the same time, the total economic volume also has a significant impact on energy demand. The higher the economic volume of a region, the higher its energy demand. This is because an increase in economic activity is usually accompanied by more energy consumption to support commercial activities, manufacturing, services, etc. Overall, energy demand shows regular changes under the dual effects of time and economic volume, with daytime and night being the main peak periods, and energy demand in economically developed regions being more significant.

[0186] 2. Problem Modeling

[0187] The present invention optimizes the energy management system of LEO satellites. It will predict the solar radiation energy and business demand energy during the operation of LEO satellites based on data, and carry out energy planning in advance. Under the premise of ensuring that business needs are met to the greatest extent, the average DOD of the battery between LEO satellite operations is minimized to extend the life of the LEO satellite.

[0188] The optimal use range of lithium-ion batteries is 20%-80% DOD. Following this range can significantly extend the life of the battery.

[0189] The present invention ensures that the lithium-ion batteries in the LEO satellite operation can be within the DOD of the best use range by controlling the decision-making and planning of the charging and discharging energy of the batteries in each time slot.

[0190] The battery energy status is updated as follows:

[0191] E B (t j+1 )=E B (t j )+η c-f E d-B (t j )-η f-c E B-d (t j )

[0192] The battery status is limited by the battery capacity as follows:

[0193]

[0194] in The lower limit of battery capacity is the minimum power contained in the battery hardware design. If the power exceeds this limit, the battery will stop supplying power. The maximum capacity of the battery.

[0195] In any time slot, the sum of the battery's charging energy and the battery's current energy cannot exceed the battery capacity limit, as shown in the following formula:

[0196]

[0197] The charging efficiency and discharging efficiency of the LEO satellite battery of the present invention are different, and the discharging efficiency of the battery is slightly greater than the charging efficiency. In any time slot, the discharge energy of the battery cannot exceed the difference between the energy of the battery in the current time slot and the lower limit of the battery, as follows:

[0198]

[0199] Among them, η f-c For the battery charging efficiency.

[0200] In the present invention, the energy in the system in any time slot satisfies the following relationship:

[0201] E g (t j )=E d-b (t j )+E d-B (t j )

[0202] In the present invention, the energy allocated to the business demand is limited so that it does not exceed the predicted business demand to avoid energy waste, as follows:

[0203]

[0204] In addition, in order to give full play to the business potential of LEO satellites, the present invention will give priority to meeting business needs when the energy demand for certain future businesses is very large according to the prediction results, and charge the battery to a suitable power level in advance to ensure that the business potential of LEO satellites is maximized, as follows:

[0205]

[0206] in, Indicates in the future The business demand energy in a time slot is very large, and when the battery is charged to the maximum power it can reach in advance, it still cannot meet the difference in business demand energy. If this amount exists for a period of time in the future, then the business needs of certain time slots will be chosen from a global perspective to achieve the best global goal.

[0207] The present invention models the problem P1 as follows:

[0208]

[0209] st

[0210] Then, the present invention will calculate the DOD according to the calculated energy state of the battery in each time slot, further calculate the average life after the algorithm is optimized, and accumulate the energy that the algorithm can correctly process the business to obtain the indicator BHLOI proposed by the present invention, which is as follows:

[0211]

[0212] By calculating the energy actually used to process the task, BHLOI can reflect the ability of different algorithms to meet the task requirements. The higher this part of energy, the greater the amount of tasks that the algorithm can handle. In addition, by introducing the optimized average battery life, BHLOI can also reflect the effectiveness of the algorithm in extending the battery life. The longer this part of life, the better the algorithm performs in terms of energy saving and extending the battery life.

[0213] Convexity analysis of optimization problems: Since problem P1 has an absolute value, in order to facilitate the solution and application of KKT conditions, we introduce an auxiliary variable δ(t j )≥0 such that:

[0214]

[0215] Then P1 can be resolved as:

[0216]

[0217] Since P1 is δ(t j ) is a linear combination s., t. If the linear function is convex, then P1 is about δ(t j ) is a convex function, and δ(t j ) and E B (t j ) is a linear inequality, and the objective function of the optimization problem P1 is convex.

[0218] The battery state equation is linear and is a convex set in the feasible domain. The battery capacity limit satisfies the linear inequality constraint and is a convex set in the feasible domain. The charging limit, discharging limit and business limit are similar, all of which are linear inequality constraints and are convex sets in the feasible domain.

[0219] In general, the constraints of problem P1 are all linear, and the feasible domain defined is a convex set. Problem P1 is a convex optimization problem.

[0220] Introducing Lagrange multipliers:

[0221] λ1(t j ):correspond

[0222] λ2(t j ):correspond

[0223] λ3(t j ):correspond

[0224] λ4(t j ):correspond

[0225] λ5(tj ):correspond

[0226] μ1(t j ):Corresponding to -E d-B (t j )

[0227] μ2(t j ):Corresponding to -E B-d (t j )

[0228] γ1(t j ):correspond

[0229] γ2(t j ):correspond

[0230] ν1(t j ): corresponds to -δ(t j )

[0231] The Lagrangian function is constructed as follows:

[0232]

[0233] Take the first-order derivative of the Lagrangian function and solve the (Karush-Kuhn-Tucker, KKT) condition.

