An Adaptive Prediction and Correction Method for Non-Stationary Time Series Data Distribution Drift

By constructing shuffled samples and explicitly maintaining a training distribution prototype, and using a dynamic gating network and error buffer to calculate the correction magnitude, the problems of insufficient distribution drift perception and information leakage in power time series forecasting are solved, achieving efficient adaptive prediction correction and improving the stability and accuracy of forecasts.

CN123115599APending Publication Date: 2026-11-17HUNAN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611291927.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-25
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing online adaptive prediction methods cannot accurately detect the training-test distribution shift in power time series scenarios, leading to a decline in prediction performance. Furthermore, there are issues of information leakage and inappropriate correction magnitude in multi-step prediction.

Method used

By constructing a sample training set, shuffling current and historical samples, calculating the correction magnitude using a dynamic gating network and error buffer, explicitly maintaining the training distribution prototype, generating correction input and output sequences, and performing adaptive prediction correction.

Benefits of technology

It significantly improves the long-term stability and accuracy of non-stationary power time series forecasting, is suitable for efficient deployment at the edge, and solves the defects and information leakage problems of distributed drift sensing.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115599A_ABST
    Figure CN123115599A_ABST
Patent Text Reader

Abstract

This application provides an adaptive prediction and correction method for non-stationary time-series data distribution drift, comprising: acquiring a shuffled current sample and a corresponding shuffled historical sample; determining a current correction magnitude based on the shuffled historical sample and the shuffled current sample; determining a correction input sequence based on the shuffled current sample, the shuffled historical sample, and the current correction magnitude; and determining a final prediction based on the correction input sequence, the shuffled historical sample, and the current correction magnitude. The method proposed in this application provides a quantitative benchmark for drift perception by explicitly maintaining the prototype features of the training distribution, effectively solving the deficiency of existing methods in being unable to perceive the global offset between test data and the training domain; significantly improving the long-term stability and accuracy of non-stationary power time-series prediction, and is suitable for efficient deployment at edge computing environments.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the fields of time series forecasting and machine learning technology, and in particular to an adaptive forecasting and correction method for the drift of non-stationary time series data distribution. Background Technology

[0002] Time series forecasting plays a crucial role in areas such as power load dispatching, renewable energy generation forecasting, and power grid security early warning. For example, in power transformer monitoring, it is necessary to predict future trends several hours or even tens of hours in advance based on historical data from multiple sensors, such as oil temperature and load current, to support power grid dispatching decisions. Deep neural network models typically achieve high accuracy by training with large amounts of historical offline data. However, in actual power systems, data distribution often changes slowly or abruptly over time—electricity load is affected by seasonal changes and fluctuations in economic activity, and transformer characteristics drift with aging. Traditional offline models have fixed parameters and struggle to adapt to these non-stationary characteristics, leading to a continuous decline in predictive performance over time. Therefore, there is an urgent need for online adaptive forecasting methods that can dynamically adjust based on real-time data feedback.

[0003] However, current online adaptive forecasting methods face three key problems that urgently need to be solved in the context of power time series scenarios.

[0004] First, existing methods generally ignore the systematic shift between the test data and the original training distribution. Most online adaptive strategies only fine-tune the model or apply empirical corrections to the predicted values ​​based on the prediction errors of the most recent steps, lacking explicit modeling and preservation of the data distribution characteristics during the training phase. When the power grid enters a completely new operating mode, the conceptual characteristics of the test samples may deviate significantly from the distribution center of the training set. Existing methods cannot quantitatively perceive "how far the current state is from the training domain," leading to blind adjustments based solely on local error signals, lacking a global sense of shift direction, and resulting in significantly insufficient adaptability.

[0005] Second, in multi-step prediction scenarios, the prediction model typically outputs the sequence for the next H time steps at once, while the true labels are only available after H time steps. This H-step delay makes any adaptive mechanism based on true labels inherently lag behind actual data changes. If the most recently available label is used directly for correction to simplify implementation, it actually breaks the temporal causality, leaking future information into the correction of the past. This not only distorts offline evaluation metrics but may also introduce completely wrong correction directions when the distribution shifts rapidly. On the other hand, if the model strictly waits for H-step labels, the distribution changes during the lag period have accumulated to a significant extent, and the model is always in a state of "chasing the drift." Especially when the prediction step size H is large, the performance degradation caused by the lag is more severe.

