Business service system time series data prediction method and device, storage medium and product

Abstract

The embodiment of the invention provides a business service system time series data prediction method and device, a storage medium and a product, and the method comprises the steps: receiving a prediction request for future time series data of a business service system monitoring index, and enabling the prediction request to comprise first historical time series data of the business service system monitoring index; determining a first periodic component of the first historical time sequence data and a second periodic component of the future time sequence data based on a pre-learned periodic mode of a business service system monitoring index; determining a first residual component of the first historical time sequence data according to the first historical time sequence data and the first periodic component, and predicting a second residual component of future time sequence data according to the first residual component; and fusing the second periodic component and the second residual component to obtain future time series data. Core periodic features of the monitoring indexes are reserved through double-space prediction, random fluctuation of the monitoring indexes is predicted through global context information, and prediction accuracy and disturbance robustness are improved.

Description

[0001] This application claims priority to Chinese Patent Application No. 202511439054.2, filed on October 9, 2025, entitled “Method, apparatus, storage medium and product for time series data prediction in business service system”, the entire contents of which are incorporated herein by reference. Technical Field

[0002] This disclosure relates to the field of computer technology, and in particular to a method, device, storage medium, and product for predicting time-series data in a business service system. Background Technology

[0003] In many business systems, time series forecasting plays a crucial role in ensuring system reliability. Business service systems continuously collect hundreds of monitoring indicators, including CPU utilization, memory usage, and network traffic, forming high-dimensional time series data. The accuracy of its forecasts directly affects many core business requirements. For example, fault prediction requires early identification of abnormal server states (such as abnormal CPU spikes), capacity planning requires accurate prediction of resource usage trends (such as storage space consumption), and performance optimization relies on accurate prediction of service response latency fluctuations.

[0004] Existing time series forecasting methods struggle to balance forecast accuracy with robustness to disturbances, failing to meet the demands of business systems for high-precision forecasts and strong anti-interference capabilities. Summary of the Invention

[0005] This disclosure provides a method, device, storage medium, and product for predicting time-series data in a business service system, in order to improve the accuracy and robustness to disturbances in predicting time-series data in a business service system.

[0006] In a first aspect, embodiments of this disclosure provide a method for predicting time-series data in a business service system, including:

[0007] Receive a prediction request for future time-series data of monitoring indicators of a business service system, wherein the prediction request includes first historical time-series data of the monitoring indicators;

[0008] Based on the pre-learned periodic patterns of the business service system monitoring indicators, a first periodic component of the first historical time series data and a second periodic component of the future time series data are determined; wherein, the periodic pattern is a regular change that occurs periodically in the time series data; the periodic component is a partial sequence extracted from the time series data according to the periodic pattern.

[0009] The first residual component of the first historical time series data is determined based on the first historical time series data and the first periodic component, and the second residual component of the future time series data is predicted based on the first residual component.

[0010] The future time series data is obtained by fusing the second periodic component with the second residual component.

[0011] Secondly, embodiments of this disclosure provide a time-series data prediction device for a business service system, comprising:

[0012] A periodic extraction unit is used to determine a first periodic component of the first historical time series data and a second periodic component of the future time series data based on the periodic patterns of the pre-learned monitoring indicators of the business service system; wherein, the periodic pattern is a regular change that occurs periodically in the time series data; and the periodic component is a partial sequence extracted from the time series data according to the periodic pattern.

[0013] The prediction unit is configured to determine a first residual component of the first historical time series data based on the first historical time series data and the first periodic component, and to predict a second residual component of the future time series data based on the first residual component.

[0014] A fusion unit is used to fuse the second periodic component and the second residual component to obtain the future time series data.

[0015] Thirdly, embodiments of this disclosure provide an electronic device, including: a processor and a memory;

[0016] The memory stores computer-executed instructions;

[0017] The processor executes computer execution instructions stored in the memory, causing the at least one processor to perform the business service system time-series data prediction method as described in the first aspect and various possible designs of the first aspect.

[0018] Fourthly, embodiments of this disclosure provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the time-series data prediction method for a business service system as described in the first aspect and various possible designs of the first aspect.

[0019] Fifthly, embodiments of this disclosure provide a computer program product, including a computer program that, when executed by a processor, implements the time-series data prediction method for a business service system as described in the first aspect and various possible designs of the first aspect.

[0020] The present disclosure provides a method, device, storage medium, and product for predicting time-series data of a business service system. This involves receiving a prediction request for future time-series data of a business service system monitoring indicator, wherein the prediction request includes first historical time-series data of the monitoring indicator; determining a first periodic component of the first historical time-series data and a second periodic component of the future time-series data based on a pre-learned periodic pattern of the business service system monitoring indicator; wherein the periodic pattern is a regular change that occurs periodically in the time-series data; the periodic component is a partial sequence extracted from the time-series data according to the periodic pattern; determining a first residual component of the first historical time-series data based on the first historical time-series data and the first periodic component, and predicting a second residual component of the future time-series data based on the first residual component; and fusing the second periodic component and the second residual component to obtain the future time-series data. This embodiment of the disclosure predicts the periodic and residual components of future time-series data of monitoring indicators of a business service system. While retaining the core periodic characteristics of the monitoring indicators, it also uses global context information to predict the random fluctuations of the monitoring indicators, reducing the sensitivity to noise and outliers in historical time-series data, and improving the accuracy of prediction and robustness to disturbances. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A scenario example diagram illustrating the time-series data prediction method for a business service system provided in an embodiment of this disclosure;

[0023] Figure 2 This is a schematic flowchart of a time-series data prediction method for a business service system provided in an embodiment of the present disclosure;

[0024] Figure 3 A schematic diagram illustrating the decomposition of time-series data according to an embodiment of this disclosure;

[0025] Figure 4 A schematic diagram illustrating the construction of periodic components according to an embodiment of this disclosure;

[0026] Figure 5 A schematic diagram illustrating a scenario of a time-series data prediction method for a business service system provided in an embodiment of this disclosure;

[0027] Figure 6 A schematic diagram of a training process provided in an embodiment of this disclosure;

[0028] Figure 7 A schematic diagram illustrating random time point sampling provided in an embodiment of this disclosure;

[0029] Figure 8 A structural block diagram of a time-series data prediction device for a business service system provided in an embodiment of this disclosure;

[0030] Figure 9 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present disclosure. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0032] In large-scale business service systems, time series forecasting plays a crucial role in ensuring system reliability. Business service systems continuously collect hundreds of monitoring indicators, including CPU utilization, memory usage, and network traffic, forming high-dimensional time series data. The accuracy of its forecasts directly affects many core business requirements. For example, fault prediction requires early identification of abnormal server states (such as abnormal CPU spikes), capacity planning requires accurate prediction of resource usage trends (such as storage space consumption), and performance optimization relies on accurate prediction of service response latency fluctuations.

