A time series data prediction method and related apparatus

Abstract

The present disclosure provides a prediction method of time series data and related devices, the method comprising: constructing a feature extractor based on original time series data; wherein the original time series data comprises source domain time series data and target domain time series data; processing the original time series data based on a cross-modal semantic anchoring method to obtain a semantic anchor point; optimizing the feature extractor based on the semantic anchor point to obtain an optimized feature extractor; and predicting and calibrating the target domain time series data to be predicted based on the optimized feature extractor and the semantic anchor point to obtain a prediction result. Through mode enhancement, cross-modal semantic anchoring, shared space contrastive learning and adaptive calibration in the reasoning stage, the present disclosure effectively improves the generalization performance and robustness of the time series prediction model in a heterogeneous environment.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and in particular to a method and apparatus for predicting time series data. Background Technology

[0002] With the rapid development of big data, artificial intelligence, and distributed sensing technologies, massive amounts of multivariate time-series data have accumulated during the operation of complex dynamic systems (such as smart grids, intelligent transportation, and large-scale environmental monitoring systems). Accurate prediction of time-series data is not only a core prerequisite for realizing intelligent resource scheduling and preventive maintenance, but also a key link in ensuring the stability of knowledge-driven decision-making systems.

[0003] However, in real-world applications, time-series prediction models face significant challenges from multi-source, heterogeneous environments. While existing deep learning models have achieved some success in processing stationary, single-domain data, they still present the following challenges when dealing with cross-domain migration and dynamic environmental changes:

[0004] Significant cross-domain feature distribution shift: Most existing time series prediction models are based on the stationarity assumption. When data originates from different acquisition devices, geographical regions, or operating environments, there are significant statistical distribution differences between the source and target domains. Traditional alignment methods only focus on fitting statistical quantities and lack the ability to capture the underlying logical patterns of time series data, resulting in a significant decrease in the model's prediction accuracy in the target domain.

[0005] Lack of cross-modal semantic association mechanisms: Time series data itself has rich physical meaning and trend characteristics, but existing methods usually treat time series data as simple numerical values, ignoring the deep semantic information behind them. This neglect of semantics prevents models from using external prior knowledge (such as generalization knowledge in large-scale pre-trained models) to help identify complex evolutionary patterns, resulting in limited model representation capabilities when dealing with sudden fluctuations or nonlinear trends.

[0006] The inference process lacks adaptive correction and uncertainty measurement: Traditional time series models often use fixed mapping logic during the deployment phase, making it difficult to dynamically adjust prediction results based on real-time input feedback. Especially when the environment fluctuates drastically, the model lacks a mechanism to evaluate the confidence of the output, which can easily lead to overfitting or incorrect prediction trends, and lacks robustness in complex dynamic environments.

[0007] Therefore, there is an urgent need to develop a time-series data prediction method based on cross-modal semantic anchoring and adaptive correction to solve the problems of feature distortion and generalization failure in heterogeneous environments.

[0008] It should be noted that the above introduction to the technical background is only for the purpose of providing a clear and complete explanation of the technical solutions disclosed herein, and for facilitating understanding by those skilled in the art. It should not be assumed that these technical solutions are known to those skilled in the art simply because they have been described in the background section of this disclosure. Summary of the Invention

[0009] In view of the above, the purpose of one or more embodiments of this disclosure is to provide a method and related apparatus for predicting time series data, so as to solve or partially solve the problems raised in the background art.

[0010] To achieve the above objectives, this disclosure provides a method for predicting time series data, the method comprising: A feature extractor is constructed based on the original time-series data; wherein, the original time-series data includes source domain time-series data and target domain time-series data; Based on the cross-modal semantic anchoring method, the original time-series data is processed to obtain semantic anchor points; Based on the semantic anchor points, the feature extractor is optimized to obtain an optimized feature extractor; Based on the optimized feature extractor and the semantic anchor point, the target domain time series data to be predicted is predicted and calibrated to obtain the prediction result.

[0011] Based on the same inventive concept, this disclosure also provides a prediction system for time series data, the system comprising: The construction module is configured to build a feature extractor based on the original time-series data; wherein the original time-series data includes source domain time-series data and target domain time-series data; The processing module is configured to process the original time-series data based on a cross-modal semantic anchoring method to obtain semantic anchor points; The optimization module is configured to optimize the feature extractor based on the semantic anchor point to obtain an optimized feature extractor. The prediction module is configured to predict and calibrate the target domain time series data to be predicted based on the optimized feature extractor and the semantic anchor point, and obtain the prediction result.

