Ocean multi-element time series prediction method for strict causal online scenarios

Abstract

The present application relates to the technical field of marine environment monitoring, and more particularly to a marine multi-element time series prediction method for strict causal online scenarios. The method comprises obtaining a normalized input sequence through a marine multi-element historical input window at the current time; decomposing the normalized input sequence into a trend component and a non-trend component through a multi-scale adaptive decomposition module; constructing a sparse cross-variable interaction mask according to a cross-variable interaction mask generation module to obtain a trend prediction result and a non-trend prediction result; fusing the trend prediction result and the non-trend prediction result to obtain a normalized predicted output; performing inverse normalization processing on the normalized predicted output to obtain a final prediction result; and in an online updating stage, obtaining an updated prediction model. The method realizes continuous prediction of future changes of marine multi-elements to improve the accuracy, stability and deployment consistency of online prediction of marine multi-elements.

Description

Technical Field

[0001] This invention relates to the field of marine environmental monitoring technology, and in particular to a method for predicting marine multi-factor time series data in strictly causal online scenarios. Background Technology

[0002] Ocean buoys, shore-based stations, nearshore observation platforms, and other marine monitoring equipment can continuously collect data on various elements of the marine environment, such as temperature, salinity, turbidity, chlorophyll concentration, dissolved oxygen, pH, current velocity, and current direction. These marine element data typically arrive continuously over time and collectively reflect the dynamic changes in the marine environment. Therefore, forecasting future time periods based on multi-element marine monitoring data is of great significance for marine environmental early warning, nearshore ecological monitoring, marine ranching management, red tide risk analysis, and marine engineering operation and maintenance.

[0003] Unlike conventional offline prediction tasks, ocean monitoring systems are often in a continuous operational state during actual deployment. The model needs to receive data, output future prediction results, and update the model as new observation data continuously arrives. Ocean multi-factor prediction is an online prediction task in many scenarios. Online prediction requires that the model can only use information that is already available before the current time to make predictions and updates, and cannot use future observation labels that have not yet been fully revealed.

[0004] Existing online time series forecasting methods, under overlapping sliding window conditions, often employ the following approach: after completing multi-step forecasts for a given historical window, the model is immediately updated using the complete ground truth labels corresponding to the forecast interval, and then forecasts are performed for the next adjacent window. However, in multi-step forecasting tasks, adjacent forecast windows typically overlap in their target intervals. If, before the current physical moment, the model has already updated using ground truth labels that partially overlap with the subsequent forecast target interval, label leakage will occur, causing the evaluation results to deviate from the actual deployment scenario.

[0005] Furthermore, ocean multi-element data exhibits significant multi-timescale characteristics and dynamic coupling features. On the one hand, ocean elements typically encompass long-term background evolution, periodic variations, and short-term local disturbances simultaneously. For example, sea surface temperature may be affected by seasonal background changes, diurnal heat exchange, and short-term weather processes; turbidity may be affected by background transport trends and short-term disturbances such as wind, waves, and runoff; chlorophyll concentration may exhibit both seasonal variations and rapid local anomalies. Using a single-scale unified modeling approach can easily lead to the mixing of variations at different scales, reducing prediction accuracy and stability. On the other hand, the correlations between different ocean elements and between different stations are not fixed but dynamically change with sea state, time period, spatial location, and local processes. For example, the correlation between temperature and chlorophyll concentration is stronger in some periods, while the coupling between salinity and turbidity is more pronounced in others; the correlations between different stations also change with tidal currents, wind fields, runoff transport, and mixing and diffusion conditions. Using a fixed, fully connected, intervariate modeling approach can easily introduce ineffective interactions and even noise propagation.

[0006] In online scenarios with delayed feedback, if the results of cross-variable interactions are directly and forcibly used for trend correction, it can easily lead to trend prediction bias when sea state changes rapidly, local disturbances increase, or observation noise is large, and errors will accumulate in subsequent prediction processes. Summary of the Invention

[0007] In view of this, the present invention provides a method for predicting marine multi-factor time series in strictly causal online scenarios, so as to achieve continuous prediction of future changes of marine multi-factors and improve the accuracy, stability and deployment consistency of marine multi-factor online prediction.

[0008] In a first aspect, the present invention provides a method for ocean multi-factor time series forecasting in strictly causal online scenarios, the method comprising:

[0009] Step 1: Obtain the historical input window of the ocean multi-element at the current moment, and use it as the current prediction sample to perform instance-level normalization processing to obtain the normalized input sequence; Step 2: Decompose the normalized input sequence into trend components and non-trend components using the multi-scale adaptive decomposition module; Step 3: Construct a sparse intervariate interaction mask based on the intervariate interaction mask generation module, so that the trend prediction branch and the non-trend prediction branch can call it together to obtain the trend prediction result and the non-trend prediction result. Step 4: Fuse the trend forecast results with the non-trend forecast results to obtain the normalized forecast output; perform inverse normalization processing on the normalized forecast output to obtain the final forecast result. Step 5, Online Update Phase: Only historical samples whose target prediction interval has been fully revealed before the current physical time are used to update the parameters of the prediction model to obtain the updated prediction model.

[0010] Optionally, step 1 includes symbol definition: Let the time series of multiple marine elements continuously collected by the marine monitoring system be denoted as: ; in, Indicates time Ocean multi-element observation vectors, Indicates the number of observed variables; For the current physical moment The construction length is History input window: ; Length is The future target window is defined as: ; In batch processing, the input tensor is defined as: ; in, Indicates batch size, Indicates the length of the playback. Indicates the number of variables; Let the overall prediction model be: ; in, Represents the current physical time. The model parameters below, Indicating a view on the future Prediction results of multiple ocean elements within a time step; The core constraint is: at the current physical moment Before making a prediction, online updates are only allowed using historical samples that have been fully revealed for the target interval. Step prediction task, current physical time The latest sample used for updating is: ; At the current physical moment All are now observable.

