An online prediction method, device, medium, and product for multivariate time series data.
By combining reversible instance normalization and a multi-scale time series fusion model, the problems of prediction error and computational complexity of multivariate time series data in non-stationary environments are solved, and efficient and adaptive online prediction is achieved.
Patent Information
- Application Number
- CN202611323108.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-28
- Publication Date
- 2026-11-17
AI Technical Summary
Existing multivariate time series data prediction models suffer from large prediction errors, high computational complexity, difficulty in balancing multi-scale feature extraction and real-time requirements, and lack of adaptive capabilities when faced with non-stationary characteristics and concept drift.
A reversible instance normalization method is used to preprocess multidimensional time series data, which is then trained in conjunction with a multi-scale time series fusion model. In the real-time prediction stage, the model is adjusted online through drift detection and incremental update mechanisms.
It significantly improves the prediction accuracy and computational efficiency of multi-scale time series fusion models, and has anti-interference and adaptive capabilities, meeting the needs of industrial-grade real-time online prediction.
Smart Images

Figure CN123115673A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of computer data processing and artificial intelligence, and in particular to an online prediction method, device, medium, and product for multivariate time series data. Background Technology
[0002] With the deepening of power market reform and the high proportion of renewable energy integration, extremely high demands are being placed on online forecasting technologies for multivariate time series data (e.g., spot market electricity prices). In practical scenarios such as industrial monitoring, system scheduling, and resource allocation, multidimensional time series data containing target variables and multiple related characteristics (e.g., the stationary characteristics caused by the intermittency of renewable energy output) are typically collected. High-precision online forecasting of these non-stationary time series data can provide reliable data support for downstream automation control and decision support systems.
[0003] Currently, most mainstream time series data prediction methods are based on deep learning models. However, in practical online prediction applications, existing technologies still have the following significant drawbacks: First, real-world multivariate time series data often exhibits strong non-stationary characteristics. Due to dynamic changes in the external environment or fluctuations in the data source itself, the underlying statistical distribution of time series data will dynamically shift over time (i.e., "concept drift"). Existing prediction models mostly adopt a static deployment mode of "offline training and online inference," where the model remains fixed once deployed. This static mechanism significantly increases the model's prediction error when faced with shifts in data distribution, and lacks the ability to automatically detect concept drift and perform online adaptive correction, leading to a severe decline in long-term prediction accuracy.
[0004] Secondly, multivariate time series typically involve a mix of short-term local fluctuations and long-term macroeconomic trends, making it difficult for existing models to balance feature extraction capabilities with computational efficiency. On the one hand, while existing complex deep learning models possess strong global feature extraction capabilities, their computational complexity is extremely high, resulting in large memory overhead and making it difficult to meet the low latency and high throughput requirements of industrial-grade real-time online prediction. On the other hand, while existing lightweight models offer fast inference speeds, they struggle to effectively decouple, align, and fuse multi-scale features at different time resolutions, leading to limited prediction accuracy when faced with complex fluctuations.
[0005] Therefore, how to provide a multivariate time series prediction method that can efficiently extract multi-scale time series features, overcome input distribution offset interference, and have online concept drift perception and low-cost incremental update capabilities, so as to significantly improve the computational efficiency and adaptability of the model while ensuring long-term prediction accuracy, has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0006] The purpose of this application is to provide an online prediction method, device, medium, and product for multivariate time series data, which can significantly improve the computational efficiency and adaptability of the model while ensuring long-term prediction accuracy.
[0007] To achieve the above objectives, this application provides the following solution: Firstly, this application provides an online prediction method for multivariate time series data, including: Acquire multidimensional time series data; the multidimensional time series data includes the target time series and its corresponding associated features; The multidimensional time series data is subjected to adaptive preprocessing to obtain preprocessed data; the adaptive preprocessing includes: missing data completion and reversible instance normalization. Using the preprocessed data, the multi-scale time series fusion model is trained to obtain a trained multi-scale time series fusion model. In the real-time prediction phase, concept drift detection is performed based on changes in the prediction error distribution. When concept drift is detected, an online incremental update mechanism is triggered to update the trained multi-scale time series fusion model using newly added multi-dimensional time series data. The updated multi-scale time series fusion model is used to obtain normalized prediction results, and the normalized prediction results are then denormalized to obtain the predicted value of the target time series.