[0012] Based on the same inventive concept, this disclosure also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a method for predicting time-series data as described in any of the preceding claims.

[0013] Based on the same inventive concept, this disclosure also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute any of the time-series data prediction methods described above.

[0014] Based on the same inventive concept, this disclosure also provides a computer program product, including one or more computer programs, which, when executed by one or more processors, implement any of the above-described time-series data prediction methods.

[0015] This disclosure provides a prediction method for time-series data. It constructs a stable source-domain representation of the time-series data through a pattern-guided enhancement mechanism, reducing the impact of negative transfer. It utilizes a pre-trained large language model to extract high-level semantic features as anchor points, aligning the numerical features of the source and target domains with the semantic space. Cross-domain comparative learning is performed in the shared semantic space to eliminate domain differences. Residual correction and uncertainty adjustment mechanisms are introduced during the inference stage in the target domain, enabling adaptive prediction in dynamic environments. This disclosure effectively improves the generalization performance and robustness of time-series prediction in heterogeneous environments. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in one or more embodiments of this disclosure or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only one or more embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating a method for predicting time-series data provided in an embodiment of this disclosure; Figure 2 A schematic diagram of the structure of a time series data prediction system provided in an embodiment of this disclosure; Figure 3 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this disclosure clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0019] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of this disclosure should have the ordinary meaning understood by one of ordinary skill in the art to which this disclosure pertains. The terms "first," "second," and similar words used in one or more embodiments of this disclosure do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0020] As described in the background section, the relevant technologies have the following problems: I. Existing methods are difficult to effectively address the significant feature distribution shift between the source and target domains in cross-domain migration tasks. Due to the lack of in-depth capture of the underlying physical logic and evolution patterns of time series data, the generalization ability of the models is insufficient and the prediction accuracy is significantly reduced. Second, existing methods generally ignore the rich semantic information contained in time series data and fail to establish an effective cross-modal semantic association mechanism, thus failing to make full use of external prior knowledge to enhance the ability to identify and represent complex evolutionary laws and sudden fluctuations. Third, existing methods lack adaptive correction mechanisms based on real-time feedback and the ability to quantify prediction uncertainty during the inference stage, resulting in poor robustness in dynamically changing environments and a tendency to overfit or make incorrect trend judgments.

[0021] Therefore, there is an urgent need to develop a technical solution for predicting time-series data to address the issues of feature distortion and generalization failure in heterogeneous environments.

[0022] Based on some implementations of this disclosure, a scheme for predicting time-series data is provided, applicable to complex scenarios such as energy management, traffic scheduling, and weather forecasting. In this scheme, a stable time-series data representation in the source domain is constructed through a pattern-guided enhancement mechanism to reduce the impact of negative transfer. Subsequently, a pre-trained large language model is used to extract high-level semantic features as anchor points to align the numerical features of the source and target domains with the semantic space. Next, cross-domain comparative learning is performed in the shared semantic space to eliminate domain differences. Finally, residual correction and uncertainty adjustment mechanisms are introduced in the inference stage of the target domain to achieve adaptive prediction in dynamic environments. This disclosure effectively improves the generalization performance and robustness of time-series prediction in heterogeneous environments.

[0023] refer to Figure 1 The present disclosure provides a method for predicting time series data, comprising: Step S101: Construct a feature extractor based on the original time series data; wherein the original time series data includes source domain time series data and target domain time series data.

[0024] In some embodiments, this disclosure employs a multi-modal augmented stable feature encoding method to address the heterogeneous inconsistency between the source and target domains. By constructing a robust feature extractor with high pattern awareness and building diverse pattern augmentations in the temporal data space, the model is forced to learn domain-independent common features.

[0025] In some embodiments, constructing a feature extractor based on the original time-series data includes: Step S1011: Obtain and enhance the original time series data to obtain an enhanced sequence set.

[0026] The original time series data is constructed into a multivariate time series sequence, and a sample sequence is represented as follows:

[0027] in, Indicates the length of the input window. Indicates the feature dimension.