[0033] Existing methods for time series prediction include time domain models (such as linear models, attention models, convolutional models, etc.), frequency domain models, and mathematical transform domain models.

[0034] In the time domain, linear models typically rely on simple linear layers or multilayer perceptrons for prediction. Attention models capture long-term dependencies through attention mechanisms, while convolutional models convert time series data into 2D convolutional processing for prediction. Frequency domain models can extract frequency domain features based on Fourier transforms or use a pure frequency domain multilayer perceptron architecture for prediction. Mathematical transform domain models can incorporate Legendre-Fourier domain projections for prediction or map to the basis function domain for prediction.

[0035] However, the high reliability requirements of business service systems necessitate that time series forecasts possess both high accuracy and strong anti-interference capabilities, while existing methods struggle to balance forecast accuracy with robustness to disturbances.

[0036] To address the aforementioned technical issues, this disclosure provides a method for predicting time-series data of a business service system. This method predicts the periodic and residual components of future time-series data of monitoring indicators of the business service system. While preserving the core periodic characteristics of the monitoring indicators, it also utilizes global context information to predict the random fluctuations of the monitoring indicators, reducing the sensitivity to noise and outliers in historical time-series data, and improving the accuracy of prediction and robustness to disturbances.

[0037] The application scenarios of the time-series data prediction method for the business service system in this disclosure embodiment are as follows: Figure 1 As shown, the system receives a prediction request for future time-series data of a business service system monitoring indicator, wherein the prediction request includes first historical time-series data of the monitoring indicator; based on a pre-learned periodic pattern of the business service system monitoring indicator, it determines a first periodic component of the first historical time-series data and a second periodic component of the future time-series data; wherein the periodic pattern is a regular change that occurs periodically in the time-series data; the periodic component is a partial sequence extracted from the time-series data according to the periodic pattern; based on the first historical time-series data and the first periodic component, it determines a first residual component of the first historical time-series data, and predicts a second residual component of the future time-series data based on the first residual component; and it fuses the second periodic component and the second residual component to obtain the future time-series data.

[0038] The following will provide a detailed description of the time-series data prediction method for the business service system disclosed herein, with reference to specific embodiments.

[0039] refer to Figure 2 , Figure 2 This is a flowchart illustrating a time-series data prediction method for a business service system according to an embodiment of this disclosure. The method of this embodiment can be applied to electronic devices such as terminal devices or servers. The time-series data prediction method for the business service system includes:

[0040] S201. Receive a prediction request for future time-series data of the business service system monitoring indicators, wherein the prediction request includes the first historical time-series data of the business service system monitoring indicators.

[0041] In this embodiment, in the business service system, if it is necessary to predict future time-series data of monitoring indicators such as CPU utilization, memory usage, and network traffic, historical time-series data of these indicators must be collected. This prediction of future time-series data needs to be based on historical time-series data preceding the future time-series data. Users can trigger a prediction request for future time-series data of the business service system's monitoring indicators, carrying the historical time-series data of these indicators in the prediction request; this is referred to here as the first historical time-series data.

[0042] It should be noted that the method in this embodiment can be used not only for predicting future time series data of monitoring indicators of business service systems, but also for predicting future time series data in other scenarios, such as power, meteorology, environment, transportation, medical care, and economy.

[0043] S202. Based on the periodic patterns of the business service system monitoring indicators learned in advance, determine the first periodic component of the first historical time series data and the second periodic component of the future time series data.

[0044] The periodic pattern is a regular change that occurs periodically in time series data; the periodic component is a partial sequence extracted from the time series data based on the periodic pattern.

[0045] In this embodiment, all time-series data of the business service system monitoring indicators can be considered as a superposition of periodic components and residual components. That is, all time-series data of the business service system monitoring indicators can be decomposed into periodic components and residual components. The periodic component is the part of the business service system monitoring indicator time-series data that exhibits periodic changes, showing periodic repetition. This periodic repetition may be caused by changes in periodic factors (such as periodically executed tasks in the business service system). The periodic pattern is a regular change that occurs periodically in the time-series data; it is a regular fluctuation that recurs after a periodic interval in the periodic components of the business service system monitoring indicator time-series data. In this embodiment, the periodic pattern includes the periodic components within a periodic range of the periodic components of the business service system monitoring indicator time-series data (in other words, the periodic component can be understood as the waveform of the periodic components of the business service system monitoring indicator time-series data within a single period). Based on pre-learning the periodic pattern, the periodic component can be extracted from the time-series data according to the periodic pattern. The residual component is the difference between the time-series data and the periodic component, such as... Figure 3 As shown, the time-series data of the business service system monitoring indicators generally exhibit fixed periodic changes (i.e., periodic components), but also have certain random fluctuations (i.e., residual components) with respect to the periodic changes.

[0046] It should be noted that the length of the period in this embodiment can be determined based on the inherent periodicity of the monitoring indicators of the business service system. For example, the monitoring indicators of the business service system can usually adopt a daily period, that is, a 24-hour period. When the method of this embodiment is applied to other scenarios, for scenarios with natural periodicity such as electricity and weather, the length of the period can adopt its natural period (such as a daily period); for scenarios without natural periodicity such as the economy, the length of the period can be greater than the sampling period in order to capture a wider range of contextual patterns.

[0047] In this embodiment, the periodic patterns of the business service system monitoring indicators can be learned in advance. By learning to separate the periodic patterns from the original time series, the key periodic information can be completely preserved. Then, the periodic patterns can be applied to the periodic components of the first historical time series data (denoted as the first periodic component) and the periodic components of the future time series data (the second periodic component). That is, the first periodic component of the first historical time series data and the second periodic component of the future time series data can be composed of periodic patterns.