[0008] Secondly, this application provides an online prediction device for multivariate time series data, comprising: The data acquisition module is used to acquire multidimensional time series data, which includes the target time series and its corresponding associated features. The data preprocessing module is used to perform adaptive preprocessing on the multidimensional time series data to obtain preprocessed data; the adaptive preprocessing includes: missing data completion and reversible instance normalization; The model training module is used to train the multi-scale time series fusion model using the preprocessed data to obtain the trained multi-scale time series fusion model. The concept drift detection module is used to detect concept drift based on changes in the prediction error distribution during the real-time prediction phase. The model update module is used to trigger an online incremental update mechanism when concept drift is detected, and to update the trained multi-scale time series fusion model with newly added multi-dimensional time series data. The prediction module is used to obtain normalized prediction results using the updated multi-scale time series fusion model, and to perform inverse normalization processing on the normalized prediction results to obtain the predicted value of the target time series.
[0009] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the online prediction method for multivariate time series data.
[0010] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the online prediction method for multivariate time series data.
[0011] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the online prediction method for multivariate time series data.
[0012] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides an online prediction method, device, medium, and product for multivariate time series data. By employing Reversible Instance Normalization (RevIN) to dynamically adjust the statistics of multidimensional time series data, it effectively aligns the data distribution across different time periods at the instance level. This protects the multi-scale time series fusion model (TimeMixer neural network model) from input distribution shifts during training and prediction. This application not only preserves the trend information of the original multidimensional time series data but also automatically adapts to data fluctuations in non-stationary environments, significantly improving the model's robustness to out-of-distribution samples and laying a solid foundation for subsequent high-precision predictions. Offline training based on the multi-scale time series fusion model fully utilizes its multi-scale time series decomposition and mixing mechanism, enabling the simultaneous capture of short-term fluctuation patterns and long-term dependencies in the data. This application significantly improves the training efficiency and inference speed of the multi-scale time series fusion model while ensuring sufficient prediction accuracy, meeting the market's real-time response requirements. In the real-time prediction stage, this application introduces a drift detection method based on prediction error rate and constructs an online adaptive closed loop of "drift perception-lightweight repair" to realize the online adaptation of the multi-scale time series fusion model to market changes, significantly improving the generalization ability of the multi-scale time series fusion model. Attached Figure Description
[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0014] Figure 1 This is a schematic flowchart of an online prediction method for multivariate time series data in one embodiment of this application; Figure 2 This is a schematic diagram illustrating the normalization and denormalization operations performed on the data using the RevIN method; Figure 3 This is a schematic diagram of the structure of a multi-scale temporal fusion model; Figure 4 This is a diagram illustrating seasonal and trend mixing. Figure 5 This is a schematic diagram of the drift detection-online update method. Detailed Implementation
[0015] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0016] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0017] In one exemplary embodiment, such as Figure 1 As shown, an online prediction method for multivariate time series data is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is described using a server as an example, including the following steps S101 to S106. Wherein: S101, Obtain multidimensional time series data; the multidimensional time series data includes the target time series and its corresponding associated features; As a specific example, the target time series can be historical electricity prices, and the corresponding associated features include meteorological factors, calendar factors, and historical load factors for the corresponding period; that is, multidimensional time series data includes data from multiple dimensions such as historical electricity prices, meteorological factors, calendar features, and historical load. Since day-ahead electricity prices are affected by multiple external factors such as supply and demand, fluctuations in renewable energy output, seasonal temperature changes, holiday arrangements, and grid operation status, incorporating relevant influencing factors as comprehensively as possible before making accurate predictions can effectively improve the model's ability to characterize electricity price fluctuation patterns, providing reliable support for electricity market participants' bidding decisions and system dispatch operations. Furthermore, reversible instance normalization can avoid errors in electricity prediction caused by different data units and eliminate the performance degradation caused by time-varying non-stationary information. As a specific example, the target time series can be an electricity load series; S102, perform adaptive preprocessing on the multidimensional time series data to obtain preprocessed data; the adaptive preprocessing includes: missing data completion and reversible instance normalization. S102 specifically includes: S21, The random forest method is used to complete missing data in multidimensional time series data; S22, Determine the mean and standard deviation of each variable channel in the multidimensional time series data after missing data completion; S23. Using the learnable affine parameters and the mean and standard deviation, the multidimensional time series data after missing data completion is normalized to obtain preprocessed data.