[0006] Third, existing online adaptive methods often use fixed-amplitude or uniform-proportion corrections for forecast drift. However, the impact of drift in real power systems is highly non-uniform—larger corrections are needed during periods of severe system fluctuation to avoid serious deviations, while only slight adjustments are required during stable operation. Uniform correction amplitudes are prone to overcorrection or undercorrection: excessive correction may amplify noise and compromise forecast stability, while insufficient correction may fail to keep up with changes in distribution, severely affecting the robustness of forecast accuracy.

[0007] The three issues mentioned above collectively make it difficult for existing online adaptive methods to accurately detect training-test distribution shifts and apply appropriate granular dynamic corrections while strictly adhering to temporal causality and maintaining low computational overhead. This has become a key bottleneck restricting the long-term reliable deployment of power time-series forecasting models in real non-stationary environments. Summary of the Invention

[0008] This application provides an adaptive prediction and correction method for the distribution drift of non-stationary time-series data. To solve the above-mentioned technical problems, this application adopts the following technical methods: This application provides an adaptive prediction and correction method for the distribution drift of non-stationary time series data, including: Obtain the shuffled current sample and the corresponding shuffled historical sample; Based on the shuffled historical samples and the shuffled current samples, determine the current correction magnitude; Based on the shuffling of the current sample, the shuffling of the historical sample, and the current correction magnitude, a corrected input sequence is determined; The final prediction is determined based on the corrected input sequence, the shuffled historical samples, and the current correction magnitude.

[0009] Optionally, the step of obtaining the shuffled current sample and the corresponding shuffled historical sample includes: A sample training set is constructed based on historical power time-series data; Based on the training set, sample pairs are truncated at a preset step size, invalid sample pairs in the truncated sample set are discarded, and valid sample pairs are retained; The valid sample pairs are randomly shuffled at the batch level to generate the current shuffled sample and the corresponding shuffled historical sample.

[0010] Optionally, determining the current correction magnitude based on the shuffled historical samples and the shuffled current samples includes: Based on the shuffled historical samples and the shuffled current samples, determine the dynamic adaptation index; Based on the aforementioned dynamic adaptation index, the current correction magnitude is determined.

[0011] Optionally, determining the dynamic adaptation index based on the shuffled historical samples and the shuffled current samples includes: The shuffled historical samples are input into the backbone network to calculate a preliminary historical prediction. The backbone network is a time prediction deep neural network. The shuffled current samples are input into the backbone network to calculate the current preliminary prediction. Based on the aforementioned preliminary historical predictions, the historical mean square error of the shuffled historical samples is calculated. Based on the current preliminary prediction, the current mean square error of shuffling the current samples is calculated; The historical mean square error is inserted into the error buffer, and the mean and standard deviation of the error in the error buffer are calculated. Based on the current mean square error, the mean error, and the standard deviation, a dynamic adaptation index is determined.

[0012] Optionally, determining the current correction magnitude based on the dynamic adaptation index includes: The dynamic adaptation index is input into a multilayer perceptron for nonlinear feature extraction and mapping. After superimposing a learnable bias, it is normalized by the Sigmoid activation function and then multiplied by a preset maximum correction magnitude to obtain the current correction magnitude.

[0013] Optionally, determining the correction input sequence based on the shuffled current sample, the shuffled historical samples, and the current correction magnitude includes: The original input in the shuffled current sample is input to the current concept projector at the input end to obtain the current input concept vector; The shuffled historical samples are input into the input historical concept projector to obtain a historical input concept vector. Obtain a training distribution prototype, which is the distribution center to be optimized, and is used to perceive the global offset of the training domain. Based on the current input concept vector, the historical input concept vector, and the training distribution prototype, a corrected input sequence is determined.

[0014] Optionally, determining the corrected input sequence based on the current input concept vector, the historical input concept vector, and the training distribution prototype includes: The first complementary information is obtained by subtracting the current input concept vector from the training distribution prototype. The difference between the current input concept vector and the historical input concept vector is used to obtain the second complementary information; The first complementary information and the second complementary information are fused to generate an input enhanced offset signal; The input enhancement drift signal is input to the input-side adapter and calculated with the current correction amplitude to generate an input correction mask; The input correction mask is multiplied element-wise with the original input in the scrambled current sample to generate the corrected input sequence.

[0015] Optionally, the update process of the training distribution prototype includes: Obtain the input concept vectors for all samples from the current concept projector at the input end; The batch mean is determined based on the input concept vectors of all the samples. Based on the batch mean, the training distribution prototype is updated step by step using an exponential moving average.

