Basic model construction method for time sequence prediction task

Abstract

The invention provides a time sequence prediction task-oriented basic model construction method, which comprises the following steps of: constructing a Patch division layer which is used for cutting an input time sequence by adopting a context window adaptive mechanism combining time domain and frequency domain statistical characteristics to generate overlapped time periods; an embedded layer is constructed and used for converting the processed time sequence into embedded representation and generating an embedded Token through linear projection processing; constructing a forward and reverse frequency domain enhanced bidirectional state space module which is used for outputting forward and reverse modeling results by taking the input sequence and the time reversal sequence thereof as input, and fusing the forward and reverse modeling results to obtain a final fusion result; constructing a feed-forward layer which is used for taking the final fusion result as input to encode a time dependency relationship; and constructing an output and projection layer for mapping the coded time sequence to a predicted future value through linear projection. The method can adapt to various time series data under zero sample setting, and has good modeling robustness and cross-variable modeling capability.

Description

Technical Field

[0001] This application relates to the field of time series forecasting technology, and in particular to a method for constructing a basic model for time series forecasting tasks. Background Technology

[0002] Currently, time series forecasting technology is widely used in physical systems that are time-dependent and periodic, such as weather forecasting systems, industrial control systems, transportation systems, and energy demand systems. These physical systems typically generate a large amount of time series data during operation from sensors, controllers, or terminal devices. Time series forecasting technology can analyze this data to predict the future operating state of the physical system.

[0003] Currently, there are several mainstream implementation schemes for time series forecasting:

[0004] (1) Traditional statistical modeling methods, (2) Deep learning-based sequence models, (3) Pre-trained large-scale time series basic models.

[0005] However, existing time series forecasting techniques still have the following key shortcomings:

[0006] (1) Insufficient generalization ability. Time series data generated by the operation of different physical systems vary significantly in terms of frequency, periodicity, and dependency structure. Existing time series forecasting techniques cannot adapt to different time series forecasts under multiple tasks and scenarios. For example, temperature data in meteorological forecasting systems, traffic flow data in transportation systems, and even operating data of machinery and equipment in industrial control all have unique time dependencies and noise characteristics. Existing methods tend to exhibit poor generalization ability when faced with these heterogeneous time series data.

[0007] (2) Lack of robustness to high-frequency noise and local disturbances. Physical systems cannot avoid external interference in actual operation. The time series data generated during operation often contains high-frequency noise with short-term jumps or periodic superpositions. This makes the predictions of existing models susceptible to noise interference and prone to misleading dependency modeling. For example, in traffic flow prediction of traffic systems, sudden events (such as traffic accidents) may have a significant impact on the prediction results.

[0008] (3) Limited cross-channel modeling capability in multivariate sequences. Physical systems typically involve multiple factors in actual operation. Traditional models often treat different variables as independent input channels, lacking an effective mechanism for information fusion between variables, making it difficult to model the "interaction relationship between input variables" (such as the correlation between readings from multiple sensors or traffic flow indicators). For example, in energy demand forecasting, different types of energy demand data may be affected by both seasonality and weather changes, and existing models struggle to effectively capture these cross-effects. Summary of the Invention

[0009] This application aims to at least partially address one of the technical problems in the related art.

[0010] Therefore, the first objective of this application is to propose a basic model construction method for time series forecasting tasks, which solves the technical problems of existing methods, can adapt to various time series data under zero-sample settings, and has good modeling robustness and cross-variable modeling capabilities.

[0011] The second objective of this application is to propose a basic model building system for time series forecasting tasks.

[0012] The third objective of this application is to propose a computer device.

[0013] The fourth objective of this application is to provide a non-transitory computer-readable storage medium.

[0014] To achieve the above objectives, the first aspect of this application proposes a method for constructing a basic model for time series prediction tasks. The time series data is data generated by the operation of a physical system. The time series data is acquired through a signal acquisition device, which includes at least one of a sensor, a controller, or a terminal device. The operating state of the physical system exhibits time dependence and periodicity. The basic model for time series prediction tasks is used to analyze the time series data generated by the operation of the physical system and predict the operating state of the physical system at future times. The method includes:

[0015] A patch partitioning layer is constructed to segment the input time series using a context window adaptive mechanism that combines time-domain and frequency-domain statistical features, generating overlapping time periods.

[0016] An embedding layer is constructed to convert the processed time series into an embedded representation and generate embedded tokens through linear projection processing;

[0017] A forward and backward frequency-domain enhanced bidirectional state-space module, FBMamba, is constructed to take the input sequence and its time-reversed sequence as input, output the forward and backward modeling results, and fuse them to obtain the final fused result.

