A time series prediction method and system based on cross-period phase alignment
By using a cross-cycle phase-aligned time series forecasting method, the problems of low efficiency and insufficient forecast reliability in existing technologies are solved, achieving more efficient time series forecasting that is suitable for large-scale complex datasets.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIHANG UNIV
- Filing Date
- 2026-03-30
- Publication Date
- 2026-06-30
AI Technical Summary
Existing time series forecasting methods are limited in efficiency and generalization when dealing with large-scale complex datasets, and the distribution of small time blocks drifts over time, leading to a decrease in forecast reliability and high computational resource consumption.
A time series prediction method based on cross-cycle phase alignment is adopted. Phase representation is obtained through normalization and cross-cycle alignment. The phase characterization is obtained by aggregation and context distribution using a cross-phase routing layer, and then prediction and reconstruction are performed.
It reduces computational and memory requirements, improves the stability and efficiency of predictions, and is applicable to larger models and a wider range of data representations.
Smart Images
Figure CN122309989A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data prediction, and in particular to a time series prediction method and system based on cross-cycle phase alignment. Background Technology
[0002] Currently, segmented modeling methods for time series forecasting typically involve first processing the original sequence into several smaller time segments based on local periods or fixed windows, with each time segment serving as an independent modeling unit. Then, deep models are used to model temporal correlations and cross-period relationships at the time segment level. These modeling methods can be further extended to cross-dimensional and cross-scale interactions, leveraging periodic priors and improving the predictive power of long sequences. Their core assumption is that short-range patterns within a period are relatively stable, and the process of "segmenting into time segments + attention modeling" serves as the main framework.
[0003] While the aforementioned segmented modeling method improves the effectiveness of utilizing periodicity, in real data, periodic patterns drift with external factors, leading to an expansion of the representation space dimension, a large number of parameters and computational costs, and limited efficiency and generalization when extended to large-scale complex datasets. Furthermore, the distribution of small time segments continues to drift over time, reducing the reliability of out-of-sample training and failing to meet the needs of broader coverage, higher-dimensional representations, and larger models. Summary of the Invention
[0004] Existing piecewise modeling methods for time series forecasting are prone to dimensional expansion of the representation space and drift in prediction results, requiring significant computational resources and reducing prediction reliability. In view of these problems, this invention is proposed to provide a time series forecasting method based on cross-period phase alignment that overcomes or at least partially solves these problems, comprising: Acquire time series data, perform standardization and cross-period alignment processing on the time series data, and obtain the phase representation of the time series data; The phase representation is mapped to obtain the latent space representation of the phase representation; The latent space representation is aggregated through a cross-phase routing layer to obtain aggregated information; the aggregated information is then subjected to context distribution through the cross-phase routing layer to obtain a phase representation. The phase representation is subjected to prediction processing to obtain future prediction results; the future prediction results are then subjected to time-by-time prediction and restoration processing to obtain the final prediction results.
[0005] Optionally, time series data is acquired, and the time series data is standardized and aligned across periods to obtain the phase representation of the time series data, including... Retrieve time-series data from a database, arranged in chronological order. The time-series data contains N variables, and each variable contains data from T time steps. The data for each variable... as follows: ; In the above formula, This represents the value of the j-th variable at time step t, where T represents the number of time steps and N represents the number of variables. Represents the real number field; The time series data is standardized according to the following formula: ; In the above formula, express The value after standardization Let represent the mean of the j-th variable. This represents the standard deviation of the j-th variable; Get each variable at the time interval The autocorrelation value and the estimated principal period length are obtained. By aligning points of the same phase and performing cross-period phase changes based on the main period length, the phase representation of the time series data is obtained.
[0006] Optionally, the phase representation is mapped to obtain the latent space representation of the phase representation, including: The phase function is mapped into the latent space by a mapping function, and a position code is added to each phase position in the latent space to obtain the latent space representation of the phase representation.
[0007] Optionally, the latent space representation is aggregated through a cross-phase routing layer to obtain aggregated information; the aggregated information is then subjected to context distribution through the cross-phase routing layer to obtain a phase representation, including: Several routers are constructed into a cross-phase routing layer using an attention mechanism; the latent space representation is then subjected to multi-head attention aggregation processing through the cross-phase routing layer to obtain aggregated information. The aggregated information is distributed to the residual-feedforward neural network via the router context for processing to obtain the phase representation.
