Time sequence prediction system and method based on multi-scale information dynamic modeling

Abstract

The invention provides a time series prediction system and method based on multi-scale information dynamic modeling, and belongs to the technical field of time series prediction.The time series prediction method comprises the steps that firstly, an input sequence is processed through an instance normalization module, then a sequence decomposition module decomposes time series information into different scales, and the energy ratio of each scale is calculated; then, the multi-scale information regulation and control module performs energy enhancement on the low-energy scale information to balance the attention of the model on the scale information; the enhanced multi-scale sequence is sent to a prediction module, prediction values of all scales in the future are generated respectively, and weighted fusion is carried out according to the previously calculated energy proportions of all the scales; and finally, the prediction data are recovered to original distribution through inverse instance normalization. According to the method, the prediction precision is remarkably improved, and through a multi-scale sequence decomposition and energy enhancement mechanism, the perception and modeling capability of the method on hidden key features in a low-energy scale is remarkably improved. The calculation efficiency is high, and the universality is high.

Description

Technical Field

[0001] This invention belongs to the field of time series forecasting technology, specifically, it relates to a time series forecasting system and method based on dynamic modeling of multi-scale information. Background Technology

[0002] Time series forecasting is an important research area in statistics, economics, and engineering, aiming to scientifically infer future trends and patterns by systematically analyzing the evolution patterns of historical data. Efficient and accurate time series forecasting not only helps reveal the internal operating mechanisms of complex systems but also plays a crucial role in various fields such as economic management, public services, and environmental science, including market demand analysis and forecasting, optimal allocation of public resources, and the formulation of ecological protection strategies. With the deepening of digital transformation, the time series data accumulated by various industries exhibits new characteristics of being massive in scale, diverse in type, and frequently collected. Against this backdrop, how to effectively extract information from massive amounts of time series data and achieve efficient modeling and analysis has become a key challenge that urgently needs to be addressed in the current era of data science.

[0003] Time series forecasting is a type of technical method that uses analysis and modeling to infer and predict future data trends based on the patterns of historical data changes over time.

[0004] Given a time series ,in Indicates the number of variables (channels). This represents the length of the historical sequence. Let... Indicates the first The goal of time series forecasting is to predict the values ​​of each variable at each time step. , Let be the future time step to be predicted. Consider a model. ,in Given the model parameters, the entire process of time series prediction can be abstracted as follows:

[0005] Model output results It can be used to predict .

[0006] Time series data, as direct observations of real-world dynamic systems, often contain complex patterns driven by different processes and existing at multiple time scales. Especially under high-sampling-frequency observation environments, these patterns (such as macroscopic cycles on an annual basis, seasonal fluctuations on a daily basis, and even microscopic fluctuations on a minute or second basis) intertwine and superimpose, forming a non-stationary and highly complex series data. To extract predictive information from such complex signals and overcome the limitations of traditional single-scale modeling methods in processing this type of data, researchers have introduced the concept of multi-scale analysis into the field of time series forecasting, proposing a series of innovative multi-scale decomposition modeling methods. The core idea of ​​these methods is to decouple the original time series into multiple components that vary at different time scales through specific signal processing or network structures, thereby reducing the modeling difficulty of each sub-component and capturing its inherent patterns more precisely.

[0007] Current multi-scale time series modeling methods mainly present the following technical paths: Scale decomposition based on moving averages: represented by AutoFormer, this method embeds the decomposition structure into the bottom layer of the model, extracts seasonal and trend components through moving averages, and achieves separation and focused modeling of core time patterns; Scale construction based on resolution adjustment: such as Pathformer, which constructs multi-scale representations by dividing segments of different sizes and uses a dual attention mechanism to collaboratively capture local details and global dependencies; Two-dimensional modeling based on frequency domain transformation: TimesNet uses Fourier transform to reshape one-dimensional time series into a two-dimensional structure according to the dominant frequency, thereby effectively capturing two-dimensional change patterns within and between periods using two-dimensional convolution; Multi-scale feature extraction based on filtering and sampling: models such as SCINet use hierarchical downsampling and convolutional filter groups to extract and fuse time series features from subsequences of different resolutions in an iterative manner.

