Earth rotation parameter prediction method and system based on dynamic tracking and segmentation model

Abstract

The application belongs to the technical field of polar motion prediction, and provides a method and system for predicting earth rotation parameters based on dynamic tracking and a segmented model. A plurality of preset prediction models are used to respectively predict historical earth rotation parameters, historical effective angular momentum and angular momentum prediction values in different segments to obtain prediction results. The plurality of prediction models are regression prediction models with different parameters. In the regression prediction model, the frequency band boundary of the different frequency band cut-off periods is analyzed based on the historical residual power spectral density to dynamically track and lock the frequency band boundary. The autoregressive order and input feature space are dynamically trimmed for different lead days. For the prediction interval intersection boundary of adjacent prediction models, the prediction value is determined by weighted smoothing stitching. The problem of being unable to follow the drift of the residual spectrum peak in the sliding window to adjust is solved, the step or discontinuity phenomenon of the prediction results between different prediction models is avoided, the divergence speed of the synthesis error is significantly reduced, and the overall prediction accuracy is improved.

Description

Technical Field

[0001] This invention belongs to the field of polar motion prediction technology, and in particular relates to a method and system for predicting Earth rotation parameters based on dynamic tracking and a segmented model. Background Technology

[0002] Earth Rotation Parameters (ERPs) are core physical quantities describing changes in the Earth's rotational state. Existing polar motion prediction methods, when incorporating external excitation information such as Effective Angular Momentum (EAM), generally employ a fixed-parameter modeling strategy across the entire timeframe, failing to dynamically adjust the model structure based on differences in the number of days leading the forecast. To further enhance physical interpretability, fusing multi-source external excitation information with advanced data-driven models has become an effective way to improve the accuracy of polar motion predictions.

[0003] However, existing fusion schemes mostly use fixed-parameter modeling throughout the entire time period. Short-term predictions rely on high-order delay terms to characterize high-frequency jitter, while long-term predictions require low-order terms to prevent overfitting and divergence. The model structure cannot be dynamically adjusted according to the number of days ahead, affecting prediction accuracy. Specifically, the frequency band boundaries of the filter bank or wavelet decomposition remain unchanged in the forecast process and cannot be adjusted with the drift of the residual spectral peak within the sliding window, which easily introduces spurious periods or phase delays. For sub-models with different numbers of days ahead, step or discontinuity in prediction results often appear at their jurisdictional boundaries. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes a method and system for predicting Earth's rotation parameters based on dynamic tracking and a segmented model. This invention integrates dynamic spectral density tracking, multi-segment dynamic feature fusion, and boundary stitching smoothing, enabling the full extraction of high-frequency exogenous variable feedback value in the short term and effectively suppressing divergence in the long term through adaptive decay control and baseline constraints. This solves the problem of not being able to adjust according to the drift of the residual spectral peak within the sliding window, avoiding step jumps or discontinuities in prediction results between different prediction models. Over longer time spans, the divergence rate of the synthesis error is significantly reduced, improving the overall prediction accuracy.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: In a first aspect, the present invention provides a method for predicting Earth rotation parameters based on dynamic tracking and a segmented model, comprising: Obtain historical Earth rotation parameters, historical effective angular momentum, and predicted angular momentum values; Historical Earth rotation parameters, historical effective angular momentum, and predicted angular momentum values ​​are segmented according to different numbers of days ahead. Using multiple pre-set prediction models, the historical Earth rotation parameters, historical effective angular momentum, and predicted angular momentum values ​​in different segments are predicted to obtain the prediction results. Among them, multiple prediction models are regression prediction models with different parameters; in the regression prediction models, the frequency band boundaries of different frequency band cutoff periods are dynamically tracked and locked based on historical residual power spectral density analysis; the autoregression order and input feature space are dynamically trimmed for different lead days; and the predicted value is determined by weighted smoothing stitching for the cross boundaries of the forecast intervals of adjacent prediction models.

[0006] Furthermore, the historical effective angular momentum and the predicted angular momentum values ​​are preprocessed before being input into the prediction model: all externally predicted exogenous variables are translated and aligned online using a preset dynamic bias correction vector to determine the aligned fluid angular momentum sequence. The construction of the dynamic bias correction vector includes: extracting a historical sliding window in the current element initiated by the rolling forecast, determining the mean difference between the external forecast and the measured values ​​within the window, and constructing the dynamic bias correction vector.

