Multi-element time series data prediction method and device based on hierarchical frequency model

Abstract

The invention relates to the technical field of data prediction, and provides a multivariate time series data prediction method and device based on a hierarchical frequency model, and the method comprises the steps: obtaining multivariate time series data; determining a frequency band dynamic division parameter based on the multivariate time sequence data; determining a low-frequency and high-frequency signal demarcation point based on the frequency band dynamic division parameter and the multivariate time sequence data, and further determining a signal set which corresponds to the multivariate time sequence data and comprises low-frequency signals and high-frequency signals; and determining a prediction result corresponding to the multivariate time series data based on the hierarchical frequency model, the low-frequency signal and the high-frequency signal. According to the method, the low-frequency signal and the high-frequency signal in the multivariate time series data are adaptively determined based on the frequency band dynamic division parameters, the trained hierarchical frequency model is utilized to predict the signals, and low-frequency features and high-frequency details in the data are captured, so that the adaptability to complex characteristics of the multivariate time series data can be improved; and the prediction precision and robustness of the model in processing complex time sequence data are enhanced.

Description

Technical Field

[0001] The present disclosure relates to the technical field of data prediction, and in particular to a multivariate time series data prediction method and device based on a hierarchical frequency model. Background Art

[0002] As a key branch of data science and machine learning, time series forecasting plays a central role in practical applications, such as weather trend analysis in meteorology and load forecasting in the energy industry. With the growing volume of data across various industries and the increasing demands for forecast accuracy and timeliness, accurately predicting valuable information from complex time series data has become a key research issue.

[0003] Statistical methods are often used for short-term forecasting. However, they are often subject to strict assumptions when faced with complex characteristics of real-world time series data, such as high noise, nonlinearity, and long-term trends. Meanwhile, frequency-domain and time-domain methods combine to bridge the gap between the time and frequency domains to accommodate the diversity and complexity of time series. However, because their frequency band division relies on fixed thresholds, they lack adaptability and cannot flexibly address the characteristics of different time series. Summary of the Invention

[0004] The embodiments of the present disclosure at least provide a multivariate time series data prediction method and device based on a hierarchical frequency model, which adaptively determines the low-frequency signals and high-frequency signals in the multivariate time series data based on the dynamic division parameters of the frequency bands, and uses the trained hierarchical frequency model to predict them, thereby capturing the low-frequency features and high-frequency details in the data, thereby improving the adaptability to the complex characteristics of the multivariate time series data, and thereby enhancing the prediction accuracy and robustness of the model when processing complex time series data.

[0005] The present disclosure provides a multivariate time series data prediction method based on a hierarchical frequency model, including: Acquire multivariate time series data; and determine a frequency band dynamic division parameter based on the multivariate time series data; and determine a low- and high-frequency signal demarcation point based on the frequency band dynamic division parameter and the multivariate time series data; Determining a signal set corresponding to the multivariate time series data based on the multivariate time series data and the low- and high-frequency signal demarcation points; wherein the signal set includes a low-frequency signal and a high-frequency signal; A prediction result corresponding to the multivariate time series data is determined based on the trained hierarchical frequency model, the low-frequency signal and the high-frequency signal.

[0006] In some possible embodiments, determining a signal set corresponding to the multivariate time series data based on the multivariate time series data and the low-high frequency signal demarcation point includes: Determine a frequency domain signal corresponding to the multivariate time series data based on a first time-frequency transformation method and the multivariate time series data; Determine a low-frequency spectrum and a high-frequency spectrum based on the frequency domain signal and the low- and high-frequency signal dividing points respectively; The low-frequency signal is determined based on a second time-frequency transform method and the low-frequency spectrum; and the high-frequency signal is determined based on a second time-frequency transform method and the high-frequency spectrum.

[0007] In some possible embodiments, the hierarchical frequency model includes a low-frequency signal processing module, a high-frequency signal processing module, a feature fusion module, and a result prediction module; and determining a prediction result corresponding to the multivariate time series data based on the trained hierarchical frequency model, the low-frequency signal, and the high-frequency signal includes: Inputting the low-frequency signal into the low-frequency signal processing module to obtain a low-frequency signal feature vector; Inputting the high-frequency signal into the high-frequency signal processing module to obtain a high-frequency signal feature vector; Determine a fused feature vector corresponding to the multivariate time series data based on the feature fusion module, the feature fusion ratio coefficient of each feature vector, the low-frequency signal feature vector, and the high-frequency signal feature vector; A prediction result corresponding to the multivariate time series data is determined based on the result prediction module and the fused feature vector.