[0008] However, despite significant progress in scale separation and feature extraction, existing methods still suffer from common shortcomings in multi-scale information fusion. This leads to three core problems: (1) After multi-scale decomposition, the current scale still carries information from other scales, which will cause interference to the modeling of the current scale; (2) In the modeling process, it is easy to overemphasize high-energy patterns and ignore low-energy but important scale components that contain key turning point information; (3) The coarse-grained fusion mechanism cannot dynamically evaluate and weigh the differences in contributions of components at different scales according to the prediction target, which limits the effective use of multi-scale information in complex time series systems.

[0009] The aforementioned problems together constitute the main bottleneck limiting further performance improvements of existing methods. Summary of the Invention

[0010] To overcome the above problems, this invention proposes a time series forecasting system and method based on dynamic modeling of multi-scale information. The research focuses on the dynamic trade-offs and modeling of multi-scale information, and proposes a targeted improved model, DETS.

[0011] This invention is achieved through the following technical solution: A time series forecasting system based on multi-scale information dynamic modeling: The prediction system includes: an instance normalization and de-instance normalization module, a multi-scale sequence decomposition module, a multi-scale information regulation module, and a multi-scale prediction module; The instance normalization and de-instance normalization modules are described, wherein instance normalization is used to standardize the distribution of the input historical time series data; and is connected to the multi-scale prediction module, which restores the prediction results to the original data distribution through de-instance normalization. The multi-scale sequence decomposition module, connected to the instance normalization module, is used to decompose the standardized time-series data into subsequences of different scales. The multi-scale information control module is connected to the multi-scale sequence decomposition module and is used to perform energy equalization processing on the decomposed subsequences at each scale. The multi-scale prediction module is connected to the multi-scale information control module and is used to independently predict and weightedly fuse subsequences at each scale.

[0012] Furthermore, the multi-scale sequence decomposition module includes a frequency domain transformation unit, a frequency band division unit, and an energy calculation unit; The frequency domain conversion unit is used to convert time-domain data into a frequency-domain representation; The frequency band division unit is used to uniformly divide the frequency domain data into multiple non-overlapping scales according to the frequency range; The energy calculation unit is used to calculate the energy value and energy percentage weight at each scale.

[0013] Furthermore, the multi-scale information control module enhances the information at low energy scales to a level comparable to that at high energy scales through energy enhancement operations.

[0014] Furthermore, the multi-scale prediction module includes multiple independent linear predictors, each of which specializes in processing time-series data at one scale.

[0015] Furthermore, the multi-scale prediction module dynamically weights and fuses the prediction results according to the energy proportion weight of each scale.

[0016] A time series forecasting method based on multi-scale information dynamic modeling, the method specifically includes the following steps: Step 1: Perform instance normalization and de-instance normalization on the time series input to eliminate data distribution bias. Step 2: Perform Fourier transform on the input time series information through the multi-scale sequence decomposition module, convert it to the frequency domain and divide it into different scales, and calculate the energy proportion of each scale. Step 3: Based on the energy proportion of each scale, the low-energy scale information is enhanced by the multi-scale information control module to adjust the energy of each scale to the same level. Then, the time domain sequence is restored by frequency band segmentation and inverse Fourier transform to balance the model's attention to each scale information. Step 4: The enhanced multi-scale sequence is fed into the multi-scale prediction module. The linear predictor is used to independently predict the time-domain information of each scale, generate the predicted values ​​of each scale in the future, and perform weighted fusion based on the previously calculated energy proportion of each scale. Step 5: Finally, the predicted data is restored to its original distribution through inverse instance normalization by the instance normalization and inverse instance normalization modules.