[0028] In some embodiments, the simulation addresses typical scenarios common in industrial IoT and environmental monitoring, such as equipment aging, sensor accuracy shifts, and operating condition switching. This disclosure utilizes the multivariate time series... Perform diverse data augmentations to generate corresponding augmented sequence sets. The enhancement is achieved through one or more of the following perturbation strategies: Frequency masking: In industrial monitoring, to address periodic signal loss caused by machine downtime or inconsistent sampling frequencies, a real-valued Fourier transform is performed on the signal. After random sampling and binary masking in the frequency domain, the signal is restored to simulate periodic anomalies. The calculation method is as follows:

[0029] in, For input multivariate time series, For real Fourier transform, For real numbers, the inverse Fourier transform is... A binary mask for random sampling. This method employs element-wise multiplication. It enables the encoder to reconstruct the core timing pattern without relying on complete frequency domain information.

[0030] Amplitude jitter: To simulate the effects of electromagnetic interference, quantization error, or random pulse noise on a real sensor, this disclosure injects... Distributed Gaussian noise is used to quantify the model's tolerance to sensor measurement errors. The calculation method is as follows:

[0031] in, For input multivariate time series, This is a Gaussian noise term. This represents the noise variance.

[0032] Correlation decoupling: To simulate the loss of inter-variable synergistic relationships due to local component failures in a multivariable system, this disclosure forcibly influences the dependency and synergistic relationships between features by linearly mixing the original sequence (i.e., the input multivariable time series sequence) and its sequence after random rearrangement along the feature dimension. The calculation method is as follows:

[0033] in, For input multivariate time series, The sequence after shuffling the feature dimensions. This is the mixing coefficient.

[0034] Trend Shift: To simulate feature shift caused by concept drift or long-term physical degradation of the system, this disclosure combines time-averaged pooling technology to introduce a slow drift of the trend term, simulating the long-term degradation mode of the system. The calculation method is as follows:

[0035] in, For input multivariate time series, This indicates a time-averaged pooling operation. This is the mixing coefficient.

[0036] Step S1012: Based on the dual encoder architecture, the original time series data and the enhanced sequence set are extracted and optimized to obtain a feature extractor.

[0037] In some embodiments, in order to maintain the smoothness of the temporal representation under the four types of enhanced perturbations described above, this disclosure employs a dual encoder architecture to extract features from the original temporal data.

[0038] The dual-encoder architecture includes a main encoder and a momentum encoder. The main encoder is responsible for extracting the feature vectors of the original time-series data. The momentum encoder is maintained as a moving average version of the parameters of the master encoder, aiming to generate a more stable representation of the target features. The exponential moving average update rule is as follows:

[0039] in, The parameter set of the main encoder. This is the parameter set for the momentum encoder. This is the momentum coefficient.

[0040] In the aforementioned momentum consistency learning process, the robustness of the feature encoder is ensured by co-optimizing the two objective functions: the contrast consistency loss and the pattern classification loss.

[0041] The Contrastive Consistency Loss eliminates domain-specific noise by narrowing the spatial distance between the original features extracted by the master encoder and the enhanced view features extracted by the momentum encoder. The calculation formula is as follows:

[0042] in: This represents the number of samples in the current training batch. The original sequence embedding vector output by the main encoder; It is a set of positive samples, containing stable representations of the momentum encoder output and corresponding enhanced view embeddings; The cosine similarity metric function is used. This is the preset temperature adjustment coefficient.

[0043] The Pattern Classification Loss, by introducing a lightweight pattern recognition head, forces the model to identify what kind of artificial perturbation (such as frequency masking or trend shift) was used on the current sample, thereby enhancing the model's ability to perceive subtle temporal deviations. The calculation formula is as follows:

[0044] in, The total number of preset enhancement mode categories (such as frequency masking, amplitude jitter, correlation decoupling, and trend shift). For the sample Experience the first Real labels for enhanced modes; The fused feature vector of the main encoder output and the enhanced view output; The pattern recognition head is based on features Predict this sample Belongs to the pattern category The conditional probability distribution.

[0045] By analyzing the contrast consistency loss and the pattern classification loss Perform joint optimization to determine the overall training objective. :

[0046] in, To compare the loss of consistency, For pattern classification loss, and These are adjustable weighting coefficients, used to balance feature stability and pattern sensitivity.

[0047] The overall training objective obtained through joint optimization Parameters of the main encoder Perform iterative optimization. Once the training process converges, the trained feature extractor is obtained.