[0008] Optionally, the phase representation is subjected to prediction processing to obtain a future prediction result; the future prediction result is subjected to time-by-time prediction and reconstruction processing to obtain a final prediction result, including: The phase representation is predicted by a shared prediction head to obtain the phase point prediction result for each phase in the future several periods, which is then used as the future prediction result. The phase point prediction results are mapped onto a time axis to obtain time-by-time prediction results; the time-by-time prediction results are then subjected to denormalized dimensional restoration processing to obtain the final prediction results.
[0009] As one aspect of the present invention, embodiments of the present invention also provide a time series prediction system based on cross-period phase alignment, comprising: The data acquisition module is used to acquire time series data; A phase representation generation module is used to perform standardization and cross-period alignment processing on the time series data to obtain the phase representation of the time series data. A mapping module is used to perform mapping processing on the phase representation to obtain the latent space representation of the phase representation; The aggregation processing module is used to perform aggregation processing on the latent space representation through the cross-phase routing layer to obtain aggregated information; The distribution processing module is used to perform context distribution processing on the aggregated information through the cross-phase routing layer to obtain a phase representation; The prediction and restoration module is used to perform prediction processing on the phase representation to obtain future prediction results; and to perform time-by-time prediction and restoration processing on the future prediction results to obtain the final prediction result.
[0010] Optionally, the data acquisition module is used to acquire time series data, including: Retrieve time-series data from a database, arranged in chronological order. The time-series data contains N variables, and each variable contains data from T time steps. The data for each variable... as follows: ; In the above formula, This represents the value of the j-th variable at time step t, where T represents the number of time steps and N represents the number of variables. Represents the real number field; The phase representation generation module is used to perform standardization and cross-period alignment processing on the time series data to obtain the phase representation of the time series data, including: The time series data is standardized according to the following formula: ; In the above formula, express The value after standardization Let represent the mean of the j-th variable. This represents the standard deviation of the j-th variable; Get each variable at the time interval The autocorrelation value and the estimated principal period length are obtained. By aligning points of the same phase and performing cross-period phase changes based on the main period length, the phase representation of the time series data is obtained.
[0011] Optionally, the mapping module is used to perform mapping processing on the phase representation to obtain the latent space representation of the phase representation, including: The phase function is mapped into the latent space by a mapping function, and a position code is added to each phase position in the latent space to obtain the latent space representation of the phase representation.
[0012] Optionally, the aggregation processing module is used to perform aggregation processing on the latent space representation through a cross-phase routing layer to obtain aggregated information, including: Several routers are constructed into a cross-phase routing layer using an attention mechanism; the latent space representation is then subjected to multi-head attention aggregation processing through the cross-phase routing layer to obtain aggregated information. The distribution processing module is used to perform context distribution processing on the aggregated information through the cross-phase routing layer to obtain phase representation, including: The aggregated information is distributed to the residual-feedforward neural network via the router context for processing to obtain the phase representation.
[0013] Optionally, the prediction and restoration module is used to perform prediction processing on the phase representation to obtain a future prediction result; and to perform time-by-time prediction and restoration processing on the future prediction result to obtain a final prediction result, including: The phase representation is predicted by a shared prediction head to obtain the phase point prediction result for each phase in the future several periods, which is then used as the future prediction result. The phase point prediction results are mapped onto a time axis to obtain time-by-time prediction results; the time-by-time prediction results are then subjected to denormalized dimensional restoration processing to obtain the final prediction results.
[0014] The beneficial effects of the above-mentioned technical solutions provided in the embodiments of the present invention include at least the following: This invention provides a time series prediction method and system based on cross-period phase alignment. The method involves acquiring time series data, standardizing and aligning the data across periods to obtain a phase representation; mapping the phase representation to obtain a latent space representation; aggregating the latent space representation through a cross-phase routing layer to obtain aggregated information; context-distributing the aggregated information through the cross-phase routing layer to obtain a phase characterization; predicting the phase characterization to obtain future prediction results; and performing time-by-time prediction and reconstruction on the future prediction results to obtain the final prediction result. The cross-phase routing mechanism reduces pairwise phase interactions to a two-hop attention process from phase to router and back to phase. The complexity increases linearly with the number of phases, reducing memory and computational requirements. Phase units are more stable over time, facilitating the use of lower intrinsic dimensions and improving operational efficiency.