[0048] Optionally, the periodic pattern can be periodically replicated (i.e., one periodic pattern followed by another) to cover the full time range of the first historical time series data and the future time series data, resulting in a periodic component of global time series data that covers the full time range of the first historical time series data and the future time series data. A portion of the time range of the first historical time series data is extracted from the periodic component of the global time series data as the first periodic component of the first historical time series data, and a portion of the time range of the future time series data is extracted from the periodic component of the full time range as the second periodic component of the future time series data.

[0049] In another alternative embodiment, such as Figure 4 As shown, the first time point corresponding to the starting time point of the first historical time series data can also be determined in the periodic pattern, that is, t mod W, where W is the length of the periodic pattern (i.e., the length of one period). The starting point of the periodic pattern is calibrated to the starting time of the first historical time series data to obtain the periodic pattern starting from the first time point, which changes the starting point of the periodic pattern. For example, assuming the periodic pattern is a periodic component from 0:00 to 24:00, and the starting time point of the first historical time series data is 15:00, then the periodic pattern is adjusted to a periodic component from 15:00 to 15:00 the next day. The adjustment process actually moves the portion from 0:00 to 15:00 after 24:00. Furthermore, the periodic pattern starting from the first time point is periodically replicated to construct a periodic component of global time series data covering the full time range of the first historical time series data and future time series data, where the repetition count is... That is, the length of the periodic component over the entire time span is Where H is the length of the entire time range of the first historical time series data and future time series data; then, through truncation, in Figure 4 The first and second periodic components can be represented as P. t:t+H-1 The first periodic component of the first historical time series data is defined as the portion within the range from the start time point to the end time point of the first historical time series data in the periodic component of the full time range. The second periodic component of the future time series data is defined as the portion within the range from the start time point to the end time point of the future time series data in the periodic component of the full time range. This enables the accurate construction of the periodic components of the first historical time series data and the future time series data from the periodic pattern, and reuses the learned key periodic information.

[0050] S203. Determine the first residual component of the first historical time series data based on the first historical time series data and the first periodic component, and predict the second residual component of the future time series data based on the first residual component.

[0051] In this embodiment, the first historical time series data consists of a periodic component and a residual component. The first periodic component of the first historical time series data can be determined in the aforementioned steps. Therefore, the residual component of the first historical time series data (denoted as the first residual component) can be obtained by subtracting the first periodic component from the first historical time series data.

[0052] The aforementioned steps can determine the second periodic component of the future time series data. Since the future time series data is also composed of periodic components and residual components, it is also necessary to predict the residual component of the future time series data (denoted as the second residual component).

[0053] In this embodiment, the second residual component of future time series data can be predicted based on the first residual component of the first historical time series data. It should be noted that any feasible method can be used to predict the second residual component of future time series data based on the first residual component of the first historical time series data.

[0054] For example, a prediction model can be trained and predictions can be made based on the prediction model, where the input of the prediction model is the first residual component of the first historical time series data and the output is the second residual component of the future time series data.

[0055] Alternatively, the first residual component of the first historical time series data can be subjected to a certain mathematical transformation, such as transforming it to another mathematical transformation domain (e.g., frequency domain, trigonometric basis function domain, etc.), and a pre-trained prediction model can be used to make predictions in this mathematical transformation domain. Then, the prediction results can be transformed back to the time domain to obtain the second residual component of the future time series data.

[0056] Of course, other prediction methods can also be used, but this embodiment does not impose any limitations.

[0057] S204. The second periodic component and the second residual component are fused to obtain the future time series data.

[0058] In this embodiment, after obtaining the second periodic component and the second residual component of the future time series data, the second periodic component and the second residual component can be fused together, that is, the second periodic component and the second residual component are added together to obtain the prediction result of the future time series data.

[0059] The business service system time-series data prediction method provided in this embodiment receives a prediction request for future time-series data of a business service system monitoring indicator, wherein the prediction request includes first historical time-series data of the monitoring indicator; based on a pre-learned periodic pattern of the business service system monitoring indicator, a first periodic component of the first historical time-series data and a second periodic component of the future time-series data are determined; wherein the periodic pattern is a regular change that occurs periodically in the time-series data; the periodic component is a partial sequence extracted from the time-series data according to the periodic pattern; a first residual component of the first historical time-series data is determined based on the first historical time-series data and the first periodic component, and a second residual component of the future time-series data is predicted based on the first residual component; the second periodic component and the second residual component are fused to obtain the future time-series data. This embodiment predicts the periodic and residual components of future time-series data of business service system monitoring indicators. While retaining the core periodic characteristics of the business service system monitoring indicators, it also utilizes global context information to predict the random fluctuations of the business service system monitoring indicators, reducing the sensitivity to noise and outliers in historical time-series data, and improving the accuracy of prediction and robustness to disturbances.

[0060] Based on any of the above embodiments, the prediction of the second residual component of the future time series data according to the first residual component in S203 may specifically include:

[0061] Based on the first residual component, a first basis function is generated through a basis function transformation operation, and the first basis function is used to characterize the first residual component.

[0062] Based on the coefficients of the first basis function, the coefficients of the second basis function are predicted using a multilayer perceptron base model. The second basis function is used to characterize the second residual component of the future time series data, and the second residual component of the future time series data is constructed based on the second basis function.

[0063] In this embodiment, based on existing mathematical principles, it is shown that any continuous function in a function space can be identified as a linear combination of basis functions, similar to vector decomposition in linear algebra. For example, DAM (Deep Data-dependant Approximate Analytical Model) uses this principle to transform event sequences into a basis function domain to support any prediction range. Inspired by this, this embodiment treats the time series as a continuous function (mapping time indices to residual components), and projects irregular residual components onto a pre-constructed basis function domain through function fitting.

[0064] Considering the inherent characteristics of time series monitoring indicators in business service systems, this embodiment selects Fourier trigonometric basis functions to capture time series patterns. A preset dimension of trigonometric basis functions can be defined, with 182 selectable frequencies corresponding to the periodicity of minutes, hours, days, and weeks. Each frequency shares two orthogonal basis functions (sine and cosine), generating a total of 364 basis functions to capture short-term and long-term time series dynamics. Specifically, the 182 frequencies may include:

[0065] Minute increments (5-minute intervals): 1 minute, 6 minutes, 11 minutes, ..., 56 minutes;

[0066] Hourly intervals (15-minute intervals): 60 minutes, 75 minutes, 90 minutes, ..., 1425 minutes;

[0067] Daily timeframes (6-hour intervals): 24 hours, 30 hours, 36 hours, ..., 162 hours;

[0068] Weekly levels (1-week interval): 1 week, 2 weeks, 3 weeks, ..., 51 weeks.

