<?xml-model href='http://www.tei-c.org/release/xml/tei/custom/schema/relaxng/tei_all.rng' schematypens='http://relaxng.org/ns/structure/1.0'?><TEI xmlns="http://www.tei-c.org/ns/1.0">
	<teiHeader>
		<fileDesc>
			<titleStmt><title level='a'>A Global-Local Probsparse Self-Attention Transformer for LEO Satellite Orbit Prediction</title></titleStmt>
			<publicationStmt>
				<publisher>IEEE</publisher>
				<date>12/18/2024</date>
			</publicationStmt>
			<sourceDesc>
				<bibl> 
					<idno type="par_id">10600884</idno>
					<idno type="doi">10.1109/ICMLA61862.2024.00019</idno>
					
					<author>Guan Huang</author><author>Tao Shu</author>
				</bibl>
			</sourceDesc>
		</fileDesc>
		<profileDesc>
			<abstract><ab><![CDATA[In recent years, the proliferation of LEO (Low-Earth Orbit) satellites and the accumulation of space debris have made Near-Earth space more and more crowded, and hence significantly increased the risk of collisions in this space. As a result, precise orbit prediction becomes essential for LEO satellites to avoid collision, maintain the right constellation, and perform normal space operations. Although machine learning(ML)-based satellite orbit prediction methods have been extensively explored, most existing methods are trained on simulated/synthetic orbit data, which essentially assumes a stationary orbit process and cannot reflect the non-stationary dynamic orbit changes in realworld LEO satellite constellations that are caused by satellite flight status adjustment (e.g., for the purpose of collision avoidance). In this study, we propose a novel multi-range (globallocal) self-attention transformer-based ML model, the GloLoSAT (Global-Local Probsparse Self-Attention Transformer), and train the model over real-world LEO satellite orbit data crawled from N2YO.com to give more precise orbit prediction in case of a series of orbit adjustments. Theoretical analysis demonstrates our Global-Local Probsparse Self-Attention mechanism can achieve O(L log(L)) computational complexity with respect to the input sequence length. Extensive experiments conducted on real satellite orbit tracking datasets demonstrate the efficacy of GloLoSAT in achieving consistent performance improvements across various prediction scenarios compared to their counterparts.]]></ab></abstract>
		</profileDesc>
	</teiHeader>
	<text><body xmlns="http://www.tei-c.org/ns/1.0" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xmlns:xlink="http://www.w3.org/1999/xlink">
<div xmlns="http://www.tei-c.org/ns/1.0"><head>I. INTRODUCTION</head><p>Traditionally, space exploration has focused on deep space discovery, such as seeking to identify new astronomical objects and unravel the secrets of the universe <ref type="bibr">[1]</ref>. However, a recent shift in focus has emerged towards the practical utilization of near-earth space through low-Earth orbit (LEO) satellites <ref type="bibr">[2]</ref>. These satellites, orbiting at altitudes below 2,000 km, are progressively utilized across a spectrum of applications, including terrestrial surveillance, global ubiquitous communication, cataclysm surveillance, navigation, and ecological assessment <ref type="bibr">[3]</ref>. The Starlink project, for example, currently operates over 5,800 active satellites and has plans to increase this number upwards to 42,000 in the near future <ref type="bibr">[4]</ref>. With the nearearth space becoming more and more crowded, the swift deployment and operation of LEO satellites raises potential risks of collision. To mitigate these risks, there is an urgent need for more precise LEO satellite orbit prediction to ensure more effective collision avoidance.</p><p>Conventionally, satellite orbit prediction approaches can be classified into two distinct categories: physics-based approaches and machine learning (ML)-based approaches.</p><p>Physics-based methods, which include general perturbation methods <ref type="bibr">[5]</ref>, special perturbation methods <ref type="bibr">[6]</ref>, and semianalytical methods <ref type="bibr">[7]</ref>, have been studied for many decades. These methods are based on the classical laws of motion and principles of gravity, enabling the prediction of satellite trajectories by solving differential equations of motion, either analytically or numerically <ref type="bibr">[8]</ref>. However, these methods cannot consistently yield high-precision orbit predictions for LEO satellites due to the difficulty in taking into account a large number of trajectory-perturbing random factors such as environmental influences like atmospheric drag and forces of solar radiation, as well as particular satellite attributes, including mass, shape, and information regarding maneuvers. Consequently, the prediction errors can be substantial (e.g., in the order of hundreds of kilometers), potentially rendering them less effective for practical applications.