[0018] Construct a feedforward layer to encode time dependencies using the final fusion result as input;

[0019] Construct an output and projection layer to map the encoded time series to predicted future values ​​through linear projection.

[0020] To achieve the above objectives, a second aspect of the present invention proposes a basic model construction system for time series prediction tasks. The time series data is generated by the operation of a physical system and is acquired through a signal acquisition device, which includes at least one of a sensor, a controller, or a terminal device. The operating state of the physical system exhibits time dependence and periodicity. The basic model for time series prediction tasks is used to analyze the time series data generated by the operation of the physical system and predict the operating state of the physical system at future times. The basic model for time series prediction tasks includes:

[0021] The Patch partitioning layer is used to segment the input time series using a context window adaptive mechanism that combines time-domain and frequency-domain statistical features, generating overlapping time periods.

[0022] The embedding layer is used to convert the processed time series into an embedded representation and generate embedded tokens through linear projection processing.

[0023] The forward and backward frequency domain enhanced bidirectional state space module FBMamba is used to take the input sequence and its time-reversed sequence as input, output the forward and backward modeling results, and fuse them to obtain the final fused result;

[0024] The feedforward layer is used to encode time dependencies, taking the final fusion result as input.

[0025] The output and projection layers are used to map the encoded time series to predicted future values ​​through linear projection.

[0026] To achieve the above objectives, a third aspect of the present invention provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described method for constructing a basic model for time series forecasting tasks.

[0027] To achieve the above objectives, a fourth aspect of the present invention provides a non-transitory computer-readable storage medium that, when the instructions in the storage medium are executed by a processor, enables the execution of the aforementioned basic model construction method for time series prediction tasks.

[0028] The basic model construction method and system for time series prediction tasks in this application proposes an adaptive context window mechanism that combines time-domain and frequency-domain statistical features. This ensures that an appropriate context length can be automatically selected for different frequency data, avoiding the decline in generalization ability caused by manually setting or fixing windows, and significantly improving the modeling adaptability for multiple tasks and multiple domains. It also constructs a frequency-domain enhanced bidirectional state space module, FBMamba, to suppress noise and short-term jumps at the frequency domain level, improve model robustness, retain core periodic / trend components, improve long-term dependency modeling ability, and enhance the stability and prediction accuracy of the model in complex real-world scenarios. Furthermore, it constructs forward and backward FBMamba modules to form a bidirectional modeling structure and a cross-channel dependency capture mechanism, significantly improving cross-channel modeling capabilities.

[0029] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0030] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0031] Figure 1 This is a flowchart illustrating a basic model construction method for time series forecasting tasks provided in Embodiment 1 of this application.

[0032] Figure 2 This is a schematic diagram of the basic model structure for time series prediction tasks in an embodiment of this application;

[0033] Figure 3 This is a schematic diagram of the structure of a basic model construction system for time series prediction tasks provided in an embodiment of this application. Detailed Implementation

[0034] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0035] The following describes, with reference to the accompanying drawings, a method and apparatus for building a basic model for time series prediction tasks according to embodiments of this application.

[0036] Figure 1This is a flowchart illustrating a basic model construction method for time series prediction tasks provided in Embodiment 1 of this application. The time series data is data generated by the operation of a physical system. The time series data is acquired through a signal acquisition device, which includes at least one of a sensor, a controller, or a terminal device. The operating state of the physical system has time dependence and periodicity. The basic model for time series prediction tasks is used to analyze the time series data generated by the operation of the physical system and predict the operating state of the physical system at future times.

[0037] like Figure 1 As shown, the basic model construction method for this time series forecasting task includes the following steps:

[0038] Step 101: Construct a Patch partitioning layer, which uses a context window adaptive mechanism that combines time-domain and frequency-domain statistical features to segment the input time series and generate overlapping time periods.

[0039] Specifically, to address the issue of insufficient generalization ability in existing time series forecasting techniques, an adaptive context window selection mechanism is proposed, including:

[0040] First, perform ACF (autocorrelation function) and PACF (partial autocorrelation function) analysis on the original time series to determine whether it has significant periodicity, and discard time series that do not have significant periodicity.

[0041] The FFT (Fast Fourier Transform) is used to accurately calculate the dominant frequency of the time series and estimate the initial period;

[0042] Using the Bayesian optimization algorithm with the model's predicted MSE as the objective function, the context window length is fine-tuned within a certain fluctuation range to obtain the final optimal window Lopt.

