On-Orbit Life Prediction and Analysis of Triple-Junction Gallium Arsenide Solar Arrays for MEO Satellites

Abstract

This paper focuses on the triple-junction gallium arsenide solar array of a MEO (Medium Earth Orbit) satellite that has been in orbit for 7 years. Through a combination of theoretical and data-driven methods, it conducts research on anti-radiation design verification and life prediction. This study integrates the Long Short-Term Memory (LSTM) algorithm into the full life cycle management of MEO satellite solar arrays, providing a solution that combines theory and engineering for the design of high-reliability energy systems. Based on semiconductor physics theory, this paper establishes an output current calculation model. By combining radiation attenuation factors obtained from ground experiments, it derives the theoretical current values for the initial orbit insertion and the end of life. Aiming at the limitations of traditional physical models in addressing solar performance degradation under complex radiation environments, this paper introduces an LSTM algorithm to deeply mine the high-density current telemetry data (approximately 30 min per point) accumulated over 7 years in orbit. By comparing the prediction accuracy of LSTM with traditional models such as Recurrent Neural Network (RNN) and Feedforward Neural Network (FNN), the significant advantage of LSTM in capturing the long-term attenuation trend of solar arrays is verified. This study integrates deep learning technology into the full life cycle management of solar arrays, constructs a closed-loop verification system of “theoretical modeling–data-driven intelligent prediction”, and provides a solution for the long-life and high-reliability operation of the energy system of MEO orbit satellites.

1. Introduction

Solar cells (photovoltaic cells) are semiconductor devices that directly convert solar energy into electrical energy using the photovoltaic effect [1]. Due to the accumulation of space radiation effects, the atomic structure of semiconductor materials changes [2]. As solar arrays are deployed outside satellites and directly exposed to the space environment, their power output capability gradually degrades with the extension of satellite on-orbit time [3].

Medium Earth Orbit (MEO) is located between Low Earth Orbit and Geostationary Orbit [4]. Currently, most global MEO satellites operate at an altitude of approximately 20,000 km. This region is adjacent to the Van Allen radiation belts—the harshest radiation environment in Earth’s space [5], where protons of 0.1 MeV–400 MeV, electrons of 40 keV–7 MeV, and a small amount of heavy ions gather [6]. The longer a satellite remains in orbit, the greater the total radiation dose from charged particles of different energy levels, leading to more significant performance degradation [7]. The design life of medium-to-high-orbit satellites is typically 10–15 years [8]. The strong radiation environment exacerbates the displacement damage effect of solar cells: the shortening of minority carrier lifetime reduces current density, and electrical performance parameters such as open-circuit voltage, short-circuit current, and output power continuously decline, seriously affecting the on-orbit reliability and service life of spacecraft [9,10,11]. Therefore, compared with satellites in other orbits, MEO satellites need to pay more attention to the impact of radiation on the output power of solar arrays [7,12].

Deep learning technologies have promoted innovations in the field of anomaly detection [13]. As an improved version of Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) solves the problems of gradient vanishing and long-term dependency through a gating mechanism [14], showing outstanding performance in time series analysis for industrial fault diagnosis, financial risks, and other fields [15]. However, in spacecraft energy systems, the application of LSTM in predicting the long-term degradation of solar arrays still needs to be expanded.

In existing studies, ref. [16] analyzes the life of solar arrays based on physical models, and [17] proposes a data-driven surrogate model, which improves the prediction accuracy of output power to 98.65% and is 10 million times faster than traditional simulations, but the data originates from simulations rather than on-orbit measurements. Ref. [18] combines physical models with deep learning to construct a hybrid prediction model using on-orbit data from stratospheric airships. Refs. [19,20] predict the performance degradation of solar arrays by simplifying physical models combined with evolutionary algorithms and comprehensive environmental simulation. However, traditional physical models mostly rely on empirical formulas, making them difficult to adapt to the nonlinear changes in radiation degradation and failing to fully utilize the advantages of LSTM in processing long-period sequences.