</p><p>Recently, ML-based approaches have emerged as a complementary framework to traditional physics-based models in satellite orbit prediction. At a high level, these methods can be classified into two types: (1) Physics-Enhancing approaches <ref type="bibr">[9]</ref>, <ref type="bibr">[10]</ref> and (2) Error-Correction approaches <ref type="bibr">[11]</ref>- <ref type="bibr">[14]</ref>. In the first type, ML-based methods are often used to enhance the representation of unmodeled random variables in physics-based models or serve as solvers for motion-related differential equations. In the second type, ML models are specifically trained to characterize prediction errors of physicsbased methods. The predictions generated by the physicsbased approaches are then re-tuned using the error corrections generated by the trained ML model. Regardless of their types; however, rather than employing real orbit data for model training, existing ML-based approaches often rely on synthetic datasets that are either entirely simulated based on some space parameters such as gravity field, solar radiation, and atmosphere model <ref type="bibr">[10]</ref>, <ref type="bibr">[13]</ref>, <ref type="bibr">[15]</ref> or calculated from the well-known SGP (Simplified Perturbations) models, utilizing TLEs (Two-Line Element sets) or CPFs (Consolidated Prediction Format files) <ref type="bibr">[9]</ref>, <ref type="bibr">[11]</ref>, <ref type="bibr">[14]</ref>. A major limitation of these ML-based methods is that the random processes and factors simulated in their synthetic/simulated dataset are assumed to be stationary. As a result, ML models trained on such datasets fail to capture the non-stationary dynamic orbit changes in real-world LEO satellite constellations. These changes are due to necessary satellite status adjustments and maneuvers to maintain the constellation or avoid collisions, making the accuracy of these ML-based methods questionable in real-world scenarios.</p><p>In this study, we propose a novel multi-range (global-local) self-attention transformer-based ML model for accurate LEO satellite orbit prediction. In contrast to existing ML-based approaches, the proposed model is trained over real-world LEO satellite orbit data and is capable of giving more precise orbit prediction in case of a series of orbit adjustments. In particular, we first propose a novel web crawler, called N2YO-crawler, designed to collect real-time satellite orbit tracking data from N2YO.com, which provides live real-time trajectory tracking data for over 28,751 objects in outer space, including satellites and space debris. The data for LEO satellites, such as the 5,814 StartLink satellites and the 628 OneWeb satellites, are updated every second on this website. N2YO.com is a highly reputable provider of real-time satellite tracking services. Notably, the space surveillance data used by N2YO.com are diligently collected by the US Space Surveillance Network (SSN), which operates under the rigorous oversight of the US Air Force Space Command (AFSPC) <ref type="bibr">[16]</ref>.</p><p>Secondly, trained over the real orbit tracking data collected by our N2YO-crawler, we propose a novel three-stage transformer-based framework for predicting LEO satellite trajectories, named GloLoSAT. An illustrative overview of GloLoSAT is provided in Fig. <ref type="figure">1</ref>. To overcome the limitations of existing transformer-based LSTF (Long-sequence Timeseries Forecasting) models <ref type="bibr">[17]</ref>- <ref type="bibr">[21]</ref> that often focus solely on either global or local context, lacking a comprehensive view of the input sequence, we introduce an innovative Global-Local Probsparse Self-Attention mechanism. This mechanism contains both local and global Probsparse self-attention modules, enabling GloLoSAT to capture crucial and non-stationary contextual information at various scales.