[0018] As a specific example, Figure 2 middle Represents the original input data. Represents the normalized data. The model's predictions in the normalized space This represents the final prediction after denormalization. The probability distribution of data at each stage. The representative data are the mean and standard deviation. These are the learnable parameters of the model; such as Figure 2 As shown, the steps for normalization using RevIN are as follows: 1. Calculate instance-level statistics for the target time series. Each instance in (in , These represent the number of variables and the length of the input sequence, respectively. (For the set of real numbers), calculate the mean of each variable channel separately. with standard deviation (Calculation method as follows) Figure 2 (as shown in part (a-3)). (1); (2); in, For the first In the nth sample, the nth The variable channel in the first... The value at each time step, For the first In the nth sample, the nth The variable channel in the first... The value at each time step, and They are all time step indexes. As the output index Used as an index for summation traversal.
[0019] 2. Determine the learnable affine parameters; affine parameters include scaling parameters. Translation parameters These two parameters are automatically optimized through backpropagation during model training to enhance the expressive power of the normalization operation.
[0020] 3. Establish a reversible instance normalization module, defined as: (3); (4); in, Represents variable channels, Represents a time-series index. Represents the sample number. This represents a small constant to prevent division by zero. Equation (3) is the normalization operation (such as...). Figure 2 (As shown in the left part), the original input Converted to normalized Equation (4) is the inverse normalization operation (e.g. Figure 2 (As shown in the right part), the normalized predicted value of the model output. Restore the prediction results to the original distribution , This represents the predicted value output under normalization. This represents the prediction result after inverse normalization.
[0021] 4. By eliminating the distribution differences between instances through normalization, the model input has a more consistent mean and variance, weakening the non-stationary characteristics of the data, so that the TimeMixer model can more accurately focus on the local dynamic changes within the sequence. After obtaining the model output, inverse normalization is performed using equation (4) to restore the prediction results to the actual value range of the original time series, ensuring that the prediction results have actual physical meaning.
[0022] S103, Using the preprocessed data, the multi-scale temporal fusion model is trained to obtain a trained multi-scale temporal fusion model; S103 and above also includes: dividing the preprocessed data into training and testing sets; and using the training set to train the multi-scale temporal fusion model; As a specific implementation, when the target time series is historical electricity prices, a multi-scale time series fusion model is used to model the series, considering that the price series is driven by multiple scale factors such as intraday fluctuations, intraweek cycles, and sudden spikes. This multi-scale time series fusion model employs a sophisticated architecture based entirely on a multilayer perceptron, avoiding the high computational overhead caused by the self-attention mechanism of complex models such as the Transformer. It achieves extremely high inference efficiency while ensuring powerful time series feature extraction capabilities. Through cross-scale decomposition and hybridization mechanisms, the multi-scale time series fusion model deeply extracts multi-scale features of electricity prices at different time resolutions, taking into account both macro trends and sudden spikes, effectively alleviating the limitations of single-scale models in balancing stability and sensitivity. Thanks to its lightweight pure MLP design and efficient multi-scale hybrid structure, the multi-scale time series fusion model not only excels in predicting day-ahead electricity prices in complex market environments in terms of prediction accuracy but also boasts significant advantages in computational efficiency, providing a foundation for applications in the electricity spot market. like Figure 3 As shown, the multi-scale temporal fusion model includes an input embedding layer, multiple stacked Past Decomposition Mixing (PDM) modules, and Future Multipredictor Mixing (FMM) modules. The input embedding layer is used to perform multi-scale downsampling on the preprocessed data, generating a multi-scale time series set from fine-grained to coarse-grained. The stacked past decomposable mixing module is used to decompose and mix the multi-scale time series set, extracting seasonal and trend features respectively. The future multi-predictor mixing module is used to integrate feature prediction results from multiple scales to obtain the normalized prediction results. The multi-scale time series fusion model can effectively capture the fluctuation patterns of the target time series at different time scales and make high-precision predictions of the current day's target. The feature prediction results include seasonal and trend features. Furthermore, the specific modeling process of the multi-scale temporal fusion model for features is as follows: 1) Perform multi-scale downsampling on multi-dimensional time series data; To analyze complex temporal variation patterns, the multi-scale time series fusion model employs a multi-scale decomposition and hybrid architecture. The multi-scale time series fusion model first uses average pooling to decompose multi-dimensional time series data. downsampling to Multiple scales are used to obtain a multi-scale time series set. ,in, , , Indicates the number of variables. Underlying sequence. For multidimensional time series data, the most refined temporal changes are preserved; top-level sequence This captures macroscopic trends. All scale sequences are mapped to deep features through an embedding layer. Multi-scale partitioning enables multi-scale time series fusion models to simultaneously observe the micro-details (fine scale) and macro-trends (coarse scale) of time series, laying the foundation for subsequent multi-scale mixing.