[0234] For E B (t j ) The partial derivative is as follows:

[0235]

[0236] For E d-B (t j )η f-c The partial derivative is as follows:

[0237]

[0238] For E B-d (t j )The partial derivative is as follows:

[0239]

[0240] For δ(t j )The partial derivative is as follows:

[0241]

[0242] By solving the above optimization variables to obtain the possible optimal solution, the complementary slackness condition is verified:

[0243] st

[0244]

[0245]

[0246]

[0247]

[0248]

[0249] μ1(t)E d-B (t j )=0

[0250] μ2(t j )-E B-d (t j )=0

[0251]

[0252]

[0253] v1(t j )δ(t j )=0

[0254] Confirm feasibility conditions (Primal and Dual Feasibility)

[0255] Primal Feasibility:

[0256] All original constraints must be satisfied, including the introduced auxiliary variable constraints.

[0257] Dual Feasibility Condition:

[0258] All Lagrange multipliers must be non-negative, as follows:

[0259] λ i (i=1,2,3,4,5),μ1(t j )≥0,μ2(t j )≥0,γ1(t j ),γ2(t j ),ν1(t j ).

[0260] 3. Green Energy Management Strategy

[0261] The green energy management strategy proposed in the present invention consists of three parts, namely solar radiation energy prediction, business demand energy prediction and life optimization decision.

[0262] The present invention will respectively predict the solar radiation energy and business demand energy in the LEO satellite orbit, and then make a life optimization decision based on these two energies. In addition, the life optimization problem is modeled using Lyapunov theory, and then the AMPC algorithm is used to optimize and solve it, and finally an optimization decision is made to increase the operating life of the LEO satellite.

[0263] A. Solar radiation energy prediction

[0264] In the energy management of low earth orbit (LEO) satellites, accurate prediction of solar radiation energy is crucial. The present invention uses a hybrid model (CNN-LSTM) of convolutional neural network (CNN) and long short-term memory network (LSTM) to predict solar radiation energy. The CNN-LSTM model has been shown to effectively capture the fluctuation pattern of power generation and improve prediction accuracy. This model achieves higher prediction accuracy than using CNN or LSTM alone by combining the feature extraction ability of CNN and the time-dependent capture ability of LSTM.

[0265] First, historical solar radiation data are collected, including sunshine duration, radiation intensity, etc., and then missing values ​​and outliers are removed to ensure data quality.

[0266] Secondly, the historical data of solar radiation energy is preprocessed. g (t) is the historical data set of solar radiation energy. The historical data set is discretized with a unit time length of Δt. The sequence of the discretized historical data set is x g (t) = [x g (t1),···,x g (t j ),···,x g (t Jg )], where the total number of time slots of historical data is J g .

[0267] Then divide the historical data set into training sets according to a certain ratio and test set The training set serves as the input for model training, and the test set will be used to verify the accuracy of model predictions.

[0268] Then the training set is normalized as shown in the formula:

[0269]

[0270] where max(x g ) and min(x g ) are the maximum and minimum values ​​in the historical data respectively.

[0271] Import the normalized data into the model for training.

[0272] After the model training is completed, the test set is used for verification. If the model training error is stable within the error limit L, the model parameters are saved.

[0273] The trained model is used to predict the solar radiation energy of the LEO satellite, and the data predicted by the model is denormalized to obtain the prediction result, as shown in the following formula:

[0274]

[0275] The energy sequence of the predicted solar radiation from LEO satellite is: In addition, after conversion according to formula (2-8), the solar energy sequence obtained by the LEO satellite is

[0276] In order to effectively evaluate the accuracy of the prediction model, the coefficient of determination (R 2 ) to evaluate the prediction accuracy. 2 As shown below:

[0277]

[0278] in is the average value of solar energy.

[0279] B. Business Demand Energy Forecast

[0280] The present invention uses the LSTM-RF algorithm to predict the business demand energy. LSTM is good at capturing long-term dependencies in time series and can effectively handle the time correlation in business demand. RF is good at processing high-dimensional and nonlinear data and can capture complex feature relationships. By combining the advantages of LSTM and RF, the accuracy of business demand prediction can be improved, which is suitable for complex time series and multi-dimensional feature data.

[0281] The historical data set of business demand energy is x b (t u )x b (t) = [x b (t1),···,x b (t u ),···,x b (t J)] Then preprocess and divide the training set according to a certain ratio and test set The data is normalized and then used as the input for model training and verification. If the model training is successful, the model is saved and the business demand energy of the LEO satellite is predicted. The data predicted by the model is then denormalized to obtain the predicted data.