[0048] In some embodiments, the multi-objective optimization strategy ensures that the model can still extract low-level data representations that are both discriminative and generalizable even in complex industrial disturbance environments.

[0049] In some embodiments, this disclosure constructs a pattern-guided representation learning mechanism to address the negative transfer phenomenon that easily occurs in cross-domain prediction. First, diverse enhancement results are generated by applying perturbation techniques such as frequency masking, amplitude jittering, correlation decoupling, and trend shifting to the original time-series data. Then, consistency constraints are constructed using the master encoder and momentum encoder to capture key temporal evolution features while removing noise interference specific to the source domain. This method ensures that the model can extract temporal data features that are domain-invariant and sensitive to pattern changes, laying a stable representation foundation for subsequent semantic alignment.

[0050] Step S102: Based on the cross-modal semantic anchoring method, the original time series data is processed to obtain semantic anchor points.

[0051] In some embodiments, this disclosure employs a cross-modal semantic anchoring method in a large language model space. Through cross-modal mapping, numerical patterns that are difficult to align across domains in traditional time series analysis are transformed into embedded representations with high-level semantics, thereby overcoming the limitation of insufficient robustness of simply relying on numerical statistical modeling when facing heterogeneous distributions.

[0052] In some embodiments, the cross-modal semantic anchoring method processes the original time-series data to obtain semantic anchor points, including: Step S1021: Semantic extraction is performed on the original time-series data to generate a language description.

[0053] In some embodiments, in response to the physical evolution logic contained in the time series, this disclosure does not directly process the original numerical values, but instead converts the numerical sequence into a standardized natural language description through a semantic extraction layer.

[0054] For the multivariate time series The system calculates local statistical characteristics (such as instantaneous slope, local extrema, energy distribution, and fluctuation frequency) within a time window, and uses a preset mapping method to convert these local statistical characteristics into standardized natural language descriptions. .

[0055] Wherein, the language description The generation process is as follows:

[0056] in, For the input of the first A multivariate time series sample; This is a fixed prefix for prompts, used to structure the final generated description; As a semantic mapper, the function can abstract the underlying numerical values ​​into text labels such as "sharp rise", "periodic oscillation" or "stable trend".

[0057] In some embodiments, through numerical abstraction, natural language is used as an intermediary bridge to unify heterogeneous numerical sequences with different dimensions and frequencies into the same logical description framework.

[0058] Step S1022: Based on the pre-trained large language model, extract the language description to obtain semantic anchor points.

[0059] In some embodiments, to incorporate generalized prior knowledge inherent in the large language model, this disclosure generates a structured language description. The input is fed into a pre-trained, parameter-frozen large language model. Its deep self-attention mechanism is used to encode features of the text description. The word embedding vectors at the end of the corresponding sequence in the last hidden state of the model are extracted and used as stable reference points for this temporal pattern in the semantic space, called semantic anchor points. The semantic anchor point Represented as:

[0060] in, Represents the raw time series data input. Coordinates or feature representations in the semantic space understood by large language models; For language description.

[0061] In some embodiments, during this process, all parameters of the large language model remain frozen (not involved in training), ensuring... As a stable semantic anchor, it avoids semantic shifts caused by parameter updates during model training.

[0062] Step S1023: Based on the semantic anchor point, perform semantic alignment constraints on the feature extractor, including: In some embodiments, the feature vectors generated by the temporal encoder (the main encoder in a dual-encoder architecture) With the semantic anchor points from the large language model They exist in different distribution spaces and cannot be directly compared.

[0063] Therefore, to achieve cross-modal alignment, this disclosure establishes a projection network composed of a multilayer sensing mechanism to project the feature vectors. Mapped to the semantic anchor point Projected features are obtained in the same semantic embedding space. :

[0064] in, For real-time feature vectors, For the learnable projection weights of the projection network, This is a non-linear activation function. Through this projection network, cross-modal dimensional alignment and feature projection are achieved, enabling numerical features to be used for measurement calculations alongside semantic features.

[0065] In some embodiments, to ensure that the numerical features learned by the temporal encoder can truly capture deep semantic logic, this disclosure introduces a semantic alignment loss function. The projection feature is required to be To the stable semantic anchor point in the semantic space To move closer.