[0023] 2) Extract multi-scale features using previously decomposable hybrid modules; Even after downsampling, interwoven temporal patterns still exist within sequences at different scales. Therefore, a multi-scale temporal fusion model addresses this by using stacked past decomposable mixing modules. In the... In a previously decomposable hybrid module, the decomposition blocks First of all Decomposed into trend components With periodic components Subsequently, the multi-scale time-series fusion model employs differentiated mixing strategies for multi-scale interaction, taking into account the different characteristics of seasonal and trend components. Overall, the... A past decomposable hybrid module can be formally represented as: (5); (6); in, This represents the total number of floors. This represents a decomposition block, where the trend term is obtained through a moving average, and the residuals are used as the seasonal term. For deep feature representation at different time scales, the number of channels is ; It contains two linear layers and an intermediate GELU activation function for inter-channel information exchange; and These represent seasonal mixing and trend mixing, respectively. Figure 4 This diagram illustrates the seasonal and trend mixing within the previously decomposable mixing module. It employs bottom-up and top-down information flow directions to effectively integrate micro-seasonal and macro-trend information at different scales.
[0024] In seasonal analysis, longer periods can be viewed as aggregates of shorter periods. Therefore, bottom-up residual mixing can supplement detailed information for coarser-scale seasonal modeling. Technically, for multi-scale seasonal component sets... Utilizing a finer scale (the first) Seasonal details (layer) to enhance the current scale (the first layer) Modeling of layers: (7); in, Specifically, it is implemented as two linear layers along the time dimension and the GELU activation function between them, with an input dimension of... The output dimension is .
[0025] In trend analysis, coarser-scale upper-level time series provide clearer macroscopic information compared to finer-scale series. Therefore, multi-scale time series fusion models employ a top-down mixing approach. Technically, this involves addressing the multi-scale trend components. Using a coarser scale (the first) Macro trends at the first level (level 1) guide the development of finer-scale (level 2) strategies. Modeling of layers: (8); in, It consists of two linear layers and a GELU activation function in between, with an input dimension of... The output dimension is The specific structure is as follows Figure 4 As shown, the decomposable mixing module effectively integrates fine-scale seasonal details with coarse-scale trend background information, ultimately achieving multi-scale mixing in historical information extraction.
[0026] 3) Generate the final prediction using the future multi-predictor hybrid module.
[0027] After processing by the aforementioned L-layer decomposable hybrid module, the multi-scale temporal fusion model obtains a comprehensive set of multi-scale historical information, denoted as . To fully leverage their complementarity, the future multi-predictor hybrid module will be equipped with a dedicated predictor for each scale, allowing each to tap into its unique predictive potential. The predictions generated at each scale will then be integrated to form a unified prediction output. (9); (10); in, This represents a predictor for a sequence at scale m, which maps historical sequences to predicted outputs for the next F steps. The output of the future multi-predictor hybrid module is , representing the target prediction result.
[0028] Model training is performed by minimizing the prediction error. Mean squared error (MSE) is used as the loss function to measure the normalized predicted value output by the multi-scale temporal fusion model. Compared with the true value The differences between the historical target values (i.e., those that have undergone the same normalization process) are analyzed, and the parameters of the multi-scale temporal fusion model are iteratively updated using the backpropagation algorithm and the Adam optimizer until the loss function converges, thus obtaining the trained multi-scale temporal fusion model. When making real-time predictions, for new input samples, preprocessing is first performed according to S102, and then the normalized prediction value is obtained by inputting it into the trained multi-scale temporal fusion model. Then, the final prediction result is restored to the original target range through an inverse normalization operation. To achieve precise delivery of the current objectives.