This study focuses on the triple-junction gallium arsenide solar array of a MEO satellite that has been in orbit for 7 years, with significant data advantages: compared with the two-week sampling interval [21], the data collection frequency of this study reaches 30 min per point; compared with the 5-year and 8-month period [22], the 7-year long-term sequence in this paper provides richer information for model training. The 7-year data span (2 more years than the 5-year dataset in previous studies) directly brings a larger volume of training data. Such extended temporal coverage captures more complete degradation cycles and cumulative environmental impacts, such as space radiation (protons and electrons) and temperature (with fluctuations ranging approximately from −120 °C to 60 °C), which are crucial for modeling long-term performance trends. Meanwhile, the 7-year data used in this paper enables the neural network model, through training on extended sequences, to better capture the correlation between early-stage performance parameters and late-stage degradation, improving the predictive capability for power output decay over a 10-year lifespan. It also reduces extrapolation errors. Shorter datasets may lead to overfitting or extrapolation bias in long-term predictions, whereas the 7-year data provides a more robust basis for trend extrapolation. By converting on-orbit time into seconds as input and combining data normalization processing, the model training efficiency is improved. Comparative experiments show that the LSTM model performs best in capturing the degradation trend of solar arrays and predicting performance degradation.

2. Materials and Methods

2.1. Design of Triple-Junction Gallium Arsenide Solar Arrays

For a MEO orbit satellite, the solar array is required to deliver an output power of no less than 2000 W at the 42 V bus voltage by the end of its 10-year design life. The power generation units of the satellite’s solar array adopt triple-junction gallium arsenide solar cells with an average photoelectric conversion efficiency of 30%, including two specifications: 39.8 mm × 60.4 mm × 0.175 mm and 30.6 mm × 40.3 mm × 0.175 mm. A 0.12 mm-thick cerium-doped glass cover slip is pasted on the cell surface, with specific parameters shown in

Table 1. Electrical parameters of GaInP2/GaAs/Ge solar cells.

Parameters	Value
Short-circuit Isc	17.67 mA/cm2
Open-circuit voltage Voc	2.7 V
Optimum operating point current Imp	17.13 mA/cm2
Optimum operating point voltage Vmp	2.396 V
Fill factor FF	0.86
Photoelectric conversion efficiency η	30%
Absorptivity: αS	≤0.92
Hemispherical emissivity: εH	0.82 ± 0.03

. The design parameters of the solar cell circuit are listed in

Table 2. The relevant parameter table adopted for the design of the solar cell circuit.

Parameters	Value
Circuit and diode voltage drop	3 V
Voltage irradiation degradation factor	0.92
Voltage temperature coefficient	−6.4 mV/°C~−7.2 mV/°C
Combined loss factor (Voltage/Current)	0.98
Electrical performance test error factor (Voltage/Current)	0.98
Current particle irradiation degradation factor	0.91
Current ultraviolet radiation loss factor	0.98
Current temperature coefficient	0.006 mA/cm2 °C~0.014 mA/cm2 °C

, and 26 cell pieces are connected in series.

During product manufacturing, we used the dedicated Large Area PV Sunlight Simulator (LAPSS Ⅱ) (Spectrolab, Sylmar, CA, USA) testing equipment to quantitatively test the electrical performance output characteristics of each circuit under the standard test conditions of AM0, 1353 W/m², and 25 °C. The test data results are listed in

Table 3. Statistical table of output characteristics test data for solar array circuits.

	Sub-Array	Voc (V)	Vmp (V)	Isc (A)	Imp (A)	I46 (A)	Pmax (W)
Ground Test Results of +Y Solar Wing	1	69.69	62.48	5.894	5.477	5.733	342.2
	3	69.81	62.24	6.806	6.342	6.628	394.7
	5	69.77	62.54	6.365	5.899	6.175	368.9
	7N	69.73	62.01	4.698	4.316	4.555	267.6
	7Z	69.84	62.09	1.707	1.607	1.667	99.80
	9W	69.82	63.16	4.276	3.955	4.163	249.8
	9Z	69.76	61.79	2.139	2.020	2.078	124.8
Total Power of +Y Solar Wing							1847.8
Ground Test Results of −Y Solar Wing	2	69.65	62.9	5.920	5.484	5.754	344.9
	4	69.77	62.08	6.801	6.318	6.579	392.2
	6	69.73	62.40	6.386	5.907	6.182	368.6
	8N	69.78	62.53	4.667	4.344	4.535	271.6
	8Z	69.72	61.96	1.701	1.608	1.667	99.65
	10W	69.85	62.48	4.287	3.982	4.167	248.8
	10Z	69.82	62.70	2.125	1.976	2.075	123.9
Total Power of −Y Solar Wing							1849.65
Total Power of +Y and −Y Solar Wing							3697.45