[0015] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.
[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the time series prediction method based on cross-cycle phase alignment provided in an embodiment of the present invention. Figure 2 It is an architecture based on a time series forecasting method with cross-cycle phase alignment.
[0018] Figure 3 This is a comparison diagram of segmented units and phase units.
[0019] Figure 4 This is a schematic diagram of the time series prediction method based on cross-period phase alignment provided in an embodiment of the present invention; Detailed Implementation
[0020] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.
[0021] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," "outer," "far," "near," "front," and "rear," etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the accompanying drawings and are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.
[0022] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.
[0023] Please see Figures 1 to 3 As shown, an embodiment of this application provides a time series prediction method based on cross-period phase alignment. This time series prediction method based on cross-period phase alignment includes: Acquire time series data, perform standardization and cross-period alignment on the time series data to obtain the phase representation of the time series data; The phase representation is mapped to obtain the latent space representation of the phase representation; The latent space representation is aggregated through a cross-phase routing layer to obtain aggregated information; the aggregated information is then subjected to context distribution through a cross-phase routing layer to obtain the phase representation. The phase representation is processed to obtain the future prediction result; the future prediction result is then processed to predict and restore it time by time to obtain the final prediction result.
[0024] The beneficial effects of the above embodiments are that the time series prediction method based on cross-cycle phase alignment uses a cross-phase routing mechanism to reduce the interaction between pairs of phases to a two-hop attention from phase to router and then to phase. The complexity increases linearly with the number of phases, reducing memory and computational requirements. The phase units are more stable as they evolve over time, making it easier to use lower intrinsic dimensions and improve operating efficiency.
[0025] In another embodiment, time series data is acquired, and the time series data is standardized and aligned across periods to obtain a phase representation of the time series data, including... Retrieve time-series data from the database, arranged in chronological order. The time-series data contains N variables, and each variable contains data from T time steps; where the data for each variable... as follows: ; In the above formula, This represents the value of the j-th variable at time step t, where T represents the number of time steps and N represents the number of variables. Represents the real number field; The time series data is standardized according to the following formula: ; In the above formula, express The value after standardization Let represent the mean of the j-th variable. This represents the standard deviation of the j-th variable; Get each variable at the time interval The autocorrelation value and the estimated principal period length are determined using the following formula; specifically, the autocorrelation value of each variable at the time interval is determined using the following formula. Autocorrelation value : ; In the above formula, express The value after standardization; Estimate the length of the principal period using the following formula. : ; In the above, argmax represents the argument of the maximum function. This represents an optional set of periods.
[0026] By aligning points of the same phase and performing cross-period phase changes based on the main period length, the phase representation of the time series data is obtained; specifically, Points at the same phase (e.g., the same time of day) can be aligned, and phase evolution across periods can be performed by focusing on the main period length, thus obtaining the phase representation of the time series data, as shown in the following formula: ; ; In the above formula, This represents the phase difference corresponding to the current time step t. The phase difference mentioned above represents the time difference between two phases, i.e., the time interval mentioned earlier. Same meaning, Represents the modulo function , Indicates phase and the main period The standardized value is given, and the parameter symbol is used. Equivalent representation.
[0027] In another embodiment, the phase representation is mapped to obtain the latent space representation of the phase representation, including: The phase function is mapped into the latent space using a mapping function, and a positional code is added to each phase position in the latent space to obtain the latent space representation of the phase representation. The specific mapping process is as follows: ; ; ; In the above formula, Indicates phase The latent space representation. Represents a mapping function. Indicates positional embedding, This indicates that the phase of the position embedding has been taken into account. The latent space representation. The latent space representation of all phases, where phase .