[0011] Optionally, the multi-scale adaptive decomposition module in step 2 is used to normalize the input sequence. Decomposed into trend components Non-trend components The multi-scale adaptive decomposition module includes multiple smooth scales, a scale weight generation network, a weighted trend fusion process, trend base terms, non-trend base terms, and decomposition strength parameters. set up Each of the following smoothing scales corresponds to a smoothing kernel or moving average window. , obtained the Candidate trend sequences at various scales: ; in, Indicates window size as Moving average operation; For the The first sample The scale weight generation network generates weights for each smooth scale based on its historical sequence of variables: ; in, This represents a scale weight generation network; Based on weighted trend fusion, it can be written as: ; in, For trend-based items, non-trend-based items are represented as follows: ; Introducing decomposition strength parameters ,in Then the first The first sample The trend and non-trend components of each variable are constructed as follows: ; .

[0012] Optionally, in step 3, the trend component is input into the trend prediction branch, and variable-wise trend extrapolation is performed to obtain the basic trend prediction result; under the constraint of sparse intervariate interaction mask, the intervariate correction is performed on the trend prediction branch features corresponding to the basic trend prediction result, and the trend side reliability gating coefficient is generated according to the current sample trend context information to constrain the correction process and obtain the trend prediction result. The non-trend component is input into the time-domain / frequency-domain expert collaboration module. The routing module selects the target expert from multiple time-domain experts with different segment resolutions and multiple frequency-domain experts with different segment resolutions. The outputs of the target experts are then fused. Under the constraint of sparse intervariate interaction mask, the interaction between variables is performed to obtain the non-trend prediction result.

[0013] Optionally, the intervariate interaction mask generation module includes real-valued fast Fourier transform, Mahalanobis distance calculation, and Gumbel-Bernoulli sampling to generate sparse intervariate interaction masks that can be called by both the trend prediction branch and the non-trend prediction branch. The non-trend prediction branch includes a dimension rearrangement module, a routing module, a time-domain / frequency-domain expert collaboration module, a superimposed variable embedding module, a feature fusion module, a cross-variable interaction module, and a non-trend prediction module; the trend prediction branch includes a relative trend representation module, a channel-independent trend extrapolation module, and a trend-side cross-variable correction module; among them, the dimension rearrangement module is used to transform the input representation by... Rearranged as The time-domain expert set and frequency-domain expert set in the time-domain / frequency-domain expert collaboration module each include multiple expert modules with different segment resolutions. The time-domain expert set is represented as follows: The frequency domain expert set is represented as The routing module is used to select a target expert from experts of different resolutions to participate in the prediction based on the current sample features.

[0014] Optionally, under the constraint of sparse intervariate interaction mask, the interaction between variables is performed. The variable feature input, query projection, key projection and value projection together constitute the input of the interaction correlation calculation. The interaction correlation calculation, mask constraint, weight normalization, interaction weight, feature weighted aggregation and output mapping together constitute the variable feature generation process after the interaction. For the The first sample First, construct a frequency domain description based on the normalized input for each variable: ; in, Represents the real-valued Fast Fourier Transform; Then calculate the relationship distance based on the differences between the two variables: ; in, For learnable projection matrix; Convert relational distance to affinity: ; The interaction probability is obtained after normalization. And generate sparse interaction masks. : ; ; in, This indicates Gumbel-Bernoulli sampling; the retention probability of diagonal elements, i.e., variables that are self-connected, is set to be greater than that of off-diagonal elements to ensure that each variable retains its own information path; set up The variable feature input in the branch is Query projection, key projection, and value projection are represented as follows: ; in, , , These are the query projection matrix, the key projection matrix, and the value projection matrix, respectively. The interaction correlation calculation is expressed as: ; in, Indicates the scaling dimension; Subsequently, a sparse intervariate interaction mask is used to mask the interaction correlation, and the interaction weights are obtained by weight normalization. The interaction weights are expressed as: ; ; in, Interaction weights; Based on feature weighted aggregation and output mapping, the interactive variable features for: ; in, This indicates an output mapping operation; the result of cross-variable interaction is defined as: ; in, This represents a module for calculating interactions between variables that are subject to mask constraints.

[0015] Optionally, the trend prediction branch first constructs a relative trend representation using the relative trend representation module. This module relativizes the trend components, using the trend value at the last moment within the current window as a reference to highlight the change in the future trend relative to the current state. Subsequently, the channel-independent trend extrapolation module performs variable-wise extrapolation on the trend components, for the first... The first sample Given 1 variable, the basic trend extrapolation is defined as: ; in, Indicates the first Trend mapping matrix of variables; Indicates the bias term; Indicates the first The first sample The trend value of each variable at the last moment within the current window; The trend-side intervariate correction module, under the constraint of sparse intervariate interaction mask, obtains the constrained trend correction term. Among them, constraint processing includes constraints based on sparse intervariate interaction masks and norm pruning or magnitude limiting of the correction term; Introducing a trend-side reliability gating factor: ; The basic gating is represented as follows: ; The sample-related prior gating is represented as: ; in, For the Sigmoid function; For learnable global gating parameters; This is the lower limit of the gate; , The learnable coefficient; A risk score calculated based on the current trend context; The final trend prediction result is: ; in, This represents the trend correction term for the nth variable in the bth sample.

[0016] Optionally, step 4 includes: Perform instance-level normalization for each sample and each variable separately, for the input tensor , its first The first sample The mean and standard deviation of each variable are defined as follows: ; ; The output of the normalized prediction is: ; After the prediction is completed, the original dimensions are restored by inverse normalization: ; in, Indicates the prediction step index; Given a history input window The overall prediction process is then described as follows: ; ; ; ; ; ; in, This indicates an instance-level normalization operation; This represents a multi-scale adaptive decomposition module; Indicates a non-trend forecast branch; Indicates the trend prediction branch; This represents the sparse intervariate interaction mask corresponding to the current sample; The branch fusion result corresponds to the fusion output of trend prediction results and non-trend prediction results in the normalized space, that is... The branch fusion results are processed by the inverse normalization module to obtain the final prediction results.

[0017] Optionally, step 5 includes: A strict causal online update mechanism is adopted. In the strict causal unlabeled leakage update method, the evaluation and model update performed at the current physical moment correspond to: historical samples that have fully revealed the target interval before the current physical moment, that is, evaluating the historical prediction interval that has been fully revealed and updating using the historical samples that have been fully revealed. First, at the current physical moment The update rules for the model parameters are as follows: ; in, This indicates an online parameter update operation; Next, the updated parameters are used to predict the current sample: ; because Includes those not yet in the current physical time The previously fully revealed information, and its relation to the current forecast target There is overlap; if used directly... A full update would result in tag leakage, therefore it is not allowed at the current physical moment. Use directly before prediction Update the model.