[0007] Furthermore, the dynamic bias correction vector for: ; in: The length of the historical sliding window; For time index variables; The simplified Julian day count value corresponding to the current element for which the rolling forecast was initiated; for The exogenous variable values ​​of the external forecast at that time; for Locally measured historical values ​​of exogenous variables at any given time.

[0008] Furthermore, the aligned fluid angular momentum sequence is transformed into a theoretical change in Earth's rotation parameters: a continuous state transition model based on Chandler oscillation period and quality factor is established. ; in, State vector The derivative; The aligned fluid angular momentum excitation function; system matrix and control matrix for: ; ; in, This is the Chandler natural angular frequency. The damping ratio is given; the system matrix is... and control matrix Discretize into state transition matrix With the gain matrix : ; ; in, This represents the actual time interval between two adjacent data epochs. These are dummy variables for integration; from this, the fluid-induced sequence of Earth's rotation parameters can be derived.

[0009] Furthermore, the dynamic tracking based on historical residual power spectral density analysis includes: estimating the power spectral density of the historical residual sequence using the Welch method to obtain the frequency response spectrum; and dynamically searching for the period corresponding to the maximum energy spectral peak within the search sub-intervals corresponding to different frequency bands. ; in, Candidate period; This represents the power spectral density at the corresponding frequency. Set the value for the preset cycle; This is the dominant cycle with the highest energy within the interval.

[0010] Furthermore, the regression prediction model is a vector autoregressive model with exogenous variables in the baseline, utilizing historical measured sequences. With fluid excitation sequence , build with Regularized regressor, regressor measurement prediction equation for: ; in, , These are the autoregression order and the lag order of the exogenous variable, respectively. , This is the corresponding weight matrix; , It is a lagged term; Harmonic terms; Constant bias; residual sequence for: ; in, These are the single-step predictions from the baseline model.

[0011] Furthermore, the prediction model includes a first prediction model, a second prediction model, and a third prediction model, with the number of days leading increasing sequentially. The first prediction model is responsible for short-term fine-tuning, using a higher order, and the regression input includes the original term, the residual three-subband term, exogenous variables, and harmonic components. The second prediction model is responsible for medium-term smooth transition, using a medium order. The third prediction model is responsible for long-term trend extrapolation, with the regression mode switched to a purely autonomous extrapolation mode, high-order disturbances of exogenous variables turned off, and the order reduced to the minimum.

[0012] Furthermore, ahead of schedule The virtual residual feedback value of the step input is subject to the following damping constraints: ; ; in, For the first Virtual residual feedback after step damping; This is the current predicted value; This is the baseline reference value; The basic damping coefficient; This represents the attenuation rate.

[0013] Furthermore, the weighted smooth stitching to determine the predicted value includes: in the overlapping region Inside, final predicted value Linear weighting of the predictions from the left and right models: ; ; in, Index for days ahead; , These are the predicted values ​​for the left and right segmented models; and For weights.

[0014] Secondly, the present invention also provides a system for predicting Earth rotation parameters based on a dynamic tracking and segmented model, comprising: The data acquisition module is configured to acquire historical Earth rotation parameters, historical effective angular momentum, and predicted angular momentum values. The prediction module is configured to: segment the historical Earth rotation parameters, historical effective angular momentum, and predicted angular momentum values ​​according to different lead days; and use multiple preset prediction models to predict the historical Earth rotation parameters, historical effective angular momentum, and predicted angular momentum values ​​in different segments to obtain prediction results. Among them, multiple prediction models are regression prediction models with different parameters; in the regression prediction models, the frequency band boundaries of different frequency band cutoff periods are dynamically tracked and locked based on historical residual power spectral density analysis; the autoregression order and input feature space are dynamically trimmed for different lead days; and the predicted value is determined by weighted smoothing stitching for the cross boundaries of the forecast intervals of adjacent prediction models.

[0015] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the Earth rotation parameter prediction method based on dynamic tracking and segmentation model described in the first aspect.

[0016] Fourthly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the steps of the Earth rotation parameter prediction method based on dynamic tracking and segmentation model described in the first aspect.