[0043] By using an adaptive context window selection mechanism, the model can automatically select an appropriate context length for data of different frequencies, avoiding the decline in generalization ability caused by manually setting the window, and significantly improving the adaptability of multi-task and multi-domain modeling.

[0044] Step 102: Construct an embedding layer to convert the processed time series into an embedded representation and generate embedded tokens through linear projection processing;

[0045] Step 103: Construct forward and backward frequency domain enhanced bidirectional state space FBMamba modules, which take the input sequence and its time-reversed sequence as input, output forward and backward modeling results, and fuse them to obtain the final fusion result;

[0046] Specifically, a frequency domain enhancement and noise suppression mechanism is proposed, and a frequency domain enhanced bidirectional state space module FBMamba is constructed, including:

[0047] Perform an FFT transformation on the patched input sequence;

[0048] The top-k main frequency with the largest amplitude is retained, and IFFT is performed to reconstruct the signal to achieve frequency domain noise reduction;

[0049] Further applying the Blackman window function smooths the spectral edges and reduces leakage;

[0050] The above denoised sequence is embedded into the state-space modeling process (SSM) to form two FBMamba branches, forward and backward, and the final temporal feature representation is obtained through element-level weighted fusion.

[0051] By employing frequency domain enhancement and noise suppression mechanisms, noise and short-term jumps are suppressed at the frequency domain level, improving model robustness, preserving core periodic / trend components, enhancing long-term dependency modeling capabilities, and improving the model's stability and prediction accuracy in complex real-world scenarios.

[0052] Specifically, a bidirectional modeling structure and a cross-channel dependency capture mechanism are proposed, and mutually independent forward and reverse FBMamba modules are constructed, including:

[0053] Simultaneously, the input sequence and its time-reversed sequence are fed into the corresponding FBMamba modules for processing;

[0054] The outputs are Z_forward and Z_backward, representing the forward and backward modeling results, respectively.

[0055] The final fusion result Z_fbm is obtained through element-level addition operations;

[0056] Subsequent dependency enhancement and output prediction are performed by a lightweight feedforward network.

[0057] By using a bidirectional modeling structure and a cross-channel dependency capture mechanism, forward and backward information are integrated to complete the modeling context, clarify the interaction and synchronization trends among multiple modeling variables, and improve the overall prediction accuracy.

[0058] Step 104: Construct a feedforward layer to encode time dependencies using the final fusion result as input;

[0059] Step 105: Construct an output and projection layer to map the encoded time series to predicted future values ​​through linear projection.

[0060] Specifically, if the physical system is a weather forecasting system, then the time series data is the meteorological data acquired through sensors, including the temperature, humidity, wind direction, wind speed, and rainfall of a certain area. By analyzing the time series data of the weather forecasting system, the temperature, humidity, wind direction, wind speed, and rainfall of the area at future times can be predicted.

[0061] It also includes: proposing a high-quality pre-trained dataset construction method (TSR-1B), including:

[0062] Raw time-series data is collected from public sources of multiple physical systems (meteorology, power, economy, transportation, Internet, etc.). When existing physical systems are in operation, sensors, controllers or terminal devices within them usually generate a large amount of time-series data. This embodiment acquires existing time-series data to ensure the authenticity of the data.

[0063] Applications include sliding window segmentation, zero-value filtering, MAD stability filtering, and seasonal decomposition.

[0064] High-quality subsequences are then selected through cross-correlation analysis and Hurst index analysis;

[0065] Construct a balanced subset TSR-1B containing 1 billion time points for pre-training the base model.

[0066] By constructing high-quality pre-trained datasets, we can cover short, medium, and long-term dependency structures, ensuring the richness of the periodic structure of the pre-trained data and improving the model's generalization ability and performance consistency under zero-shot settings.

[0067] It also includes: proposing a unified and efficient model structure and training process, including:

[0068] The Patch+LinearEmbedding mechanism is used to standardize long sequences into fixed-length token representations;

[0069] Modeling token sequences and incorporating temporal dependencies using FBMamba;

[0070] The output layer uses matrix projection to predict multi-step future values;

[0071] The entire model maintains linear time complexity and can be scaled up to long sequence / multivariate tasks.

[0072] The basic model structure for time series prediction tasks constructed in this application is as follows: Figure 2 As shown, it contains four main levels:

[0073] Embedding Layer: This layer converts the original time series into an embedded representation and generates embedded tokens through linear projection (W_emb).