[0018] The technical solution provided by this invention includes the following steps: acquiring the historical input window of multiple ocean elements at the current moment, using it as the current prediction sample to perform instance-level normalization processing to obtain a normalized input sequence; decomposing the normalized input sequence into trend components and non-trend components through a multi-scale adaptive decomposition module; constructing a sparse intervariate interaction mask based on an intervariate interaction mask generation module for joint use by the trend prediction branch and the non-trend prediction branch to obtain trend prediction results and non-trend prediction results; fusing the trend prediction results and non-trend prediction results to obtain the output of the normalized prediction; performing inverse normalization processing on the output of the normalized prediction to obtain the final prediction result; and in the online update stage, using only historical samples whose target prediction interval has been fully revealed before the current physical moment to update the parameters of the prediction model to obtain the updated prediction model. This method enables continuous prediction of future changes in multiple ocean elements, thereby improving the accuracy, stability, and deployment consistency of online prediction of multiple ocean elements. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart of a method for ocean multi-factor time series prediction in strictly causal online scenarios provided in an embodiment of the present invention; Figure 2 A flowchart of another method for ocean multi-factor time series prediction in strictly causal online scenarios provided by an embodiment of the present invention; Figure 3 This is a schematic diagram of a multi-scale adaptive decomposition module provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a time-domain / frequency-domain expert collaboration module provided in an embodiment of the present invention, wherein (a) is a time-domain expert module and (b) is a frequency-domain expert module; Figure 5 This is a schematic diagram of a cross-variable interaction module provided in an embodiment of the present invention; Figure 6 This is a comparative diagram of the online update mechanism provided in the embodiments of the present invention, wherein (a) is an update method with tag leakage and (b) is an update method with strict causality and no tag leakage. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” used in the embodiments of this invention are also intended to include the plural forms unless the context clearly indicates otherwise.

[0023] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0024] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0025] This invention provides a method for predicting multi-factor time series data of the ocean in strictly causal online scenarios, such as... Figure 1 and Figure 2 As shown, the method includes: Oceanographic data such as temperature, salinity, turbidity, and chlorophyll concentration collected by ocean buoys, stations, and other ocean monitoring platforms typically encompass long-term background evolution, periodic variations, and short-term local disturbances. The coupling relationships between different elements and different stations also dynamically change with sea state and time. To address these characteristics, this invention separates trend and non-trend components using a multi-scale adaptive decomposition module, models local dynamics and periodic spectral structures using a time-domain / frequency-domain expert collaboration module, dynamically filters more valuable element interactions using a sparse intervariate interaction mask, and suppresses unreliable intervariate corrections under drastic sea state changes through trend-side reliability gating. This improves the accuracy, stability, and causal consistency of online predictions for multiple oceanographic elements.

[0026] Specifically, changes in marine elements are constrained by natural laws such as air-sea exchange, tidal processes, runoff transport, mixing and diffusion, biochemical processes, light variations, and local hydrodynamic processes. Changes at different time scales and the coupling relationships between multiple variables have clear physical and environmental significance. Therefore, the data processed in this invention belongs to monitoring data with definite technical meaning in the technical field, and the prediction and updating methods adopted in this invention also belong to a set of technical means oriented towards marine technology scenarios.

[0027] Step 1: Obtain the historical input window of the ocean multi-element at the current moment, and use it as the current prediction sample to perform instance-level normalization processing to obtain the normalized input sequence.

[0028] In this embodiment of the invention, marine multi-element refers to marine environmental observation data such as temperature, salinity, turbidity, chlorophyll concentration, dissolved oxygen, pH value, and current velocity continuously collected by marine buoys, shore-based stations, nearshore observation platforms, or other marine monitoring equipment.

[0029] In this embodiment of the invention, symbol definitions are included before step 1: Let the time series of multiple marine elements continuously collected by the marine monitoring system be denoted as: ; in, Indicates time Ocean multi-element observation vectors, This indicates the number of observed variables; variables include marine monitoring elements such as seawater temperature, salinity, turbidity, and chlorophyll concentration, and can also be further extended to observations collected synchronously at other stations.

[0030] In this embodiment of the invention, the variable can represent either different marine monitoring elements at the same site, or the joint observation dimension after expanding the site dimension and the element dimension. Therefore, the cross-variable interaction can correspond to the interaction between different marine elements, or the interaction between different observation dimensions after jointly expanding the site and the element.

[0031] For the current physical moment The construction length is History input window: ; Length is The future target window is defined as: ; In batch processing, the input tensor is defined as: ; in, Indicates batch size, Indicates the length of the playback. Indicates the number of variables; Let the overall prediction model be: ; in, Represents the current physical time. The model parameters below, Indicating a view on the future Prediction results of multiple ocean elements within a time step; The core constraint is: at the current physical moment Before making a prediction, online updates are only allowed using historical samples that have been fully revealed for the target interval. Step prediction task, current physical time The latest sample used for updating is: ; At the current physical moment All are now observable.

[0032] Step 2: Decompose the normalized input sequence into trend components and non-trend components using the multi-scale adaptive decomposition module.

[0033] In embodiments of the present invention, such as Figure 3 As shown, in step 2, the multi-scale adaptive decomposition module is used to normalize the input sequence. Decomposed into trend components Non-trend components The multi-scale adaptive decomposition module includes multiple smoothing scales (such as...). Figure 3 The components include smoothing scale 1, smoothing scale 2, smoothing scale 3, scale weight generation network, weighted trend fusion process, trend base terms, non-trend base terms, and decomposition strength parameters. set up Each of the following smoothing scales corresponds to a smoothing kernel or moving average window. , obtained the Candidate trend sequences at various scales: ; in, Indicates window size as Moving average operation; For the The first sample The scale weight generation network generates weights for each smooth scale based on its historical sequence of variables: ; in, This represents a scale weight generation network; Based on weighted trend fusion, it can be written as: ; in, For trend-based items, non-trend-based items are represented as follows: ; Introducing decomposition strength parameters ,in Then the first The first sample The trend and non-trend components of each variable are constructed as follows: ; .