[0282] If the model prediction error is stable within the error limit If the forecast result is within , the forecasted business demand energy sequence is as follows

[0283] C. Solving optimization problems

[0284] In the present invention, although CNN-LSTM and LSTM-RF are used to predict solar energy and business demand, the prediction accuracy of these models is limited, and the prediction results still contain certain errors and uncertainties. In addition, the charging and discharging behavior of satellite batteries and the dynamic characteristics of the entire system make the problem more complicated, random and dynamic. Therefore, problem P1 is a stochastic optimization problem. The present invention will use Lyapunov theory to transform the stochastic optimization problem into a deterministic optimization problem for each time slot. Accordingly, the deterministic optimization problem is decomposed into several sub-problems, which are solved in an iterative manner.

[0285] According to the Lyapunov optimization framework, the present invention defines the Lyapunov function as follows:

[0286]

[0287] Where Φ(t)=E B (t). In addition, the Lyapunov drift plus penalty function in any time slot can be written as:

[0288]

[0289] The present invention proposes to use ΔΦ(t j ) as a new minimization objective, rather than solving a complex stochastic optimization problem P1 by optimizing ΔΦ(t j ) upper bound, ensuring that the deviation between the battery energy in each time slot and the planned target battery energy is minimized. Therefore, in time slot t j The optimization problem P1 can be reformulated as P2:

[0290]

[0291] st

[0292] Where V is the system control parameter. The present invention transforms the complex stochastic optimization problem P1 into a common optimization problem P2. P2 is constructed by adding a Lyapunov drift term on the basis of P1. The relationship between P2 and P1 can be expressed as follows:

[0293] P2=[VP1+ΔΦ(t j )|Φ(t j )]

[0294] When ΔΦ(t j ) is small enough, the result of P2 is the same as the result of solving P1. In this case, the objective function P2 is simplified to P1. Thus, P2 is reduced to P1. Therefore, under the stability condition, P2 is mathematically equivalent to P1.

[0295] Since the deviation between the predicted result and the actual result and the dynamic nature of the battery state still exist, the present invention uses an improved AMPC algorithm to solve the problem. The algorithm is a method that combines adaptive control and model predictive control (MPC). It is suitable for control problems in dynamic and uncertain environments. The main idea of ​​AMPC is to better reflect the dynamic behavior of the system by updating the system model in real time, and to predict and optimize based on the updated model, so as to achieve more accurate control. The specific process is as follows:

[0296] System modeling: Establish a dynamic model of the system. This model is used to describe the relationship between the system state and the control input. In AMPC, the system model can be updated over time. In the present invention, the system state is the state of the battery, as shown in Formula X.

[0297] Initialization: Set the initial battery charge E B (t0), and read the service demand energy of each time slot Energy obtained by converting solar radiation energy

[0298] Model update: Different from the traditional MPC algorithm, this invention introduces the function of adaptive improvement. According to the error between the actual result and the predicted result, the subsequent prediction results will be automatically adjusted. First, the solar radiation energy prediction error is defined and business energy prediction error The two errors in any time slot satisfy the following relationship:

[0299]

[0300] Corrections are divided into short-term corrections and long-term corrections. Short-term corrections are local optimizations within a smaller time window to ensure that the system behavior in the current time slot and the last few time slots meets expectations. The details are as follows:

[0301]

[0302] Where u is the time slot scale of the short-term correction, λ corr is the short-term error weight, which is used to indicate the weight of error correction. The short-term correction of business demand energy is similar.

[0303] Long-term correction is a global optimization within a longer time window. It will statistically accumulate errors over a period of time and adjust subsequent results based on the accumulated errors to ensure the stability and performance of the system over a longer period of time.

[0304]

[0305] Where U is the time slot scale of the long-term correction, λ lcorr is the long-term error weight. The same is true for the long-term correction of business demand energy.

[0306] Rolling optimization: Rolling optimization is an online optimization method that solves the optimization problem in each time slot, applies the decision of the first time slot, and then rolls to the next time slot and repeats this process. P2 is continuously solved. Finally, the optimal charging and discharging decision for each time slot is obtained.

[0307] A. Network parameter settings

[0308] The simulation scenario is set to a single LEO satellite operating in its orbit, where the LEO satellite orbit height h is 550 km and the orbit period T s The satellite operation time is set to 1 day, a total of 24 hours, including 2880 time slots. The total number of ground time periods M divided in the business model is 24, the total number of regions N is 14, and the specific coverage time of LEO satellites in each region is obtained through simulation experiments designed by STK software. The battery capacity is obtained by considering the satellite's energy acquisition and business needs as well as the satellite's volume and weight. The specific simulation parameters are shown in Table 1:

[0309] Table 1 Related parameter settings

[0310]

[0311] Among them, the STK simulation part, the present invention adopts, after setting the relevant parameters of the LEO satellite, generating a simulated LEO satellite orbit, setting the number of ground service stations, and counting the relevant time of the LEO satellite serving the ground station.