Remarks: 7N represents the part of Sub-array 7 on the inner panel, and 7Z represents the part of Sub-array 7 on the middle panel; 9W represents the part of Sub-array 9 on the outer panel, and 9Z represents the part of Sub-array 9 on the middle panel; 8N represents the part of Sub-array 8 on the inner panel, and 8Z represents the part of Sub-array 8 on the middle panel; 10W represents the part of Sub-array 10 on the outer panel, and 10Z represents the part of Sub-array 10 on the middle panel. I46 represents the output current of the solar cell when the output voltage is 46 V. (The bus voltage of the satellite described in this paper is 42.2 V, and the voltage drop across the circuit and diodes is 3 V. The actual operating voltage is approximately 46 V, so the output current of the solar array at this voltage was tested).

below. Taking Sub-array 1 as an example, its IV curve is shown in Figure 1.

As the number of on-orbit years increases, the IV curve of solar cells exhibits a gradual decrease in short-circuit current, a slow reduction in open-circuit voltage, and a decline in fill factor. This manifests as an overall steady downward shift of the IV curve, with specific characteristics:

Horizontal leftward shift: The open-circuit voltage plateau decreases.

Vertical downward shift: The peak current reduces.

The maximum power point (MPP) region shrinks.

These changes indicate a decline in the operating output current corresponding to the working point voltage (between 45 and 46 V) of the on-orbit satellite. According to the prediction in this paper, the on-orbit output current of the battery array will decay by 12.6% at the end of its 10-year life, meaning that the operating output current corresponding to the working point voltage (45–46 V) of the on-orbit satellite will decrease by 12.6%(the degradation trend is reflected by the decay of the on-orbit solar array output current discussed in Chapter 3 of this paper). Since the satellite described in this paper is in orbit, ground instruments cannot be used to measure its post-decay IV curve. However, the degradation trend can be reflected through on-orbit telemetry data, specifically by monitoring the degradation of the solar array’s operating output current.

The satellite power system described in this paper consists of a solar array, a lithium-ion battery pack, a power controller, etc. The main function of the power subsystem is to achieve energy transmission and power balance between the solar array and the battery pack, ensuring the normal power supply of the primary power system.

The power controller adopts a bus scheme with three-domain full-regulation control based on the voltage error signal (MEA). MEA uniformly manages the power regulation modules inside the power controller to maintain the bus voltage constant during both the illumination period and the shadow period. The bus voltage of the satellite described in this paper is 42.2 V. Considering a 3 V voltage drop from the circuit and diodes, the output voltage of the solar array at the end of its life needs to be greater than 45.2 V.

The calculation formula for the optimal operating point voltage of a single cell at the end of its life is shown in Formula (1).

$$Vcell=[V_{mp}+{\beta }_v\cdot (T-25\left)\right]\cdot K_{Vzh}\cdot K_{Vcs}\cdot K_{Vfz}$$

(1)

In the formula:

Vcell: The optimal operating point voltage of a single cell at the end of its life;

$K_{Vzh}$: Voltage ultraviolet radiation loss factor;

$K_{Vcs}$: Voltage combination loss factor;

$K_{Vfz}$: Attenuation factor under the combined effect of Voltage particle irradiation and ultraviolet irradiation.

According to Table 1 and Table 2, the optimal operating point voltage of the solar cell at the end of its life can be calculated as 1.83 V. To meet the output voltage of 45.2 V, the number of series-connected solar cells is designed to be 26 strings, resulting in an optimal operating point voltage of 47.58 V at the end of the life.

The solar array is designed with a total of 10 Sub-arrays. Sub-arrays 1 to 10 correspond to shunt stages 1 to 10 in sequence, with 5 Sub-arrays arranged on each wing. The +Y Wing includes Sub-arrays 1, 3, 5, 7, and 9, while the −Y Wing includes Sub-arrays 2, 4, 6, 8, and 10. The parallel connection modes of the cells in each Sub-array are as follows:

+Y Wing:

Sub-array 1: 14 large cells in parallel.

Sub-array 3: 16 large cells in parallel.

Sub-arrays 5, 7, 9: 14 large cells + 2 small cells in parallel.

Sub-array 7 is distributed across the inner and middle panels.

Sub-array 9 is distributed across the middle and outer panels.

−Y Wing:

Sub-array 2: 14 large cells in parallel.

Sub-array 4: 16 large cells in parallel.

Sub-arrays 6, 8, 10: 14 large cells + 2 small cells in parallel.