[0069] In this embodiment, a basis transformation is used to unify any sampling pattern into a representation of the same dimension, ensuring that the model always learns in a consistent function space. Of course, the preset dimension is not limited to the 182 frequencies mentioned above, and can also be other frequencies.

[0070] Furthermore, based on the first residual component, a first basis function is constructed through basis function transformation operations, namely, a first triangular basis function of a preset dimension. The first triangular basis function is used to represent the first residual component, and its coefficients are obtained. Using a pre-trained multilayer perceptron base model, the coefficients of a second triangular basis function are predicted based on the coefficients of the first triangular basis function. The second triangular basis function is the aforementioned second basis function, which is a triangular basis function of a preset dimension used for the second residual component of future time-series data. The second triangular basis function is constructed based on the coefficients of the second triangular basis function output by the multilayer perceptron base model, and the second residual component of future time-series data is constructed based on the second triangular basis function. The specific process is as follows:

[0071] A first triangular basis function of a preset dimension can be constructed based on the first residual component through basis function transformation operations. That is, a triangular basis function of a preset dimension is fitted based on the first residual component and the corresponding sampling time step. The fitting process is actually solving the coefficient set of the first triangular basis function. This coefficient set includes the coefficient θ of each basis function (sine and cosine), denoted as the first coefficient set.

[0072] Alternatively, the basis function coefficients θ can be solved using the regularized least squares method:

[0073]

[0074] in, Indicates the sampling time step The basis function matrix is ​​calculated at the location. It is a regularization parameter that ensures numerical stability by mitigating ill-conditioned systems. The sampled time-series data (first residual component) has three main advantages: 1) avoiding temporal interpolation artifacts; 2) ensuring consistent functional representation at different sampling locations; and 3) projecting time-series data of different lengths onto a fixed-dimensional coefficient space.

[0075] After constructing a first triangular basis function of a preset dimension based on the first residual component and obtaining the first set of coefficients of the first triangular basis function, the second set of coefficients of the second triangular basis function can be predicted based on the first set of coefficients of the first triangular basis function corresponding to the first residual component. The second triangular basis function is a triangular basis function of a preset dimension used for the second residual component of future time-series data. The prediction method can be any feasible method. In this embodiment, a multilayer perceptron base model is pre-trained. Inputting the first set of coefficients into the multilayer perceptron base model can output the predicted second set of coefficients. Optionally, this embodiment is not limited to the multilayer perceptron base model, or other feasible deep learning models can also be used; this embodiment is not limited to any particular model.

[0076] Furthermore, based on the second set of coefficients, a second triangular basis function representing the second residual component of future time series data can be constructed, and the second residual component of future time series data can be constructed based on the second triangular basis function. In other words, the sampling time step of future time series data is input into the second triangular basis function, and the corresponding value of the second residual component is output through the second triangular basis function.

[0077] After obtaining the second residual component of the future time series data, the second periodic component of the future time series data can be fused with the second residual component to obtain the predicted value of the future time series data.

[0078] Combination Figure 5 The following uses mathematical formulas to explain in detail the time-series data prediction method for the business service system in the above embodiments:

[0079] First, the first historical time-series data of the business service system monitoring metrics are provided. , where H represents the H sampling time step of the first historical time series data, and C is the channel (dimension) of the first historical time series data, which may include different channels such as CPU utilization, memory usage, and network traffic. The business service system time series data prediction method in the above embodiment can be executed for one or more dimensions. This is the sampling time step for the first historical time series data, and For future time steps of future time series data (what needs to be predicted is the value of the time series data at each future time step);

[0080] By using the periodic patterns (LPEs) of pre-learned business service system monitoring metrics, the first periodic component of the first historical time-series data can be generated. and the second periodic component of future time series data :

[0081] ,

[0082] The difference between the first historical time series data and the first periodic component is determined as the first residual component of the first historical time series data: A basis function transformation operation is performed on the first residual component sequence to construct a first triangular basis function of a preset dimension, resulting in the first set of coefficients of the first triangular basis function. ;

[0083] Using a multilayer perceptron base model (e.g., MLP or other base depth models), according to Predict the second set of coefficients for the second trigonometric basis functions. The second triangular basis function is the triangular basis function corresponding to the second residual component of the future time series data. The prediction process can be expressed as:

[0084]

[0085] It should be noted that all channels of the first historical time series data can share a single multilayer perceptron base model. .

[0086] The second residual component of future time series data can be reconstructed using the second trigonometric basis function. get:

[0087]

[0088] Here This represents the basis function matrix to be computed at future time steps.

[0089] Finally, the second residual component and the second periodic component of the future time series data are fused to obtain the future time series data. :

[0090]

[0091] Through the above process, it is clearly understood that directly reconstructing the second residual component of future time-series data from the first triangular basis function is avoided. The main reasons include: 1) The first triangular basis function specifically fits the unique pattern of the first residual component of the first historical time-series data; direct extrapolation would lead to overfitting of historical data patterns and a lack of ability to generalize to future time-series data; 2) The fluctuations in real-world time-series data do not conform to strict mathematical functional relationships, and their inherent non-functional characteristics hinder the direct extrapolation of coefficients. Multilayer perceptron base model The core value of (such as MLP and other foundational deep models) lies in learning the dynamic transformation between historical and future coefficient spaces, effectively capturing the evolutionary patterns of residual patterns.

[0092] Based on any of the above embodiments, it is necessary to pre-learn the periodic patterns of the business service system monitoring indicators and pre-train the multilayer perceptron base model. Considering that directly adopting an end-to-end training method may result in the learned periodic patterns and the multilayer perceptron base model failing to accurately reflect the inherent characteristics of the data, in order to ensure the accuracy and stability of the learning, this embodiment can adopt a two-stage training method, such as... Figure 6 As shown, after obtaining the second historical time-series data of the business service system monitoring indicators as training data, the first stage extracts periodic patterns based on the second historical time-series data; the second stage performs joint optimization of the periodic patterns and the multilayer perceptron base model based on the second historical time-series data, that is, collaborative fine-tuning.