[0017] Fifthly, the present invention also provides a computer program product, the computer program product comprising a computer program, which, when executed by a processor, implements the steps of the method for predicting Earth rotation parameters based on a dynamic tracking and segmentation model as described in the first aspect.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention utilizes multiple pre-set prediction models to predict historical Earth rotation parameters, historical effective angular momentum, and predicted angular momentum values ​​within different segments, yielding prediction results. These multiple prediction models are regression prediction models with different parameters. Within the regression prediction models, frequency band boundaries for different frequency band cutoff periods are dynamically tracked and locked based on historical residual power spectral density analysis. The autoregressive order and input feature space are dynamically trimmed for different lead days. For the intersection boundaries of prediction intervals between adjacent prediction models, weighted smoothing is used to determine the predicted value. By integrating dynamic spectral density tracking, multi-segment dynamic feature fusion, and boundary smoothing, the invention fully extracts the feedback value of high-frequency exogenous variables in the short term, and effectively suppresses divergence in the long term through adaptive attenuation control and baseline constraints. This solves the problem of not being able to adjust according to the drift of residual spectral peaks within the sliding window, avoiding step jumps or discontinuities in prediction results between different prediction models. Over longer time spans, the divergence rate of the synthesis error is significantly reduced, improving overall prediction accuracy. Attached Figure Description

[0019] The accompanying drawings, which form part of this embodiment, are used to provide a further understanding of this embodiment. The illustrative embodiments and their descriptions are used to explain this embodiment and do not constitute an improper limitation of this embodiment.

[0020] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0023] Example 1: Earth's rotation parameters are core physical quantities describing changes in the Earth's rotational state, primarily including polar motion and diurnal variation. Polar motion reflects the periodic oscillation of the instantaneous rotational axis relative to the Earth's crust, while diurnal variation reflects the non-uniformity of the Earth's rotation rate. These parameters arise from complex processes such as angular momentum exchange between the atmosphere, oceans, and terrestrial water reserves, as well as core-mantle coupling and solid tides. High-precision monitoring and forecasting of these parameters are crucial for maintaining the celestial reference frame, deep space navigation, and precise satellite orbit determination.

[0024] Existing polar motion forecasting methods generally employ a fixed-parameter modeling strategy throughout the entire forecast period when incorporating external excitation information such as effective angular momentum (EAM), failing to dynamically adjust the model structure based on differences in the number of days leading the forecast. Short-term forecasts rely on higher-order delay terms to characterize high-frequency jitter, while long-term forecasts require lower-order terms to prevent overfitting and divergence. The fixed-parameter architecture struggles to simultaneously meet the accuracy requirements of different forecast spans. Furthermore, existing signal decomposition methods maintain fixed frequency band boundaries throughout the forecast process, failing to track the time-varying drift of residual spectral peaks; multiple forecast sub-models are prone to step discontinuities at their boundaries; and the initial biases of external effective angular momentum forecast sequences are not adequately corrected online.

[0025] Research on polar motion prediction has evolved from traditional time series analysis to data-driven modeling. Early works often employed autoregressive moving averages and least squares combined autoregressive models. More recently, statistical methods such as vector autoregression and dynamic mode decomposition have been developed, which can capture the linear evolution characteristics of polar motion sequences to some extent, but error accumulation over medium- to long-term forecast periods remains significant. To overcome the limitations of linear extrapolation, neural networks and hybrid models have been extensively introduced. Existing techniques attempt to combine signal decomposition or least squares fitting with nonlinear networks, such as replacing traditional AR residual predictors with LSTM networks to construct LS+LSTM, CSLS+LSTM, and other frameworks, improving medium- to long-term residual modeling capabilities. Building upon this, architectures such as BiLSTM, CNN-LSTM-attention, and LSVMD-Informer further enhance time series feature extraction and improve forecast accuracy.

[0026] To further enhance physical interpretability, effective angular momentum data was incorporated into the forecast framework. Introducing atmospheric, oceanic, and hydrological angular momentum functions improved 1–30 day forecast accuracy by up to 38.4%. Subsequent work developed various fusion strategies, including genetic algorithm-optimized neural networks, CSLS hybrid models, and ConvLSTM, and corrected for systematic errors in the EAM forecast data itself, effectively enhancing the forecast performance of polar motion and diurnal variation. The conclusions of the Second Earth Orientation Parameter Forecast Comparison Activity also confirmed that assimilating EAM forecast information is a key strategy for improving polar motion forecast accuracy. Furthermore, exploration of denoising and preprocessing EAM data to enhance its predictive utility is ongoing. These advances indicate that fusing multi-source external excitation information with advanced data-driven models has become an effective way to further improve polar motion forecast accuracy.