</p><p>The main contributions of this work are four-fold: (i) We introduce N2YO-crawler, an efficient tool for collecting and building real-time satellite orbit tracking datasets from N2YO.com. (ii) We propose a general three-stage transformerbased framework, GloLoSAT, for predicting LEO satellite trajectories, featuring a Global-Local Probsparse Self-Attention mechanism that captures both global and local contexts to handle non-stationary orbit changes, improving prediction accuracy. (iii) We conduct a theoretical analysis of the Global-Local Probsparse Self-Attention mechanism, revealing its computational complexity to be O(L log(L)), highlighting its scalability and practical application of the mechanism in tasks that involve long sequences. (iv) We conducted extensive experiments on our crawled real-world satellite (STARLINK-30506) orbit dataset, demonstrating our proposed methods significantly improve the accuracy of LEO satellite orbit prediction, consistently outperforming the counterparts on MGDE (Mean Geocentric Distance Error) by reducing prediction errors by up to 60.2% among various practical scenarios.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>II. RELATED WORK</head></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>A. ML-based Satellites Orbit Prediction</head><p>Over the past several years, ML-based approaches have been recognized as valuable tools to assist conventional physics-based approaches for satellite orbit prediction. These methods can be categorized into two distinct types: (1) Physics-Enhancing approaches. The work presented in <ref type="bibr">[9]</ref> introduces a differentiable SGP4 model called &#948;SGP4, which employs the neural networks as the solver for the motion-related differential equations within the conventional SGP4 model. Similarly, the study in <ref type="bibr">[10]</ref> proposes a physics-informed machine learning algorithm that conceptualizes the orbital trajectory as partial differential equations and utilizes deep neural operators to capture the system's dynamics. (2) Error-Correction approaches. In <ref type="bibr">[11]</ref>, <ref type="bibr">[12]</ref>, ensemble learning methods based on BT (Boosting Tree) algorithms are adopted to identify patterns in the prediction errors of physics-based models, and error corrections are then applied to the physics-based orbit predictions. In <ref type="bibr">[13]</ref>, <ref type="bibr">[14]</ref>, the authors employ an SVM (Support Vector Machine) as the error correction model to learn the SGP4-based prediction errors and then refine the SGP4-based orbit prediction based on an error correction frame.</p><p>While these ML-based methods significantly contribute to orbit prediction, they rely on simulated satellite orbit datasets generated by orbital numerical simulators or simplified perturbation models. These datasets lack the full dynamic complexity of actual satellite orbits, making their errors more pronounced in the increasingly crowded LEO space with high launch rates and abundant space debris. Considering the inherent limitations of simulated datasets, one can foresee the margin of their orbit prediction error will increase progressively. Therefore, it becomes imperative to incorporate real satellite orbit tracking data into the study of LEO satellite orbit prediction to strengthen the safety and sustainability of LEO space activities.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>B. Transformer-based LSTF Models</head><p>In recent years, inspired by the effectiveness of transformers in CV (Computer Vision tasks) <ref type="bibr">[22]</ref> and NLP (Natural Language Processing <ref type="bibr">[23]</ref>, there has been increasing interest in exploring transformer-based models for time-series forecasting tasks. In Informer <ref type="bibr">[20]</ref>, the authors propose an innovative extension of the conventional Transformer architecture by incorporating the probabilistic sparsity function into the conventional self-attention module, termed the ProbSparse Self-Attention mechanism. This design aims to streamline the complexity inherent in standard self-attention and crossattention mechanisms. In FEDformer <ref type="bibr">[21]</ref>, the authors opt for applying the Transformer in the frequency domain rather than the time domain. This strategic shift allows FEDformer to better capture the global properties of time series data. The authors of ETSformer <ref type="bibr">[18]</ref> introduce a novel approach to timeseries Transformer architecture by replacing the vanilla selfattention module with exponential smoothing attention and frequency attention modules, aiming to diminish time complexity. The work Informer(stack) <ref type="bibr">[19]</ref> presents a multi-stack informer architecture within its encoder blocks to broaden the model's perspective. The initial stack processes the entire input sequence. Subsequent stacks then focus on progressively smaller halves, with the second stack handling the right half and the third stack processing the left half of the remaining sequence. The recent work <ref type="bibr">[17]</ref> introduces the Fourier-Mixed Window Attention approach, named FWin-Transformer. This method segments sequences into shorter spans, applying window attention layers to these segments, followed by a Fourier layer to merge all the window attention outputs. The intent is to enhance the model's understanding of micro-level details within the input sequence.