[0008] In some possible embodiments, after determining the prediction result corresponding to the multivariate time series data based on the result prediction module and the fused feature vector, the method includes: Obtaining a true result corresponding to the multivariate time series data with the same time length as the predicted result, and constructing a proportional coefficient optimization sample set based on the low-frequency signal and the high-frequency signal corresponding to the multivariate time series data and the true result; The hierarchical frequency model is optimized and trained based on the proportional coefficient optimization sample set, and the feature fusion proportional coefficients of each feature vector corresponding to the feature fusion module in the hierarchical frequency model are optimized and iterated, so as to realize the prediction of new multivariate time series data by using the hierarchical frequency model after the feature fusion proportional coefficient optimization iteration.

[0009] In some possible embodiments, the low-frequency signal processing module includes a Bi-LSTM module and a CNN module, and the low-frequency signal feature vector includes a first low-frequency signal feature vector and a second low-frequency signal feature vector; Inputting the low-frequency signal into the low-frequency signal processing module to obtain a low-frequency signal feature vector includes: Using the Bi-LSTM module to capture the global temporal dependency of the low-frequency signal, and generate the first low-frequency signal feature vector; The CNN module is used to extract local spatial features of the low-frequency signal to generate the second low-frequency signal feature vector.

[0010] In some possible embodiments, inputting the high-frequency signal into the high-frequency signal processing module to obtain a high-frequency signal feature vector includes: The high-frequency signal is subjected to noise suppression and feature compression by the lightweight linear transformation layer of the high-frequency signal processing module to generate the high-frequency signal feature vector.

[0011] In some possible embodiments, the hierarchical frequency model is trained by the following steps: Acquire a multivariate time series training data set; wherein the multivariate time series training data set includes multiple sets of multivariate time series training data; Determining a training sample set based on the multivariate time series training data set; wherein the training sample set includes a plurality of training sample subsets, each training sample subset corresponds to different multivariate time series training data, and includes low-frequency signal training samples and high-frequency signal training samples corresponding to the multivariate time series training data; Acquiring a hierarchical frequency model to be trained; and training the hierarchical frequency model to be trained based on the training sample set, including: For each training sample subset, inputting the low-frequency signal training sample corresponding to the training sample subset into a low-frequency signal processing module to obtain a low-frequency signal training feature vector; and inputting the high-frequency signal training sample corresponding to the training sample subset into a high-frequency signal processing module to obtain a high-frequency signal training feature vector; and, based on the feature fusion module, the low-frequency signal training feature vector, and the high-frequency signal training feature vector, determining a fused training feature vector corresponding to the multivariate time series training data; and, based on the result prediction module and the fused training feature vector, determining a prediction result sample corresponding to the multivariate time series training data; Determining a target loss between the prediction result sample and the sample label corresponding to the multivariate time series training data based on a preset loss function, and adjusting the model parameters of the hierarchical frequency model to be trained based on the target loss; Repeat the above steps until the training results meet the preset requirements, and obtain the trained hierarchical frequency model; The step of determining the fused training feature vector corresponding to the multivariate time series training data based on the feature fusion module, the low-frequency signal training feature vector, and the high-frequency signal training feature vector includes: Determining a feature fusion ratio coefficient for each feature vector based on the multivariate time series training data and the learnable weight parameter; Based on the feature fusion module, the feature fusion ratio coefficient of each feature vector, the low-frequency signal training feature vector and the high-frequency signal training feature vector, a fusion training feature vector corresponding to the multivariate time series training data is determined.

[0012] The present disclosure provides a multivariate time series data prediction device based on a hierarchical frequency model, comprising: a demarcation point determination module, configured to obtain multivariate time series data; determine a frequency band dynamic division parameter based on the multivariate time series data; and determine a low- and high-frequency signal demarcation point based on the frequency band dynamic division parameter and the multivariate time series data; A signal partitioning module, configured to determine a signal set corresponding to the multivariate time series data based on the multivariate time series data and the low- and high-frequency signal demarcation points; wherein the signal set includes a low-frequency signal and a high-frequency signal; A data prediction module is used to determine a prediction result corresponding to the multivariate time series data based on the trained hierarchical frequency model, the low-frequency signal and the high-frequency signal.

[0013] In some possible embodiments, the signal division module is specifically configured to: Determine a frequency domain signal corresponding to the multivariate time series data based on a first time-frequency transformation method and the multivariate time series data; Determine a low-frequency spectrum and a high-frequency spectrum based on the frequency domain signal and the low- and high-frequency signal dividing points respectively; The low-frequency signal is determined based on a second time-frequency transform method and the low-frequency spectrum; and the high-frequency signal is determined based on a second time-frequency transform method and the high-frequency spectrum.