[0312] In the research object of the present invention, 14 ground stations are set up, which are distributed in different regions. The specific coverage time of LEO satellites to these regions is obtained by STK software simulation.

[0313] B. Solar radiation energy prediction

[0314] In order to verify the performance advantages predicted by the CNN-LSTM algorithm and LSTM-RF algorithm of the present invention, the present invention considers the following benchmark algorithms.

[0315] Recurrent Neural Network (RNN): This algorithm introduces recurrent connections in the network to remember the information of previous time steps, thereby capturing the dynamic characteristics in the time series, building an RNN model, and predicting the solar energy and business demand energy of LEO satellites.

[0316] Long Short-Term Memory (LSTM): This algorithm captures long-term dependencies by introducing memory units and gating mechanisms, constructs an LSTM model, and predicts the solar energy and business demand energy of LEO satellites.

[0317] The prediction results of different algorithms for LEO satellite solar energy are shown in the figure below. Figure 5 As shown:

[0318] In the figure: the horizontal axis is the satellite's operating time in hours, and the vertical axis is energy in kilowatt-hours (kw / h). From the figure, it can be found that the CNN-LSTM prediction value has the highest accuracy for global solar radiation energy, while the RNN and LSTM prediction values ​​have lower accuracy. The specific errors are shown in Table 2:

[0319] Table 2 Solar radiation energy prediction accuracy

[0320]

[0321] C. Business demand energy forecast

[0322] The comparison chart of different algorithms for business demand energy prediction is as follows Figure 6 As shown in the figure: The horizontal axis is the satellite's operating time in hours, and the vertical axis is energy in kilowatt-hours (kw / h). The present invention selects RNN and LSTM as benchmark algorithms to compare the prediction accuracy of different algorithms for business demand energy. It can be found from the figure that the energy predicted by LSTM-RF is basically consistent with the actual energy in the whole world, and the prediction accuracy is significantly higher than that of RNN and LSTM algorithms. The specific errors are shown in Table 3:

[0323] Table 3 Business demand energy prediction accuracy

[0324]

[0325] D. Life Optimization Decision

[0326] In order to compare the performance of the algorithm proposed in the present invention, the present invention selects the following algorithms as comparison benchmark algorithms:

[0327] The changes in LEO satellite lifespan under different algorithms are as follows: Figure 7 As shown: The battery life curve in the figure shows the changes in battery life of four algorithms (AMPC, Simple MPC, Game Theory and Greedy) over the entire time period. The battery life of the AMPC algorithm fluctuates greatly, but generally remains at a high level, indicating that it comprehensively considers short-term and long-term energy requirements and charging and discharging strategies, and the optimization effect is good. The battery life performance of the Simple MPC and Game Theory algorithms is relatively stable, but slightly lower than that of AMPC, showing deficiencies in energy management. Although the Game Theory algorithm is more efficient at certain moments, it has large overall fluctuations and a lower lifespan, indicating that it is too aggressive and lacks long-term considerations. Overall, the AMPC algorithm has more advantages in extending battery life, while simple or aggressive strategies may lead to life fluctuations and shortened overall lifespan.

[0328] The energy comparison of accumulated business processing under different algorithms is shown in the figure below: Figure 8 As shown: Figure 8 The cumulative energy usage of the four algorithms in the entire time slot is shown. Combined with the analysis of the battery life change graph, it can be seen that the improved MPC algorithm (AMPC) performs best between energy utilization and battery life, balancing the two. The simple MPC algorithm has low energy utilization efficiency and is more conservative, but it effectively extends battery life. The game theory algorithm is more balanced in energy distribution, but does not perform well when dealing with peak demand. The greedy algorithm has high energy utilization efficiency, close to AMPC, but excessive focus on current demand may lead to long-term efficiency decline and affect battery life. Selecting the AMPC algorithm can extend battery life while meeting business needs, improving overall system performance and reliability.

[0329] Comparison of BHLOI under different algorithms Fig. 9 As shown: Fig. 9The comparison of four different optimization algorithms on BHLOI is shown. This indicator takes into account both cumulative energy usage and battery life, reflecting the effectiveness and efficiency of the algorithm in actual operation. The results show that the AMPC (Adaptive Model Predictive Control) algorithm has the highest overall performance, followed by the unmodified simple MPC algorithm and the greedy algorithm, while the game theory algorithm has the lowest overall performance. The AMPC algorithm can better balance energy usage and battery life, so it performs well in practical applications. Although the unmodified simple MPC algorithm is simple, it is relatively robust when dealing with complex environments. Although the greedy algorithm has a high effect, it does not consider the charging and discharging efficiency, and its practical application may be limited. The game theory algorithm has the lowest overall performance indicator, which may be because it fails to fully optimize the balance between battery life and energy usage when simulating multiple participants' games.