[0093] In the first stage, periodic patterns are extracted based on the second historical time series data, which may specifically include:

[0094] S301. Divide the second historical time series data into a first group of time series data and a second group of time series data that are sequentially linked in time; wherein, the second group of time series data serves as the actual value of the future time series data of the first group of time series data;

[0095] S302. Determine the predicted values ​​of future time series data of the first group of time series data through an initial periodic pattern and a single-layer linear model; wherein, the single-layer linear model is used to map historical residual components to future residual components, and the future residual components are used to generate predicted values ​​of future time series data.

[0096] S303. Determine the loss value based on the actual value of the future time series data of the first group of time series data and the predicted value of the future time series data of the first group of time series data, and adjust the initial periodic pattern and single-layer linear model based on the loss value.

[0097] In this embodiment, considering that a periodic pattern is essentially a series of time-series data within a single period, the values ​​at each time point in the periodic pattern can be used as parameters to be trained, and can be jointly optimized with the parameters of a single-layer linear model. The single-layer linear model is a model that predicts the residual components of future time-series data based on the residual components of historical time-series data in the time domain. Therefore, in this embodiment, using the first set of time-series data, the periodic components of the first set of time-series data and the second set of time-series data can be determined using an initial periodic pattern, thereby determining the residual components of the first set of time-series data. Based on the residual components of the first set of time-series data, the single-layer linear model predicts the residual components of the second set of time-series data. By fusing the periodic components and residual components of the second set of time-series data, the predicted value of the second set of time-series data is obtained, and the loss compared to the actual value is calculated. The parameters of the initial periodic pattern and the single-layer linear model are then adjusted to achieve joint optimization. The first stage can undergo multiple iterations, and the initial periodic pattern can be updated after each iteration. The specific process of determining the predicted value of future time-series data of the first set of time-series data using the initial periodic pattern and the single-layer linear model in the first stage may include:

[0098] S3021. Determine the third periodic component of the first group of time series data and the fourth periodic component of the second group of time series data according to the initial periodic pattern.

[0099] S3022. Determine the third residual component of the first set of time series data based on the first set of time series data and the third periodic component, and predict the fourth residual component of the second set of time series data based on the third residual component using the single-layer linear model.

[0100] S3023. The fourth periodic component and the fourth residual component are fused to obtain the predicted value of the second set of time series data, which is used as the predicted value of the future time series data of the first set of time series data.

[0101] In this embodiment, a single-layer linear model is used to directly map the third residual component to future time to obtain the fourth residual component. However, this embodiment is not limited to using a single-layer linear model; other deep learning models can also be used. The model input is the third residual component, and the output is the fourth residual component, thus avoiding complex basis function transformations or sampling operations.

[0102] As can be seen, the first stage is a simplified architecture that operates directly in the time domain, without involving complex basis function transformations or sampling operations, and can accurately learn the inherent periodicity of the time series data used for training. It is understandable that in the first stage, the periodic pattern and the parameters of the single-layer linear model are used as optimization objectives; the periodic pattern can be viewed as a series of parameters over the time domain.

[0103] This embodiment learns the periodic patterns of the business service system monitoring indicators in the first stage, instead of using classic period-trend decomposition methods such as Moving Average (MOV). This is because classic period-trend decomposition methods like MOV calculate trend components through sliding window aggregation and then derive periodic components through residual analysis, which has two key limitations: 1) Boundary distortion, requiring filler at sequence boundaries (zero-fill / forward / backward fill), introducing artificial artifacts that propagate to the prediction results; 2) Perturbation entanglement, failing to effectively separate perturbations from periodic patterns, reducing robustness. The method in this embodiment for learning the periodic patterns of the business service system monitoring indicators, compared to classic period-trend decomposition methods like MOV, avoids boundary distortion and perturbation entanglement, thus improving robustness. This embodiment separates the periodicity, avoiding information loss due to sampling. Furthermore, the periodic pattern is a learnable periodic term, improving the accuracy of periodic pattern extraction while isolating perturbations within the periodic pattern.

[0104] In the second stage, end-to-end training is used to jointly optimize the periodic pattern and the multilayer perceptron base model based on the training time-series data, i.e., collaborative fine-tuning, to simultaneously improve the accuracy of the periodic pattern and the multilayer perceptron base model. Specifically, this may include:

[0105] S401. Divide the second historical time series data into a first group of time series data and a second group of time series data that are sequentially linked in time; wherein, the second group of time series data serves as the actual value of the future time series data of the first group of time series data;

[0106] S402. Determine the predicted value of the future time series data of the first group of time series data through the periodic pattern and the multilayer sensor base model;

[0107] S403. Determine the loss value based on the actual value of the future time series data of the first group of time series data and the predicted value of the future time series data of the first group of time series data, and adjust the periodic pattern and the multilayer perceptron base model in a coordinated manner based on the loss value.

[0108] In this embodiment, based on the periodic pattern of the first-stage training and combined with the multilayer perceptron (MPB) base model, the predicted values ​​of the second set of training time-series data can be determined according to the third set of time-series data. The MPB base model is a model that predicts the coefficients of basis functions in the basis function domain, as detailed in the above embodiment. After determining the predicted values ​​of the second set of training time-series data, the loss between the predicted values ​​and the actual values ​​can be calculated. The parameters of the periodic pattern and the MPB base model are adjusted based on the loss to achieve joint optimization. The specific process of determining the predicted values ​​of future time-series data of the first set of time-series data using the periodic pattern and the MPB base model may include:

[0109] S4021. Determine the fifth periodic component of the first group of time series data and the sixth periodic component of the second group of time series data according to the periodic pattern.

[0110] S4022. Determine the fifth residual component of the first set of time series data based on the first set of time series data and the fifth periodic component, and predict the sixth residual component of the second set of time series data based on the fifth residual component using the multilayer sensor base model.

[0111] S4023. The sixth periodic component and the sixth residual component are fused to obtain the predicted value of the second set of time series data, which is used as the predicted value of the future time series data of the first set of time series data.