[0014] In some possible embodiments, the hierarchical frequency model includes a low-frequency signal processing module, a high-frequency signal processing module, a feature fusion module, and a result prediction module; the data prediction module is specifically used to: Inputting the low-frequency signal into the low-frequency signal processing module to obtain a low-frequency signal feature vector; Inputting the high-frequency signal into the high-frequency signal processing module to obtain a high-frequency signal feature vector; Determine a fused feature vector corresponding to the multivariate time series data based on the feature fusion module, the feature fusion ratio coefficient of each feature vector, the low-frequency signal feature vector, and the high-frequency signal feature vector; A prediction result corresponding to the multivariate time series data is determined based on the result prediction module and the fused feature vector.

[0015] In some possible embodiments, the data prediction module is further configured to: Obtaining a true result corresponding to the multivariate time series data with the same time length as the predicted result, and constructing a proportional coefficient optimization sample set based on the low-frequency signal and the high-frequency signal corresponding to the multivariate time series data and the true result; The hierarchical frequency model is optimized and trained based on the proportional coefficient optimization sample set, and the feature fusion proportional coefficients of each feature vector corresponding to the feature fusion module in the hierarchical frequency model are optimized and iterated, so as to realize the prediction of new multivariate time series data by using the hierarchical frequency model after the feature fusion proportional coefficient optimization iteration.

[0016] In some possible embodiments, the low-frequency signal processing module includes a Bi-LSTM module and a CNN module, and the low-frequency signal feature vector includes a first low-frequency signal feature vector and a second low-frequency signal feature vector; The data prediction module is specifically used for: Using the Bi-LSTM module to capture the global temporal dependency of the low-frequency signal, and generate the first low-frequency signal feature vector; The CNN module is used to extract local spatial features of the low-frequency signal to generate the second low-frequency signal feature vector.

[0017] In some possible embodiments, the data prediction module is specifically configured to: The high-frequency signal is subjected to noise suppression and feature compression by the lightweight linear transformation layer of the high-frequency signal processing module to generate the high-frequency signal feature vector.

[0018] In some possible embodiments, the device further includes: A data acquisition module is used to acquire a multivariate time series training data set; wherein the multivariate time series training data set includes multiple sets of multivariate time series training data; a sample determination module, configured to determine a training sample set based on the multivariate time series training data set; wherein the training sample set includes a plurality of training sample subsets, each training sample subset corresponds to different multivariate time series training data, and includes low-frequency signal training samples and high-frequency signal training samples corresponding to the multivariate time series training data; The model training module is used to obtain a hierarchical frequency model to be trained; and train the hierarchical frequency model to be trained based on the training sample set, including: For each training sample subset, inputting the low-frequency signal training sample corresponding to the training sample subset into a low-frequency signal processing module to obtain a low-frequency signal training feature vector; and inputting the high-frequency signal training sample corresponding to the training sample subset into a high-frequency signal processing module to obtain a high-frequency signal training feature vector; and, based on the feature fusion module, the low-frequency signal training feature vector, and the high-frequency signal training feature vector, determining a fused training feature vector corresponding to the multivariate time series training data; and, based on the result prediction module and the fused training feature vector, determining a prediction result sample corresponding to the multivariate time series training data; Determining a target loss between the prediction result sample and the sample label corresponding to the multivariate time series training data based on a preset loss function, and adjusting the model parameters of the hierarchical frequency model to be trained based on the target loss; Repeat the above steps until the training results meet the preset requirements, and obtain the trained hierarchical frequency model; The step of determining the fused training feature vector corresponding to the multivariate time series training data based on the feature fusion module, the low-frequency signal training feature vector, and the high-frequency signal training feature vector includes: Determining a feature fusion ratio coefficient for each feature vector based on the multivariate time series training data and the learnable weight parameter; Based on the feature fusion module, the feature fusion ratio coefficient of each feature vector, the low-frequency signal training feature vector and the high-frequency signal training feature vector, a fusion training feature vector corresponding to the multivariate time series training data is determined.

[0019] An embodiment of the present disclosure provides a computer device, including: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor and the memory communicate through the bus, and when the machine-readable instructions are executed by the processor, a multivariate time series data prediction method based on a hierarchical frequency model as described in any possible implementation manner described above is performed.

[0020] An embodiment of the present disclosure provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for predicting multivariate time series data based on a hierarchical frequency model as described in any of the possible implementations described above is implemented.

[0021] The multivariate time series data prediction method and device based on the hierarchical frequency model provided in the embodiments of the present disclosure include: first, obtaining multivariate time series data; and determining the frequency band dynamic division parameters based on the multivariate time series data; and, determining the low- and high-frequency signal demarcation points based on the frequency band dynamic division parameters and the multivariate time series data; second, determining the signal set corresponding to the multivariate time series data based on the multivariate time series data and the low- and high-frequency signal demarcation points; wherein the signal set includes low-frequency signals and high-frequency signals; finally, determining the prediction results corresponding to the multivariate time series data based on the trained hierarchical frequency model, low-frequency signals and high-frequency signals.