[0027] However, as described in the background section, existing fusion schemes mostly employ fixed-parameter modeling throughout the entire forecast period. Short-term predictions rely on high-order delay terms to characterize high-frequency jitter, while long-term predictions require low-order terms to prevent overfitting and divergence. The model structure cannot be dynamically adjusted according to the number of days ahead. Furthermore, the frequency band boundaries of signal decomposition remain unchanged throughout the forecast process, making it difficult to track the time-varying drift of the residual spectral peaks. Step shifts easily occur at the boundaries between different forecast sub-models, and the initial biases of external EAM forecast sequences are not adequately corrected online. Specifically, the frequency band boundaries of filter banks or wavelet decomposition remain unchanged throughout the forecast process, failing to adjust to the drift of residual spectral peaks within the sliding window, easily introducing spurious periods or phase delays. For sub-models with different numbers of days ahead, step shifts or discontinuities in prediction results often occur at their respective jurisdictional boundaries. Externally input EAM forecast sequences are not dynamically aligned, and initial biases directly enter the numerical integration process, accelerating error divergence.

[0028] To solve at least one of the above problems, such as Figure 1 As shown, this embodiment provides a method for predicting Earth rotation parameters based on dynamic tracking and a segmented model. It integrates adaptive spectral density tracking, multi-segment dynamic feature fusion, and a time-varying segmented regression rolling forecast framework with smoothed boundary stitching. Experiments conducted on actual ERP / EAM datasets with a 180-day lead time and 70-period rolling forecast demonstrate that the method can fully extract the feedback value of high-frequency exogenous variables in the short term, and effectively suppress divergence in the long term through adaptive decay control and baseline constraints. Over a span of 90 to 180 days, the divergence rate of the synthesis error is significantly reduced, and the overall prediction accuracy is greatly improved. The specific content and explanation of the method in this embodiment are as follows: S1. Initial alignment of the excitation response with the measured reference: Due to differences in computational models, exogenous variable forecast files exhibit systematic biases compared to historical observed sequences. The current meta-count (Modified Julian Date, MJD) at the time of the rolling forecast initiation is... ), cut length is Using a historical sliding window, determine the mean difference between the forecasts inside and outside the window and the local measured values, and construct a dynamic bias correction vector: ; in, This is the dynamic bias correction vector; Historical sliding window length (unit: days); Use a time index variable (to iterate through each day within the history sliding window); The MJD count value corresponding to the current element initiated by the rolling forecast; for Exogenous variable values ​​from external forecasts at time (from EAM forecast products); for The exogenous variable observations are local historical measurements at each time point. The summation and averaging of these values ​​across all time points within the window yields the systematic average bias of the external forecast relative to the local measurements.

[0029] The constructed dynamic bias correction vector is used to perform online translation and alignment of all externally predicted exogenous variables during the lead time period to eliminate systematic drift and obtain the aligned fluid angular momentum sequence.

[0030] S2, Numerical integration of fluid angular momentum excitation response: The aligned fluid angular momentum sequence (excitation term) This is transformed into theoretical changes in the Earth's rotation parameters (ERP). A continuous state transition model based on the Chandler oscillation period Tc and the quality factor Q is established: ; in, State vector The derivative, It is a two-dimensional state vector consisting of polar shifts and their rates of change; The aligned fluid angular momentum excitation function; system matrix and control matrix for: ; ; in, This is the Chandler natural angular frequency. Let be the damping ratio. It is discretized into a state transition matrix using the precise exponential matrix method (expm). With the gain matrix : ; ; This leads to the deduction of the ERP fluid excitation sequence for purely fluid-induced excitation. and the response sequence for the future leading period. .

[0031] S3. Vector Autoregressive (VARX) model construction and residual extraction with baseline containing exogenous variables: Using historical measured sequences With fluid excitation sequence , build with The regularized baseline VARX regressor (Ridge Regression) has the following measurement prediction equation: ; in, for Time-based ERP prediction vector; , These are the autoregression order and the lag order of the exogenous variable, respectively. , This is the corresponding weight matrix; , It is a lagged term; For Chandler, annual, and semi-annual harmonic terms; This is a constant bias.

[0032] ; in, This is the matrix of all weight combinations; The training set is the matrix of measured values. The characteristic matrix is ​​composed of lag terms and harmonic terms; Ridge parameters (e.g.) ); This is the identity matrix. Further calculation of the single-step prediction residuals: ; in, These are the single-step predictions from the baseline model. It is a residual sequence.