[0066] The semantic alignment loss is calculated as follows:

[0067] in, For projection features, For semantic anchor points.

[0068] Through this loss constraint, the model can learn a "semantic-aware" representation capability. This means that even if the two time series in the source and target domains have significant deviations in numerical distribution, as long as their underlying physical meanings are consistent, they will have extremely high similarity in the projected space. This mechanism provides a feature guarantee of semantic consistency for subsequent cross-domain transfer.

[0069] In some embodiments, this disclosure introduces a large language model as a semantic reference anchor to extract the deeper semantic meaning of numerical features. First, the local statistical characteristics of the time series are transformed into a structured text description and mapped to the embedding space of the large language model to generate semantic anchor points. Then, by calculating the alignment loss between the time series embedding and the semantic anchor points, the numerical representation is forced to converge towards semantic features with logical interpretability. This mechanism effectively alleviates the semantic drift problem of numerical patterns and realizes the fusion and correction of cross-modal knowledge.

[0070] Step S103: Based on the semantic anchor point, optimize the feature extractor to obtain the optimized feature extractor.

[0071] In some embodiments, this disclosure utilizes a shared anchor-based cross-domain feature alignment method, employing a pre-constructed unified semantic space as a cross-modal bridge, aiming to overcome the distributional differences between the source and target domains and achieve highly reliable cross-domain knowledge transfer. By mapping temporal features from different domains to the same set of semantic anchors, this disclosure can effectively suppress negative transfer effects caused by differences in sensor specifications, operating environments, etc.

[0072] In some embodiments, optimizing the feature extractor based on the semantic anchor point to obtain an optimized feature extractor includes: Step S1031: Based on the semantic pre-correction method, process the source domain time series data and the target domain time series data respectively to obtain the corresponding semantic correction features.

[0073] In some embodiments, in order to eliminate the residual bias of temporal features after projection onto the semantic space, this disclosure introduces a semantic pre-correction method.

[0074] By calculating projection features With semantic anchors The difference between them is used to generate a compensation signal, and the feature vector is fine-tuned to obtain semantic correction features with higher purity. .

[0075] The semantic correction features Represented as

[0076] in, For feature vectors, For projection features, As semantic anchor points, These are learnable correction weights. This method allows the features of each sample to be pre-initialized with semantic alignment in the latent space before entering the prediction layer, thus enhancing the depth of feature representation.

[0077] Semantic pre-correction processing is performed on the source domain time series data and the source domain time series data respectively to obtain source domain correction features. and target domain correction features .

[0078] Step S1032: Calculate the inter-domain alignment loss based on the semantic correction features and the semantic anchor points.

[0079] In some embodiments, in order to further reduce the statistical distribution differences between the source domain and the target domain, this disclosure adopts a local alignment strategy centered on a unified semantic anchor point.

[0080] Calculate the source domain correction features respectively Correction features with the target domain to the semantic anchor point The distance is used to determine the inter-domain alignment loss. This eliminates distributional differences.

[0081] in, For source domain correction features, Correct features for the target domain. This is the semantic anchor point.

[0082] In some embodiments, this alignment method ensures that regardless of whether the data originates from the source domain or the target domain, as long as its internal physical logic is consistent, its projection position in the latent space will converge toward the same semantic center (anchor point), thereby establishing a domain-independent pattern association.

[0083] Step S1033: Calculate the cross-domain contrast loss based on the semantic correction features.

[0084] In some embodiments, to preserve subtle anomaly discrimination features while aligning distributions, this disclosure introduces a cross-domain contrastive learning mechanism. By constructing positive and negative sample pairs within a shared semantic space, the distance between semantically identical cross-domain samples is further reduced, while semantically dissimilar samples are pushed apart.

[0085] The positive and negative sample pairs include positive sample pairs and negative sample pairs. Specifically, the positive sample pairs consist of source domain samples and target domain samples with the same semantic category; the negative sample pairs consist of cross-domain samples with different semantic categories.

[0086] By optimizing cross-domain contrast loss This brings positive sample pairs closer together and pushes negative sample pairs further apart.

[0087] in, For source domain correction features, Correct features for the target domain. The cosine similarity function is used. For temperature parameters, Other target domain samples within the batch (as negative samples). This mechanism can effectively suppress negative transfer effects, ensuring that the model can still maintain a high sensitivity to key temporal fluctuations when transferring to the target domain, thereby achieving stable and accurate predictions in cross-domain scenarios.