[0022] The present invention adaptively determines the low-frequency signals and high-frequency signals in multivariate time series data based on the dynamic division parameters of frequency bands, and predicts them using a trained hierarchical frequency model to capture the low-frequency features and high-frequency details in the data. This can improve the adaptability to the complex characteristics of multivariate time series data, and further enhance the prediction accuracy and robustness of the model when processing complex time series data, thereby providing more accurate and reliable data support for practical application scenarios such as meteorological weather trend analysis and energy industry load forecasting.

[0023] In order to make the above-mentioned objectives, features and advantages of the present disclosure more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following briefly introduces the drawings that need to be cited in the embodiments. The drawings herein are incorporated into and constitute a part of the specification. These drawings illustrate embodiments consistent with the present disclosure and, together with the specification, are used to illustrate the technical solutions of the present disclosure. It should be understood that the following drawings only illustrate certain embodiments of the present disclosure and should not be regarded as limiting the scope. For those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without inventive effort.

[0025] Figure 1 A flowchart of a multivariate time series data prediction method based on a hierarchical frequency model provided by an embodiment of the present disclosure is shown; Figure 2 A flow chart of a signal division method provided by an embodiment of the present disclosure is shown; Figure 3 A flowchart of a method for determining a prediction result based on a hierarchical frequency model provided by an embodiment of the present disclosure is shown; Figure 4 A flow chart of a method for dynamically adjusting a feature fusion ratio coefficient provided by an embodiment of the present disclosure is shown; Figure 5A flowchart of a method for training a hierarchical frequency model provided by an embodiment of the present disclosure is shown; Figure 6 A schematic diagram of the structure of a multivariate time series data prediction device based on a hierarchical frequency model provided by an embodiment of the present disclosure is shown; Figure 7 A schematic diagram of the structure of another multivariate time series data prediction device based on a hierarchical frequency model provided by an embodiment of the present disclosure is shown; Figure 8 A schematic structural diagram of a computer device provided by an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solutions and advantages of the embodiments of the present disclosure clearer, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of the embodiments. The components of the embodiments of the present disclosure generally described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present disclosure provided in the drawings is not intended to limit the scope of the disclosure for which protection is sought, but merely represents selected embodiments of the present disclosure. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present disclosure.

[0027] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings.

[0028] The term "and / or" herein simply describes an association relationship, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, and the existence of B alone. In addition, the term "at least one" herein refers to any combination of at least two of any one or more of a plurality of items. For example, "at least one of A, B, and C" can represent any one or more elements selected from the set consisting of A, B, and C.

[0029] As a key branch of data science and machine learning, time series forecasting plays a central role in practical applications, such as weather trend analysis in meteorology and load forecasting in the energy industry. With the growing volume of data across various industries and the increasing demands for forecast accuracy and timeliness, accurately predicting valuable information from complex time series data has become a key research issue.

[0030] Research has found that among traditional time series analysis methods, statistical models represented by ARIMA and SARIMA rely on strict stationary assumptions and linear relationship assumptions, making it difficult to effectively model complex nonlinear patterns, and the capture of high-frequency noise or long-period trends is easily interfered by outliers.

[0031] In recent years, joint modeling of frequency domain and time domain has become a research focus. For example, the signal can be decomposed into low-frequency trends and high-frequency disturbances through Fourier or wavelet transform. However, existing methods mostly use fixed frequency band division (such as preset thresholds), which makes it difficult to adapt to the spectral characteristics of different time series.

[0032] Based on the above research, the embodiments of the present disclosure provide a multivariate time series data prediction method and device based on a hierarchical frequency model. By adaptively determining the low-frequency signals and high-frequency signals in the multivariate time series data based on the dynamic division parameters of the frequency bands, and using the trained hierarchical frequency model to predict them, the low-frequency features and high-frequency details in the data are captured, which can improve the adaptability to the complex characteristics of the multivariate time series data, and further enhance the prediction accuracy and robustness of the model when processing complex time series data, thereby providing more accurate and reliable data support for practical application scenarios such as meteorological weather trend analysis and energy industry load forecasting.

[0033] To facilitate understanding of this embodiment, the execution subject of the multivariate time series data prediction method based on the hierarchical frequency model provided by the embodiment of the present disclosure is first introduced in detail. The execution subject of the multivariate time series data prediction method based on the hierarchical frequency model provided by the embodiment of the present disclosure is a computer device. The computer device can be a terminal device or a server. Among them, the terminal device can also be a mobile device, a user terminal, a terminal, a handheld device, a computing device, a vehicle-mounted device, a wearable device, etc. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud storage, big data and artificial intelligence platforms. Optionally, the method can also be applied to an implementation environment composed of computer devices and servers.