[0033] S4. Residual Spectrum Adaptive Tracking and Causal Filter Bank Segmentation: To ensure that the filter bank closely matches the residual energy distribution in real time, the historical residual sequence is analyzed online during each rolling forecast. Welch method power spectral density estimation is performed to obtain the frequency response spectrum. Within the three search sub-intervals of low frequency, mid frequency, and high frequency, the period corresponding to the maximum energy spectral peak is dynamically sought. ; in, Candidate period (days); Set the value for the preset cycle; This represents the power spectral density at the corresponding frequency. This is the dominant period with the highest energy within the interval. Based on this, the low-frequency cutoff period of the current epoch is determined. Intermediate frequency cutoff cycle and high-frequency cutting cycle Then, it is substituted into a causal Butterworth filter bank to decompose the residual into low-frequency subbands in real time. Intermediate frequency subband and high-frequency subband .

[0034] S5. Multi-time-segment time-varying autoregressive modeling: To address the signal characteristics with varying lead times, taking 180 days as an example, three complementary regression sub-models are constructed: S1 (lead time 1-7 days), S2 (lead time 2-7 days), and S3 (lead time 3-7 days). S1 is responsible for fine-tuning in the short term, employing a higher order, and its regression input comprehensively includes the original terms, residual three-subband terms, exogenous variable three-subband terms, and harmonic components. S2 (lead time 8-15 days) is responsible for smooth transition in the medium term, employing a medium order, simplifying high-frequency delay features, and retaining basic subband feedback. S3 (lead time 16-180 days) is responsible for extrapolating long-term trends. To prevent overfitting due to high-frequency noise, the regression mode is switched to a purely autonomous extrapolation mode, high-order perturbations of exogenous variables are disabled, and the order is minimized. Each segment independently calculates its feature mapping weight matrix. .

[0035] S6. Closed-loop multi-step rolling calculation with adaptive decay control: In the future time period, each segmented model will undergo closed-loop recursive iteration within its jurisdiction of days. To prevent uncontrollable oscillations in the residual feedback during multi-step recursion, a geometric decay control mechanism is introduced. The virtual residual feedback value leading the input at the k-th step is subject to the following damping constraints: ; ; in, This is the virtual residual feedback after the k-th step of damping; This is the current predicted value; This is the baseline reference value; The basic damping coefficient; This represents the decay rate. As k increases, The residual feedback tends to zero, and the long-term forecast safety is attached to the baseline trend.

[0036] S7, Smooth Boundary Sealing Technique: The step transition at the intersection of piecewise models is eliminated by using a linear weight transition technique (Seam Smoothing) at the boundary of different prediction time periods. Let S1 end at... Heavens, S2 begins The transition half-bandwidth is [number] days. (e.g., 2.5 days), in the overlapping area Within this model, the final predicted value is a linear weighted average of the predictions from both the left and right models: ; ; in, Index for days ahead; , These are the predicted values ​​for the left and right segmented models; weights Incrementing linearly from 0 to 1, It decreases linearly from 1 to 0, achieving a seamless and smooth transition.

[0037] In summary, the adaptive filter bank segmentation method based on online analysis of historical residual power spectral density in this embodiment can dynamically track and lock the low-frequency cutoff period. Intermediate frequency cutoff cycle and high-frequency cutting cycle Frequency band boundaries; a multi-time-segment VARX time-varying feature regression architecture that dynamically prunes the autoregression order and input feature space for different lead days, enabling full sub-band feedback in the short term and automatically reducing the order and converting to a purely autonomous calculation mode in the long term; a closed-loop feedback iteration mechanism with geometric series damping constraints, through... Attenuation control avoids long-period error divergence; linear weighted smoothing stitching technique at the intersection of adjacent forecast intervals eliminates curve abrupt changes caused by switching of multi-level sub-regressors; and a dynamic bias correction method for exogenous variables based on historical fixed windows ensures initial alignment between the excitation response and the measured baseline.

[0038] In other embodiments, to achieve equivalent or broader protection, the time-domain causal filter bank can be replaced with non-stationary signal processing methods such as online wavelet packet decomposition, empirical mode decomposition, or variational mode decomposition; the ridge regression for solving multi-segment coefficients can be replaced with Lasso regression, support vector machine regression, Gaussian process regression, or shallow feedforward neural networks, while maintaining the equivalent overall architecture of multi-segment rolling extrapolation and residual adaptive feedback; the baseline model in step three can also be replaced by a combination of least squares harmonic period fitting and multi-order polynomials. All these alternatives achieve the objectives of this invention.