Sub-array 8 is distributed across the inner and middle panels.

Sub-array 10 is distributed across the middle and outer panels.

The Sub-array configuration of each solar wing is shown in Figure 2 and Figure 3.

The satellite operates in a MEO orbit at an altitude of 21,000 km, adjacent to the Van Allen radiation belts, and is subjected to a harsh charged particle radiation environment (with a design life of 10 years). Its main environmental parameters are as follows:

In terms of particle radiation, the cumulative flux of charged particles is 4.7 × 10¹⁴ e/cm² for electron flux (VOC) and 4.4 × 10¹⁴ e/cm² for ion flux (ISC), with a total radiation dose of 1.2 × 10⁹ rad. Such radiation can cause displacement damage to solar cells [23], shorten the minority carrier lifetime, and thus lead to the continuous decline of electrical performance parameters such as open-circuit voltage, short-circuit current, and output power [24].

In terms of ultraviolet (UV) radiation, the spectral range is 10–400 nm, with a total radiation dose of 3.05 × 10¹⁰ J/m². The equivalent dose calculated based on the full illumination period throughout the service life is 6275 ESH (Equivalent Sun Hours), which can reduce the light transmittance of glass cover slips and adhesives, thereby affecting the power output of the solar array [25].

The power output of the solar array is primarily influenced by factors such as temperature, illumination, solar incidence angle, and space environment. Ref. [26] has analyzed its on-orbit degradation characteristics and influencing factors. The satellite in this paper uses yaw control to keep the solar array always oriented toward the sun [27], so the influence of the solar incidence angle is ignored in the analysis. The daily solar–Earth distance factors during the satellite’s orbit are obtained through STK12.0 software simulation, and the output current of the solar array is fitted to eliminate the influence of the solar–Earth distance factors [28,29]. Temperature mainly affects short-term fluctuations, and the interference is reduced by sampling data at stable temperature moments during the illumination period. For the particle radiation environment, a radiation degradation factor of 0.91 at the end of 10 years is introduced in the ground design based on past experience. Zeng Yi et al. from the Beijing Institute of Spacecraft System Engineering explained the selection of this parameter in their article [30].

For the ultraviolet (UV) radiation environment, the equivalent UV radiation dose during the service life is calculated as 6275 ESH [31]. Ground tests show that when the glass cover slips and cover adhesives are subjected to radiation exceeding 10,000 ESH, the light transmittance change rate is less than 2%. Therefore, a UV degradation redundancy factor of 0.98 at the end of 10 years is introduced in the circuit design to ensure the power demand at the end of life. Based on the above battery design, ground tests, and theoretical analysis, the current of a single cell can be obtained.

The output current of a single large cell is shown in Formula (1) [32,33]:

$$Ib=[I_{mp}\cdot S+S\cdot {\beta }_i\cdot (T-25\left)\right]\cdot K_{Azh}\cdot K_{Acs}\cdot K_{Afz}=0.362\mathrm{A}$$

(2)

In the formula:

$I_{mp}$: Optimum operating point current of a single solar cell at the beginning of life;

S: Area of the solar cell;

${\beta }_i$: Current temperature coefficient;

T: Operating temperature for current;

$K_{Azh}$: Current ultraviolet radiation loss factor;

$K_{Acs}$: Current combination loss factor;

$K_{Afz}$: Attenuation factor under the combined effect of current particle irradiation and ultraviolet irradiation.

The output current of a small, large cell is shown in Formula (2):

$$Il=\left[I_{mp}\cdot S+S\cdot {\beta }_i\cdot \left(T-25\right)\right]\cdot K_{Azh}\cdot K_{Acs}\cdot K_{Afz}=0.181\mathrm{A}$$

(3)

A single-wing solar panel is connected with 72 large cells and 6 small cells in parallel. Therefore, the output current of the satellite’s solar array after 10 years in orbit is 27.15 A, as shown in Formula (3).

$$Ia=Ib\cdot 72+Il\cdot 6=27.15\mathrm{A}$$

(4)

2.2. On-Orbit Telemetry Data Processing

On-orbit satellites downlink telemetry data at a certain rate. The satellite described in this paper transmits telemetry through the measurement and control channel, including multiple links such as signal acquisition, coding processing, modulation transmission, and ground reception decoding. Its core is to achieve reliable transmission of satellite telemetry data from space to the ground. The telemetry data of the solar array current transmitted by the satellite is derived from the voltage signal, real-time collected by the temperature sensor (thermistor) in the single machine, and the ground converts it into a current value through a conversion coefficient. Different sampling rates are set according to the transmission mechanism. The transmitted telemetry is data-encoded according to a certain protocol to form a telemetry frame (including frame synchronization header, data field, check code, etc.).