[0034] The following describes in detail the multivariate time series data prediction method based on the hierarchical frequency model provided by the embodiment of the present application in conjunction with the accompanying drawings. Figure 1 FIG. 1 is a flowchart of a multivariate time series data prediction method based on a hierarchical frequency model provided by an embodiment of the present disclosure. The method includes the following steps S101 to S103: S101, obtaining multivariate time series data; and determining a frequency band dynamic division parameter based on the multivariate time series data; and determining a low- and high-frequency signal demarcation point based on the frequency band dynamic division parameter and the multivariate time series data.

[0035] It's understandable that multivariate time series data refers to data on the changes of multiple variables over the same time dimension. This data may come from various sources, such as sensor networks and meteorological observation stations. For example, in the meteorological field, this data may include temperature, humidity, wind speed, and other meteorological data for a certain region over a period of time. In the energy field, this data may include the load, temperature, and electricity demand of multiple power grids in a certain region over a period of time.

[0036] In some possible embodiments, after obtaining multivariate time series data, preprocessing operations such as data cleaning, noise reduction, and redundant information removal can be performed to improve data quality. For example, removing outliers, filling missing values, or standardizing data from different sources to make them comparable can help further improve the accuracy and reliability of subsequent analytical models.

[0037] Here, since multivariate time series data in different scenarios (such as different periodic intensities and noise levels) may be different and complex, if a fixed frequency band threshold is used to divide its low and high frequency signals, it may not be able to adapt to the frequency characteristics of different time series, thereby limiting the separation effect of low-frequency trends and high-frequency disturbances. Therefore, the present disclosure proposes that the corresponding frequency band dynamic division parameters can be determined based on the multivariate time series data. Specifically, the time series characteristics of the data and the frequency characteristics of the signal, such as the periodic changes, sudden changes and noise effects of the signal in different time periods, are comprehensively considered to determine the frequency band dynamic division parameters corresponding to the multivariate time series data. Furthermore, the dividing point between low-frequency signals and high-frequency signals (i.e., the low-frequency and high-frequency signal dividing point) can be determined based on the frequency band dynamic division parameters corresponding to the multivariate time series data and the multivariate time series data. This dividing point determines which frequency components belong to low-frequency signals and which belong to high-frequency signals. Among them, the low-frequency and high-frequency signal dividing point is usually set within the range of the spectrum (i.e., from 0 to the Nyquist frequency), and can be dynamically adjusted according to the frequency band dynamic division parameters and the characteristics and needs of the data.

[0038] In the embodiment of the present disclosure, the process of determining the low and high frequency signal demarcation points according to the frequency band dynamic division parameters and multivariate time series data is mainly expressed as follows: , high frequency band coverage (N is the sequence length of the multivariate time series data.) The parameter 𝛽 serves as a dynamic frequency band partitioning parameter. Its value range and variation pattern are flexibly set based on the specific application scenario and data characteristics of the multivariate time series data. This allows for more flexible and accurate signal partitioning to accommodate the characteristics and needs of different data, thereby improving the accuracy and reliability of subsequent signal processing and analysis.

[0039] In some other embodiments, the dynamic frequency band partitioning parameters may be obtained through other calculation methods, which are not specifically limited herein. For example, an optimization model may be constructed based on the statistical characteristics of the data (such as the mean, variance, and autocorrelation function), and the specific values of the dynamic frequency band partitioning parameters may be determined by finding the optimal solution to the model. Furthermore, the values may be determined by combining machine learning algorithms (such as neural networks and support vector machines).

[0040] In this disclosed embodiment, dynamic frequency band partitioning parameters are used to adaptively adjust the low-frequency / high-frequency boundary, overcoming the limitations of related spectrum analysis techniques that rely on manually set fixed thresholds. This mechanism dynamically optimizes frequency band partitioning based on the characteristics of input time series data, significantly improving its applicability to diverse scenarios.

[0041] S102: Determine a signal set corresponding to the multivariate time series data based on the multivariate time series data and the low- and high-frequency signal demarcation points.

[0042] Specifically, after obtaining the low- and high-frequency signal demarcation points, the multivariate time series data can be segmented based on the low- and high-frequency signal demarcation points to obtain a signal set consisting of low-frequency and high-frequency signals. Here, low-frequency signals refer to the components of time series data that change more slowly and have longer periods, typically reflecting the long-term trend or basic pattern of the data; high-frequency signals refer to the components of time series data that change more quickly and have shorter periods, potentially including data noise, short-term fluctuations, or abnormal events.

[0043] For example, in a weather forecast scenario, multivariate time series data could be temperature, humidity, wind speed, and other data for a specific region over a period of time. By dividing these data based on the low- and high-frequency signal demarcation points, we can identify low-frequency signals (such as long-term seasonal variations in weather data) and high-frequency signals (such as short-term weather fluctuations) in the weather series.