[0028] In another embodiment, the latent space representation is aggregated through a cross-phase routing layer to obtain aggregated information; the aggregated information is then subjected to context distribution through a cross-phase routing layer to obtain a phase representation, including: Several routers are constructed into a cross-phase routing layer using an attention mechanism; the latent space representation is then aggregated using multi-head attention aggregation processing through the cross-phase routing layer to obtain aggregated information; Specifically, an attention mechanism can be used to construct a cross-phase routing layer with a small number of routers. By using a two-hop structure, information is first concentrated on the routers, avoiding the secondary complexity caused by full connectivity between phases. The process is as follows: ; ; The aggregated information is distributed to the residual-feedforward neural network via the router context for processing, resulting in a phase representation. In the above equation, This represents the aggregated information output by the cross-phase routing layer. This indicates multi-head attention aggregation processing. This represents the router matrix corresponding to the cross-phase routing layer. Indicates the number of routers. Represents the dimension of the latent space. The latent space representation of all phases. , , These represent linear layers corresponding to the matrix W multiplied by the three key intermediate matrices Q, K, and V within the attention mechanism. This represents the transpose of matrix K. This represents the softmax function.
[0029] After contextualizing the router, the aggregated information is distributed to the residual-feedforward neural network for processing, thereby returning the global context to each phase and obtaining the phase representation. The above phase representation forms a cross-phase-dependent contextualized representation, and the process is as follows: ; ; In the above formula, This indicates the phase information that is distributed back by the router after being contextualized. This represents the aggregated information output by the cross-phase routing layer, i.e., the router output obtained from the previous hop. , , These represent the other linear layer corresponding to the matrix W multiplied by the three key intermediate matrices Q, K, and V within the attention mechanism. This represents all phase representations in the final calculation. Presentation layer normalization processing, This indicates feedforward neural network processing. This indicates multi-head attention aggregation processing.
[0030] In another embodiment, the phase representation is subjected to prediction processing to obtain a future prediction result; the future prediction result is then subjected to time-by-time prediction and reconstruction processing to obtain a final prediction result, including: By using a shared prediction head to perform prediction processing on the phase representation, the prediction results of the phase points of each phase in the future several periods are obtained, and these results are used as future prediction results. The phase point prediction results are mapped onto the time axis to obtain the time-by-time prediction results; the time-by-time prediction results are then denormalized to restore the dimensions, resulting in the final prediction results.
[0031] In practice, a shared prediction head can be used to predict the phase points of each phase in the future for several cycles, and the corresponding future prediction results can be obtained. The process is as follows: ; In the above formula, Indicates future prediction results. Represents the prediction function. express, This indicates the number of periods (e.g., weeks) for the output prediction. This indicates the principal period of the above estimate. The corresponding real master cycle.
[0032] The phase point prediction results are then laid back onto the time axis according to the phase-generated grid to obtain the time-by-time prediction results. The process is as follows: ; In the above formula, Indicates the first Each variable at time step The predicted value, Indicates period, The above future prediction results The three-dimensional tensor representation of .
[0033] Furthermore, the time-by-time prediction results undergo de-standardization and dimensional restoration to obtain the final prediction results. The process is as follows: ; In the above formula, This indicates the final prediction result. This indicates the prediction results at each time step.
[0034] Please see Figure 4 As shown, an embodiment of this application provides a time series prediction system based on cross-period phase alignment. This time series prediction system based on cross-period phase alignment includes: The data acquisition module is used to acquire time series data; The phase representation generation module is used to perform standardization and cross-period alignment on time series data to obtain the phase representation of the time series data. The mapping module is used to perform mapping processing on the phase representation to obtain the latent space representation of the phase representation; The aggregation processing module is used to aggregate the latent space representation through the cross-phase routing layer to obtain aggregated information. The distribution processing module is used to perform contextual distribution processing on aggregated information through the cross-phase routing layer to obtain phase representation; The prediction and restoration module is used to perform prediction processing on the phase representation to obtain future prediction results; and to perform time-by-time prediction and restoration processing on the future prediction results to obtain the final prediction result.
[0035] The beneficial effects of the above embodiments are that the time series prediction system based on cross-cycle phase alignment adopts a cross-phase routing mechanism to reduce the interaction between pairs of phases to a two-hop attention from phase to router and then to phase. The complexity increases linearly with the number of phases, reducing memory and computing requirements. The phase unit is more stable as it evolves over time, which makes it easier to use a lower intrinsic dimension and improves operating efficiency.