[0074] Patching Layer: By segmenting the input time series, multiple overlapping time periods are generated, enabling the model to handle local temporal dynamics.

[0075] The FBMamba module is divided into two sub-modules: forward and backward. The forward and backward sub-modules model time dependencies from the context, respectively, and incorporate frequency domain denoising and smoothing mechanisms to improve robustness against noise. The outputs of both modules are merged and then further processed.

[0076] Feedforward layer (FFN TD Encoding Layer): Further encodes time dependencies to optimize model performance.

[0077] Output and Projection Layer: Maps the encoded time series to the predicted future values ​​through a linear projection (W_proj).

[0078] Figure 2 (d) is the Frequency Domain Enhancement Module (FBMamba Block). This figure shows the structure of the FBMamba module:

[0079] FFT and IFFT processing: First, the input is converted to the frequency domain using FFT, and the Top-k frequency with the largest amplitude is selected for processing. Then, the signal is reconstructed using IFFT to remove noise.

[0080] Blackman window smoothing: Smooths the spectrum, reduces frequency domain leakage, and further clarifies the signal.

[0081] State-space modeling (SSM): Linear state-space modeling is performed using Selective SSM, combined with nonlinear activation functions (SiLU).

[0082] To achieve the above embodiments, this application also proposes a basic model construction system for time series prediction tasks. The time series data is data generated by the operation of a physical system. The time series data is acquired through a signal acquisition device, which includes at least one of a sensor, a controller, or a terminal device. The operating state of the physical system has time dependence and periodicity. The basic model for time series prediction tasks is used to analyze the time series data generated by the operation of the physical system and predict the operating state of the physical system at future times.

[0083] Figure 3 This is a schematic diagram of the structure of a basic model construction system for time series prediction tasks provided in an embodiment of this application.

[0084] like Figure 3As shown, the basic model building system for time series forecasting tasks includes:

[0085] The Patch partitioning layer is used to segment the input time series using a context window adaptive mechanism that combines time-domain and frequency-domain statistical features, generating overlapping time periods.

[0086] The embedding layer is used to convert the processed time series into an embedded representation and generate embedded tokens through linear projection processing.

[0087] The forward and backward frequency-domain enhanced bidirectional state-space FBMamba module is used to take the input sequence and its time-reversed sequence as input, output the forward and backward modeling results, and fuse them to obtain the final fused result.

[0088] The feedforward layer is used to encode time dependencies, taking the final fusion result as input.

[0089] The output and projection layers are used to map the encoded time series to predicted future values ​​through linear projection.

[0090] Furthermore, in this embodiment of the application, the adoption of the context window adaptive mechanism combining time-domain and frequency-domain statistical features includes:

[0091] Autocorrelation function (ACF) and partial autocorrelation function (PACF) analyses are performed on time series to determine whether they exhibit significant periodicity, and time series that do not exhibit significant periodicity are discarded.

[0092] The dominant frequency of the time series is calculated using FFT, the initial period is estimated, and the context window length is set.

[0093] Using the Bayesian optimization algorithm with the model's predicted MSE as the objective function, the length of the context window is fine-tuned within a floating range to obtain the final optimal window Lopt.

[0094] Specifically, in this embodiment of the application, constructing the forward FBMamba module includes:

[0095] The patched input sequence is converted to the frequency domain using FFT.

[0096] The top-k main frequency with the largest amplitude is retained, and IFFT is performed to reconstruct the signal to achieve frequency domain noise reduction;

[0097] Apply a Blackman window function to the denoised sequence to smooth the spectral edges;

[0098] The denoised sequence is embedded into the state-space modeling process (SMM) to form forward and backward FBMamba branches, and the final temporal feature representation is obtained through element-level weighted fusion.

[0099] Furthermore, in this embodiment of the application, the step of fusing the forward and backward modeling results to obtain the final fusion result includes:

[0100] The final fusion result is obtained through element-level addition operations.

[0101] It should be noted that the explanation of the above-described embodiment of the basic model construction method for time series forecasting tasks also applies to the basic model construction system for time series forecasting tasks in this embodiment, and will not be repeated here.

[0102] To implement the above embodiments, the present invention also proposes a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the methods described in the above embodiments.

[0103] To implement the above embodiments, the present invention also proposes a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method of the above embodiments.