Sensors in the satellite telemetry system degrade over time, mainly due to the extreme conditions of the space environment (space irradiation leading to performance degradation of semiconductor devices) and natural aging of materials. Sensor degradation can lead to reduced measurement accuracy, increased noise, prolonged response time, or even complete failure. To address sensor degradation, the key parameters of this satellite (the solar array output current used in this paper is a key parameter) are configured with three temperature sensors to improve reliability through majority voting; additionally, the telemetry output circuit adopts triple-mode redundancy, enabling the system to continue normal operation when a single circuit fails due to radiation. Meanwhile, all components selected for the satellite described in this paper use devices whose radiation resistance equivalent meets the 10-year life equivalent under the space environment conditions of this orbit. Moreover, the satellite described in this paper is a commissioned satellite that is operating normally. Therefore, the above can demonstrate that the data used in this paper is accurate and reliable.

For the band of the transmission channel, a certain carrier frequency is selected to modulate the data. The satellite transmits to the ground measurement and control station through the antenna for signal reception, frequency conversion, demodulation, and decoding processing. Error data will be generated during this transmission period. Some individual points in Figure 4a of this paper are roughly error data, so when performing machine learning, the error data needs to be eliminated.

In this paper, sampling points are extracted at 30 min intervals to obtain seven years of data from May 2018 to May 2025 for this MEO orbit satellite. The original data is shown in Figure 4a. Since MEO orbit satellites experience two shadow seasons each year, each lasting approximately 45 days, the array current decreases during the penumbra period and drops to 0 during the umbra period. To analyze the long-term degradation of the on-orbit satellite’s solar array, it is necessary to exclude data where the current is lower than the normal value during the penumbra and umbra periods. In addition, occasional bit errors may occur during the demodulation, decoding, and transmission of satellite telemetry data, which need to be eliminated. The specific processing flow is as follows:

First step: The Threshold method is used to remove error code points and sampling points with values lower than the normal value. The processed data is shown in Figure 4b.

Second step: Fitting the solar–Earth distance factor for the data obtained in Step 1, with the results shown in Figure 4c.

Third step: As the array degradation causes the data to exhibit a long-term downward trend, traditional anomaly detection methods for stationary data (such as the 3σ principle, Local Outlier Factor (LOF), isolation forest, etc.) fail when processing penumbra solar array current data within the threshold range. Therefore, a moving window anomaly detection method suitable for slowly changing trend scenarios is adopted, specifically implemented as follows:

• Sliding Window Statistic Calculation:

The statistical characteristics of local data are calculated point-by-point through a fixed-size sliding window to adapt to the slow trend changes in the data. The window size is set to $W=130$. For the i-th data point, its corresponding window is $\left(\right[i-W+1,i\left]\right)$ (including the current point and the previous $(W-1)$ points). A large window $(W=130)$ is insensitive to trend changes and suitable for capturing long-term degradation, while a small window is vulnerable to short-term noise interference, which may misidentify normal fluctuations as anomalies. The window size of 130 in this paper matches the data characteristics. The statistical formulas used in this method are as follows:

• Local Mean: ${\mu }_i={\displaystyle \frac{1}{n_i}}\sum _{k\in \mathrm{w}\mathrm{indow}}{x}_k^{’}$

Where $x_k^{’}$ is the valid value after removing outliers, and $n_i$ is the number of valid samples in the window.

B.

Local Standard Deviation: ${\sigma }_i=\sqrt{{\displaystyle \frac{1}{n_i-1}}\sum _{k\in \mathrm{w}\mathrm{indow}}(x_k^{’}-{\mu }_i{)}^2}$

This method identifies outliers based on the standardized deviation $(Z-\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e})$ of local statistics. The $Z-\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}$ formula is ${\mathrm{Z}-\mathrm{score}}_i={\displaystyle \frac{x_i-{\mu }_i}{{\sigma }_i}}$.

An outlier is determined when $\left(\right|{\mathrm{Z}-\mathrm{score}}_i|>T)$, where the threshold $(T=2.9)$ is determined by experience and test results. Physically, if a data point deviates from the local mean by more than $2.9\sigma$, it is considered a significant outlier (corresponding to a confidence interval of approximately 99.8%).