</p><p>While models like <ref type="bibr">[18]</ref>, <ref type="bibr">[20]</ref> excel at capturing global context, their reliance on self-attention mechanisms that operate solely on the entire input sequence can overlook local contexts, fine-grained details, and short-term dependencies. This limitation may lead to inefficiencies when processing long sequences. Conversely, <ref type="bibr">[17]</ref> emphasizes local spans with a local-only window approach, but this may limit the comprehensive capture of the entire input context. Additionally, <ref type="bibr">[19]</ref>'s multi-stack architecture aims to balance local and global contexts but inherently overlooks the left half of the input sequence, potentially biasing self-attention modules and affecting forecasting performance.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>III. DATASETS</head><p>In this section, we first introduce N2YO-crawler. We then outline the acquisition process for the Real-time satellite orbit tracking dataset (D Real) , TLEs dataset (D T LE ), and the SGP4-based orbit prediction dataset (D SGP 4 ).</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>A. N2YO-crawler</head><p>To dynamically interact with N2YO.com, we use Selenium as our automated web browser. We then develop a tailored URL generator for our N2YO-crawler, which accepts Satellite NORAD IDs and maps them to corresponding real-time orbit tracking URLs, eliminating the need for manual URL inputs. we reconfigure N2YO-crawler's web driver to extract data from the dynamic "tabledata" element on N2YO.com, ensuring focused data acquisition. To avoid delays and missed updates, we optimize the web driver to retrieve real-time satellite tracking data without page refreshes. To prevent data loss, we introduced a csv writer function that transfers data from RAM to the local hard drive every second, ensuring data persistence even during large-scale crawls.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>B. Real-time Satellite Orbit Tracking Dataset Collection</head><p>To construct our real orbit tracking dataset D Real , we first utilize N2YO-crawler to gather data from N2YO.com over an 11-day period, from December</p><p>8, 2023, at 17:01:06 UTC to December 19, 2023, at 14:16:16 UTC, resulting in 940,511 data points. It is also important to notice that during this period, our target satellite, STARLINK-30506, executed two maneuvers: one on December 16, 2023, at 18:02:09 UTC and another on December 18, 2023, at 05:13:32 UTC. Each collected data point initially comprised 14 features, including NORAD ID, Local Time, UTC, latitude, longitude, Altitude [Km], Altitude [Mi], Speed [Km/s], Speed [Mi/s], Azimuth, Elevation, Right Ascension, Declination, and Local Sidereal Time. Noting that the values of the last five features are contingent on the data collector's IP location, and to concentrate on the satellite trajectory prediction task, we trimmed our dataset to include only five essential features: UTC, Altitude [Mi], latitude, longitude, and Altitude [Km]. This data trimming was performed to ensure that the data set is specifically relevant to satellite trajectories and is independent of the geographical location of the data collector.</p><p>To further facilitate the learning process for the LSTF, we augmented each record with four additional attributes: sintime, costime, index, and sinindex. According to <ref type="bibr">[24]</ref>, the inclusion of these features can provide a consistent distance measure for time-related variables, capturing periodicity and removing discontinuities, thus potentially aiding the model in learning time-based patterns. Precisely, the values of sintime and costime are computed by applying the sine function to the corresponding date, represented as the number of seconds since January 1, 1970, 00:00:00 UTC. The index is a sequential identifier, commencing with 1 for the first record. The sinindex is derived by applying the sine function to the index.