[0017] An electronic device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the above method.

[0018] A computer-readable storage medium for storing computer instructions that, when executed by a processor, implement the steps of the above-described method.

[0019] Beneficial effects of the invention Compared with the prior art, the present invention has the following significant advantages: 1. Significantly Improved Prediction Accuracy: Experiments on multiple publicly available time-series prediction datasets (such as ETTh1, ETTm1, and Electricity) demonstrate that our method significantly outperforms the current state-of-the-art models in prediction performance. Specifically, across the two core evaluation metrics for all seven datasets, we achieved 17 best results, comprehensively outperforming all existing comparative methods.

[0020] 2. Balanced multi-scale information focus: This invention effectively alleviates the potential bias of the model towards high-energy scales through multi-scale sequence decomposition and energy enhancement mechanisms, and significantly improves its ability to perceive and model key features hidden in low-energy scales.

[0021] 3. High computational efficiency: While ensuring high prediction accuracy, this invention effectively controls the time and space complexity of the model, achieving a good balance between accuracy and efficiency.

[0022] 4. Versatility: This invention adopts a modular design, has good scalability, and can be flexibly integrated with other time series prediction models in a plug-and-play manner. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the overall structure of the framework of the present invention; Figure 2 Example graph showing the frequency of prediction results for each model on ETTH2; Figure 3 Efficiency experiments were conducted for each model. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] All models in the following examples are trained using the AdamW optimizer, and the loss function is defined as the predicted value. Compared with the true value The mean absolute error (MAE) was used to evaluate the model. Batch size was set according to dataset characteristics: 256 for the Electricity dataset, 32 for the Exchange dataset, and 2048 for the remaining datasets. The datasets were divided into training, test, and validation sets in a 7:2:1 ratio. The maximum number of epochs during training was set to 100, and early stopping was used to prevent overfitting. The model embedding dimension was uniformly set to 256. The input sequence length was fixed at 96, and model performance was evaluated at four prediction horizons: {96, 192, 336, 720}. All reported results are the average of five independent experiments under the same configuration, and the evaluation metrics are mean squared error (MSE) and mean absolute error (MAE). The experimental environment was an NVIDIA RTX 5070ti GPU (16GB VRAM).

[0026] This invention proposes a time series forecasting method (DETS) based on multi-scale information dynamic modeling for time series forecasting.

[0027] (1) To address the problem of multi-scale information extraction, an independent modeling method based on frequency band decomposition is proposed. This method first transforms the original time-series data from the time domain to the frequency domain through Fourier transform, and then divides it into multiple independent frequency band components. By modeling the information of each frequency band separately, the mutual interference of noise between different frequency bands is effectively avoided, while enhancing the model's ability to capture key information within a specific frequency band.

[0028] (2) To address the problem that low-energy scale information is easily overlooked, an energy-based multi-scale information control module was designed. This module uses an amplitude compensation strategy to enhance the intensity of low-energy scale information to the level of high-energy components, thereby increasing the model's sensitivity and attention to low-energy components that contain key signals.

[0029] (3) To achieve an adaptive balance of information contribution at each scale, an adaptive fusion mechanism based on the original energy is proposed. This mechanism dynamically adjusts the fusion weight of each scale component in the final prediction by evaluating the energy distribution of each scale component in the original sequence, thereby achieving effective integration of multi-scale features based on contribution.

[0030] Table 1 Commonly Used Symbols

[0031] The overall structure and process are as follows Figure 1 As shown, it mainly includes four modules: instance normalization and deinstance normalization module, frequency band-based multi-scale sequence decomposition module, energy-based multi-scale information regulation module, and energy-based multi-scale prediction module.