2.

Rel-time Update:

To ensure that outliers do not affect the calculation of subsequent statistics and enhance the algorithm’s accuracy and reliability, data is processed point-by-point. Once an outlier is detected, it is immediately marked, and subsequent statistics are updated. For the $i$-th point, the window statistics $({\mu }_i,{\sigma }_i)$ are calculated. If it is identified as an outlier, $\left(x_i\right)$ is marked as $\mathrm{NaN}$ (Not a Number). When calculating the ($i+1$)-th point, the window automatically includes the updated $\mathrm{NaN}$ value, and the statistics automatically ignore $\mathrm{NaN}$ (i.e., the denominator in the statistic calculation excludes the number of $\mathrm{NaN}$ values). After the above processing, abnormal data during the penumbra period is effectively removed, as shown in Figure 4d. This method avoids the interference of trend degradation on outlier detection by dynamically updating window statistics, making it suitable for outlier identification in long-term non-stationary data.

3. Results and Discussion

3.1. Assessment of Satellite Initial Orbit Insertion Current

The satellite described in this paper was launched in May 2018. After the satellite entered orbit and the solar panels were deployed, the output current of the solar array is shown in Figure 5 below. The calculation for the output current of the solar array is as shown in Formula (5):

$$I=[I_{mp}\cdot S+S\cdot {\beta }_i\cdot (T-25\left)\right]\cdot K_{Azh}\cdot K_{Acs}\cdot K_{Afz}$$

(5)

According to the parameters of the battery cells and the number of parallel cells in each shunt stage, the output current of a single-wing solar panel in the early on-orbit stage is shown in

Table 4. Configuration and output current for a single solar array during initial orbit.

Shunt Stage		Output Current/A (25°)	Large Cell Parallel Number	Small Cell Parallel Number	${\mathit{I}}_{\mathit{m}\mathit{p}}$	Large Cell Area (cm2)	Small Cell Area (cm2)
1	2	5.54	14	0	17.13 mA/cm2	24.0392 =3.98 × 6.04	12.3318 =3.6 × 4.03
3	4	6.33	16	0
5	6	5.94	14	2
7	8	5.94	14	2
9	10	5.94	14	2
Total		29.69	72	6	/	/	/

as follows:

Taking the first shunt stage as an example, the calculation process of its initial on-orbit output current is as follows: In the initial stage of life, the degradation of ultraviolet radiation and particle radiation is not considered, and only the combined loss factor (0.98) and the current electrical performance test error factor (0.98) are introduced. Then, the single-stage output current is $Is1=\left[I_{mp}\right]\cdot 14\cdot 24.0392\cdot 0.98\cdot 0.98=5.54\mathrm{A}$, corresponding to a single-wing current of 29.69 A.

The actual solar irradiance (S) of the on-orbit satellite needs to consider the solar-earth distance factor [17].

$$S=S_0\cdot f^2\cdot \mathrm{c}\mathrm{o}\mathrm{s}\theta$$

(6)

where $S_0=1361$ W/m² is the solar constant, and $\theta$ is the angle between the solar rays and the plane of the solar array (the satellite described in this paper uses yaw control to face the sun, and the solar array is perpendicular to the solar vector, so $\theta =0$).

The solar–Earth distance factor (Solar Distance Factor) is the ratio of the actual solar–Earth distance to the average solar–Earth distance (1 astronomical unit, approximately 1.496 × 10⁸ km), which is used to correct the variation of solar irradiance [34,35]. The calculation of the solar–Earth distance factor is shown in Formula (6) [36]:

$$f={\displaystyle \frac{\mathrm{r}}{\mathrm{r}0}}$$

(7)

where $\mathrm{r}$ is the actual solar–Earth distance, and $\mathrm{r}0$ is the average solar–Earth distance. Considering the solar–Earth distance factor on the launch day in May, the theoretically corrected output current of the actual single-wing solar panel in the early on-orbit stage is 29.11 A. The measured current data of the satellite in the early orbit insertion stage is shown in Figure 5. It can be seen that the actual on-orbit current is highly consistent with the theoretical calculation value, verifying the reliability of the initial working state of the solar array.

By taking the average of the daily telemetry data of the solar array current, the annual degradation of the satellite over 7 years in orbit is shown in

Table 5. Statistical table of solar array output attenuation in orbit for 7 years.