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>C. Other Datasets Collection</head><p>Following the methods used in <ref type="bibr">[14]</ref>, we build the TLEs dataset (D T LE ) by collecting TLEs from the Space-Track.com website throughout December 8, 2023, to December 19, 2023, collecting 30 TLEs. We then constructed an SGP4based orbit prediction dataset (D SGP 4 ) through a two-step process. First, we generate orbital position predictions using the SGP4 model <ref type="bibr">[25]</ref> with the most recent TLEs from D T LE , ensuring each prediction utilized the most current TLE for its corresponding timeframe. Second, since SGP4 outputs are in the TEME (True Equator Mean Equinox) frame <ref type="bibr">[26]</ref>, which differs from N2YO.com's geodetic coordinates, we use Astropy to convert TEME values to geodetic coordinates.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>IV. PROPOSED METHOD: GLOLOSAT</head><p>This section defines the task, details the GloLoSAT's three stages, analyzes the computational complexity of the Global-Local Probsparse Self-Attention Mechanism, and describes architecture variants of the error correction model.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>A. Task Definition</head><p>In this paper, the real orbit tracking dataset (D Real ), TLE dataset (D T LE ), and SGP4-based orbit prediction dataset (D SGP 4 ) for the satellite STARLINK-30506 are used. It is imperative to notice that the original trajectory values (such as altitude, longitude, and latitude) in D Real exhibit near periodic behavior, presenting challenges for ML models because the statistical properties of these series, such as mean or variance, change over time. In order to address this learning challenge, instead of directly learning the satellite trajectories, our proposed method attempts to learn the error between the trajectories predicted by the SGP4 model and the actual trajectories provided by N2YO. This error turns out to be less periodic but more stochastic and, hence, more learnable by our ML model. Hence, our primary objective is to improve the accuracy of predicting the aforementioned errors, thereby enhancing the accuracy of orbit predictions.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>B. Data Preprocessing Stage</head><p>We first introduce the notations. The datapoint in D Real at a given time t i is denoted as GEO X Real (t i ), where GEO means the last three features of the datapoint are in the geodetic coordinate frame. The datapoint in D SGP 4 at time t i is denoted as GEO X SGP 4 (t i ). We define D e as the SGP4 prediction error dataset, which incorporates the positional discrepancies between D Real and D SGP4 . Similarly, the datapoint in D e at time t i is denoted as GEO X e (t i ).</p><p>The GloLoSAT begins with a data preprocessing stage. Initially, it takes the datasets D Real and D SGP4 as input , and subsequently constructs the SGP4 prediction error dataset D e by calculating the positional discrepancies between D Real and D SGP4 . Specifically, the computation is detailed by the following equation:</p><p>where T is the set of timestamps for which data points exist in both D Real and D SGP4 . It is essential to recognize that both D Real and D SGP 4 encompass three geodetic attributes: latitude, longitude, and altitude <ref type="bibr">[Km]</ref>. These common attributes are leveraged to evaluate the positional discrepancies, thereby constituting the error dataset D e . In particular, this error dataset D e retains the structural integrity of D Real , substituting the values of the three geodetic features with the respective positional differences, while the values of the remaining features remain consistent with those in D Real . Following the acquisition of the error dataset, D e is partitioned into distinct subsets to support the subsequent stages. Specifically, the initial 75% of D e is allocated for the training dataset, 10% is reserved for the validation dataset, and the remaining 15% is dedicated to the evaluation of orbit prediction accuracy.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>C. Error Correction Model Training Stage</head><p>We introduce a multi-range self-attention mechanism for the error correction model in the GloLoSAT framework, trained on dataset D e to characterize SGP4-based prediction errors. This mechanism enables the model to provide a multi-scope perspective on the input sequence. Specifically, our proposed model is inspired by Informer. We modify Informer's encoder block's Probsparse self-attention module, replacing it with our Global-Local Probsparse self-attention modules. A. Global-Local Probsparse Self-Attention Mechanism.