[0036] In another embodiment, the data acquisition module is used to acquire time series data, including: Retrieve time-series data from the database, arranged in chronological order. The time-series data contains N variables, and each variable contains data from T time steps; where the data for each variable... as follows: ; In the above formula, This represents the value of the j-th variable at time step t, where T represents the number of time steps and N represents the number of variables. The phase representation generation module is used to perform standardization and cross-period alignment on time series data to obtain the phase representation of the time series data, including: The time series data is standardized according to the following formula: ; In the above formula, express The value after standardization Let represent the mean of the j-th variable. This represents the standard deviation of the j-th variable; Get each variable at the time interval The autocorrelation value and the estimated principal period length are obtained. By aligning points of the same phase and performing phase changes across periods based on the length of the main period, the phase representation of the time series data is obtained.
[0037] In another embodiment, the mapping module is used to perform mapping processing on the phase representation to obtain the latent space representation of the phase representation, including: The phase function is mapped into the latent space by a mapping function, and a position code is added to each phase position in the latent space to obtain the latent space representation of the phase representation.
[0038] In another embodiment, the aggregation processing module is used to perform aggregation processing on the latent space representation through a cross-phase routing layer to obtain aggregated information, including: Several routers are constructed into a cross-phase routing layer using an attention mechanism; the latent space representation is then aggregated using multi-head attention aggregation processing through the cross-phase routing layer to obtain aggregated information; The distribution processing module is used to perform context distribution processing on aggregated information through the cross-phase routing layer to obtain phase representations, including: The aggregated information is distributed to the residual-feedforward neural network via the router context for processing to obtain the phase representation.
[0039] In another embodiment, the prediction and restoration module is used to perform prediction processing on the phase representation to obtain future prediction results; and to perform time-by-time prediction and restoration processing on the future prediction results to obtain the final prediction result, including: By using a shared prediction head to perform prediction processing on the phase representation, the prediction results of the phase points of each phase in the future several periods are obtained, and these results are used as future prediction results. The phase point prediction results are mapped onto the time axis to obtain the time-by-time prediction results; the time-by-time prediction results are then denormalized to restore the dimensions, resulting in the final prediction results.
[0040] The operation and effect of the time series prediction system based on cross-period phase alignment of the present invention are consistent with the above-mentioned time series prediction method based on cross-period phase alignment, and the description of the time series prediction system based on cross-period phase alignment will not be repeated here.
[0041] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. This disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims. Thus, if these modifications and variations of the invention fall within the scope of the claims of the invention and their equivalents, the invention is also intended to include these modifications and variations.
Claims
1. A time series prediction method based on cross-period phase alignment, characterized in that, include: Acquire time series data, perform standardization and cross-period alignment processing on the time series data, and obtain the phase representation of the time series data; The phase representation is mapped to obtain the latent space representation of the phase representation; The implicit space representation is aggregated by a cross-phase routing layer to obtain aggregated information; The aggregated information is processed through context distribution by the cross-phase routing layer to obtain a phase representation. The phase representation is subjected to prediction processing to obtain future prediction results; the future prediction results are then subjected to time-by-time prediction and restoration processing to obtain the final prediction results.
2. The time series prediction method based on cross-period phase alignment as described in claim 1, characterized in that: Acquire time series data, perform standardization and cross-period alignment processing on the time series data to obtain the phase representation of the time series data, including... Retrieve time-series data from a database, arranged in chronological order. The time-series data contains N variables, and each variable contains data from T time steps. The data for each variable... as follows: ； In the above formula, This represents the value of the j-th variable at time step t, where T represents the number of time steps and N represents the number of variables. Represents the real number field; The time series data is standardized according to the following formula: ； In the above formula, express The value after standardization Let represent the mean of the j-th variable. This represents the standard deviation of the j-th variable; Get each variable at the time interval The autocorrelation value and the estimated principal period length are obtained. By aligning points of the same phase and performing cross-period phase changes based on the main period length, the phase representation of the time series data is obtained.
3. The time series prediction method based on cross-period phase alignment as described in claim 1, characterized in that: The phase representation is mapped to obtain the latent space representation of the phase representation, including: The phase function is mapped into the latent space by a mapping function, and a position code is added to each phase position in the latent space to obtain the latent space representation of the phase representation.