[0044] In some possible embodiments, reference Figure 2 As shown, when performing signal segmentation on multivariate time series data according to the low-high frequency signal dividing point, the following steps S201 to S203 may be included: S201, determining a frequency domain signal corresponding to the multivariate time series data based on a first time-frequency transformation method and the multivariate time series data.

[0045] It can be understood that the time-frequency transformation method is a method that converts time series data into the frequency domain, allowing the data to be analyzed in the frequency domain to reveal the distribution characteristics of the signal in frequency. Common time-frequency transformation methods include Fast Fourier Transform (FFT) and wavelet transform. Here, using the first time-frequency transformation method to transform the multivariate time series data, the frequency domain signal corresponding to the multivariate time series data can be obtained. The frequency domain signal reveals the distribution of the data in different frequency components.

[0046] S202: Determine a low-frequency spectrum and a high-frequency spectrum based on the frequency domain signal and the low- and high-frequency signal dividing point.

[0047] It is understandable that after the low- and high-frequency signal dividing point is determined, the low-frequency spectrum and the high-frequency spectrum can be determined respectively according to the frequency domain signal and the low- and high-frequency signal dividing point.

[0048] Here, the task of separating the low-frequency spectrum and the high-frequency spectrum can be achieved through frequency domain masking. Specifically, a binary mask matrix is generated according to the dividing point of the low and high-frequency signals, where the frequency positions below the dividing point are marked as 1 (retained) and the frequency positions above the dividing point are marked as 0 (suppressed); the mask is multiplied point by point with the frequency domain signal, so as to retain only the low-frequency component (corresponding to the low-frequency spectrum) or only the high-frequency component (corresponding to the high-frequency spectrum).

[0049] S203 : Determine the low-frequency signal based on a second time-frequency transform method and the low-frequency spectrum; and determine the high-frequency signal based on a second time-frequency transform method and the high-frequency spectrum.

[0050] Here, the second time-frequency transform method may be an inverse transform method corresponding to the first time-frequency transform method, such as an inverse Fourier transform or an inverse wavelet transform. Through the inverse transform, the frequency domain signal can be converted back to a time domain signal, thereby obtaining a low-frequency signal and a high-frequency signal, thereby completing the spectrum separation of the time series signal.

[0051] S103: Determine a prediction result corresponding to the multivariate time series data based on the trained hierarchical frequency model, the low-frequency signal and the high-frequency signal.

[0052] Here, after obtaining the low-frequency signal and the high-frequency signal, the low-frequency signal and the high-frequency signal can be predicted based on the trained hierarchical frequency model to obtain the prediction results corresponding to the multivariate time series data. Among them, the hierarchical frequency model includes a low-frequency signal processing module, a high-frequency signal processing module, a feature fusion module and a result prediction module. The low-frequency signal processing module is mainly responsible for processing the low-frequency signal extracted from the multivariate time series data to reflect the long-term trend or regularity of the data. The high-frequency signal processing module is mainly responsible for processing the high-frequency signal extracted from the multivariate time series data to reflect the short-term fluctuations or noise in the data. The feature fusion module is used to fuse the feature vectors extracted from the low-frequency signal and the high-frequency signal according to the feature fusion ratio coefficient of each feature vector to generate a fused feature vector corresponding to the multivariate time series training data. The result prediction module is used to determine the final prediction result based on the fused feature vector. The training method of the hierarchical frequency model will be explained in detail later.

[0053] For example, in a weather forecasting scenario, the challenge lies in the multidimensionality and complexity of the forecast data. Data such as temperature, humidity, and wind speed often exhibit significant seasonality and diurnal fluctuations. A hierarchical frequency model can predict future weather conditions based on long-term seasonal variations (low-frequency signals) and short-term weather fluctuations (high-frequency signals) in meteorological data. For example, by analyzing meteorological data from a specific region over a period of time, low-frequency signals can reveal seasonal temperature trends, such as long-term patterns of higher summer temperatures and lower winter temperatures. High-frequency signals can capture short-term changes in weather systems, such as sudden cold fronts and heavy rainfall in localized areas. Using a hierarchical frequency model, detailed weather information, including daily temperature ranges and precipitation probabilities, can be predicted for the next week, helping relevant departments and the public prepare for emergencies.

[0054] For example, in the energy industry, load forecasting is crucial for power system scheduling. Multivariate time series data may include factors such as regional electricity demand, weather changes, and industrial production. Based on a hierarchical frequency model, accurate future power load forecasts can be obtained based on long-term trends (low-frequency signals) and short-term fluctuations (high-frequency signals) in load data. For example, by analyzing power load data from the past few years, low-frequency signals can reveal the long-term upward trend in power load due to population growth, while high-frequency signals can capture short-term load fluctuations caused by differences in power consumption between weekdays and holidays, and between day and night. Using a hierarchical frequency model, power load can be predicted for each hour of the day ahead, helping power dispatchers rationally plan power generation, ensure stable and reliable power supply, avoid power shortages or surpluses, and improve energy efficiency.