[0112] In this embodiment, the predicted value of the future time series data of the first set of time series data is determined by the periodic pattern and the multilayer perceptron base model, as described in the above embodiment. The specific process of predicting the sixth residual component of the second set of time series data based on the fifth residual component using the multilayer perceptron base model includes:

[0113] Based on the fifth residual component, a third basis function is generated through a basis function transformation operation, and the third basis function is used to characterize the fifth residual component;

[0114] Based on the coefficients of the third basis function, the coefficients of the fourth basis function are predicted using a multilayer perceptron base model. The fourth basis function is used to characterize the sixth residual component of the second set of time series data, and the sixth residual component of the second set of time series data is constructed based on the fourth basis function.

[0115] In this embodiment, the specific process of predicting the sixth residual component of the second set of time-series data based on the fifth residual component using the multilayer perceptron base model is the same as the time-series data prediction method for the business service system described in the above embodiment. The difference lies in the input data, which can be found in the detailed explanation of the time-series data prediction method for the business service system using mathematical formulas described above. Specifically, in the case of using trigonometric basis functions, a third trigonometric basis function is generated based on the fifth residual component through basis function transformation operations; based on the third set of coefficients of the third basis function, the fourth set of coefficients of the fourth trigonometric basis function is predicted using the multilayer perceptron base model, and the sixth residual component of the second set of time-series data is constructed based on the fourth basis function.

[0116] Optionally, the loss values ​​of the first stage and the second stage can be calculated using any feasible loss function. In this embodiment, the loss values ​​of the first stage and the second stage are calculated using the MAE loss function (Mean Absolute Error), which is more robust to anomalies than the MSE loss function (Mean Squared Error). The MAE loss function reduces the impact of outliers through absolute error, thereby improving the model's anti-interference ability.

[0117] Based on the above embodiments, in the second stage, in order to dynamically utilize the entire context window, a basis function aligned sampling method is proposed. That is, when generating the third basis function based on the fifth residual component through basis function transformation operations, the fifth residual component is sampled at random time points. The third basis function is constructed based on the partial data of the fifth residual component obtained by random time point sampling, that is, the third triangular basis function model of the preset dimension. This forces the model to learn and predict based on randomly selected points. However, in the inference stage, the first residual component of the first historical time series data does not need to be sampled at random time points to avoid over-reliance on specific points.

[0118] like Figure 7 As shown, when sampling at random time points, the sampling rate can be controlled using the hyperparameter. Control sampling at random time points. Given (H time steps and C channels), the sampled subsequence contains Each time step is sampled while preserving the channel dimension. Randomly ignoring a portion of time steps in each training iteration prevents the model from becoming overly reliant on specific locations.

[0119] Traditional methods for handling unsampled points typically employ time-domain interpolation techniques (such as zero-padding, forward / backward padding, or linear interpolation). However, these methods introduce human bias, leading to prediction distortion. Existing mathematical principles demonstrate that any continuous function in a function space can be represented as a linear combination of basis functions, similar to vector decomposition in linear algebra. For example, DAM utilizes this principle to transform time series to the basis function domain to support arbitrary prediction ranges. In this embodiment, a portion of the data from the fifth residual component, sampled at random time points, is treated as a continuous function (mapping the sampling time steps to observations). Function fitting projects the irregular sampling sequence onto the basis function domain, avoiding the need for time-domain interpolation and thus preventing distortion caused by interpolation.

[0120] More specifically, after introducing random time point sampling, a basis function transformation operation is performed on the fifth residual component to construct the third basis function, that is, the third triangular basis function of the preset dimension, thus obtaining the third coefficient set of the third triangular basis function model. It can be represented as follows:

[0121]

[0122] By using a random sampling strategy during training, the model learns to fully utilize global temporal input information for prediction. Combined with a basis function alignment strategy, noise introduced by interpolation or padding of unsampled points in the time domain is avoided.

[0123] It should be noted that the training phase described above can be executed on the same device as the inference phase, or on different devices.

[0124] Corresponding to the time-series data prediction method for the business service system in the above embodiment, Figure 8 This is a structural block diagram of a time-series data prediction device for a business service system provided in an embodiment of this disclosure. For ease of explanation, only the parts relevant to the embodiments of this disclosure are shown. (Refer to...) Figure 8 The business service system time series data prediction device 800 includes: a receiving unit 801, a periodic extraction unit 802, a prediction unit 803, and a fusion unit 804.

[0125] The receiving unit 801 is configured to receive a prediction request for future time-series data of monitoring indicators of the business service system, wherein the prediction request includes the first historical time-series data of the monitoring indicators.

[0126] The periodic extraction unit 802 is used to determine the first periodic component of the first historical time series data and the second periodic component of the future time series data based on the periodic patterns of the pre-learned monitoring indicators of the business service system.

[0127] The periodic pattern is a regular change that occurs periodically in time series data; the periodic component is a partial sequence extracted from the time series data based on the periodic pattern.

[0128] The prediction unit 803 is configured to determine a first residual component of the first historical time series data based on the first historical time series data and the first periodic component, and to predict a second residual component of the future time series data based on the first residual component.

[0129] The fusion unit 804 is used to fuse the second periodic component with the second residual component to obtain the future time series data.

[0130] The business service system time-series data prediction device provided in this embodiment receives a prediction request for future time-series data of a business service system monitoring indicator. The prediction request includes first historical time-series data of the monitoring indicator. Based on a pre-learned periodic pattern of the business service system monitoring indicator, a first periodic component of the first historical time-series data and a second periodic component of the future time-series data are determined. The periodic pattern is a regular change that occurs periodically in the time-series data; the periodic component is a partial sequence extracted from the time-series data according to the periodic pattern. A first residual component of the first historical time-series data is determined based on the first historical time-series data and the first periodic component, and a second residual component of the future time-series data is predicted based on the first residual component. The second periodic component and the second residual component are fused to obtain the future time-series data. This embodiment predicts the periodic and residual components of the future time-series data of the business service system monitoring indicator. It retains the core periodic characteristics of the monitoring indicator while also utilizing global context information to predict random fluctuations in the monitoring indicator. This reduces sensitivity to noise and outliers in historical time-series data, improving prediction accuracy and robustness to disturbances.

[0131] In one or more embodiments of this disclosure, when the periodic extraction unit 802 determines the first periodic component of the first historical time-series data and the second periodic component of the future time-series data based on the periodic patterns of the pre-learned business service system monitoring indicators, it is used to:

[0132] The periodic start point of the periodic pattern is calibrated to the start time of the first historical time series data;

[0133] Periodic components of global time-series data covering historical and forecast ranges are constructed by periodically replicating the periodic patterns described above.