[0034] Therefore, oceanographic variables with more pronounced trends can be analyzed through larger... A more thorough decomposition can be performed, while variables that change rapidly or have weak trends can be decomposed through smaller decompositions. Avoid excessive decomposition.

[0035] Step 3: Construct a sparse intervariate interaction mask based on the intervariate interaction mask generation module, so that the trend prediction branch and the non-trend prediction branch can call it together to obtain the trend prediction result and the non-trend prediction result.

[0036] In this embodiment of the invention, in step 3, the trend component is input into the trend prediction branch, and variable-wise trend extrapolation is performed to obtain the basic trend prediction result. Under the constraint of sparse intervariate interaction mask, the intervariate correction is performed on the trend prediction branch features corresponding to the basic trend prediction result, and the trend side reliability gating coefficient is generated according to the current sample trend context information to constrain the correction process and obtain the trend prediction result. The non-trend component is input into the time-domain / frequency-domain expert collaboration module. The routing module selects the target expert from multiple time-domain experts with different segment resolutions and multiple frequency-domain experts with different segment resolutions. The outputs of the target experts are then fused. Under the constraint of sparse intervariate interaction mask, the interaction between variables is performed to obtain the non-trend prediction result.

[0037] In this embodiment of the invention, the non-trend prediction branch is used to process local fluctuations, oscillations, and periodic residual information in the input sequence, such as... Figure 4 As shown in (a) and (b), both the time-domain expert module and the frequency-domain expert module receive non-trend component inputs and output time-domain expert features and frequency-domain expert features, respectively. In one embodiment, the time-domain expert set and the frequency-domain expert set are not single-resolution expert sets, but are composed of multiple experts with different segment resolutions to correspond to the ocean multi-element change patterns at different time granularities.

[0038] In the context of multi-element marine scenarios, time-domain features are used to describe local short-term changes and event-driven processes, while frequency-domain features are used to describe spectral structures such as tidal cycles, diurnal cycles, and periodic mixing and transport processes. Therefore, the aforementioned time-domain / frequency-domain collaborative modeling is used to simultaneously characterize local dynamics and periodic processes.

[0039] In this embodiment of the invention, for the non-trend component of each variable The input is divided into segments of varying lengths to form a multi-resolution input representation. Based on this multi-resolution input representation, both a time-domain expert set and a frequency-domain expert set are constructed. Let the time-domain expert output be: ; The frequency domain expert output is: ; in, Indicates the first A time-domain expert; Indicates the first Each frequency domain expert; the expert output dimension is... , Indicates the dimension of the hidden features; exist Figure 4 In the time-domain expert module shown in (a), the non-trend component input is first processed by segmentation, then the time-domain representation is constructed through the time-domain feature mapping module, then enters the two-stage feature mixing module, and finally the corresponding time-domain expert output is generated through output mapping. Figure 4 (a) The segmentation process in the example employs different segment length settings to support multi-resolution temporal feature extraction. In some embodiments, the two-stage feature blending module includes intra-segment feature blending and inter-segment feature blending, used to model local variation patterns within a single segment and correlation information between different segments, respectively.

[0040] exist Figure 4 In the frequency domain expert module shown in (b), the non-trend component input first undergoes segmentation processing, then a frequency domain representation is constructed through the frequency domain feature mapping module, and further segment frequency domain transformation processing is performed to form a spectral feature representation for use by the frequency domain expert. Subsequently, the spectral features enter the two-stage feature mixing module, and the corresponding frequency domain expert output is generated through output mapping. Figure 4 The segment segmentation process in (b) also employs different segment length settings to support multi-resolution frequency domain feature extraction. The segment frequency domain transformation process may include performing a real-valued fast Fourier transform on the segment and combining the real and imaginary parts of the transform result.

[0041] The routing module generates routing scores for both the time-domain expert set and the frequency-domain expert set. ; ; The corresponding routing probability is: ; ; in, This represents the temperature parameter.

[0042] Suppose that experts are selected from the time-domain expert set and the frequency-domain expert set respectively. Individual and If there are 1 target expert, the corresponding target expert index set is denoted as: ; Subsequently, the selected expert outputs are normalized and weighted, and the joint normalization constant is: ; The corresponding weights are: ; ; The result of fusion: ; ; In this embodiment of the invention, under the action of the dimension rearrangement module and the routing selection module, the time-domain expert output and the frequency-domain expert output are respectively sent to the fusion path of the time-domain expert set and the frequency-domain expert set; the superimposed variable embedding module is used to introduce learnable identification information for different variables to enhance the distinguishability of different marine elements or different station variables under shared expert conditions; let the first... The learnable embedding vectors corresponding to the variables are: Then, the time-domain features and frequency-domain features after superimposed variable embedding are respectively: ; ; The feature fusion module is used to fuse the time-domain features and frequency-domain features after embedding the superimposed variables. The fused features are represented as follows: ; in, Indicates the first The first sample The fusion features corresponding to each variable. All samples and their corresponding variables. Stacking is performed to obtain the overall fusion features of the non-trend prediction branches. To maintain consistency with the overall representation of the aforementioned non-trend prediction branch, the non-trend prediction branch features are represented as follows: ; Subsequently, the fused features interact with each other under the constraints of the intervariate interaction module, and finally obtain the non-trend prediction result through the non-trend prediction module: ; in, This represents the forecast mapping operation for non-trend forecast branches.

[0043] In this embodiment of the invention, the intervariate interaction mask generation module includes real-valued fast Fourier transform, Mahalanobis distance calculation and Gumbel-Bernoulli sampling, used to generate a sparse intervariate interaction mask that can be called by both the trend prediction branch and the non-trend prediction branch. The non-trend prediction branch includes a dimension rearrangement module, a routing module, a time-domain / frequency-domain expert collaboration module, a superimposed variable embedding module, a feature fusion module, a cross-variable interaction module, and a non-trend prediction module; the trend prediction branch includes a relative trend representation module, a channel-independent trend extrapolation module, and a trend-side cross-variable correction module; among them, the dimension rearrangement module is used to transform the input representation by... Rearranged as The time-domain expert set and frequency-domain expert set in the time-domain / frequency-domain expert collaboration module each include multiple expert modules with different segment resolutions. The time-domain expert set is represented as follows: The frequency domain expert set is represented as The routing module is used to select a target expert from experts of different resolutions to participate in the prediction based on the current sample features.