[0330] Overall, the AMPC algorithm has the most practical application value, while the simple MPC and greedy algorithms also have certain advantages in specific scenarios. Although the game theory algorithm has advantages in theory, it needs further optimization in practical applications to realize its potential.

[0331] The complexity analysis of different algorithms is as follows:

[0332] Table 4 Complexity of different algorithms

[0333]

[0334] Where X GT Y is a participant in the Game Theory algorithm. GT For decision-making space.

[0335] The complexity of the AMPC algorithm is O(T), and it performs best when dealing with complex time series problems, and can effectively balance battery life and energy usage. The MPC algorithm also has an O(T) complexity and performs well in short-term prediction and control problems. The Greedy algorithm is also O(T), and it works well in simple situations, but lacks a global perspective in complex situations. The GameTheory algorithm has a higher complexity of O(T·X·Y). Although it can consider the game of multiple participants, its calculations are complex and the effect is average. It does not perform as well as the previous two in terms of battery life and energy balance.

[0336] In order to intuitively demonstrate the influence of the Lyapunov control parameter V on the battery state in the present invention, Fig.10 The variation of battery status with Lyapunov control parameter V is shown: Fig.10The effect of different V values ​​on the average battery state and the deviation from the target battery state is shown. As the V value continues to rise, the battery state will get closer and closer to the target battery state, and the deviation from it will continue to decrease, and the improvement in battery life will become more and more obvious. However, a larger V value will cause the battery state of the LEO satellite to always remain around the target battery state during operation, which will affect the LEO satellite's ability to handle business. Therefore, the V value can affect the LEO satellite's battery life and ability to handle business. In problem solving, the battery life of the LEO satellite and its ability to handle business will be weighed, and the V value will be adjusted to change the weight of the two.

[0337] In addition, the present invention conducts a verification experiment on the effect of different battery capacities on the life optimization of LEO satellites to compare the performance of the algorithm proposed by the present invention under different battery capacities. Fig.11 As shown: Fig.11 The effect of different battery capacities on the average life after optimization is demonstrated. The results show that the battery life is the highest at 642.5 cycles when the battery capacity is 10kWh, and then the battery life decreases as the battery capacity increases. This shows that under certain conditions, medium-capacity batteries can more effectively utilize charging and discharging strategies to extend battery life, while excessive battery capacity may reduce life due to charging loss or failure to fully discharge. These findings provide important guidance for battery selection and energy management strategy optimization.

[0338] The specific life span of LEO satellites with different battery capacities varies as follows: Fig.12 As shown: Fig.12 The figure shows how the battery life changes over time for different battery capacities (5kWh to 50kWh). The figure clearly shows that the battery capacity has a significant impact on the battery life. The life of smaller battery capacities (such as 5kWh and 10kWh) decreases rapidly at the beginning of operation, but soon stabilizes. Conversely, larger battery capacities (such as 40kWh and 50kWh) show a lower life in the end, although their life decreases slowly in the early stage. This shows that smaller capacity batteries are more susceptible to wear and tear in early use, while larger capacity batteries have increased long-term management complexity, resulting in a decrease in overall life, despite stable initial performance. This reflects the combined impact of battery capacity and energy management strategy on battery life.

[0339] The energy consumed by LEO satellites with different battery capacities to process business needs is as follows: Fig.13 As shown: Fig.13It shows how the cumulative energy usage changes over time under different battery capacity conditions. As the battery capacity increases, the cumulative energy used to process the business also gradually increases. This shows that larger battery capacities can better meet energy needs during long-term operation. In the figure, each curve shows the cumulative energy usage of different target battery capacities from 5kWh to 50kWh during operation. It can be observed that the energy usage accumulation of batteries with larger capacity is more obvious in the later time period, especially in high-demand periods, the energy accumulation curve is steeper. This reflects the advantage of large-capacity batteries during peak demand periods.

[0340] The present invention comprehensively considers objective conditions such as solar radiation energy, existing solar panel conversion efficiency and area, volume and weight of LEO satellite, and selects a battery capacity of 10kWh as the battery capacity of the LEO satellite of the present invention.

[0341] V. Conclusion

[0342] In view of the limited energy of LEO satellites and the substantial increase in business volume in the 6G scenario, this paper proposes a LEO satellite life optimization strategy based on data prediction and matching. It aims to improve the battery life and business processing capabilities of LEO satellites and reduce the operating costs of LSNs.

[0343] The strategy proposed in the present invention is roughly divided into three parts, namely, a solar energy prediction part, a business demand energy prediction part and a life optimization part.