</p><p>Our proposed Global-Local Probsparse self-attention mechanism, as depicted on the right-hand side in Fig. <ref type="figure">1</ref>, enables the model to have a multi-scale perspective on input sequence processing by capturing both global and local contexts, surpassing the capabilities of its predecessor, which primarily focused on either global or local information. Specifically, the global Probsparse self-attention module is configured to capture the entire sequence input, while the local Probsparse self-attention modules are intended to concentrate on distinct segments of the sequence input. In the following, we will define some notations and present a formal modeling of the Global-Local Probsparse self-attention mechanism, along with an analysis of its computational complexity.</p><p>Notations: The time sequence starting at t and concluding at t+&#8710; is represented by the notation (t : t+&#8710;). The sequence input for the encoder is denoted as X(t : t + &#8710;) &#8712; R LIn&#215;dem , where L In = &#8710; represents the input sequence length, and d em denotes the embedded dimension of the model. The queries (Q), keys(K) , and values(V ) for the input sequence are then defined as</p><p>Global Probsparse Self-attention: The global Probsparse self-attention mechanism is designed to capture the entire sequence input. Utilizing the ProbSparse Self-attention formula in <ref type="bibr">[20]</ref>, the calculation of it is defined as:</p><p>where Q g , K g , V g &#8712; R LIn&#215;dem represent the queries, keys, and values for the entire sequence input, and Qg &#8712; R u&#215;dem represents a filtered sparse matrix comprising the top u Q g s selected based on the sparsity measurement formula <ref type="bibr">[20]</ref>. The selection parameter u is defined by the equation u = c&#8226;ln L In , where c is a constant sampling factor. The sparsity measurement formula is given by:</p><p>where the q i , k j represent the ith and jth row in Q, K respectively.</p><p>Local Probsparse Self-attention: Since our local Probsparse self-attention modules are tailored to focus on distinct segments of the sequence input, we divide the entire input X(t : t + &#8710;) into N subsequences: X(t : t + &#8710; l ), X(t + &#8710; l t + 2&#8710; l ),..., X(t + (n -1)&#8710; l : t + n&#8710; l ), where &#8710; l stands for each local subsequence input length. Consequently, the queries, keys, and values are also segmented accordingly:</p><p>.., V lN ] &#166; . Thus, by using the ProbSparse Selfattention formula Eq.( <ref type="formula">3</ref>), we compute local Probsparse selfattention for each subsequence as follows:</p><p>where Q li , K li , V li &#8712; R L In/N &#215;dem signifying the queries, keys, and values for the ith subsequence input, and the filtered sparse matrix Qli &#8712; R m&#215;dem contains the top m Q li s selected based on the sparsity measurement formula Eq.( <ref type="formula">3</ref>). The parameter m is controlled by the relation m = c &#8226; ln L N . Global-Local Probsparse Self-Attention: After computing the local Probsparse self-attention and the global Probsparse self-attention for the input sequence, we concatenate these attentions to form the Global-Local Probsparse Self-Attention:</p><p>Computational Complexity: Given a sequence input X(t : t + &#8710;) &#8712; R L&#215;dem , the computational complexity of Global-Local Probsparse Self-Attention is as follows:</p><p>Proof: for the computational complexity of Global-Local Probsparse Self-Attention</p><p>By applying the time complexity formula of ProbSparse Attention <ref type="bibr">[20]</ref> to Eq.( <ref type="formula">7</ref>), we obtain:</p><p>RHS of Eq.( <ref type="formula">10</ref>)</p><p>It is clear that Informer's Probsparse Self-Attention mechanism exhibits computational complexity on the order of O(L log(L)), which aligns with our proposed Global-Local Probsparse Self-Attention mechanism. This indicates that our method can achieve processing efficiency similar to that of Informer while also enhancing the model's ability to capture global and local context.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>D. Architecture Variants</head><p>For the error correction models, we build two variants of GloLoSAT, denoted as GloLoSAT-OF (Orbital Features Focused) and GloLoSAT-TL (Triple Independent). The GloLoSAT-OF variant is achieved by modifying the final FCN (fully connected layer) of the original GloLoSAT so that it only outputs orbit errors under three essential orbital features: latitude, longitude, and altitude <ref type="bibr">[Km]</ref>. In contrast, the GloLoSAT-TL variant adopts a more fine-grained approach. Specifically, it is composed of three separate models. Each model diverges from the conventional methods that predict all features concurrently. Instead, their final FCN layer is modified to predict orbit errors under each orbital feature independently. As a result, this approach yields three specialized models, each uniquely responsible for predicting latitude errors, longitude errors, and altitude errors[Km] respectively. These individual predictions are then synthesized to generate the final satellite orbit error corrections. Following a similar modification strategy, we also introduce two variants for Informer <ref type="bibr">[20]</ref> and Informer(stack) <ref type="bibr">[19]</ref>, respectively: Informer-OF, Informer-TL, Informer(stack)-OF, and Informer(stack)-TL.