4. The time series prediction method based on cross-period phase alignment as described in claim 1, characterized in that: The implicit space representation is aggregated by a cross-phase routing layer to obtain aggregated information; The aggregated information is processed through context distribution at the cross-phase routing layer to obtain a phase representation, including: Several routers are constructed into a cross-phase routing layer using an attention mechanism; The latent space representation is subjected to multi-head attention aggregation processing through the cross-phase routing layer to obtain aggregated information; The aggregated information is distributed to the residual-feedforward neural network via the router context for processing to obtain the phase representation.
5. The time series prediction method based on cross-period phase alignment as described in claim 1, characterized in that: The phase representation is subjected to predictive processing to obtain future prediction results; The future prediction results are processed through time-by-time prediction and reconstruction to obtain the final prediction results, including: The phase representation is predicted by a shared prediction head to obtain the phase point prediction result for each phase in the future several periods, which is then used as the future prediction result. The phase point prediction results are mapped onto a time axis to obtain time-by-time prediction results; the time-by-time prediction results are then subjected to denormalized dimensional restoration processing to obtain the final prediction results.
6. A time series prediction system based on cross-period phase alignment, characterized in that, include: The data acquisition module is used to acquire time series data; A phase representation generation module is used to perform standardization and cross-period alignment processing on the time series data to obtain the phase representation of the time series data. A mapping module is used to perform mapping processing on the phase representation to obtain the latent space representation of the phase representation; The aggregation processing module is used to perform aggregation processing on the latent space representation through the cross-phase routing layer to obtain aggregated information; The distribution processing module is used to perform context distribution processing on the aggregated information through the cross-phase routing layer to obtain a phase representation; The prediction and restoration module is used to perform prediction processing on the phase representation to obtain future prediction results; and to perform time-by-time prediction and restoration processing on the future prediction results to obtain the final prediction result.
7. The time series prediction system based on cross-period phase alignment as described in claim 6, characterized in that: The data acquisition module is used to acquire time series data, including: Retrieve time-series data from a database, arranged in chronological order. The time-series data contains N variables, and each variable contains data from T time steps. The data for each variable... as follows: ； In the above formula, This represents the value of the j-th variable at time step t, where T represents the number of time steps and N represents the number of variables. Represents the real number field; The phase representation generation module is used to perform standardization and cross-period alignment processing on the time series data to obtain the phase representation of the time series data, including: The time series data is standardized according to the following formula: ； In the above formula, express The value after standardization Let represent the mean of the j-th variable. This represents the standard deviation of the j-th variable; Get each variable at the time interval The autocorrelation value and the estimated principal period length are obtained. By aligning points of the same phase and performing cross-period phase changes based on the main period length, the phase representation of the time series data is obtained.
8. The time series prediction system based on cross-period phase alignment as described in claim 6, characterized in that: The mapping module is used to perform mapping processing on the phase representation to obtain the latent space representation of the phase representation, including: The phase function is mapped into the latent space by a mapping function, and a position code is added to each phase position in the latent space to obtain the latent space representation of the phase representation.
9. The time series prediction system based on cross-period phase alignment as described in claim 6, characterized in that: The aggregation processing module is used to aggregate the latent space representation through a cross-phase routing layer to obtain aggregated information, including: Several routers are constructed into a cross-phase routing layer using an attention mechanism; the latent space representation is then subjected to multi-head attention aggregation processing through the cross-phase routing layer to obtain aggregated information. The distribution processing module is used to perform context distribution processing on the aggregated information through the cross-phase routing layer to obtain phase representation, including: The aggregated information is distributed to the residual-feedforward neural network via the router context for processing to obtain the phase representation.
10. The time series prediction system based on cross-period phase alignment as described in claim 6, characterized in that: The prediction and restoration module is used to perform prediction processing on the phase representation to obtain future prediction results; The future prediction results are processed through time-by-time prediction and reconstruction to obtain the final prediction results, including: The phase representation is predicted by a shared prediction head to obtain the phase point prediction result for each phase in the future several periods, which is then used as the future prediction result. The phase point prediction results are mapped onto a time axis to obtain time-by-time prediction results; the time-by-time prediction results are then subjected to denormalized dimensional restoration processing to obtain the final prediction results.