[0039] Example 2: This embodiment provides a system for predicting Earth's rotation parameters based on a dynamic tracking and segmented model, including: The data acquisition module is configured to acquire historical Earth rotation parameters, historical effective angular momentum, and predicted angular momentum values. The prediction module is configured to: segment the historical Earth rotation parameters, historical effective angular momentum, and predicted angular momentum values ​​according to different lead days; and use multiple preset prediction models to predict the historical Earth rotation parameters, historical effective angular momentum, and predicted angular momentum values ​​in different segments to obtain prediction results. Among them, multiple prediction models are regression prediction models with different parameters; in the regression prediction models, the frequency band boundaries of different frequency band cutoff periods are dynamically tracked and locked based on historical residual power spectral density analysis; the autoregression order and input feature space are dynamically trimmed for different lead days; and the predicted value is determined by weighted smoothing stitching for the cross boundaries of the forecast intervals of adjacent prediction models.

[0040] The working method of the system is the same as that of the Earth rotation parameter prediction method based on dynamic tracking and segmentation model in Example 1, and will not be repeated here.

[0041] Example 3: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the Earth rotation parameter prediction method based on dynamic tracking and segmentation model described in Embodiment 1.

[0042] Example 4: This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the program, it implements the steps of the Earth rotation parameter prediction method based on dynamic tracking and segmentation model described in Embodiment 1.

[0043] Example 5: This embodiment provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the steps of the Earth rotation parameter prediction method based on dynamic tracking and segmentation model described in Embodiment 1.

[0044] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.

Claims

1. A method for predicting Earth's rotation parameters based on dynamic tracking and a piecewise model, characterized in that, include: Obtain historical Earth rotation parameters, historical effective angular momentum, and predicted angular momentum values; Historical Earth rotation parameters, historical effective angular momentum, and predicted angular momentum values ​​are segmented according to different numbers of days ahead. Using multiple pre-set prediction models, the historical Earth rotation parameters, historical effective angular momentum, and predicted angular momentum values ​​in different segments are predicted to obtain the prediction results. Among them, multiple prediction models are regression prediction models with different parameters; in the regression prediction model, the frequency band boundary of different frequency band cutoff periods is dynamically tracked and locked based on historical residual power spectral density analysis; the autoregression order and input feature space are dynamically trimmed for different advance days; and the predicted value is determined by weighted smooth stitching for the cross boundary of the prediction interval of adjacent prediction models. The dynamic tracking based on historical residual power spectral density analysis includes: estimating the power spectral density of the historical residual sequence using the Welch method to obtain the frequency response spectrum; and dynamically searching for the period corresponding to the maximum energy spectral peak within the search sub-intervals corresponding to different frequency bands. ； in, Candidate period; This represents the power spectral density at the corresponding frequency. Set the value for the preset cycle; This is the dominant cycle with the highest energy within the interval; The regression prediction model is a vector autoregressive model with exogenous variables in the baseline, utilizing historical measured sequences. With fluid excitation sequence , build with Regularized regressor, regressor measurement prediction equation for: ； in, , These are the autoregression order and the lag order of the exogenous variable, respectively. , This is the corresponding weight matrix; , It is a lagged term; Harmonic terms; Constant bias; residual sequence for: ； in, These are the single-step predictions from the baseline model; The prediction model includes a first prediction model, a second prediction model, and a third prediction model, with the number of days leading increasing sequentially. The first prediction model is responsible for short-term fine-tuning, using a higher order, and the regression input includes the original term, the three sub-band terms of the residual, exogenous variables, and harmonic components. The second prediction model is responsible for medium-term smooth transition, using a medium order. The third prediction model is responsible for long-term trend extrapolation, with the regression mode switched to a purely autonomous extrapolation mode, high-order disturbances of exogenous variables turned off, and the order reduced to the minimum.

2. The method for predicting Earth rotation parameters based on dynamic tracking and a segmented model as described in claim 1, characterized in that, The historical effective angular momentum and the predicted angular momentum values ​​are preprocessed before being input into the prediction model: the exogenous variables of all external forecasts are translated and aligned online using a preset dynamic bias correction vector to determine the aligned fluid angular momentum sequence. The construction of the dynamic bias correction vector includes: extracting a historical sliding window in the current element initiated by the rolling forecast, determining the mean difference between the external forecasts and the measured values ​​within the window, and constructing the dynamic bias correction vector.