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>E. Orbit Prediction Stage</head><p>As illustrated on the left-hand side in Fig. <ref type="figure">1</ref>, the final stage generates the final satellite orbit prediction by integrating the error corrections with the corresponding SGP4 predictions. Specifically, this process first leverages the trained error model to process an input sequence GEO X e (t i : t i + &#8710; 1 ), where &#8710; 1 represents the length of the input sequence and obtains the predicted errors GEO X e (t i + &#8710; 1 : t i + &#8710; 1 + &#8710; 2 ), where &#8710; 2 indicating the size of the prediction window. The resulting orbit prediction is then calculated by:</p><p>TABLE I THE PERFORMANCE RESULTS ON THE SHORT-TERM PREDICTION TASK. &#8225; MEANS OUR PROPOSED MODEL. Short-Term Prediction (ECEF, unit in meters) Metric MAE RMSE MGDE SGP4 321.989 417.923 637.902 LSTM + SGP4 324.264 422.208 642.029 Informer + SGP4 535.77 648.83 1039.76 Informer(stack) + SGP4 545.47 658.43 1056.71 &#8225; Informer-OF + SGP4 295.924 403.545 583.802 &#8225; Informer(stack)-OF + SGP4 303.149 420.524 595.315 &#8225; GloLoSAT-OF + SGP4 273.481 526.086 540.527 &#8225; Informer-TL + SGP4 260.285 355.646 510.162 &#8225; Informer(stack)-TL + SGP4 264.407 365.838 520.202 &#8225; GloLoSAT-TL + SGP4 253.732 346.548 495.715 TABLE II THE PERFORMANCE RESULTS ON THE LONG-TERM PREDICTION TASK. &#8225; MEANS OUR PROPOSED MODEL. Long-term Prediction (ECEF, unit in meters) Metric MAE RMSE MGDE SGP4 321.989 417.923 637.902 LSTM + SGP4 326.446 425.655 644.288 Informer + SGP4 639.434 751.176 1228.982 Informer(stack) + SGP4 578.423 690.78 1117.3 &#8225; Informer-OF + SGP4 283.445 386.59 557.269 &#8225; Informer(stack)-OF + SGP4 308.734 411.637 610.936 &#8225; GloLoSAT-OF + SGP4 265.856 385.471 522.81 &#8225; Informer-TL + SGP4 253.576 346.116 495.361 &#8225; Informer(stack)-TL + SGP4 256.685 373.633 503.23 &#8225; GloLoSAT-TL + SGP4 250.898 341.99 489.712 improve the accuracy of satellite orbit prediction under various scenarios compared with the baselines. Our proposed models outperform the SGP4 on MGDE by decreasing the errors up to 22.3% (in Table I: Short-Term Prediction), 23.2% (in Table II: Long-term Prediction), and 18.9% (in Table III: Extended Long-Term Prediction). This implies that the positional discrepancies between D SGP 4 and D REAL are learnable, and our proposed methods make a positive contribution to the orbit prediction problems. Additionally, when compared to traditional multivariate time-series forecasting models combined with SGP4, such as LSTM + SGP4, Informer + SGP4, and Informer(stack) + SGP4, our proposed models show a potential reduction in MGDE by up to 53.1%, 60.2%, and 58.9% for short-term, long-term, and extended long-term predictions, respectively. This observation underscores the importance of the output feature dimensionality in the final FCN layer for time-series forecasting tasks. Specifically, the original LSTM, Informer, and Informer(stack) models generate outputs for an extensive range of features, which may lead to a dispersion of focus on many task-irrelevant features, such as sintime, costime, index, sinindex, and Altitude [Mi]. Conversely, our proposed models are tailored to prioritize key orbital features: latitude, longitude, and altitude [Km]. By concentrating on these orbital features, our models can minimize the dilution of attention across less relevant features, thereby leading to more accurate predictions. Performance Comparison Between Different Self-attention Mechanisms. The results across tables I through III show</p><p>TABLE III THE PERFORMANCE RESULTS ON THE EXTENDED LONG-TERM PREDICTION TASK. &#8225; MEANS OUR PROPOSED MODEL. Extended Long-Term Prediction(ECEF, unit in meters) Metric MAE RMSE MGDE SGP4 320.575 413.403 634.079 LSTM + SGP4 642.56 806.622 1251.47 Informer + SGP4 543.92 692.17 1057.39 Informer(stack) + SGP4 632.319 788.58 1207.04 &#8225; Informer-OF + SGP4 302.785 405.393 596.42 &#8225; Informer(stack)-OF + SGP4 309.561 411.01 606.83 &#8225; GloLoSAT-OF + SGP4 