[0016] Optionally, determining the final prediction based on the corrected input sequence, the shuffled historical samples, and the current correction magnitude includes: The correction input sequence is input into the backbone network to obtain the preliminary correction prediction; The corrected preliminary prediction is input to the current concept projector at the output end and mapped to obtain the current output concept vector. The shuffled historical samples are input into the output historical concept projector to obtain the historical output concept vector. The final prediction is determined based on the current output concept vector, the historical output concept vector, and the training distribution prototype.

[0017] Optionally, determining the final prediction based on the current output concept vector, the historical output concept vector, and the training distribution prototype includes: The third complementary information is obtained by subtracting the current output concept vector from the training distribution prototype. The fourth complementary information is obtained by subtracting the current output concept vector from the historical output concept vector. The third complementary information and the fourth complementary information are fused to generate an output enhanced offset signal; The output enhancement offset signal is input to the output-side adapter and calculated with the current correction magnitude to generate an output correction mask. The output correction mask is multiplied element-wise with the initial correction prediction to determine the final prediction.

[0018] This application has the following beneficial effects: The method proposed in this application provides a quantitative benchmark for drift perception by explicitly maintaining the prototype features of the training distribution, effectively solving the defect of existing methods that cannot perceive the global offset between test data and training domain; it significantly improves the long-term stability and accuracy of non-stationary power time series prediction, and is suitable for efficient deployment at the edge. Brief Description of the Drawings

[0019] Figure 1 is a schematic flowchart of an adaptive prediction and correction method oriented to distribution drift of non-stationary time series data provided by the embodiments of the present application. Detailed Description of the Embodiments

[0020] To facilitate the understanding of those skilled in the art, the present application will be further described below with reference to the embodiments and the accompanying drawings, and the content mentioned in the embodiments is not intended to limit the present application.

[0021] To solve the above technical problems, as Figure 1 shown, the present application proposes an adaptive prediction and correction method oriented to distribution drift of non-stationary time series data, comprising: Step S101: Obtain shuffled current samples and corresponding shuffled historical samples; In the present application, a sample training set is first constructed using power time series data, and each sample includes an input window X∈R C×L and a future window Y∈R C×H , where C is the number of sensor channels, L is the length of the look-back window, and H is the length of the prediction window. Any deep neural network for time series prediction is selected as the backbone network f θ (such as Transformer, NLinear, etc.), standard training is performed with the mean square error as the loss function, the parameter θ is saved after convergence, and f is frozen in all subsequent stages θ so that it is only used for forward inference and does not participate in any parameter update.

[0022] To train the subsequent concept drift perception module, paired current samples and historical samples H steps before need to be constructed. That is, based on the above sample training set, for any time t, sample pairs are intercepted according to a preset step size H: for any time t, the current sample (X t0 ,Y t0 ) and the corresponding historical sample are obtained, invalid sample pairs with t<H in the intercepted sample set are discarded, valid sample pairs are retained, all valid sample pairs are collected and then randomly shuffled at the batch level to generate shuffled current samples (X t ,Y t ) and corresponding shuffled historical samples . The purpose of shuffling is to break the original temporal proximity, so that samples of different seasons and different operation modes may be included in the same batch, generating rich and diverse concept drift patterns, forcing the adapter to learn a general drift-correction mapping, and improving the generalization ability.

[0023] Step S102: Determine the current correction amplitude based on the shuffled historical samples and the shuffled current samples; To replace the fixed correction amplitude and achieve adaptive behavior of large correction for large drifts and small correction for small drifts, this application introduces a dynamic gating network and an error buffer. The error buffer B is a fixed-capacity K first-in-first-out queue used to store the prediction mean square error of the K most recent completed samples. When the historical samples are shuffled... After the labels are fully revealed, preliminary historical predictions are calculated using the frozen backbone network. The mean square error e of the shuffled historical samples is obtained by performing a mean square calculation on the preliminary historical prediction. t-H Insert the mean square error into the error buffer B, and calculate the mean value μ of the buffer error. B and standard deviation This will shuffle the current sample (X). t ,Y t Input the backbone network and calculate the current preliminary forecast. The mean square error e of the current preliminary prediction is obtained by performing a mean square calculation on the current sample. t This will further reduce the current preliminary prediction mean square error e t The dynamic adaptation index z is defined by calculating the mean and standard deviation of the shuffled historical samples mentioned above. t for: (1) To prevent division by zero for extremely small positive numbers. When the number of samples in the buffer is less than 2, let z t =0. z t The degree of anomaly in the current prediction error relative to recent historical levels is quantified: z t ≈0 indicates that the error is within the normal range and no additional correction is needed; z t A significant increase implies a worsening of the distribution drift, requiring a more substantial correction.