[0029] After the trained multi-scale time series fusion model is deployed to the online prediction environment, it makes real-time predictions for each newly arrived daytime target sample and calculates the prediction error to form a continuous time error sequence, which is then used as the input data stream for drift detection.
[0030] S104, In the real-time prediction stage, concept drift detection is performed based on changes in the prediction error distribution; S104 specifically includes: S41, Establish a statistical test mechanism based on a sliding window to calculate the average prediction error within the current sliding window; That is, to track the test loss, assuming it follows a normal distribution. Then, statistical tests were performed. The sliding window length was set. (In this embodiment, the value is set to 20), used to store the most recent The loss value for each sample. Calculate the average loss for this test window. ; S42, Based on the average value and the mean and standard deviation of the historical error distribution, determine the significance test value; Using formula Determine the significance test value z; z is used to quantify the confidence level of the new loss deviating from the given loss distribution; in, and These are the mean and standard deviation of the historical loss distribution, respectively. For window size; S43, convert the significance test value into the observed significance level; Specifically, the calculated z-value is compared with the percentile of the standard normal distribution to obtain the observed significance level (p-value), which is used to determine whether the loss of the current window significantly deviates from the historical distribution.
[0031] S44, when the observed significance level is lower than the warning threshold (In this embodiment, it is set to 0.08), that is... Enter drift warning state; when the observed significance level is lower than the drift confirmation threshold. (In this embodiment, it is set to 0.03), that is... If a concept drift is detected, the drift confirmation threshold is less than the warning threshold. Otherwise, if there is no significant change, the model remains in its current state.
[0032] To avoid catastrophic forgetting of the model due to over-focusing on a single new data instance during online learning, S104 and later also include: S1, when entering the drift warning state, the short-time memory area is activated. M Collect subsequent arriving multidimensional time series data and corresponding prediction errors; the short-term memory region M Used to buffer historical multidimensional time series data; S2, if the subsequent arrival of multi-dimensional time series data meets the early warning conditions and eventually triggers drift confirmation, then the online incremental update mechanism of the multi-scale time series fusion model is triggered. S3, if no drift confirmation is triggered within the preset conditions after entering the warning state, then the short-term memory area is released and reset. M This avoids the long-term ineffective occupation of computing resources.
[0033] S105, when concept drift is detected, an online incremental update mechanism is triggered to update the trained multi-scale time series fusion model using newly added multi-dimensional time series data; S105 specifically includes: S51 uses the output layer parameters of the future multi-predictor hybrid module in the multi-scale temporal fusion model as the parameters to be updated; that is, in order to avoid the huge computational overhead caused by full retraining, incremental updates only fine-tune some parameters of the model; the output layer parameters of the future multi-predictor hybrid module are most sensitive to changes in data distribution, and fine-tuning these parameters can significantly reduce computational costs while ensuring model adaptability. S52, using the newly added multidimensional time series data as training data, the parameters to be updated are updated using a gradient with a preset learning rate.
[0034] Specifically, using samples within the short-term memory region M as training data, the selected parameters are updated with gradients in a small number of rounds, enabling the multi-scale temporal fusion model to quickly adapt to new data distributions. A low learning rate is used during the update process to avoid over-influencing existing knowledge.
[0035] After the incremental update is completed, the short-term memory region M is released, and the multi-scale temporal fusion model reverts to its normal prediction state. The drift detection mechanism continues to monitor newly arriving samples in real time, forming a closed-loop adaptive system of "monitoring-detection-update". The overall "drift detection-online update" method flow is as follows: Figure 5 As shown.
[0036] S106. Using the updated multi-scale time series fusion model, a normalized prediction result is obtained, and the normalized prediction result is denormalized to obtain the predicted value of the target time series.