Year	2019	2020	2021	2022	2023	2024	2025
Annual attenuation percentage	1.96%	1.44%	1.23%	1.36%	1.56%	1.75%	1.50%

as follows:

As can be seen from Table 5, the solar array shows significant degradation in the first year of orbit, mainly due to the initial concentration effect: the first year is when the solar array is first exposed to a strong radiation environment for a long time, and the damage accumulates the fastest. The reason for the significant degradation in 2024 is that this year is a high solar activity year, and multiple M9.0-class solar flares occurred, leading to greater degradation. The space environment conditions of this year are shown in Figure 6 [37].

In the figure, the “” symbol denotes proton events, while the “” symbol represents high-energy electron storms, the “” symbol denotes geomagnetic storms, while the “” symbol represents X-ray flares. The color of the symbol represents the level of the alert. Yellow indicates a yellow alert, orange an orange alert, and red a red alert. As the color deepens, the alert level increases. The main information about 2024 being a solar maximum year is as follows:

Solar activity is periodic, with an average period of about 11 years. The number of sunspots is an important indicator for measuring the intensity of solar activity, and years with a high number of sunspots are considered peak years of solar activity [38]. During a telephone conference on 15 October 2024, representatives from the National Aeronautics and Space Administration (NASA) and the National Oceanic and Atmospheric Administration (NOAA) announced that the Sun had reached the peak level of the current activity cycle, which is the 25th cycle in human detailed records. The 25th cycle is much more active than the previous one [39,40].

3.2. On-Orbit Data Prediction

After data processing, the data of the first 6 years is used as the training set, and the data of the subsequent 1 year is used as the validation set. LSTM, RNN, and FNN are respectively adopted for data training and prediction. Based on the optimal model, the output current of the battery array in the subsequent 3 years is predicted.

The prediction steps are as follows:

• Subtract the telemetry downlink time by the time of the first data point, and convert it into seconds.
• Normalize the time and current.
• Divide the training set and test set.
• Use LSTM, RNN, and FNN for training and prediction.

The architectures of the three models constructed in this paper are shown in

Table 6. Summary table of training model architectures.

Model	Number of Layers	Structure of Each Layer	Activation Function	Input Shape
LSTM	Fully Connected Layer (3 Layers)	Input Layer → Dense (64) → Dropout → Dense (32) → Dropout →Output Layer (1)	ReLU (Hidden Layer) Linear (Output Layer)	(Sample number, 1)
RNN	Recurrent Layer (2 Layers)	Input Layer → SimpleRNN (64, return_sequences=True) → Dropout → SimpleRNN (32) → Dropout →Output Layer (1)	Tanh (Hidden Layer) Linear (Output Layer)	(Sample number, time step, 1)
FNN	LSTM Layer (2 Layers)	Input Layer → LSTM (64, return_sequences=True) → Dropout → LSTM (32) → Dropout →Output Layer (1)	Tanh (Hidden Layer) Linear (Output Layer)	(Sample number, time step, 1)

as follows:

For the FNN (Fully Connected Neural Network), it is composed entirely of densely connected layers, where each layer of neurons is fully connected to the next layer. This method is simple and direct, suitable for processing independent feature inputs. However, it does not consider the time-series characteristics of the data, and has a weak ability to capture sequence patterns. It is applicable to scenarios where there is a clear nonlinear mapping relationship between inputs and outputs.

The RNN (Recurrent Neural Network) introduces recurrent connections, allowing information to be passed between time steps. It can capture short-term temporal dependencies. However, it suffers from the problem of gradient vanishing, making it difficult to learn long-distance dependencies [41,42].

LSTM is a special variant of RNN, which controls information flow through gating mechanisms (input gate, forget gate, output gate). It can effectively solve the problem of gradient vanishing and is good at learning long-term dependencies. It is suitable for long-sequence data (such as time series prediction). The 7-year current telemetry data of the on-orbit satellite in this paper belongs to long-time series data, and there is long-term attenuation and long-distance dependency. Therefore, through theoretical analysis, it is considered that the LSTM algorithm is more suitable for the prediction in this paper.

The training and optimization parameters of the three models are shown in

Table 7. Statistical table of model training and optimization parameters.