[0044] In embodiments of the present invention, such as Figure 5 As shown, it is not assumed that all ocean elements have equal importance at any given time. Intervariate interactions are performed under sparse intervariate interaction mask constraints. Variable feature inputs, query projections, key projections, and value projections together constitute the input for interaction correlation calculation. Intervariate correlation calculation, mask constraints, weight normalization, interaction weights, feature weighting aggregation, and output mapping together constitute the variable feature generation process after the interaction. The sparse intervariate interaction mask output by the sparse intervariate interaction mask generation module is denoted as... ,in, Let be a general representation of the sparse intervariate interaction mask; for the sparse intervariate interaction mask generated for the current sample, let be denoted as . In the current sample context, yes A specific example, both referring to the same type of sparse intervariate interaction mask.

[0045] In multi-element marine scenarios, the coupling relationships between different variables and different sites vary with sea state and time. For example, the correlation between temperature and chlorophyll concentration is stronger at certain times, while the coupling between salinity and turbidity is more pronounced at other times. Dynamically filtering interaction paths is used to reduce noise propagation caused by invalid interactions.

[0046] For the The first sample First, construct a frequency domain description based on the normalized input for each variable: ; in, Represents the real-valued Fast Fourier Transform; Then calculate the relationship distance based on the differences between the two variables: ; in, For learnable projection matrix; Convert relational distance to affinity: ; The interaction probability is obtained after normalization. And generate sparse interaction masks. It is represented in the form of matrix elements: ; ; in, This indicates Gumbel-Bernoulli sampling. In other embodiments, thresholding, sorting truncation, and other methods can also be used to generate sparse cross-variable interaction masks. The retention probability of diagonal elements, i.e., variable self-connections, is set to be greater than that of off-diagonal elements to ensure that each variable retains its own information path. set up The variable feature input in the branch is Query projection, key projection, and value projection are represented as follows: ; in, , , These are the query projection matrix, the key projection matrix, and the value projection matrix, respectively. The interaction correlation calculation is expressed as: ; in, Indicates the scaling dimension; Subsequently, a sparse intervariate interaction mask is used to mask the interaction correlation, and the interaction weights are obtained by weight normalization. The interaction weights are expressed as: ; ; in, Interaction weights; This is the mask matrix for the current sample; Based on feature weighted aggregation and output mapping, the interactive variable features for: ; in, This indicates an output mapping operation; the result of cross-variable interaction is defined as: ; in, This represents a module for calculating interactions between variables that are subject to mask constraints. This interaction module can be used in both non-trend prediction branches and trend prediction branches.

[0047] In this embodiment of the invention, the trend prediction branch first constructs a relative trend representation through the relative trend representation module; the relative trend representation module is used to relativize the trend components, that is, to use the trend value at the last moment within the current window as a reference benchmark to highlight the amount of change in the future trend relative to the current state; subsequently, the channel-independent trend extrapolation module performs variable-wise extrapolation on the trend components, for the first... The first sample Given 1 variable, the basic trend extrapolation is defined as: ; in, Indicates the first Trend mapping matrix of variables; Indicates the bias term; Indicates the first The first sample The trend value of each variable at the last moment within the current window; By subtracting the trend value at the last moment, the model can learn the relative amount of change in the future, rather than the absolute numerical level.

[0048] The trend-side intervariate correction module, under the constraint of sparse intervariate interaction mask, obtains the constrained trend correction term. The constraint processing includes constraints based on sparse intervariate interaction masks and norm pruning or amplitude limiting of the correction term to suppress the excessive influence of unreliable interactions on trend prediction results. Since the trend correction term directly affects future trend values, and the interaction risk corresponding to the current trend context may increase with changes in sea state; Introducing a trend-side reliability gating factor: ; The basic gating is represented as follows: ; The sample-related prior gating is represented as: ; in, For the Sigmoid function; For learnable global gating parameters; This is the lower limit of the gate; , The learnable coefficient; A risk score calculated based on the current trend context; The final trend prediction result is: ; in, This represents the trend correction term for the nth variable in the bth sample.

[0049] Through the above gating mechanism, when the current ocean state is relatively stable and the interaction confidence is high, the cross-variable correction term plays a significant role; while when local disturbances increase, sea state changes drastically, or the risk score is high, the influence of the correction term is automatically weakened to improve the stability of trend prediction.

[0050] Step 4: Combine the trend forecast results with the non-trend forecast results to obtain the normalized forecast output; perform inverse normalization processing on the normalized forecast output to obtain the final forecast result.

[0051] In this embodiment of the invention, step 4 includes: Perform instance-level normalization for each sample and each variable separately, for the input tensor , its first The first sample The mean and standard deviation of each variable are defined as follows: ; ; The output of the normalized prediction is: ; After the prediction is completed, the original dimensions are restored by inverse normalization: ; in, Indicates the prediction step index; Given a history input window The overall prediction process is then described as follows: ; ; ; ; ; ; in, This indicates an instance-level normalization operation; This represents a multi-scale adaptive decomposition module; Indicates a non-trend forecast branch; Indicates the trend prediction branch; This represents the sparse intervariate interaction mask corresponding to the current sample; The branch fusion result corresponds to the fusion output of trend prediction results and non-trend prediction results in the normalized space, that is... The branch fusion results are processed by the inverse normalization module to obtain the final prediction results.

[0052] Step 5, Online Update Phase: Only historical samples whose target prediction interval has been fully revealed before the current physical time are used to update the parameters of the prediction model to obtain the updated prediction model.

[0053] In this embodiment of the invention, strict causal online update means that before making a prediction at the current physical moment, only historical samples that have been fully revealed in the target interval are allowed to be used to update the model, and real label data that overlaps with the target interval of the subsequent prediction is not allowed to be used in advance, so as to ensure that the prediction and update process meets strict causal constraints and avoid label leakage.

[0054] In embodiments of the present invention, such as Figure 6 In (a) under the condition of adjacent overlapping prediction windows, the model update is performed using the complete true label data of the just-evaluated sample, which causes the label interval used for updating to overlap with the subsequent prediction target interval, thus forming label leakage; Figure 6 (b) of this model only evaluates fully revealed historical prediction intervals and updates model parameters using only fully revealed historical samples; therefore, it does not use real labels that overlap with subsequent prediction targets in advance. Historical input data, prediction results, and real label data correspond to the historical multi-factor data observed before the current time, the model's prediction results for the future target interval, and the subsequently revealed real observation labels, respectively. Figure 6 The label leakage in (a) comes from the direct use of the complete true label data of the just evaluated sample when updating the model, and a part of the complete true label data still overlaps with the target interval of the next prediction. Figure 6 The strict causal unlabeled leakage update method in (b) requires that the historical sample be used for evaluation and model update only after the relevant target interval has been fully revealed.