[0104] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0105] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0106] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0107] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0108] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0109] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0110] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0111] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for constructing a basic model for time series forecasting tasks, characterized in that, Time series data refers to data generated by the operation of a physical system. This time series data is acquired through a signal acquisition device, which includes at least one of a sensor, controller, or terminal device. The operating state of the physical system exhibits time dependence and periodicity. The basic model for time series prediction tasks is used to analyze the time series data generated by the operation of the physical system and predict the operating state of the physical system at future times. The method includes: A patch partitioning layer is constructed to segment the input time series using a context window adaptive mechanism that combines time-domain and frequency-domain statistical features, generating overlapping time periods. An embedding layer is constructed to convert the processed time series into an embedded representation and generate embedded tokens through linear projection processing; A forward and backward frequency-domain enhanced bidirectional state-space module, FBMamba, is constructed to take the input sequence and its time-reversed sequence as input, output the forward and backward modeling results, and fuse them to obtain the final fused result. Construct a feedforward layer to encode time dependencies using the final fusion result as input; Construct an output and projection layer to map the encoded time series to predicted future values ​​through linear projection.

2. The method as described in claim 1, characterized in that, The aforementioned context window adaptive mechanism, which combines time-domain and frequency-domain statistical features, includes: Autocorrelation function (ACF) and partial autocorrelation function (PACF) analyses are performed on time series to determine whether they exhibit significant periodicity, and time series that do not exhibit significant periodicity are discarded. The dominant frequency of the time series is calculated using FFT, the initial period is estimated, and the context window length is set. Using the Bayesian optimization algorithm with the model's predicted MSE as the objective function, the length of the context window is fine-tuned within a floating range to obtain the final optimal window Lopt.

3. The method as described in claim 1, characterized in that, Building the forward FBMamba module includes: The patched input sequence is converted to the frequency domain using FFT. The top-k main frequency with the largest amplitude is retained, and IFFT is performed to reconstruct the signal to achieve frequency domain noise reduction; Apply a Blackman window function to the denoised sequence to smooth the spectral edges; The denoised sequence is embedded into the state-space modeling process (SMM) to form forward and backward FBMamba branches, and the final temporal feature representation is obtained through element-level weighted fusion.

4. The method as described in claim 1, characterized in that, The process of fusing the forward and backward modeling results to obtain the final fusion result includes: The final fusion result is obtained through element-level addition operations.

5. A basic model construction system for time series forecasting tasks, characterized in that, Time series data refers to data generated by the operation of a physical system. This time series data is acquired through a signal acquisition device, which includes at least one of a sensor, controller, or terminal device. The operating state of the physical system exhibits time dependence and periodicity. The basic model for time series prediction tasks is used to analyze the time series data generated by the operation of the physical system and predict the operating state of the physical system at future times. The basic model for time series prediction tasks includes: The Patch partitioning layer is used to segment the input time series using a context window adaptive mechanism that combines time-domain and frequency-domain statistical features, generating overlapping time periods. The embedding layer is used to convert the processed time series into an embedded representation and generate embedded tokens through linear projection processing. The forward and backward frequency domain enhanced bidirectional state space module FBMamba is used to take the input sequence and its time-reversed sequence as input, output the forward and backward modeling results, and fuse them to obtain the final fused result; The feedforward layer is used to encode time dependencies, taking the final fusion result as input. The output and projection layers are used to map the encoded time series to predicted future values ​​through linear projection.

6. The system as described in claim 5, characterized in that, The aforementioned context window adaptive mechanism, which combines time-domain and frequency-domain statistical features, includes: Autocorrelation function (ACF) and partial autocorrelation function (PACF) analyses are performed on time series to determine whether they exhibit significant periodicity, and time series that do not exhibit significant periodicity are discarded. The dominant frequency of the time series is calculated using FFT, the initial period is estimated, and the context window length is set. Using the Bayesian optimization algorithm with the model's predicted MSE as the objective function, the length of the context window is fine-tuned within a floating range to obtain the final optimal window Lopt.

7. The system as described in claim 5, characterized in that, Building the forward FBMamba module includes: The patched input sequence is converted to the frequency domain using FFT. The top-k main frequency with the largest amplitude is retained, and IFFT is performed to reconstruct the signal to achieve frequency domain noise reduction; Apply a Blackman window function to the denoised sequence to smooth the spectral edges; The denoised sequence is embedded into the state-space modeling process (SMM) to form forward and backward FBMamba branches, and the final temporal feature representation is obtained through element-level weighted fusion.

8. The system as described in claim 5, characterized in that, The process of fusing the forward and backward modeling results to obtain the final fusion result includes: The final fusion result is obtained through element-level addition operations.

9. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the method as described in any one of claims 1-4.

10. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-4.

Images