[0032] Figure 1 The left section is the overall flowchart of the framework. First, the input sequence is processed by the instance normalization module to eliminate data distribution bias. Then, the sequence decomposition module decomposes the time-series information into different scales and calculates the energy proportion of each scale for subsequent processing. Next, the multi-scale information regulation module performs energy enhancement on the low-energy scale information to balance the model's focus on information at each scale. The enhanced multi-scale sequence is fed into the prediction module, which generates predicted values ​​for each future scale and performs weighted fusion based on the previously calculated energy proportions of each scale. Finally, inverse instance normalization restores the predicted data to its original distribution.

[0033] Figure 1The right half of the framework details the internal processes of three core modules: a frequency-band-based multi-scale sequence decomposition module, an energy-based multi-scale information regulation module, and an energy-based multi-scale prediction module. The multi-scale sequence decomposition module performs a Fourier transform on the input sequence, converting it to the frequency domain and dividing it into different scales, then calculates the energy proportion of each scale. The multi-scale energy enhancement module adjusts the energy of the information at each scale to the same level based on the energy proportion, and then restores it to the time domain sequence through frequency band segmentation and inverse Fourier transform. The multi-scale prediction module uses a linear predictor to independently predict the time domain information at each scale and fuses the prediction results based on the energy proportion.

[0034] (1) Instance normalization and de-instance normalization module: In time series forecasting tasks, data distribution may drift over time, meaning data from different time periods may exhibit different statistical characteristics. To mitigate the impact of distribution drift on model prediction performance, this framework introduces instance normalization at the input stage to unify the data distribution across different instances. After the model outputs prediction results, inverse instance normalization is performed to restore the data to its original distribution, thereby improving the model's robustness to distribution changes and ensuring the accuracy of the prediction results.

[0035] For input data The calculation process for instance normalization is as follows:

[0036] in For input In the Data for each variable (channel), The mean of this variable. This represents the standard deviation of the variable. and These are learnable scaling and translation parameters. This is a small constant used to prevent division by zero errors. Through the above calculations, data from different time periods can be adjusted to a uniform data distribution, preventing the effects of data offset.

[0037] After the model outputs its predictions, it undergoes inverse instance normalization to restore the predictions to the original data distribution. The specific calculation process is as follows:

[0038] in The first output of the prediction module Predicted values ​​for each variable. For the final inverse instance normalized, the first The predicted values ​​of each variable are obtained. Through the above calculation process, the predictions can be restored to the original distribution.

[0039] (2) Frequency band-based multi-scale sequence decomposition module The multi-scale sequence decomposition module aims to decompose the input sequence into subsequences of different scales. By independently modeling the information at each scale, it suppresses the mutual interference between information at different scales, thereby enhancing the model's ability to model information within a single scale.

[0040] For the input sequence The multi-scale sequence decomposition module first transforms it from the time domain to the frequency domain using Fourier transform, obtaining... Subsequently, Uniformly decomposed in the frequency domain into The frequency bands are divided into several non-overlapping bands, each corresponding to a specific scale. Ultimately, each band is processed independently to focus on the feature information of its respective scale. The calculation process can be represented as follows:

[0041] in, This represents the Fourier transform.

[0042] In addition, when dividing frequency bands, the energy and its relative proportion at each scale must be calculated and recorded simultaneously. This energy proportion will guide the energy enhancement operation and serve as the weighting basis for result fusion in the multi-scale prediction module. The specific energy calculation methods for each scale are as follows:

[0043] Represents the input sequence Total energy, Indicates input data The Energy at a scale , Indicates the first The starting point of each frequency band at each scale, Indicates the length of the frequency band at each scale. The first representing frequency domain information There are several frequency points. Through the above calculations, the energy at S scales can be obtained. Then, the energy proportion at each scale can be calculated using the following method:

[0044]

[0045] in Let S represent a 1×S vector consisting of S scale energies. This represents the energy percentage across S scales.

[0046] (3) Energy-based multi-scale information control module: The multi-scale information regulation module aims to address the problem of uneven energy distribution across different scales. This module amplifies low-energy information to the same level as high-energy information through energy enhancement, thereby balancing the model's focus on information at each scale, preventing the model from overemphasizing high-energy scales, and ensuring that key information at low-energy scales is effectively utilized.