[0055] For ease of understanding Figure 6 The longitudinal time progression process will Figure 6 Each set of horizontal dashed lines in the diagram can be interpreted as an online prediction, evaluation, or model update action. Figure 6 In (a) of the text, at physical time The model is based on to Using historical data as input, predictions are generated based on the historical data, and the output is... to The prediction results; when the process progresses to At that time, if immediately to The prediction result is evaluated using complete real label data, and the model is updated accordingly. , , The true label in physical moment Not yet actually arrived; further progress to At that time, the updated model will again include... to The subsequent target interval is predicted based on historical data, therefore the data used in the previous model update step is... to The label overlaps with the subsequent predicted target interval, forming Figure 6 The label (a) in the text indicates a leak. Similarly, in If model updates continue to be performed using future labels that have not yet been fully revealed, it will also lead to subsequent... , The predicted location is affected by the pre-labeled information.

[0056] In contrast, Figure 6 In (b) of the above, at physical time Still based on to Historical data prediction to The future range, that is, predictions generated based on historical data; when the process progresses to... At that time, the process shown in the figure involves evaluating fully revealed historical forecast intervals and updating using fully revealed historical samples, for example, using... to As a historical sample of complete, true label ranges, rather than an evaluation or update of just now. Output in real time to Forecast range; continue to advance to hour, to Only when all true labels are observable can the fully revealed historical prediction intervals be evaluated and updated using the fully revealed historical samples. Therefore, Figure 6 (b) time, All vertical processes, including timelines, satisfy the sequential constraint of first fully revealing and then evaluating and updating.

[0057] In this embodiment of the invention, step 5 includes: A strict causal online update mechanism is adopted. In the strict causal unlabeled leakage update method, the evaluation and model update performed at the current physical moment correspond to the historical samples that have fully revealed the target interval before the current physical moment. That is, the evaluation of the historical prediction interval that has been fully revealed and the update using the historical samples that have been fully revealed are not the samples that have just been predicted at the current moment and have not yet fully revealed the target interval. First, at the current physical moment The update rules for the model parameters are as follows: ; in, This indicates an online parameter update operation; Next, the updated parameters are used to predict the current sample: ; because Includes those not yet in the current physical time The previously fully revealed information, and its relation to the current forecast target There is overlap; if used directly... A full update would result in tag leakage, therefore it is not allowed at the current physical moment. Use directly before prediction Update the model.

[0058] In this embodiment of the invention, the loss function and offline initialization are as follows: In the offline initialization phase, this invention pre-trains the overall model using prediction loss, with the training objective expressed as: ; in, This represents the prediction error loss, including mean squared error loss, mean absolute error loss, or a combination thereof; This represents the load balancing regular expression for expert routing; Indicates the balance coefficient; Initial parameters were obtained through offline training. Then, incremental updates are performed online according to strict causal rules.

[0059] The following is a relatively complete example of a marine scenario, but the present invention is not limited thereto.

[0060] Suppose a coastal marine monitoring system continuously collects multi-element marine observation data at multiple buoys and stations. The observed variables include seawater temperature, salinity, turbidity, chlorophyll concentration, dissolved oxygen, pH value, and current velocity, totaling [data missing]. One variable. Input window length. Predicted length At that moment When making predictions, the system first takes the multivariate observations from the most recent 96 moments to form a prediction. , and perform instance-level normalization on it according to the variable.

[0061] Subsequently, settings were configured in the multi-scale adaptive decomposition module. Using smooth scales, such as scale lengths of 13, 25, and 49, three candidate trend sequences are obtained. A scale weight generation network generates three scale weights for each variable based on the temporal pattern of the current window, and then weights and merges the three candidate trend sequences into a trend base term. This is then combined with the variable-level decomposition intensity parameter. Output trend components Non-trend components .

[0062] At the same time, the strength of the relationship between variables is calculated based on the frequency domain description information of the normalized input, and a sparse intervariate interaction mask for the current sample is generated. In the intervariate interaction module, the system constructs query features, key features, and value features respectively, and obtains the interacted variable features through interaction relevance calculation, mask constraints, weight normalization, and feature weighted aggregation. The sparse intervariate interaction mask... It serves as a public mask, simultaneously used for feature interactions between variables in the non-trend prediction branch and for trend correction calls in the trend prediction branch.

[0063] Next, time-domain / frequency-domain expert collaborative prediction is performed on the non-trend components. The number of time-domain experts is set to 4, and the number of frequency-domain experts is also set to 4. These time-domain and frequency-domain experts can correspond to different segment resolutions, for example, to the non-trend input representations constructed under different segment lengths. The routing module generates time-domain expert selection scores and frequency-domain expert selection scores based on the non-trend input of each variable, selects two target experts from each set to participate in the calculation, and performs normalized weighted fusion on the target expert outputs to obtain the non-trend prediction branch fusion features. Then, in the sparse intervariate interaction mask... Under constraints, the non-trend prediction branch fusion features are subjected to variable interactions, and the non-trend prediction result is obtained through the non-trend prediction module. .

[0064] In the trend forecasting branch, trend extrapolation is performed independently for each variable to obtain the basic trend forecast results. Then, under the constraint of the interactive mask, the trend correction term is calculated. Estimate risk score based on current trend context. Regenerate the trend-side reliability gating coefficient The correction term is scaled and then merged with the basic trend prediction result to obtain the trend prediction result. .

[0065] Finally, the trend forecast results and non-trend forecast results are added in the normalized space, and the final ocean multi-factor forecast output is obtained by inverse normalization. .

[0066] Regarding online updates, since the prediction step size for the current task is... Therefore, at physical time Before making a prediction, only the following methods are allowed: Perform online updates to model parameters, without allowing the use of The complete labeling is updated. This ensures that the online prediction process meets strict causal constraints and avoids label leakage.