[0134] According to the time range of the first historical time series data, the first periodic component of the first historical time series data is extracted from the periodic component of the global time series data, and according to the time range of the future time series data, the second periodic component of the future time series data is extracted from the periodic component of the global time series data.

[0135] In one or more embodiments of this disclosure, the prediction unit 803, when predicting a second residual component of the future time-series data based on the first residual component, is configured to:

[0136] Based on the first residual component, a first basis function is generated through a basis function transformation operation, and the first basis function is used to characterize the first residual component.

[0137] Based on the coefficients of the first basis function, the coefficients of the second basis function are predicted using a multilayer perceptron base model. The second basis function is used to characterize the second residual component of the future time series data, and the second residual component of the future time series data is constructed based on the second basis function.

[0138] In one or more embodiments of this disclosure, the device further includes a training unit for:

[0139] Obtain the second historical time-series data of the monitoring indicators of the business service system;

[0140] Extract periodic patterns based on the second historical time series data;

[0141] The periodic pattern and the multilayer perceptron base model are fine-tuned in a coordinated manner based on the second historical time series data, wherein the multilayer perceptron base model is used to predict the coefficients of the basis functions characterizing the residual components in the process of predicting the second residual component of future time series data.

[0142] In one or more embodiments of this disclosure, the training unit, when extracting periodic patterns based on the second historical time-series data, is configured to:

[0143] The second historical time series data is divided into a first group of time series data and a second group of time series data that are sequentially linked in time; wherein, the second group of time series data serves as the actual value of the future time series data of the first group of time series data;

[0144] The predicted values ​​of future time series data of the first set of time series data are determined by using an initial periodic pattern and a single-layer linear model; wherein, the single-layer linear model is used to map historical residual components to future residual components, and the future residual components are used to generate predicted values ​​of future time series data;

[0145] The loss value is determined based on the actual value of the future time series data of the first group of time series data and the predicted value of the future time series data of the first group of time series data, and the initial periodic pattern and single-layer linear model are adjusted based on the loss value.

[0146] In one or more embodiments of this disclosure, the training unit, when determining the predicted values ​​of future time series data of the first set of time series data through an initial periodic pattern and a single-layer linear model, is configured to:

[0147] The third periodic component of the first group of time series data and the fourth periodic component of the second group of time series data are determined based on the initial periodic pattern.

[0148] The third residual component of the first set of time series data is determined based on the first set of time series data and the third periodic component. The fourth residual component of the second set of time series data is predicted based on the third residual component using the single-layer linear model.

[0149] The fourth periodic component and the fourth residual component are fused to obtain the predicted value of the second set of time series data, which is then used as the predicted value of the future time series data of the first set of time series data.

[0150] In one or more embodiments of this disclosure, when the training unit performs collaborative fine-tuning of the periodic pattern and the multilayer perceptron base model based on the second historical time-series data, it is used to:

[0151] The second historical time series data is divided into a first group of time series data and a second group of time series data that are sequentially linked in time; wherein, the second group of time series data serves as the actual value of the future time series data of the first group of time series data;

[0152] The predicted values ​​of future time series data of the first set of time series data are determined by the periodic pattern and the multilayer perceptron base model.

[0153] The loss value is determined based on the actual value of the future time series data of the first set of time series data and the predicted value of the future time series data of the first set of time series data. The periodic pattern and the multilayer perceptron base model are then adjusted collaboratively based on the loss value.

[0154] In one or more embodiments of this disclosure, when the training unit determines the predicted value of future time series data of the first set of time series data through the periodic pattern and the multilayer perceptron base model, it is configured to:

[0155] The fifth periodic component of the first group of time series data and the sixth periodic component of the second group of time series data are determined according to the periodic pattern.

[0156] The fifth residual component of the first set of time series data is determined based on the first set of time series data and the fifth periodic component. The sixth residual component of the second set of time series data is predicted based on the fifth residual component using the multilayer sensor base model.

[0157] The sixth periodic component and the sixth residual component are fused to obtain the predicted value of the second set of time series data, which is then used as the predicted value of the future time series data of the first set of time series data.

[0158] In one or more embodiments of this disclosure, when the training unit predicts the sixth residual component of the second set of time-series data based on the fifth residual component using the multilayer perceptron base model, it is configured to:

[0159] Based on the fifth residual component, a third basis function is generated through a basis function transformation operation, and the third basis function is used to characterize the fifth residual component;

[0160] Based on the coefficients of the third basis function, the coefficients of the fourth basis function are predicted using a multilayer perceptron base model. The fourth basis function is used to characterize the sixth residual component of the second set of time series data, and the sixth residual component of the second set of time series data is constructed based on the fourth basis function.

[0161] The device provided in this embodiment can be used to execute the technical solutions of the above method embodiments. Its implementation principle and technical effect are similar, and will not be described again here.

[0162] To implement the above embodiments, this disclosure also provides an electronic device.

[0163] refer to Figure 9 The diagram illustrates a structural schematic of an electronic device 900 suitable for implementing embodiments of the present disclosure. The electronic device 900 can be a terminal device or a server. The terminal device can include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, personal digital assistants (PDAs), tablet computers, portable media players (PMPs), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 9 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0164] like Figure 9As shown, the electronic device 900 may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 901, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 902 or a program loaded from a storage device 908 into a random access memory (RAM) 903. The RAM 903 also stores various programs and data required for the operation of the electronic device 900. The processing unit 901, ROM 902, and RAM 903 are interconnected via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.

[0165] Typically, the following devices can be connected to I / O interface 905: input devices 906 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 907 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 908 including, for example, magnetic tapes, hard disks, etc.; and communication devices 909. Communication device 909 allows electronic device 900 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 9 An electronic device 900 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.

[0166] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 909, or installed from a storage device 908, or installed from a ROM 902. When the computer program is executed by a processing device 901, it performs the functions defined in the methods of embodiments of this disclosure.

[0167] It should be noted that the computer-readable storage medium described in this disclosure can be a computer-readable signal medium, a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium may also be any computer-readable storage medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable storage medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0168] The aforementioned computer-readable storage medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0169] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the method shown in the above embodiments.