3. The method for predicting Earth rotation parameters based on dynamic tracking and a segmented model as described in claim 2, characterized in that, The dynamic bias correction vector for: ； in: The length of the historical sliding window; For time index variables; The simplified Julian day count value corresponding to the current element for which the rolling forecast was initiated; for The exogenous variable values ​​of the external forecast at that time; for Locally measured historical values ​​of exogenous variables at any given time.

4. The method for predicting Earth rotation parameters based on dynamic tracking and a segmented model as described in claim 2, characterized in that, The aligned fluid angular momentum sequence is transformed into a theoretical change in Earth's rotation parameters: a continuous state transition model based on Chandler oscillation period and quality factor is established. ； in, State vector The derivative; The aligned fluid angular momentum excitation function; system matrix and control matrix for: ； ； in, This is the Chandler natural angular frequency. The damping ratio is given; the system matrix is... and control matrix Discretize into state transition matrix With the gain matrix : ； ； in, This represents the actual time interval between two adjacent data epochs. These are dummy variables for integration; from this, the fluid-induced sequence of Earth's rotation parameters can be derived.

5. The method for predicting Earth rotation parameters based on dynamic tracking and a segmented model as described in claim 1, characterized in that, ahead of time The virtual residual feedback value of the step input is subject to the following damping constraints: ； ； in, For the first Virtual residual feedback after step damping; This is the current predicted value; This is the baseline reference value; The basic damping coefficient; This represents the attenuation rate.

6. The method for predicting Earth rotation parameters based on dynamic tracking and a segmented model as described in claim 1, characterized in that, The weighted smooth stitching method for determining predicted values ​​includes: in the overlapping region... Inside, final predicted value Linear weighting of the predictions from the left and right models: ； ； in, Index for days ahead; , These are the predicted values ​​for the left and right segmented models; and As weight.

7. A system for predicting Earth's rotation parameters based on dynamic tracking and a piecewise model, characterized in that, include: The data acquisition module is configured to acquire historical Earth rotation parameters, historical effective angular momentum, and predicted angular momentum values. The prediction module is configured to: segment the historical Earth rotation parameters, historical effective angular momentum, and predicted angular momentum values ​​according to different lead days; and use multiple preset prediction models to predict the historical Earth rotation parameters, historical effective angular momentum, and predicted angular momentum values ​​in different segments to obtain prediction results. Among them, multiple prediction models are regression prediction models with different parameters; in the regression prediction model, the frequency band boundary of different frequency band cutoff periods is dynamically tracked and locked based on historical residual power spectral density analysis; the autoregression order and input feature space are dynamically trimmed for different advance days; and the predicted value is determined by weighted smooth stitching for the cross boundary of the prediction interval of adjacent prediction models. The dynamic tracking based on historical residual power spectral density analysis includes: estimating the power spectral density of the historical residual sequence using the Welch method to obtain the frequency response spectrum; and dynamically searching for the period corresponding to the maximum energy spectral peak within the search sub-intervals corresponding to different frequency bands. ； in, Candidate period; This represents the power spectral density at the corresponding frequency. Set the value for the preset cycle; This is the dominant cycle with the highest energy within the interval; The regression prediction model is a vector autoregressive model with exogenous variables in the baseline, utilizing historical measured sequences. With fluid excitation sequence , build with Regularized regressor, regressor measurement prediction equation for: ； in, , These are the autoregression order and the lag order of the exogenous variable, respectively. , This is the corresponding weight matrix; , It is a lagged term; Harmonic terms; Constant bias; residual sequence for: ； in, These are the single-step predictions from the baseline model; The prediction model includes a first prediction model, a second prediction model, and a third prediction model, with the number of days leading increasing sequentially. The first prediction model is responsible for short-term fine-tuning, using a higher order, and the regression input includes the original term, the three sub-band terms of the residual, exogenous variables, and harmonic components. The second prediction model is responsible for medium-term smooth transition, using a medium order. The third prediction model is responsible for long-term trend extrapolation, with the regression mode switched to a purely autonomous extrapolation mode, high-order disturbances of exogenous variables turned off, and the order reduced to the minimum.

Images