[0047] The multi-scale information regulation module's energy enhancement processing of information directly relies on the energy proportion information at each scale provided by the sequence decomposition module. The calculation method for energy enhancement is as follows:

[0048]

[0049] in, Indicates the first The original frequency domain representation at each scale without energy enhancement. This corresponds to its energy percentage. After energy regulation, the frequency domain output at this scale is denoted as... The set of results after augmentation of frequency domain information at all scales is This allows us to adjust the energy of information at different scales to the same level.

[0050] After completing the above calculations, the frequency domain information at each scale needs to be separated, and the time domain data at each scale needs to be reconstructed through inverse Fourier transform. The specific calculation method is as follows:

[0051] in, The operation aims to achieve scale information segmentation by setting all frequency components outside the current target frequency band to zero. This is the inverse Fourier transform. Through the above processing, the final time-domain data at each scale after energy enhancement are obtained.

[0052] (4) Energy-based multi-scale prediction module: This module aims to predict future data using enhanced time-domain information. .

[0053] This module uses time-domain data at various scales generated after energy enhancement. As input, a linear prediction layer is then used to map the data at each scale to the prediction results at that scale. The specific calculation method is as follows:

[0054] in This is the linear predictor of this module. ). The first prediction for the linear predictor Predicted values ​​at various scales.

[0055] Subsequently, the energy proportions at each scale were calculated using the multi-scale sequence decomposition module. The prediction results at all scales are dynamically weighted and fused. The specific calculation method is as follows:

[0056] This is the model's predicted value.

[0057] Table 2 Comparison of time series forecasting performance based on MSE (mean squared error) and MAE (mean absolute error) indices =96)

[0058] like Figure 2 As shown, frequency domain analysis of the prediction results on the ETTh2 dataset reveals that the predicted output of this invention highly matches the actual values ​​in the low-energy frequency bands circled in red (this invention) and green (baseline model). In contrast, the baseline model shows significant biases in its predictions regardless of whether the low-energy components are distributed in low or high frequencies, exposing the limitations of existing methods in capturing low-energy scale data. The robust performance of this method in this regard confirms its effectiveness in addressing this technical challenge.

[0059] like Figure 3 As shown in the efficiency experiment results on the ETTh2 dataset, our method achieves the best prediction accuracy while having the fastest training speed and significantly fewer parameters than most baseline models, demonstrating excellent computational efficiency and deployment value. A time series forecasting method based on multi-scale information dynamic modeling, the method specifically includes the following steps: Step 1: Perform instance normalization and de-instance normalization on the time series input to eliminate data distribution bias. Step 2: Perform Fourier transform on the input time series information through the multi-scale sequence decomposition module, convert it to the frequency domain and divide it into different scales, and calculate the energy proportion of each scale. Step 3: Based on the energy proportion of each scale, the low-energy scale information is enhanced by the multi-scale information control module to adjust the energy of each scale to the same level. Then, the time domain sequence is restored by frequency band segmentation and inverse Fourier transform to balance the model's attention to each scale information. Step 4: The enhanced multi-scale sequence is fed into the multi-scale prediction module. The linear predictor is used to independently predict the time-domain information of each scale, generate the predicted values ​​of each scale in the future, and perform weighted fusion based on the previously calculated energy proportion of each scale. Step 5: Finally, the predicted data is restored to its original distribution through inverse instance normalization by the instance normalization and inverse instance normalization modules.

[0060] An electronic device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the above method.

[0061] A computer-readable storage medium for storing computer instructions that, when executed by a processor, implement the steps of the above-described method.

[0062] The memory in the embodiments of this application can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM). It should be noted that the memory of the methods described in this invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0063] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired means such as coaxial cable, optical fiber, digital subscriber line, DSL, or wireless means such as infrared, wireless, microwave, etc. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium such as a floppy disk, hard disk, magnetic tape; an optical medium such as a high-density digital video disc, DVD; or a semiconductor medium such as a solid-state disk, SSD, etc.