[0344] (1) Solar energy prediction: For LEO satellites, the solar energy they can receive in their orbits is limited and unstable. The present invention uses normal distribution to simulate the changes in the radiation energy of the sun to LEO satellites in the universe, and fully refers to the distribution of the sunlit area and the shadow area of ​​the LEO satellite in the existing literature to set the relevant parameters of the normal distribution, and introduces random interference terms to simulate the interference of solar energy movement in the universe on solar radiation energy, and establishes a model for LEO satellite solar radiation energy. The model can accurately simulate the changes in solar radiation energy on LEO satellites over time. In addition, the present invention uses the CNN-LSTM algorithm to predict the solar energy in the LEO satellite orbit, and compares it with LSTM and RNN. According to R 2 Compared with the values ​​of , the algorithm accuracy of CNN-LSTM increased by 8.82% and 14.90% respectively. Solar energy is the only energy that LEO satellites obtain in orbit, but its randomness and limitation bring certain adjustments to LEO operation. Accurate prediction of it can provide strong data support for LEO satellite energy decision-making.

[0345] (2) Service demand energy: For LEO satellites, they move at high speed relative to the ground, so the ground services they receive have the characteristics of dynamic changes in geographical location. Different geographical locations have a direct impact on their service demand density due to the difference in regional economic volume. Secondly, the service demand density also changes dynamically in the time dimension, which is directly related to the active habits of ground users. In view of the dual randomness of service demand in time and space, the present invention establishes a service demand energy model. In view of the variability of geographical location, the sub-satellite trajectory of LEO satellite is simulated using STK software to obtain the specific time of LEO satellite coverage of the ground. In view of the variability of time, 24 Poisson distributions are divided into 24 hours according to the different levels of activity of ground users, which are used to measure the difference in service density at different times. Then the two are matched in time, and the energy factor is set to obtain the service demand energy model of LEO satellite. This model can accurately measure the dual randomness of service demand in geography and time, and has certain guiding significance for the service model of LEO satellite. Then, the present invention uses the LSTM-RF algorithm to predict the business demand energy, and introduces the LSTM and RNN algorithms for comparison. The algorithm accuracy is improved by 14.57% and 17.10% respectively.

[0346] (3) Life optimization decision: For LEO satellites, their life is largely affected by the battery life. The battery undergoes charging and discharging cycles during the operation of the satellite. Over time, its performance will gradually decline, resulting in a reduction in battery capacity, which ultimately affects the effective operation of the satellite

[40] . Based on the prediction results of solar energy and business demand energy, the present invention uses Lyapunov theory and AMPC algorithm to manage the energy of LEO satellites, so that it can meet business needs while improving the life of LEO satellites. In addition, the present invention proposes an indicator BHLOI that can effectively measure the ability of LEO satellites to meet business needs and the effect of life improvement under the decisions of different algorithms. The simulation results show that the strategy proposed by the present invention is 10.79%, 25.22% and 58.16% higher than that of the MPC algorithm, Game Theory algorithm and Greedy Algorithm algorithm, respectively. The strategy proposed by the present invention fills the gap in the existing life optimization strategy, which ignores the business demand capacity while optimizing the life, and can effectively improve the operational efficiency of LEO satellites. It has certain guiding significance for the capacity management and operation of LEO satellites.

[0347] The above conclusions show that the strategy proposed in the present invention has certain guiding significance for the use of LEO satellites in terms of solar energy modeling, business demand capacity modeling and life optimization strategy of LEO satellites. It should be noted that the method proposed in the present invention is a strategy for a single LEO satellite, while in the LSNs scenario, decisions such as multi-satellite coordination and business aggregation will be taken into consideration. Research on network life and operational efficiency at the network level is a future research plan.