[0024] The dynamic gating network D is a lightweight multilayer perceptron, which uses the scalar z... t Mapped to the current correction magnitude δ t The specific process is as follows: The dynamic adaptation index z t Input to Multilayer Perceptron (MLP) D After nonlinear feature extraction and mapping, and superposition of learnable biases, the current correction magnitude is obtained by normalization using the Sigmoid activation function and then multiplied by a preset maximum correction magnitude. As shown in the following formula: (2) Among them, MLP DIt contains a 64-dimensional hidden layer and a ReLU activation function, where σ is the sigmoid function, g0 is a learnable bias initialized to 0, and β... max This is the preset maximum correction magnitude. The resulting δ... t ∈(0,β max The correction intensity can be dynamically adjusted according to the degree of real-time error anomaly.

[0025] Step S103: Determine the correction input sequence based on the shuffled current sample, the shuffled historical sample, and the current correction magnitude; The input adapter generates an input correction mask using the concept drift signal, applying a bounded multiplicative adjustment to the current input. The concept projector compresses the original temporal window into a compact concept vector to describe the semantic features of the temporal patterns of each variable within that window. The input adapter contains two structurally identical but parameter-independent concept projectors: The original input X in the current sample will be shuffled. t ∈R C×L Input to the current concept projector Φ x curr The mapping yields the current input concept vector c. t x ∈R C×dc This will disrupt historical patterns. Input to the historical concept projector Φ x hist The historical input concept vector is obtained by mapping. .

[0026] The internal operations of the above-mentioned conceptual projectors are all the same, with Φ x curr For example, x: (3) In this case, Linear1 maps an L-dimensional window of each variable to a d-dimensional window. c 1D concept space, Linear2 maintains the same dimensionality, GELU is the activation function, d c As a conceptual dimension, each variable is processed independently during the process, therefore c t x Each line is independently encoded, corresponding to the current state mode of the variable.

[0027] To introduce awareness of the global offset of the training domain into the drift signal at the input end, a distribution center to be optimized needs to be pre-defined. This step will train the distribution prototype c. train ∈R C×dc Initialize it as an all-zero tensor and set an uninitialized flag. At this point, c... trainAs a placeholder for drift calculation, its value will be accumulated from the output of the concept projector via EMA after joint training is initiated.

[0028] The first complementary information is obtained by subtracting the current input concept vector from the training distribution prototype; the second complementary information is obtained by subtracting the current input concept vector from the historical input concept vector; and then the first and second complementary information are fused to generate the input enhancement offset signal S. t x ∈R C×2dc As shown in the following formula: (4) Among them, the first complementary information Measuring the absolute deviation of the current concept from the training distribution prototype, used to perceive global distribution shift; second complementary information. It measures the relative drift of the current concept with respect to the history H steps prior, and is used to perceive local temporal changes.

[0029] Input enhanced drift signal S t x Input to input-side adapter G x , and the current correction magnitude Perform calculations to generate the input correction mask M. t x ∈R C×L The calculation process is as follows: (5) Among them, W1 x As a fully connected layer, it compresses the input dimension to the bottleneck dimension d. bottleneck σ is the Sigmoid activation; W2 x As a fully connected layer, the bottleneck dimension is mapped to the window length L, and its weights are initialized with zero to ensure that the mask is always 1 in the early stage of training, starting safely from the original prediction; The amplitude control factor is corrected by the dynamic gating network in the third step. Each element of the output mask falls within... Within the interval, a bounded multiplicative adjustment is performed on the input sequence.

[0030] Multiply the input mask element-wise with the original input in the shuffled current sample to obtain the corrected input sequence: (6) The above training distribution prototype needs to be updated in different training batches. The update process is as follows: Current concept projector from the input end Φ x curr Obtain the concept vector c of all samples t xThen, the batch mean is calculated based on the concept vector. ∈R C×dc Then, c is updated incrementally using the exponential moving average (EMA). train : (7) The momentum coefficient *m* is set to 0.99, causing the prototype to slowly approach the center of the entire training set's concept distribution. This update is performed in a mode with gradient computation disabled, without generating backpropagation gradients; it is only used for state maintenance of the prototype variables. In the first training batch, if the prototype is marked as uninitialized, it is used directly. For c train Assign a value and toggle the flag to its initial value. At the end of training, c train It is saved along with the other model parameters and then frozen, serving as a fixed anchor point for measuring the degree to which the current data state deviates from the training domain during the online phase.