[0170] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0171] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0172] The units described in the embodiments of this disclosure can be implemented in software or in hardware. The name of a unit does not necessarily limit the unit itself; for example, the first acquisition unit can also be described as "a unit that acquires at least two Internet Protocol addresses".

[0173] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0174] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features disclosed in this disclosure that have similar functions.

[0175] Furthermore, while the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. In certain environments, multitasking and parallel processing may be advantageous. Similarly, while several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0176] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A method for predicting time-series data in a business service system, characterized in that, include: Receive a prediction request for future time-series data of monitoring indicators of the business service system, wherein the prediction request includes first historical time-series data of the monitoring indicators of the business service system. Based on the periodic patterns of the business service system monitoring indicators learned in advance, the first periodic component of the first historical time series data and the second periodic component of the future time series data are determined. The first residual component of the first historical time series data is determined based on the first historical time series data and the first periodic component, and the second residual component of the future time series data is predicted based on the first residual component. The future time series data is obtained by fusing the second periodic component with the second residual component.

2. The method according to claim 1, characterized in that, The determination of the first periodic component of the first historical time-series data and the second periodic component of the future time-series data based on the pre-learned periodic patterns of the business service system monitoring indicators includes: The periodic start point of the periodic pattern is calibrated to the start time of the first historical time series data; Periodic components of global time-series data covering historical and forecast ranges are constructed by periodically replicating the periodic patterns described above. According to the time range of the first historical time series data, the first periodic component of the first historical time series data is extracted from the periodic component of the global time series data, and according to the time range of the future time series data, the second periodic component of the future time series data is extracted from the periodic component of the global time series data.

3. The method according to claim 1, characterized in that, The step of predicting the second residual component of the future time series data based on the first residual component includes: Based on the first residual component, a first basis function is generated through a basis function transformation operation, and the first basis function is used to characterize the first residual component. Based on the coefficients of the first basis function, the coefficients of the second basis function are predicted using a multilayer perceptron base model. The second basis function is used to characterize the second residual component of the future time series data, and the second residual component of the future time series data is constructed based on the second basis function.

4. The method according to claim 1, characterized in that, The method further includes: Obtain the second historical time-series data of the monitoring indicators of the business service system; Extract periodic patterns based on the second historical time series data; The periodic pattern and the multilayer perceptron base model are fine-tuned in a coordinated manner based on the second historical time series data, wherein the multilayer perceptron base model is used to predict the coefficients of the basis functions characterizing the residual components in the process of predicting the second residual component of future time series data.

5. The method according to claim 4, characterized in that, The step of extracting periodic patterns based on the second historical time series data includes: The second historical time series data is divided into a first group of time series data and a second group of time series data that are sequentially linked in time; wherein, the second group of time series data serves as the actual value of the future time series data of the first group of time series data; The predicted values ​​of future time series data of the first set of time series data are determined by using an initial periodic pattern and a single-layer linear model; wherein, the single-layer linear model is used to map historical residual components to future residual components, and the future residual components are used to generate predicted values ​​of future time series data; The loss value is determined based on the actual value of the future time series data of the first group of time series data and the predicted value of the future time series data of the first group of time series data, and the initial periodic pattern and single-layer linear model are adjusted based on the loss value.

6. The method according to claim 5, characterized in that, The step of determining the predicted values ​​of future time series data from the first set of time series data using an initial periodic pattern and a single-layer linear model includes: The third periodic component of the first group of time series data and the fourth periodic component of the second group of time series data are determined based on the initial periodic pattern. The third residual component of the first set of time series data is determined based on the first set of time series data and the third periodic component. The fourth residual component of the second set of time series data is predicted based on the third residual component using the single-layer linear model. The fourth periodic component and the fourth residual component are fused to obtain the predicted value of the second set of time series data, which is then used as the predicted value of the future time series data of the first set of time series data.

7. The method according to claim 4, characterized in that, The step of coordinating fine-tuning the periodic pattern and the multilayer sensor base model based on the second historical time-series data includes: The second historical time series data is divided into a first group of time series data and a second group of time series data that are sequentially linked in time; wherein, the second group of time series data serves as the actual value of the future time series data of the first group of time series data; The predicted value of future time series data of the first set of time series data is determined by the periodic pattern and the multilayer sensor base model; The loss value is determined based on the actual value of the future time series data of the first set of time series data and the predicted value of the future time series data of the first set of time series data. The periodic pattern and the multilayer perceptron base model are then adjusted collaboratively based on the loss value.

8. The method according to claim 7, characterized in that, The step of determining the predicted value of future time series data of the first set of time series data through the periodic pattern and the multilayer perceptron base model includes: The fifth periodic component of the first group of time series data and the sixth periodic component of the second group of time series data are determined according to the periodic pattern. The fifth residual component of the first set of time series data is determined based on the first set of time series data and the fifth periodic component. The sixth residual component of the second set of time series data is predicted based on the fifth residual component using the multilayer sensor base model. The sixth periodic component and the sixth residual component are fused to obtain the predicted value of the second set of time series data, which is then used as the predicted value of the future time series data of the first set of time series data.

9. The method according to claim 8, characterized in that, The step of predicting the sixth residual component of the second set of time series data based on the fifth residual component using the multilayer perceptron base model includes: Based on the fifth residual component, a third basis function is generated through a basis function transformation operation, and the third basis function is used to characterize the fifth residual component; Based on the coefficients of the third basis function, the coefficients of the fourth basis function are predicted using a multilayer perceptron base model. The fourth basis function is used to characterize the sixth residual component of the second set of time series data, and the sixth residual component of the second set of time series data is constructed based on the fourth basis function.

10. A time-series data prediction device for a business service system, characterized in that, include: The receiving unit is used to receive prediction requests for future time-series data of monitoring indicators of the business service system, wherein the prediction request includes the first historical time-series data of the monitoring indicators. A periodic extraction unit is used to determine the first periodic component of the first historical time series data and the second periodic component of the future time series data based on the periodic patterns of the pre-learned monitoring indicators of the business service system. The prediction unit is configured to determine a first residual component of the first historical time series data based on the first historical time series data and the first periodic component, and to predict a second residual component of the future time series data based on the first residual component. A fusion unit is used to fuse the second periodic component and the second residual component to obtain the future time series data.

11. An electronic device, characterized in that, include: Processor and memory; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, implement the method as described in any one of claims 1-9.

13. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-9.

Images