[0064] In implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software. The steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware processor, or by a combination of hardware and software modules in the processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are omitted here.

[0065] It should be noted that the processor in the embodiments of this application can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method embodiments can be completed by the integrated logic circuits in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied as execution by a hardware decoding processor, or as execution by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above methods.

[0066] The present invention provides a detailed description of a time series prediction system and method based on multi-scale information dynamic modeling, and elucidates the principles and implementation methods of the invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of the invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the invention. Therefore, the content of this specification should not be construed as a limitation of the invention.

Claims

1. A time series prediction system based on multi-scale information dynamic modeling, characterized in that: The prediction system comprises: an instance normalization and anti-instance normalization module, a multi-scale sequence decomposition module, a multi-scale information regulation module, and a multi-scale prediction module; The instance normalization and anti-instance normalization module, wherein the instance normalization is used to normalize the distribution of the input historical time series data; and the multi-scale prediction module is connected to restore the prediction results to the original data distribution through anti-instance normalization; The multi-scale sequence decomposition module is connected to the instance normalization module and is used to decompose the normalized time series data into sub-sequences of different scales; The multi-scale information regulation module is connected to the multi-scale sequence decomposition module and is used to perform energy equalization processing on the decomposed sub-sequences of different scales; The multi-scale prediction module is connected to the multi-scale information regulation module and is used to independently predict and weightedly fuse the sub-sequences of different scales. 2.According to the system of claim 1, characterized in that: The multi-scale sequence decomposition module comprises a frequency domain conversion unit, a frequency band division unit, and an energy calculation unit; The frequency domain conversion unit is used to convert time domain data into frequency domain representation; The frequency band division unit is used to uniformly divide the frequency domain data into multiple non-overlapping scales according to the frequency range; The energy calculation unit is used to calculate the energy value and energy proportion weight of each scale. 3.According to the system of claim 2, characterized in that: The multi-scale information regulation module enhances the information of low energy scale to the level comparable to that of high energy scale through energy enhancement operation. 4.According to the system of claim 3, characterized in that: The multi-scale prediction module comprises multiple independent linear predictors, each of which is used to process time series data of one scale. 5.According to the system of claim 4, characterized in that: The multi-scale prediction module dynamically weightedly fuses the prediction results according to the energy proportion weight of each scale.

6. A time series prediction method based on multi-scale information dynamic modeling, characterized in that: The prediction method is based on the time series prediction system based on multi-scale information dynamic modeling in any one of claims 1 to 5: The method specifically comprises the following steps: Step 1, the instance normalization and anti-instance normalization module of the time series input is normalized by the instance normalization module to eliminate data distribution deviation; Step 2, the input time series information is Fourier transformed by the multi-scale sequence decomposition module, divided into different scales in the frequency domain, and the energy proportion of each scale is calculated; Step 3, based on the energy proportion of each scale, the low energy scale information is enhanced by the multi-scale information regulation module, the energy of each scale information is adjusted to the same level, and the time domain sequence is restored by frequency band segmentation and inverse Fourier transform to balance the attention of the model to each scale information; Step 4, the enhanced multi-scale sequence is sent to the multi-scale prediction module, and linear predictors are used to independently predict the time domain information of each scale to generate prediction values of each scale, and the weighted fusion is performed according to the previously calculated energy proportion of each scale. Step 5, the predicted data is restored to the original distribution by inverse instance normalization of the last instance normalization and anti-instance normalization module. 7.An electronic device comprising a memory and a processor, the memory storing a computer program, wherein, The processor implements the steps of the method recited in claim 6 when executing the computer program.

8. A computer readable storage medium for storing computer instructions, characterized in that, The computer instructions implement the steps of the method recited in claim 6 when executed by the processor.

Images