[0088] Step S1034: Based on the inter-domain alignment loss and the cross-domain contrast loss, optimize the feature extractor to obtain the optimized feature extractor.

[0089] In some embodiments, this disclosure utilizes a unified semantic space as a bridge for cross-domain transfer. During training, feature vectors from different data sources are simultaneously brought closer to their corresponding semantic anchor points, using semantic consistency to constrain distribution alignment. Cross-domain contrastive learning further eliminates differences in sensing patterns, ensuring that the predictive knowledge learned in the source domain can be accurately transferred to the target domain, avoiding performance degradation due to statistical distribution differences.

[0090] Step S104: Based on the optimized feature extractor and the semantic anchor point, predict and calibrate the target domain time series data to be predicted to obtain the prediction result.

[0091] In some embodiments, this disclosure uses an adaptive prediction method based on residual correction and entropy weight calibration to address the timeliness problem of model migration to the target domain. By introducing a real-time feedback adjustment method during the inference stage, it ensures that the prediction system can spontaneously cope with the non-stationary environment of the dynamic evolution of the target domain.

[0092] In some embodiments, the prediction and calibration of the target domain time-series data to be predicted based on the optimized feature extractor and the semantic anchor points to obtain the prediction result includes: Step S1041: Based on the optimized feature extractor and the semantic anchor point, process the target domain time series data to be predicted to obtain adaptive features.

[0093] Based on the optimized feature extractor, feature extraction is performed on the time series data of the target domain to be predicted to obtain the corresponding time series features.

[0094] Based on the semantic anchor points, the temporal features are semantically pre-corrected to obtain the semantically corrected features of the target domain temporal data to be predicted.

[0095] In some embodiments, since the target domain environment may drift nonlinearly over time, this disclosure introduces a residual adapter. Online adjustments are performed. Using the residuals, the real-time deviation between the semantic correction features of the target domain time-series data to be predicted and the semantic anchor points is calculated. Based on this deviation, the feature distribution is dynamically compensated and re-aligned to generate adaptive features. :

[0096] in, For semantic correction features, As semantic anchor points, It is a lightweight learnable parameter matrix.

[0097] Step S1042: Determine the initial prediction result based on the adaptive features, and calibrate the initial prediction result to determine the final prediction result.

[0098] In some embodiments, to further ensure the reliability of the prediction results, especially in the face of high noise or unseen conditions, this disclosure introduces a risk control mechanism based on prediction entropy.

[0099] The adaptive feature The input is fed into the prediction layer of the model to obtain the initial predicted output probability distribution. , where C is the total number of categories.

[0100] By calculating the entropy of the initial predicted output probability distribution To quantify the semantic uncertainty of the current prediction task :

[0101] in, The initial predicted output probability distribution, In the predicted output, the sample belongs to the first... The probability values ​​of each category, The total number of categories for the prediction task. The entropy. It is the predicted output probability distribution The higher the information entropy, the more uncertain the model is about the current prediction result; conversely, the lower the entropy, the higher the confidence level of the prediction. This entropy value serves as a core indicator for quantifying the semantic uncertainty of predictions.

[0102] Introduce a calibration coefficient And utilize the semantic uncertainty to determine the probability distribution of the initial prediction output. Dynamic calibration is performed to obtain the final prediction result. To suppress the risk of overfitting under high uncertainty:

[0103] in, This is the initial predicted output probability distribution; The information entropy of the initial predicted output probability distribution represents semantic uncertainty; This represents the final predicted distribution.

[0104] In some embodiments, this weight adjustment mechanism can effectively suppress the risk of overfitting when the model faces a high uncertainty distribution.

[0105] In some embodiments, this disclosure focuses on improving the model's real-time predictive adjustment capability in dynamically changing target environments. During the inference phase in the target domain, a residual adaptor is introduced to dynamically compensate features based on the real-time deviation between the currently observed features and the semantic space. Simultaneously, the uncertainty of the prediction is quantified using the entropy value of the output state, and a decay calibration is applied to prediction branches with low confidence. This strategy enables the model to self-calibrate under non-stationary conditions, significantly enhancing the robustness and reliability of the final prediction sequence.