Claims

1. A method for optimizing the life of LEO satellites based on data prediction and matching, characterized by: include: (1) Building a network model: Using the radial distance of the LEO satellite relative to the center of the earth, the LEO satellite orbit model is established, as shown in the following formula: r(t)=(R+h)[1-ecosα(t)] Where R is the radius of the Earth, h is the altitude of the satellite, e is the eccentricity of the satellite orbit, α(t) is the eccentric anomaly angle; The orbital period of LEO satellite is as follows: Where G is the gravitational constant and M is the mass of the Earth; (2) Constructing LEO satellite energy model: Solar radiation energy is converted into electrical energy by the solar panels carried by LEO satellites, where the power of solar radiation energy is P r (t), the converted electric energy power is P g (t); The converted electric energy is distributed through the energy dispatcher; P d-b (t) is the energy power allocated to the payload module by the energy scheduler for use by various ground services; P d-B (t) is the power allocated to the battery module by the energy scheduler. When it takes a positive value, the battery is charged, and when it takes a negative value, the battery is discharged. The energy released by the discharge is used by the payload module; P b (t) Power requirements for LEO satellite processing of terrestrial services; The moment when the satellite enters the sunshine area is taken as the origin, which is the starting point of the solar radiation intensity analysis; The calculation starts from the position t of the satellite in orbit, t = 0 represents the position where the satellite just enters the sunshine area, and the solar radiation intensity P received by the LEO satellite r (t) The energy model is established as follows: Among them, P max It is the radiation intensity of the direct sunlight point. Its physical meaning is the solar radiation energy received per second per unit area perpendicular to the sunlight outside the Earth's atmosphere. α is the proportion of the sunlight area in the orbital period, 0≤t≤αT s Indicates that the satellite is operating in the sunshine area, αT s <t≤T s Indicates that it is running in the earth's shadow area; σ represents the width of the controlled radiation intensity distribution, that is, the speed at which the radiation intensity decreases from the maximum value to the smaller value; N(t) is the noise effect of solar activity and cosmic rays on the solar radiation power; Establish a discrete time system, divide an orbital period of the LEO satellite into K equal-length time slots, the length of the time slot is Δt; j represents any divided time slot, and the time slot sequence is For the t j time slot, its solar radiation energy E r (t j ) is shown in the following formula: Then the solar radiation energy sequence received by the LEO satellite in each time slot is E r (t) = {E r (t1),E r (t2),···,E r (t j ),···,E r (t K )}; For the t j time slot, the energy converted by the solar panel is: AND g (t j )=Aη r-g AND r (t j ) Where A is the area of ​​the solar panel, η r-g is the energy conversion efficiency of the satellite solar panel; then the energy sequence converted by the solar panel in each time slot is E g (t) = {E g (t1),E g (t2),···,E g (t j ),···,E g (t K )}; The energy sequence allocated by the energy scheduler to the payload module in each time slot is E d-b (t) = {E d-b (t1),E d-b (t2)···E d-b (t j )···E d-b (t K )}; The energy sequence allocated to the battery charging is E d-B (t) = {E d-B (t1),E d-B (t2)···E d-B (t j )···E d-B (t K )}; The energy sequence allocated to the battery discharge is E B-d (t) = {E B-d (t1),E B-d (t2)···E B-d (t j )···E B-d (t K )} The battery module can store the electric energy converted by the solar panel and release the energy for ground services when needed. The initial energy sequence of the battery module in each time slot is E B (t) = {E B (t1),E B (t2)···E B (t j )···E B (t K )}; The energy sequence required by terrestrial services in each time slot is E b (t) = {E b (t1),E b (t2)···E b (t j )···E b (t K )}; (3) Building a battery model: The battery is used as the energy storage system, the energy storage model is the battery system model, and the main characterization parameter of the battery energy storage system is the energy state E of any time slot. B (t j ), and the charging threshold and discharge threshold They represent the maximum and minimum energy storage of the battery respectively; when the charge threshold is exceeded, the battery will enter the overcharge state, at which time the battery will no longer be charged; conversely, when the battery energy is lower than the discharge threshold, the battery will enter the overdischarge state, at which time the battery will no longer be discharged; the battery energy sequence satisfies the following relationship: Among them, η c-f is the battery charging efficiency, η f-c is the discharge efficiency of the battery; Set the battery discharge depth DOD (t j ) is shown in the following formula: The battery life of the LEO satellite is defined as L, and the depth of discharge DOD is logarithmically related to the satellite life L, as shown in the following formula: log 10 L+A·DOD max =B Where A and B are constants that depend on the battery specifications, DOD max Indicates a complete charge and discharge, that is, the power is discharged from 100% to 0%. L is the benchmark life of the battery, which indicates the life consumed after a complete charge and discharge process; L is the average service life of LEO satellites, and the reduced average service life is ΔL, as follows: (4) Build a business model: According to the different levels of activity of ground users, a day is divided into M time periods, and its time period sequence is expressed as in m Indicates any divided time period; Assuming that the business volume in each time period follows a Poisson distribution, then the τth m The business volume in this time period is Where λ(τ m ) is the τth m The Poisson distribution parameter of a time period, k represents the number of requests for the service in that time period; In one orbital cycle, the sub-satellite trajectory is represented as a set of regional sequences according to the traffic volume. Where o represents any one of the divided regions; then the τth m The nth time period i The business models for each region are as follows: in For nth o Economic factors of different regions are used to measure the impact of the total economic volume of different regions on business volume; In the orbital period t j If it is in the τth time slot of the LEO satellite operation, m The nth time period u area, the energy required for the service in this time slot is E b (t j )=η b-e ·S(τ m ,n o ) where η b-e It is the conversion coefficient from business volume to energy required for business. Its physical meaning is the proportion of energy required to process one unit of business volume. (5) By controlling the decision-making and planning of the battery charge and discharge energy in each time slot, the lithium-ion