[0055] Specifically, refer to Figure 3 As shown, when using the trained hierarchical frequency model, low-frequency signals and high-frequency signals to determine the prediction results corresponding to the multivariate time series data, the following steps S301 to S304 may be included: S301: Input the low-frequency signal to the low-frequency signal processing module to obtain a low-frequency signal feature vector.

[0056] The low-frequency signal processing module here includes a Bi-LSTM module and a CNN module. The Bi-LSTM (Bi-directional Long Short-Term Memory) is a neural network architecture designed for processing time series data, effectively capturing global temporal dependencies in time series. Compared to traditional LSTM (Long Short-Term Memory) networks, the Bi-LSTM has a bidirectional structure, meaning it can capture information both from the past to the future (forward propagation) and from the future to the past (backward propagation). This enables a more comprehensive understanding of the contextual information in time series data. Convolutional Neural Networks (CNNs) automatically extract local patterns in time series data through convolution operations, making them suitable for processing time series data with local correlations. Specifically, the CNN module extracts local features using sliding convolution kernels, effectively capturing short-term trends or local fluctuations in time series data. Its application in low-frequency signal processing can help reveal local patterns in the data, especially for data with significant noise or pronounced local variations.

[0057] Correspondingly, the low-frequency signal eigenvectors include a first low-frequency signal eigenvector and a second low-frequency signal eigenvector, which represent the global temporal characteristics and local spatial characteristics of the low-frequency signal, respectively. Combining these two different eigenvectors during low-frequency signal analysis helps to more comprehensively and accurately describe the multidimensional characteristics of the signal, thereby improving the performance and accuracy of the analysis model.

[0058] For example, in order to effectively extract global temporal information and local spatial information and improve the expressiveness of low-frequency signal features, when determining the low-frequency signal feature vector based on the low-frequency signal processing module, the following steps (1) to (2) may be included: (1) using the Bi-LSTM module to capture the global temporal dependency of the low-frequency signal and generate the first low-frequency signal feature vector; (2) Utilizing the CNN module to extract local spatial features of the low-frequency signal and generate the second low-frequency signal feature vector.

[0059] Specifically, the Bi-LSTM module comprehensively considers the signal's long-term dependencies and changing trends over time, enabling the model to better understand the signal's evolution and identify global characteristics such as periodicity and trends. Furthermore, the CNN module effectively mines local features within the signal, such as fluctuations, anomalies, and mutations. When processing data with local structure or periodic changes, the CNN's convolution kernel is able to extract subtle changes in the signal.

[0060] In the disclosed embodiment, low-frequency signals provide richer feature representations through the synergistic effect of bidirectional LSTM (capturing long-term dependencies) and CNN convolutional networks (extracting local spatial patterns), thereby improving the model's ability to express low-frequency signals. In particular, when processing data with long-term dependencies and local complexity, the essential characteristics of the signal can be captured more accurately.

[0061] S302: Input the high-frequency signal into the high-frequency signal processing module to obtain a high-frequency signal feature vector.

[0062] It is understandable that in addition to processing low-frequency signals, the hierarchical frequency model also processes high-frequency signals to further improve the accuracy of the prediction. The high-frequency signal processing module usually includes some lightweight linear transformation layers, whose main function is to suppress noise and compress features of high-frequency signals. Among them, noise suppression refers to filtering irrelevant noise in high-frequency signals through specific algorithms or network layers such as filtering to improve the signal-to-noise ratio of the signal. Feature compression refers to reducing the characteristic dimensions of the signal through techniques such as dimensionality reduction, reducing computational complexity, thereby reducing the amount of computation and storage requirements, while retaining key signal information.

[0063] Here, the lightweight linear transformation layer can be implemented using the principal component analysis algorithm. This algorithm performs eigenvalue decomposition on the covariance matrix, selects the eigenvectors corresponding to the first few larger eigenvalues as new coordinate axes, and maps the original features to a new low-dimensional space. Alternatively, a linear discriminant analysis algorithm can be used to perform dimensionality reduction while considering category information, making the projected data more separable. After processing through the lightweight linear transformation layer of the high-frequency signal processing module, a high-frequency signal eigenvector is generated, which can more concisely represent the main characteristics of the high-frequency signal.

[0064] In this disclosed embodiment, low-frequency signals are processed through the synergy of a bidirectional LSTM (capturing long-term dependencies) and a CNN convolutional network (extracting local spatial patterns), achieving deep modeling of complex low-frequency signals. High-frequency signals are then subjected to lightweight linear transformations to suppress noise while preserving their short-term dynamic characteristics and avoiding oversmoothing. This achieves a "divide-and-conquer" approach to processing low- and high-frequency signal components, overcoming the modeling limitations of single models (such as pure LSTM or CNN).