[0106] This disclosure presents a time-series data prediction method based on cross-modal semantic anchoring and adaptive correction. It enhances feature stability through multi-modal enhancement, imparts semantic consistency to representations through language model alignment, and combines cross-domain transfer and dynamic calibration on the inference side to achieve high-precision, adaptive prediction of complex heterogeneous time-series data. This method has significant technical advantages and broad application prospects in non-stationary scenarios such as energy dispatching, traffic monitoring, weather forecasting, and industrial equipment maintenance.

[0107] This disclosure provides a prediction method and related apparatus for time series data, aiming to improve the modeling generalization and long-term prediction accuracy of time series data in multi-source heterogeneous environments. By introducing pattern-guided robust feature enhancement, semantic space alignment of large language models, and residual adaptive calibration strategies in the target domain inference process, it systematically solves the problems of distribution shift sensitivity, semantic information loss, and weak adaptability in dynamic target domains in existing technologies during cross-domain transfer.

[0108] It is understandable that this method can be executed by any device, equipment, platform, or cluster of devices with computing and processing capabilities.

[0109] It should be noted that the methods of one or more embodiments of this disclosure can be executed by a single device, such as a computer or server. The methods of this embodiment can also be applied in a distributed scenario, where multiple devices cooperate to complete the process. In such a distributed scenario, one of these devices may execute only one or more steps of the methods of one or more embodiments of this disclosure, and the multiple devices will interact with each other to complete the method described.

[0110] It should be noted that the above description pertains to specific embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in a different order than those shown in the embodiments and may still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0111] Based on the same inventive concept, corresponding to any of the methods in the above embodiments, this disclosure also provides a prediction system for time series data. For example... Figure 2 As shown, the system includes: The construction module 201 is configured to construct a feature extractor based on the original time series data; wherein the original time series data includes source domain time series data and target domain time series data. Processing module 202 is configured to process the original time-series data based on a cross-modal semantic anchoring method to obtain semantic anchor points; The optimization module 203 is configured to optimize the feature extractor based on the semantic anchor point to obtain an optimized feature extractor. The prediction module 204 is configured to predict and calibrate the target domain time series data to be predicted based on the optimized feature extractor and the semantic anchor point, and obtain the prediction result.

[0112] For ease of description, the above system is described by dividing it into various modules based on their functions. Of course, when implementing one or more embodiments of this disclosure, the functions of each module can be implemented in one or more software and / or hardware.

[0113] The system described above is used to implement the corresponding methods in the foregoing embodiments and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0114] Figure 3 This embodiment illustrates a more specific hardware structure of an electronic device, which may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0115] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this disclosure.

[0116] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this disclosure are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0117] The input / output interface 1030 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0118] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0119] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0120] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this disclosure, and not necessarily all the components shown in the figures.

[0121] The electronic devices described above are used to implement the corresponding methods in the foregoing embodiments and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0122] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0123] Based on the same inventive concept, corresponding to the time-series data prediction method in any of the above embodiments, this disclosure also provides a computer program product, which includes one or more computer programs. In some embodiments, the one or more computer programs are executable by one or more processors to cause the one or more processors to perform the time-series data prediction method. Corresponding to the execution entity for each step in each embodiment of the time-series data prediction method, the processor executing the corresponding step may belong to the corresponding execution entity. The computer program product of the above embodiments is used to cause the processor to execute the time-series data prediction method as described in any of the above embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0124] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this disclosure (including the claims) is limited to these examples; within the framework of this disclosure, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this disclosure as described above, which are not provided in detail for the sake of brevity.

[0125] Additionally, to simplify the description and discussion, and to avoid obscuring one or more embodiments of this disclosure, the provided drawings may or may not show well-known power / ground connections to integrated circuit (IC) chips and other components. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring one or more embodiments of this disclosure, and this also takes into account the fact that the details of implementation of these block diagram apparatuses are highly dependent on the platform on which one or more embodiments of this disclosure will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuitry) are set forth to describe exemplary embodiments of this disclosure, it will be apparent to those skilled in the art that one or more embodiments of this disclosure may be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0126] Although this disclosure has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0127] This disclosure includes one or more embodiments intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for predicting time series data, characterized in that, The method includes: A feature extractor is constructed based on the original time-series data; wherein, the original time-series data includes source domain time-series data and target domain time-series data; Based on the cross-modal semantic anchoring method, the original time-series data is processed to obtain semantic anchor points; Based on the semantic anchor points, the feature extractor is optimized to obtain an optimized feature extractor; Based on the optimized feature extractor and the semantic anchor point, the target domain time series data to be predicted is predicted and calibrated to obtain the prediction result.