[0031] Step S104: Determine the final prediction based on the corrected input sequence, the shuffled historical samples, and the current correction magnitude.

[0032] The above correction input sequence Input backbone network f θ A corrected preliminary prediction was obtained. Construct the current concept projector Ψ at the output end. y curr and output-side historical concept projector Ψ y hist Its internal structure is completely consistent with the conceptual projector in step S103.

[0033] The initial prediction will be corrected and input to the current concept projector Ψ at the output. y curr The mapping yields the current output concept vector c. t y ∈R C×dc This will disrupt historical patterns. Input to output historical concept projector Ψ y hist Mapping yields historical output concept vectors .

[0034] The current output concept vector c t y The third complementary information is obtained by subtracting from the training distribution prototype; the current output concept vector c is then processed. t y With historical output concept vector The fourth complementary information is obtained by subtracting the first; the third and fourth complementary information are then fused to generate the output enhanced offset signal S.t y As shown in the following formula: (8) Construct the output-side adapter G y Its structure is similar to G x The only difference is that the window length is replaced with H instead of L, and the output enhancement offset signal is input to the output-side adapter G. y The current correction magnitude is used to calculate and generate the output correction mask M. t y ∈R C×H : (9) In the formula, W1 y As a fully connected layer, it compresses the input dimension to the bottleneck dimension d. bottleneck W2 y For a fully connected layer, the bottleneck dimension is mapped to the window length L.

[0035] Output correction mask M t y Preliminary predictions with correction Element-by-element multiplication determines the final prediction: (10) The final prediction and the true label Y are calculated using Huber loss. t The error between the training set and the target set is reduced to a linear loss when the error is large, enhancing training robustness. Backpropagation is performed based on this loss, and the Adam optimizer is used to update all trainable parameters, including the parameters of the four concept projectors, W1 and W2 in the two adapters, and the parameters of the dynamic gating network, until the validation set loss converges or early stopping occurs. Finally, all module parameters and the training distribution prototype ctrain are saved. Steps S101 to S104 above are all pre-trained on the training set, providing all the necessary model components and reference prototypes for online inference.

[0036] In summary, the method proposed in this application provides a quantitative benchmark for drift perception by explicitly maintaining the prototype features of the training distribution, effectively addressing the limitation of existing methods in failing to perceive the global offset between test data and the training domain. Simultaneously, it utilizes historical samples across prediction step size H to calculate drift, accurately compensating for the inherent feedback delay of multi-step predictions without relying on future labels, thus preventing information leakage. Furthermore, a dynamic gating mechanism adaptively adjusts the correction intensity based on real-time prediction feedback, avoiding overcorrection and undercorrection. The entire online adaptation process requires only freezing the basic prediction model parameters and performing forward computation on a lightweight adapter, resulting in extremely low computational overhead. This significantly improves the long-term stability and accuracy of non-stationary power time series prediction, making it suitable for efficient deployment at edge computing.

[0037] In some embodiments, this application also provides a computer system including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0038] This application also provides a computer-readable storage medium for storing a computer program. This computer-readable storage medium can be applied to a computer device, and the computer program causes the computer device to execute the corresponding processes in the methods described above in the embodiments of this application; for brevity, further details are omitted here.

[0039] The above embodiments are preferred implementations of this application. In addition, this application can be implemented in other ways. Any obvious substitutions without departing from the concept of this technical solution are within the protection scope of this application.

[0040] To facilitate understanding by those skilled in the art of the improvements made by this application compared to the prior art, some of the accompanying drawings and descriptions have been simplified, and for clarity, some other elements have been omitted from this application. Those skilled in the art should realize that these omitted elements may also constitute the content of this application.

Claims

1. An adaptive prediction and correction method for non-stationary time-series data distribution drift, characterized in that, include: Obtain the shuffled current sample and the corresponding shuffled historical sample; Based on the shuffled historical samples and the shuffled current samples, determine the current correction magnitude; Based on the shuffling of the current sample, the shuffling of the historical sample, and the current correction magnitude, a corrected input sequence is determined; The final prediction is determined based on the corrected input sequence, the shuffled historical samples, and the current correction magnitude.