[0037] In one exemplary embodiment, an online prediction device for multivariate time series data is provided, comprising: The data acquisition module is used to acquire multidimensional time series data, which includes the target time series and its corresponding associated features. The data preprocessing module is used to perform adaptive preprocessing on the multidimensional time series data to obtain preprocessed data; the adaptive preprocessing includes: missing data completion and reversible instance normalization; The model training module is used to train the multi-scale time series fusion model using the preprocessed data to obtain the trained multi-scale time series fusion model. The concept drift detection module is used to detect concept drift based on changes in the prediction error distribution during the real-time prediction phase. The model update module is used to trigger an online incremental update mechanism when concept drift is detected, and to update the trained multi-scale time series fusion model with newly added multi-dimensional time series data. The prediction module is used to obtain normalized prediction results using the updated multi-scale time series fusion model, and to perform inverse normalization processing on the normalized prediction results to obtain the predicted value of the target time series.
[0038] This application addresses the time-series forecasting of day-ahead electricity prices in the electricity spot market, and addresses issues such as the prominent non-stationarity of electricity price data, severe concept drift, and the difficulty in achieving an effective balance between high accuracy and efficiency in forecasting models through multi-level processing. It should be noted that when forecasting day-ahead electricity prices in the electricity spot market, the forecast results provided in this application are intended to offer quantitative references on electricity price trends to electricity market participants, assisting them in optimizing their pricing strategies. This falls under the category of forecasting tools for market information services, and is not intended for administrative control or price regulation of electricity prices.
[0039] The main effects of this application are as follows: (1) It possesses anti-interference, adaptability, and robustness. The reversible instance normalization method is used to dynamically adjust the statistics of multidimensional time series data, which can effectively align the data distribution of different time periods at the instance level, so that the model is not affected by the input distribution shift during the training and prediction stages. This method not only preserves the trend information of the original time series, but also automatically adapts to the fluctuation characteristics of electricity price data in non-stationary environments, significantly improving the robustness of the model to out-of-distribution samples, and laying a solid foundation for subsequent high-precision prediction; (2) It possesses strong time series modeling capabilities and high computational efficiency. Based on the TimeMixer neural network model, offline training fully utilizes its multi-scale time series decomposition and mixing mechanism, enabling simultaneous capture of short-term fluctuation patterns and long-term dependencies in electricity prices. The TimeMixer neural network model, through its efficient time series feature extraction structure, achieves accurate modeling of the target time series while maintaining low computational complexity. This method significantly improves the model's training efficiency and inference speed while ensuring sufficient prediction accuracy, meeting the market's real-time response requirements. (3) It possesses online adaptive capabilities and an extremely low-cost incremental update mechanism. A closed-loop online adaptive mechanism of "drift perception-lightweight repair" was constructed. In the real-time prediction stage, a drift detection method based on prediction error rate was introduced. By continuously monitoring the changes in the prediction error distribution between recent and historical windows, concept drift caused by sudden changes in market supply and demand, fluctuations in new energy output, or policy adjustments was accurately identified. When a significant decline in model performance is detected, a lightweight online incremental update mechanism is automatically triggered. Only a small number of model parameters need to be fine-tuned to quickly adapt to the new data distribution, realizing online adaptation of the model to market changes and significantly improving the model's generalization ability.
[0040] This application achieves a unified approach of high precision, high efficiency, and strong generalization through the synergistic effect of the aforementioned three-layer mechanism of "active adaptation-drift detection-online update" and the application of the TimeMixer neural network model.
[0041] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal. The computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is connected to the system bus via the I / O interfaces. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The I / O interfaces of the computer device are used for exchanging information between the processor and external devices. The communication interface of the computer device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements an online prediction method for multivariate time series data.
[0042] In one exemplary embodiment, a computer device is provided, 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.
[0043] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.
[0044] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.
[0045] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.
[0046] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).
[0047] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.
[0048] In this application, all actions to acquire signals, information, or data are carried out in compliance with the relevant data protection laws and policies of the country where the location is situated, and with the authorization granted by the owner of the relevant device.
[0049] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0050] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. An online prediction method for multivariate time series data, characterized in that, include: Obtain multidimensional time series data; The multidimensional time series data includes the target time series and its corresponding associated features; The multidimensional time series data is subjected to adaptive preprocessing to obtain preprocessed data; The adaptive preprocessing includes: missing data completion and reversible instance normalization; The multi-scale time series fusion model is trained using the preprocessed data to obtain a trained multi-scale time series fusion model. In the real-time prediction phase, concept drift detection is performed based on changes in the prediction error distribution. When concept drift is detected, an online incremental update mechanism is triggered to update the trained multi-scale time series fusion model using newly added multi-dimensional time series data. The updated multi-scale time series fusion model is used to obtain normalized prediction results, and the normalized prediction results are then denormalized to obtain the predicted value of the target time series.