297.427 398.955 585.484 &#8225; Informer-TL + SGP4 268.486 363.218 526.142 &#8225; Informer(stack)-TL + SGP4 276.9 372.576 543.207 &#8225; GloLoSAT-TL + SGP4 262.916 355.085 514.392 that our Global-Local Probsparse Self-Attention-based orbit prediction methods outperform their counterparts in all tested prediction scenarios. The GloLoSAT-TL + SGP4 method exhibits improvements over the Informer-TL + SGP4 and Informer(stack)-TL + SGP4 methods, with a decrease in MGDE by up to 4.7%, 2.7%, and 5.3% for Short-Term, Long-Term, and Extended Long-Term Prediction scenarios, respectively, as shown in Tables I to III. Similarly, the GloLoSAT-OF + SGP4 method exhibits better performance compared to the Informer-OF + SGP4 and Informer(stack)-OF + SGP4 methods, achieving a reduction in MGDE by up to 9.2%, 15.8%, and 3.5% for short-term, long-term, and extended longterm predictions, respectively. These improvements underscore the essential role of the input sequence range in enhancing the self-attention module's efficacy. The Informer model processes only the input sequence's global range, potentially limiting its local contextual understanding. The Informer(stack) utilizes a tiered approach, where the first stack processes the global input sequence, the second stack addresses the right half slices of the input sequence, and the third stack focuses on the subsequent half slices, potentially overlooking the left half slices of the entire input. On the contrary, our proposed GloLoSAT, with its Global-Local Probsparse Self-Attention modules, is designed to fully encompass both global and local sequence inputs. Specifically, the global module captures the entire sequence, while the three local modules focus on the left, middle, and right thirds of the input sequence, leading to a more nuanced contextual understanding and improved orbit prediction accuracy. Performance Comparison Between Architectural Variants. The results presented in Tables I through III demonstrate that Triple Independent variants perform better than Orbital Features Focused counterparts in various orbit prediction scenarios. Informer-TL + SGP4 outperforms Informer-OF + SGP4 on MGDE by decreasing 12.6% for Short-Term Prediction, 11.1% for Long-Term Prediction, and 11.8% for Extended Long-Term Prediction. Informer(stack)-TL + SGP4 achieves a decrease in MGDE by 12.6% for Short-Term Prediction, 17.6% for Long-Term Prediction, and 10.5% for Extended Long-Term Prediction when compared to Informer(stack)-OF + SGP4. Additionally, GloLoSAT-TL + SGP4 also shows improved performance over GloLoSAT-OF + SGP4, with MGDE reductions of 8.3% for Short-Term Prediction, 6.3% for Long-Term Prediction, and 12.1% for Extended Long-Term</p><p>Prediction. These observations demonstrate the benefits of a more fine-grained approach adopted by the Triple Independent variants, where the final FCN layer is restructured to predict orbit errors for each orbital feature independently, thereby maximizing the use of additional information from other input features to enhance prediction accuracy.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>VI. CONCLUSION AND FUTURE WORK</head><p>In this paper, we attempt to address LEO satellite orbit prediction using real satellite orbit tracking data. Our approach primarily concentrates on mitigating the orbit prediction errors that arise from traditional physics-based models by implementing an ML-based error correction framework. We introduce N2YO-crawler for high-integrity, real-time data collection and present GloLoSAT, a novel three-stage transformerbased framework for real-time LEO satellite orbit prediction. To overcome the limitations of existing methods that focus only on global or local contexts, we propose a Global-Local Probsparse Self-Attention mechanism within the GloLoSAT framework, capturing both contexts with computational efficiency of O(L log(L)). Extensive experiments demonstrate GloLoSAT's effectiveness in reducing SGP4 prediction errors and improving orbit prediction performance. We note that this study focuses on applying the ML-based approach to the orbit prediction of a single satellite. In our future work, we plan to extend the proposed ML-model to the orbit prediction of multiple satellites in the same constellation, which will enable the routing research in the mobile ad hoc network formed by the constellation. ACKNOWLEDGMENT</p></div><note xmlns="http://www.tei-c.org/ns/1.0" place="foot" xml:id="foot_0"><p>Authorized licensed use limited to: Auburn University. Downloaded on June 16,2025 at 17:55:28 UTC from IEEE Xplore. Restrictions apply.</p></note>
		</body>
		</text>
</TEI>