[0067] In optional embodiments, the present invention may also have the following variations: The number of smoothing scales, the length of each smoothing window, and the scale weight generation network structure in the multi-scale adaptive decomposition module are adjusted according to the specific sea area, observation frequency, and target elements. The number of time-domain experts and frequency-domain experts, the network structure, and the input representation are configured according to the characteristics of marine monitoring data. In addition to frequency-domain amplitude, variable description information also uses time-domain statistics, spectral projection, correlation coefficient features, or combinations thereof. The generation method of the cross-variable interaction mask adopts sparsity strategies such as Gumbel-Bernoulli sampling, probability sampling, threshold screening, and sorting truncation. The trend-side reliability gating adopts gating methods based on volatility, energy distribution, rate of change, residual size, or other risk indicators. The trend prediction branch and the non-trend prediction branch share the interaction mask and generate different masks respectively. Online updates can be executed either hourly or in batches at multiple time points under strict causal constraints. In addition to being applicable to multi-element prediction of marine buoys and stations, this invention can also be extended to other environmental monitoring, nearshore ecological monitoring, or related multivariate technology monitoring scenarios.

[0068] The core of this invention lies in constructing an online prediction framework for ocean multi-element observation data, which simultaneously exhibits multi-timescale variations, dynamic intervariate coupling, and delayed feedback updates. This framework comprises a multi-scale adaptive decomposition module, a time-domain / frequency-domain expert collaboration module, a sparse intervariate interaction module, and a trend-side reliability gating module. By decomposing the input sequence into trend and non-trend components and performing targeted modeling for each, combined with a rigorous causal online update mechanism, this framework achieves continuous prediction of future changes in multiple ocean elements while avoiding label leakage.

[0069] Compared with the prior art, the present invention has the following advantages: 1. By adopting a strict causal online update mechanism, we avoid using future label information in advance in overlapping prediction scenarios, ensuring that the online prediction process of multiple marine elements meets the actual deployment conditions; 2. By using a multi-scale adaptive decomposition module, the long-term background evolution, periodic changes, and short-term local disturbances in marine elements are separated and modeled, which is beneficial to improving the ability to represent complex marine environmental change processes. 3. By using the time-domain / frequency-domain expert collaboration module, local dynamic patterns and periodic spectral structures can be captured simultaneously, improving the adaptability to scenarios where tides, diurnal variations, and local anomalies coexist. 4. By using the intervariate interaction module and sparse intervariate interaction mask, more valuable marine elements or inter-site interaction paths are dynamically filtered to reduce the redundancy and noise impact caused by fixed full-connection interaction. 5. Improve the stability and robustness of trend forecasting under conditions of rapid sea state changes by using a trend-side cross-variable correction module and a trend-side reliability gating mechanism; 6. The method of this invention can achieve online prediction of multiple marine elements under strict causal label-free leakage constraints, and has good application prospects for marine environmental monitoring and early warning.

[0070] The technical solution provided by this invention includes the following steps: acquiring the historical input window of multiple ocean elements at the current moment, using it as the current prediction sample to perform instance-level normalization processing to obtain a normalized input sequence; decomposing the normalized input sequence into trend components and non-trend components through a multi-scale adaptive decomposition module; constructing a sparse intervariate interaction mask based on an intervariate interaction mask generation module for joint use by the trend prediction branch and the non-trend prediction branch to obtain trend prediction results and non-trend prediction results; fusing the trend prediction results and non-trend prediction results to obtain the output of the normalized prediction; performing inverse normalization processing on the output of the normalized prediction to obtain the final prediction result; and in the online update stage, using only historical samples whose target prediction interval has been fully revealed before the current physical moment to update the parameters of the prediction model to obtain the updated prediction model. This method enables continuous prediction of future changes in multiple ocean elements, thereby improving the accuracy, stability, and deployment consistency of online prediction of multiple ocean elements.

[0071] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for forecasting multi-factor time series data of oceans in strictly causal online scenarios, characterized in that, The method includes: Step 1: Obtain the historical input window of the ocean multi-element at the current moment, and use it as the current prediction sample to perform instance-level normalization processing to obtain the normalized input sequence; Step 2: Decompose the normalized input sequence into trend components and non-trend components using the multi-scale adaptive decomposition module; Step 3: Construct a sparse intervariate interaction mask based on the intervariate interaction mask generation module, so that the trend prediction branch and the non-trend prediction branch can call it together to obtain the trend prediction result and the non-trend prediction result. Step 4: Fuse the trend forecast results with the non-trend forecast results to obtain the normalized forecast output; perform inverse normalization processing on the normalized forecast output to obtain the final forecast result. Step 5, Online Update Phase: Only historical samples whose target prediction interval has been fully revealed before the current physical time are used to update the parameters of the prediction model to obtain the updated prediction model.

2. The method according to claim 1, characterized in that, The step 1 is preceded by symbol definition: Let the time series of multiple marine elements continuously collected by the marine monitoring system be denoted as: ; in, Indicates time Ocean multi-element observation vectors, Indicates the number of observed variables; For the current physical moment The construction length is History input window: ； Length is The future target window is defined as: ； In batch processing, the input tensor is defined as: ; in, Indicates batch size, Indicates the length of the playback. Indicates the number of variables; Let the overall prediction model be: ; in, Represents the current physical time. The model parameters below, Indicating a view on the future Prediction results of multiple ocean elements within a time step; The core constraint is: at the current physical moment Before making a prediction, online updates are only allowed using historical samples that have been fully revealed for the target interval. Step prediction task, current physical time The latest sample used for updating is: ; At the current physical moment All are now observable.