2. The online prediction method for multivariate time series data according to claim 1, characterized in that, The adaptive preprocessing of the multidimensional time series data to obtain preprocessed data specifically includes: The random forest method was used to complete missing data in multidimensional time series data. Determine the mean and standard deviation of each variable channel in the multidimensional time series data after missing data completion; Using learnable affine parameters and the mean and standard deviation, the multidimensional time series data after missing data completion is normalized to obtain preprocessed data.
3. The online prediction method for multivariate time series data according to claim 1, characterized in that, The multi-scale temporal fusion model includes an input embedding layer, multiple stacked past decomposable fusion modules, and a future multi-predictor fusion module. The input embedding layer is used to perform multi-scale downsampling on the preprocessed data to generate a multi-scale time series set. The stacked past decomposable mixing module is used to decompose and mix the multi-scale time series set to extract seasonal features and trend features respectively. The future multi-predictor hybrid module is used to integrate feature prediction results at multiple scales to obtain the normalized prediction result; the feature prediction result includes seasonal features and trend features.
4. The online prediction method for multivariate time series data according to claim 1, characterized in that, The concept drift detection based on changes in the prediction error distribution during the real-time prediction phase specifically includes: Establish a statistical testing mechanism based on a sliding window to calculate the average prediction error within the current sliding window; The significance test value is determined based on the average value and the mean and standard deviation of the historical error distribution. The significance test values are converted into observed significance levels; When the observed significance level is lower than the warning threshold, a drift warning state is entered; when the observed significance level is lower than the drift confirmation threshold, a concept drift is determined to have occurred; the drift confirmation threshold is less than the warning threshold.
5. The online prediction method for multivariate time series data according to claim 1, characterized in that, In the real-time prediction phase, concept drift detection is performed based on changes in the prediction error distribution, followed by: When entering the drift warning state, the short-term memory area is activated to collect subsequently arriving multidimensional time series data and corresponding prediction errors; the short-term memory area is used to buffer historical multidimensional time series data. If the subsequent arrival of multidimensional time series data meets the early warning conditions and eventually triggers drift confirmation, the online incremental update mechanism of the multi-scale time series fusion model will be triggered. If drift confirmation is not triggered within the preset conditions after entering the warning state, the short-term memory area is released and reset.
6. The online prediction method for multivariate time series data according to claim 1, characterized in that, When concept drift is detected, an online incremental update mechanism is triggered to update the trained multi-scale time series fusion model using newly added multi-dimensional time series data. Specifically, this includes: The output layer parameters of the future multi-predictor hybrid module in the multi-scale temporal fusion model are used as the parameters to be updated. Using the newly added multidimensional time series data as training data, the parameters to be updated are updated using a gradient with a preset learning rate.
7. An online forecasting device for multivariate time series data, characterized in that, include: The data acquisition module is used to acquire multidimensional time series data; The multidimensional time series data includes the target time series and its corresponding associated features; The data preprocessing module is used to perform adaptive preprocessing on the multidimensional time series data to obtain preprocessed data; The adaptive preprocessing includes: missing data completion and reversible instance normalization; The model training module is used to train the multi-scale time series fusion model using the preprocessed data to obtain the trained multi-scale time series fusion model. The concept drift detection module is used to detect concept drift based on changes in the prediction error distribution during the real-time prediction phase. The model update module is used to trigger an online incremental update mechanism when concept drift is detected, and to update the trained multi-scale time series fusion model with newly added multi-dimensional time series data. The prediction module is used to obtain normalized prediction results using the updated multi-scale time series fusion model, and to perform inverse normalization processing on the normalized prediction results to obtain the predicted value of the target time series.
8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement an online prediction method for multivariate time series data according to any one of claims 1-6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the online prediction method for multivariate time series data as described in any one of claims 1-6.
10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the online prediction method for multivariate time series data as described in any one of claims 1-6.