2. The method according to claim 1, characterized in that, Before the step of obtaining the shuffled current sample and the corresponding shuffled historical sample, the following steps are included: A sample training set is constructed based on historical power time-series data; Based on the training set, sample pairs are truncated at a preset step size, invalid sample pairs in the truncated sample set are discarded, and valid sample pairs are retained; The valid sample pairs are randomly shuffled at the batch level to generate the current shuffled sample and the corresponding shuffled historical sample.

3. The method according to claim 2, characterized in that, The step of determining the current correction magnitude based on the shuffled historical samples and the shuffled current samples includes: Based on the shuffled historical samples and the shuffled current samples, determine the dynamic adaptation index; Based on the aforementioned dynamic adaptation index, the current correction magnitude is determined.

4. The method according to claim 3, characterized in that, The determination of dynamic adaptation indicators based on the shuffled historical samples and the shuffled current samples includes: The shuffled historical samples are input into the backbone network to calculate a preliminary historical prediction. The backbone network is a time prediction deep neural network. The shuffled current samples are input into the backbone network to calculate the current preliminary prediction. Based on the aforementioned preliminary historical predictions, the historical mean square error of the shuffled historical samples is calculated. Based on the current preliminary prediction, the current mean square error of shuffling the current samples is calculated; The historical mean square error is inserted into the error buffer, and the mean and standard deviation of the error in the error buffer are calculated. Based on the current mean square error, the mean error, and the standard deviation, a dynamic adaptation index is determined.

5. The method according to claim 4, characterized in that, The determination of the current correction magnitude based on the dynamic adaptation index includes: The dynamic adaptation index is input into a multilayer perceptron for nonlinear feature extraction and mapping. After superimposing a learnable bias, it is normalized by the Sigmoid activation function and then multiplied by a preset maximum correction magnitude to obtain the current correction magnitude.

6. The method according to claim 5, characterized in that, The step of determining the correction input sequence based on the shuffled current sample, the shuffled historical sample, and the current correction magnitude includes: The original input in the shuffled current sample is input to the current concept projector at the input end, and mapped to obtain the current input concept vector; The shuffled historical samples are input into the input historical concept projector to obtain a historical input concept vector. Obtain a training distribution prototype, which is the distribution center to be optimized, and is used to perceive the global offset of the training domain. Based on the current input concept vector, the historical input concept vector, and the training distribution prototype, a corrected input sequence is determined.

7. The method according to claim 6, characterized in that, The step of determining the corrected input sequence based on the current input concept vector, the historical input concept vector, and the training distribution prototype includes: The first complementary information is obtained by subtracting the current input concept vector from the training distribution prototype. The difference between the current input concept vector and the historical input concept vector is used to obtain the second complementary information; The first complementary information and the second complementary information are fused to generate an input enhanced offset signal; The input enhancement drift signal is input to the input-side adapter and calculated with the current correction amplitude to generate an input correction mask; The input correction mask is multiplied element-wise with the original input in the scrambled current sample to generate the corrected input sequence.

8. The method according to claim 6, characterized in that, The update process of the training distribution prototype includes: Obtain the input concept vectors for all samples from the current concept projector at the input end; The batch mean is determined based on the input concept vectors of all the samples. Based on the batch mean, the training distribution prototype is updated step by step using an exponential moving average.

9. The method according to claim 6, characterized in that, The step of determining the final prediction based on the corrected input sequence, the shuffled historical samples, and the current correction magnitude includes: The correction input sequence is input into the backbone network to obtain the preliminary correction prediction; The corrected preliminary prediction is input to the current concept projector at the output end and mapped to obtain the current output concept vector. The shuffled historical samples are input into the output historical concept projector to obtain the historical output concept vector. The final prediction is determined based on the current output concept vector, the historical output concept vector, and the training distribution prototype.

10. The method according to claim 9, characterized in that, The step of determining the final prediction based on the current output concept vector, the historical output concept vector, and the training distribution prototype includes: The third complementary information is obtained by subtracting the current output concept vector from the training distribution prototype. The fourth complementary information is obtained by subtracting the current output concept vector from the historical output concept vector. The third complementary information and the fourth complementary information are fused to generate an output enhanced offset signal; The output enhancement offset signal is input to the output-side adapter and calculated with the current correction magnitude to generate an output correction mask. The output correction mask is multiplied element-wise with the initial correction prediction to determine the final prediction.