3. The method according to claim 2, characterized in that, In step 2, the multi-scale adaptive decomposition module is used to normalize the input sequence. Decomposed into trend components Non-trend components ; The multi-scale adaptive decomposition module includes multiple smooth scales, a scale weight generation network, a weighted trend fusion process, trend base terms, non-trend base terms, and decomposition strength parameters. set up Each of the following smoothing scales corresponds to a smoothing kernel or moving average window. , obtained the Candidate trend sequences at various scales: ； in, Indicates window size as Moving average operation; For the The first sample The scale weight generation network generates weights for each smooth scale based on its historical sequence of variables: ； in, This represents a scale weight generation network; Based on weighted trend fusion, it can be written as: ； in, For trend-based items, non-trend-based items are represented as follows: ; Introducing decomposition strength parameters ,in Then the first The first sample The trend and non-trend components of each variable are constructed as follows: ； 。 4. The method according to claim 3, characterized in that, In step 3, the trend component is input into the trend prediction branch, and variable-wise trend extrapolation is performed to obtain the basic trend prediction result. Under the constraint of sparse intervariate interaction mask, the intervariate correction is performed on the trend prediction branch features corresponding to the basic trend prediction results, and the trend side reliability gating coefficient is generated according to the current sample trend context information to constrain the correction process and obtain the trend prediction results. The non-trend component is input into the time-domain / frequency-domain expert collaboration module. The routing module selects the target expert from multiple time-domain experts with different segment resolutions and multiple frequency-domain experts with different segment resolutions. The outputs of the target experts are then fused. Under the constraint of sparse intervariate interaction mask, the interaction between variables is performed to obtain the non-trend prediction result.

5. The method according to claim 4, characterized in that, The intervariate interaction mask generation module includes real-valued fast Fourier transform, Mahalanobis distance calculation, and Gumbel-Bernoulli sampling, used to generate sparse intervariate interaction masks that can be used by both the trend prediction branch and the non-trend prediction branch. The non-trend prediction branch includes a dimension rearrangement module, a routing module, a time-domain / frequency-domain expert collaboration module, a superimposed variable embedding module, a feature fusion module, a cross-variable interaction module, and a non-trend prediction module; the trend prediction branch includes a relative trend representation module, a channel-independent trend extrapolation module, and a trend-side cross-variable correction module; among them, the dimension rearrangement module is used to transform the input representation by... Rearranged as The time-domain expert set and frequency-domain expert set in the time-domain / frequency-domain expert collaboration module each include multiple expert modules with different segment resolutions. The time-domain expert set is represented as follows: The frequency domain expert set is represented as The routing module is used to select a target expert from experts of different resolutions to participate in the prediction based on the current sample features.

6. The method according to claim 4, characterized in that, Under the constraint of sparse intervariable interaction mask, the interaction between variables is performed. The variable feature input, query projection, key projection and value projection together constitute the input of the interaction correlation calculation. The interaction correlation calculation, mask constraint, weight normalization, interaction weight, feature weighted aggregation and output mapping together constitute the variable feature generation process after the interaction. For the first The first sample First, construct a frequency domain description based on the normalized input for each variable: ； in, Represents the real-valued Fast Fourier Transform; Then calculate the relationship distance based on the differences between the two variables: ； in, For learnable projection matrix; Convert relational distance to affinity: ； The interaction probability is obtained after normalization. And generate sparse interaction masks. : ； ； in, This indicates Gumbel-Bernoulli sampling; the retention probability of diagonal elements, i.e., variables that are self-connected, is set to be greater than that of off-diagonal elements to ensure that each variable retains its own information path; set up The variable feature input in the branch is Query projection, key projection, and value projection are represented as follows: ； in, , , These are the query projection matrix, the key projection matrix, and the value projection matrix, respectively. The interaction correlation calculation is expressed as: ； in, Indicates the scaling dimension; Subsequently, a sparse intervariate interaction mask is used to mask the interaction correlation, and the interaction weights are obtained by weight normalization. The interaction weights are expressed as: ； ； in, Interaction weights; Based on feature weighted aggregation and output mapping, the interactive variable features for: ； in, This indicates an output mapping operation; the result of cross-variable interaction is defined as: ； in, This represents a module for calculating interactions between variables that are subject to mask constraints.

7. The method according to claim 4, characterized in that, The trend prediction branch first constructs a relative trend representation through the relative trend representation module; The relative trend representation module is used to relativize trend components, that is, to use the trend value at the last moment in the current window as a reference benchmark to highlight the amount of change in the future trend relative to the current state. Subsequently, the channel-independent trend extrapolation module performs variable-wise extrapolation on the trend components, for the first... The first sample Given 1 variable, the basic trend extrapolation is defined as: ； in, Indicates the first Trend mapping matrix of variables; Indicates the bias term; Indicates the first The first sample The trend value of each variable at the last moment within the current window; The trend-side intervariate correction module, under the constraint of sparse intervariate interaction mask, obtains the constrained trend correction term. Among them, constraint processing includes constraints based on sparse intervariate interaction masks and norm pruning or magnitude limiting of the correction term; Introducing a trend-side reliability gating factor: ； The basic gating is represented as follows: ； The sample-related prior gating is represented as: ； in, For the Sigmoid function; For learnable global gating parameters; This is the lower limit of the gate; , The learnable coefficient; A risk score calculated based on the current trend context; The final trend prediction result is: ； in, This represents the trend correction term for the nth variable in the bth sample.

8. The method according to claim 4, characterized in that, Step 4 includes: Perform instance-level normalization for each sample and each variable separately, for the input tensor , its first The first sample The mean and standard deviation of each variable are defined as follows: ； ； The output of the normalized prediction is: ； After the prediction is completed, the original dimensions are restored by inverse normalization: ； in, Indicates the prediction step index; Given a history input window The overall prediction process is then described as follows: ； ； ； ； ； ； in, This indicates an instance-level normalization operation; This represents a multi-scale adaptive decomposition module; Indicates a non-trend forecast branch; Indicates the trend prediction branch; This represents the sparse intervariate interaction mask corresponding to the current sample; The branch fusion result corresponds to the fusion output of trend prediction results and non-trend prediction results in the normalized space, that is... The branch fusion results are processed by the inverse normalization module to obtain the final prediction results.

9. The method according to claim 8, characterized in that, Step 5 includes: A strict causal online update mechanism is adopted. In the strict causal unlabeled leakage update method, the evaluation and model update performed at the current physical moment correspond to: historical samples that have fully revealed the target interval before the current physical moment, that is, evaluating the historical prediction interval that has been fully revealed and updating using the historical samples that have been fully revealed. First, at the current physical moment The update rules for the model parameters are as follows: ； in, This indicates an online parameter update operation; Next, the updated parameters are used to predict the current sample: ； because Includes those not yet in the current physical time The previously fully revealed information, and its relation to the current forecast target There is overlap; if used directly... A full update would result in tag leakage, therefore it is not allowed at the current physical moment. Use directly before prediction Update the model.

Images