Model	Optimizer	Loss Function	Training Parameters	Regularization
LSTM	Adam	MSE	Epochs = 50, Batch size = 32	Dropout (0.2)
RNN	Adam	MSE	Epochs = 50, Batch size = 32	Dropout (0.2)
FNN	Adam	MSE	Epochs = 50, Batch size = 32	Dropout (0.2)

as follows:

By comparing the training/validation losses, the most suitable model is selected. The loss function curves of the three models are shown in Figure 7. The model performance comparison is shown in Figure 8 and

Table 8. Comparison of model prediction accuracy.

Model	MSE	MAE	RMSE
LSTM	0.0139	0.1025	0.1180
RNN	0.0614	0.2201	0.2478
FNN	0.0572	0.2126	0.2392

. The prediction results are shown in Figure 9.

From Figure 7, Figure 8 and Figure 9, it can be observed that LSTM shows the best loss convergence effect on the validation set. The MSE is only 1.39%, significantly improving the prediction accuracy compared with 6.14% of RNN and 5.72% of FNN. The LSTM prediction stays near the mean of the original data, while the predictions of RNN and FNN at the end period both exceed the actual maximum current. The maximum difference between LSTM and RNN predictions reaches 0.5 A. The prediction results of RNN and FNN may lead to unreasonable energy evaluation, causing the failure to meet the load requirements and resulting in an imbalance of system energy.

As shown in the actual telemetry results of the satellite sailboard current in Figure 9, the data exhibits small-amplitude fluctuations. This is because when the satellite is in orbit, the fluctuation of load current triggers the shunt regulation function of the battery array, causing the shunt circuit to frequently switch among the shunt state, adjustment state, and load supply state, thereby affecting the temperature of the S3R shunt stage circuit. According to the previous formula, the performance of a single battery will change with temperature, so the output current of the actual solar array produces small-amplitude fluctuations due to temperature factors.

This paper focuses on the long-term attenuation trend in life prediction, so the influence of short-term fluctuations is not considered. Therefore, in Figure 9, the prediction results of the three models (LSTM, RNN, and FNN) do not reflect short-term fluctuations, but only show the long-term life attenuation trend. By comparison, the prediction results of the LSTM model are most consistent with the actual situation, which fully demonstrates that the LSTM model has significant advantages in the long-sequence prediction task of the output current of the solar array.

The LSTM model is used to predict the current trend for the next three years in orbit, and the results are shown in Figure 10. The prediction shows that at the end of the life, the solar current is 26.49 A, corresponding to an attenuation of 12.56%. Through the previous design and theoretical analysis, the current of the solar array at the end of the life should be 27.15 A. The consistency with the theoretical analysis results reaches 97.6%. It can be seen that the prediction results in this chapter are close to the theoretical design results. At the summer solstice at the end of the life, the Sun–Earth distance factor is 0.9685, and the output power of the double-wing solar array can reach 2165 W at this time, which can fully meet the 2000 W load demand of the satellite.

3.3. Comparison with the Training Results of Low-Density Models

Traditional studies often use MEO orbit data at two-week or monthly intervals for solar array current prediction. To quantify the technical advantages of the high-density telemetry data in this paper, the research team down-sampled the training data to a low-density mode of one point every two weeks and conducted comparative experiments based on three algorithms: LSTM, RNN, and FFN. The prediction results are shown in Figure 11; in the low-density data scenario, LSTM’s prediction accuracy is lower than that of RNN due to the loss of its long-sequence dependency modeling advantage. Among them, RNN performs the best with a mean squared error (MSE) of 0.0351, but this value is 2.53 times higher than the prediction result of the high-density data in this paper (MSE = 0.0139). This result fully verifies the dual advantages of high-density time sampling and long on-orbit period-high-density data. High-density data significantly enhances the model’s ability to capture nonlinear current fluctuations by preserving more temporal feature details, while the long on-orbit time series provides the model with more complete samples of space environment and equipment status evolution, ultimately achieving effective reduction of prediction errors.

4. Conclusions

This paper conducts research on the anti-radiation design of a triple-junction gallium arsenide solar array for a MEO satellite that has been in orbit for 7 years. By constructing a high-density telemetry dataset, experimental results show that the prediction accuracy of the LSTM model is 4.75% higher than that of the RNN, with a coincidence rate of 97.6% with theoretical analysis results. The mean squared error (MSE) is only 0.0139, significantly reducing the prediction error compared with the scenario of low-density data (one point every two weeks). It successfully achieves precise prediction of the output power at the end of the solar array’s life. This method not only verifies the effectiveness of the anti-radiation design but also provides data support for the design optimization of future high-reliability solar arrays by revealing the nonlinear law of battery performance degradation.