battery in the LEO satellite operation can be kept in the DOD within the optimal use range; the battery energy state is updated as shown in the following formula: AND B (t j+1 )=E B (t j )+η c-f AND d-B (t j )-η f-c AND B-d (t j ) The battery status is limited by the battery capacity as follows: in The lower limit of battery capacity is the minimum power contained in the battery hardware design. If the power is lower than this, the battery will stop supplying power. is the maximum capacity of the battery; In any time slot, the sum of the battery's charging energy and the battery's current energy cannot exceed the battery capacity limit, as shown in the following formula: In any time slot, the battery discharge energy cannot exceed the difference between the battery energy in the current time slot and the battery lower limit, as follows: Among them, η f-c The charging efficiency of the battery; The energy in the system in any time slot satisfies the following relationship: E g (t j )=E d-b (t j )+E d-B (t j ) in, Indicates in the future The energy demand of the business in a time slot is very large, and when the battery is charged to the maximum power it can reach in advance, it still cannot meet the energy demand of the business. If this difference exists for a period of time in the future, the business demand of certain time slots will be traded off from a global perspective to achieve the best global goal; (6) Solar radiation energy prediction A hybrid model (CNN-LSTM) of convolutional neural network (CNN) and long short-term memory network (LSTM) is used to predict solar radiation energy. First, collect historical solar radiation data, including sunshine duration, radiation intensity, etc., and then remove missing values ​​and outliers to ensure data quality; Secondly, the historical data of solar radiation energy is preprocessed, assuming x g (t) is the historical data set of solar radiation energy. The historical data set is discretized with a unit time length of Δt. The sequence of the discretized historical data set is The total number of time slots of historical data is J g ; Then divide the historical data set into training sets according to a certain ratio and test set The training set is used as the input for model training, and the test set will be used to verify the accuracy of model predictions; Then the training set is normalized as shown in the formula: where max(x g ) and min(x g ) are the maximum and minimum values ​​in the historical data respectively; the normalized data is imported into the model for training, and after the model training is completed, it is verified with the test set. If the model training error is stable within the error limit L, the model parameters are saved; The trained model is used to predict the solar radiation energy of the LEO satellite, and the data predicted by the model is denormalized to obtain the prediction result, as shown in the following formula: The energy sequence of the predicted solar radiation from LEO satellite is: In addition, after conversion, the solar energy sequence obtained by the LEO satellite is In order to effectively evaluate the accuracy of the prediction model, the coefficient of determination (R 2 ) to evaluate the prediction accuracy, R 2 As shown below: in is the average value of solar energy; (7) Business demand energy prediction: The business demand energy is predicted using the LSTM-RF algorithm. The historical data set of business demand energy is x b (t u )x b (t) = [x b (t1),···,x b (t u ),···,x b (t J )], and then preprocess and divide the training set according to a certain ratio and test set The data is normalized and then used as the input for model training and verification. If the model training is successful, the model is saved and the service demand energy of the LEO satellite is predicted. The data predicted by the model is then denormalized to obtain the predicted data. If the model prediction error is stable within the error limit If the forecast result is within , the forecasted business demand energy sequence is as follows (8) Life optimization decision: According to the Lyapunov optimization framework, the Lyapunov function is defined as follows: Where Φ(t)=E B (t); In addition, the Lyapunov drift plus penalty function in any time slot can be written as: In time slot t j The optimization problem P1 can be reformulated as P2: Where V is the system control parameter; the complex stochastic optimization problem P1 is transformed into a common optimization problem P2. P2 is constructed by adding the Lyapunov drift term on the basis of P1. The relationship between P2 and P1 can be expressed as follows: P2=[VP1+ΔΦ(t j )|Φ(t j )] When ΔΦ(t j ) is small enough, the result of P2 is the same as the result of solving P1; in this case, the objective function P2 is simplified to P1; thereby reducing P2 to P1; therefore, under the stability condition, P2 is mathematically equivalent to P1; Due to the deviation between the predicted results and the actual results, and the dynamic nature of the battery state still exists, the improved AMPC algorithm is used to solve the problem. The specific process is as follows: System modeling: Build a dynamic model of the system; this model is used to describe the relationship between the system state and the control input; in AMPC, the system model can be updated over time; Initialization: Set the initial battery charge E B (t0), and read the service demand energy of each time slot Energy obtained by converting solar radiation energy Model update: Introduce the function of adaptive improvement, automatically adjust the subsequent prediction results according to the error between the actual results and the predicted results, first define the solar radiation energy prediction error and business energy prediction error The two errors in any time slot satisfy the following relationship: Corrections are divided into short-term corrections and long-term corrections. Short-term corrections are local optimizations within a smaller time window to ensure that the system behavior in the current time slot and the last few time slots meets expectations, as follows: Where u is the time slot scale of the short-term correction, λ corr is the short-term error weight, which is used to indicate the weight of error correction. The short-term correction of business demand energy is similar. Long-term correction is to perform global optimization over a longer time window, accumulate statistics of errors over a period of time, and adjust subsequent results based on the accumulated errors to ensure the stability and performance of the system over a longer period of time; Where U is the time slot scale of the long-term correction, λ lcorr The same is true for the long-term error weight and the long-term correction of business demand energy; Rolling optimization: By solving the optimization problem in each time slot, applying the decision of the first time slot, and then rolling to the next time slot, repeating this process; continuously solving P2; and finally obtaining the optimal charging and discharging decision for each time slot.

Images