2. The method according to claim 1, characterized in that, The feature extractor constructed based on the original time-series data includes: Acquire and enhance the original time-series data to obtain an enhanced sequence set; Based on a dual encoder architecture, the original time-series data and the enhanced sequence set are extracted and optimized to obtain a feature extractor.

3. The method according to claim 1, characterized in that, The cross-modal semantic anchoring method processes the original time-series data to obtain semantic anchor points, including: Semantic extraction is performed on the original time-series data to generate a language description; The language description The generation process is as follows: in, For the input of the first A multivariate time series sample; This is a fixed prefix for prompts, used to structure the final generated description; Functions as semantic mappers; Based on a pre-trained large language model, the language description is extracted to obtain semantic anchor points; The semantic anchor point Represented as: in, Represents the raw time series data input. Coordinates or feature representations in the semantic space understood by large language models; For language description; Based on the semantic anchor points, semantic alignment constraints are applied to the feature extractor.

4. The method according to claim 1, characterized in that, The optimization of the feature extractor based on the semantic anchor point to obtain an optimized feature extractor includes: Based on the semantic pre-correction method, the source domain time series data and the target domain time series data are processed respectively to obtain the corresponding semantic correction features; Based on the semantic correction features and the semantic anchor points, calculate the inter-domain alignment loss; Based on the aforementioned semantic correction features, calculate the cross-domain contrast loss; Based on the inter-domain alignment loss and the cross-domain contrast loss, the feature extractor is optimized to obtain an optimized feature extractor.

5. The method according to claim 1, characterized in that, The process involves predicting and calibrating the target domain time-series data based on the optimized feature extractor and the semantic anchor points to obtain prediction results, including: Based on the optimized feature extractor and the semantic anchor point, the target domain time series data to be predicted is processed to obtain adaptive features; An initial prediction result is determined based on the adaptive features, and the initial prediction result is calibrated to determine the final prediction result.

6. The method according to claim 5, characterized in that, The process of processing the target domain time-series data to be predicted based on the optimized feature extractor and the semantic anchor points to obtain adaptive features includes: Based on the optimized feature extractor, feature extraction is performed on the target domain time series data to be predicted to obtain the corresponding time series features. Based on the semantic anchor points, the temporal features are semantically pre-corrected to obtain the semantically corrected features of the target domain temporal data to be predicted. Using a residual adapter, the real-time deviation between the semantic correction features of the target domain time-series data to be predicted and the semantic anchor points is calculated in real time. Based on the deviation, the feature distribution is dynamically compensated and re-aligned to generate adaptive features. : in, For semantic correction features, As semantic anchor points, It is a lightweight learnable parameter matrix.

7. The method according to claim 6, characterized in that, The process of determining an initial prediction result based on the adaptive features, calibrating the initial prediction result, and determining the final prediction result includes: Based on the adaptive features The initial predicted output probability distribution is obtained. ; By calculating the entropy of the initial predicted output probability distribution. Quantify the semantic uncertainty of the current prediction task : in, The initial predicted output probability distribution, In the predicted output, the sample belongs to the first... The probability values ​​of each category, The total number of categories for the prediction task; Introduce a calibration coefficient And utilize the semantic uncertainty to determine the probability distribution of the initial prediction output. Dynamic calibration is performed to obtain the final prediction result. : in, This is the initial predicted output probability distribution; The information entropy of the initial predicted output probability distribution represents semantic uncertainty; This represents the final predicted distribution.

8. A prediction system for time series data, characterized in that, The system includes: The construction module is configured to build a feature extractor based on the original time-series data; wherein the original time-series data includes source domain time-series data and target domain time-series data; The processing module is configured to process the original time-series data based on a cross-modal semantic anchoring method to obtain semantic anchor points; The optimization module is configured to optimize the feature extractor based on the semantic anchor point to obtain an optimized feature extractor. The prediction module is configured to predict and calibrate the target domain time series data to be predicted based on the optimized feature extractor and the semantic anchor point, and obtain the prediction result.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executed by the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions for causing the computer to perform the method according to any one of claims 1 to 7.

Images