[0065] S303: Determine a fused feature vector corresponding to the multivariate time series data based on the feature fusion module, the feature fusion ratio coefficient of each feature vector, the low-frequency signal feature vector, and the high-frequency signal feature vector.

[0066] Here, after obtaining the low-frequency signal feature vectors (the first low-frequency signal feature vector and the second low-frequency signal feature vector) and the high-frequency signal feature vector, the feature fusion module will perform weighted fusion on the first low-frequency signal feature vector, the second low-frequency signal feature vector, and the high-frequency signal feature vector according to the feature fusion ratio coefficient of each feature vector to obtain a comprehensive feature representation, namely the fused feature vector.

[0067] Specifically, during the fusion process, each eigenvector is multiplied by its corresponding weight (i.e., the feature fusion scale coefficient), and then these weighted eigenvectors are added together to generate a fused eigenvector. This ensures that the feature information of different frequency components is properly represented and integrated in the fused eigenvector, preserving the global temporal and local spatial characteristics of the low-frequency signal while incorporating the rapid changes and detailed information of the high-frequency signal.

[0068] For example, the weighted fusion formula of weight parameters can be expressed as: ; in, 、 、 is the feature fusion ratio coefficient of each feature vector, 、 、 They are low-frequency Bi-LSTM output, low-frequency CNN output, and high-frequency output respectively.

[0069] For example, since multivariate time series data may exhibit different characteristic patterns at different stages, for example, in the trend-dominated stage, the global time series characteristics of low-frequency signals may be more important; while in the noise-significant stage, the detailed information of high-frequency signals may be more critical. Therefore, in order to more effectively fuse the information in each feature vector, the feature fusion ratio coefficient of each feature vector proposed in the present disclosure is dynamically adjusted, that is, the fusion ratio coefficient of each feature vector can be dynamically adjusted according to the characteristic pattern of the current time series data, so that the fused feature representation can better adapt to the current data characteristics. The specific adjustment process will be explained later.

[0070] S304: Determine a prediction result corresponding to the multivariate time series data based on the result prediction module and the fused feature vector.

[0071] Furthermore, after obtaining the fused feature vector, it can be input into the result prediction module, which can include a series of fully connected layers and activation functions (such as ReLU and Sigmoid) for further feature extraction and transformation of the fused feature vector. In a fully connected layer, each neuron is connected to all neurons in the previous layer. The input features are linearly transformed using the learned weight matrix, and nonlinear factors are introduced through the activation function to enhance the model's expressive power. After being processed by multiple layers of fully connected layers, the fused feature vector is mapped into a high-dimensional space that contains rich information related to the prediction task. Finally, through the output layer of the result prediction module (such as the softmax layer), the feature representation in the high-dimensional space is mapped to the specific prediction result, resulting in a prediction result corresponding to the multivariate time series training data.

[0072] For example, in order to realize the dynamic adjustment of the feature fusion ratio coefficient proposed in the present disclosure, refer to Figure 4 As shown in FIG. 1 , a method for dynamically adjusting a feature fusion ratio coefficient proposed in the present disclosure may include the following steps S3041 to S3042: S3041, obtaining a real result corresponding to the multivariate time series data with the same time length as the predicted result, and constructing a proportional coefficient optimization sample set based on the low-frequency signal and high-frequency signal corresponding to the multivariate time series data and the real result.

[0073] Here, the true result is the true value corresponding to the multivariate time series data in the actual process, and its time length is the same as the time length of the predicted result corresponding to the multivariate time series data. Then, a proportional coefficient optimization sample set is constructed based on the low-frequency signal and high-frequency signal corresponding to the multivariate time series data and the true result, so as to achieve optimized training of the feature fusion module of the hierarchical frequency model.

[0074] In some possible embodiments, after constructing the scale factor optimized sample set, it is necessary to re-partition the low-frequency and high-frequency signals. This can be done by following steps S101-S102 described above, which will not be repeated here. By re-partitioning the low-frequency and high-frequency signals in the scale factor optimized sample set, information at different frequency levels can be better extracted based on the actual characteristics and variations of the data, thereby optimizing the training process, enhancing the model's ability to handle complex time series data, and improving prediction accuracy.

[0075] S3042, optimizing and training the hierarchical frequency model based on the proportional coefficient optimization sample set, optimizing and iterating the feature fusion proportional coefficients of each feature vector corresponding to the feature fusion module in the hierarchical frequency model, and realizing the prediction of new multivariate time series data by using the hierarchical frequency model after the feature fusion proportional coefficient optimization iteration.

[0076] Specifically, during the optimization training process, an iterative algorithm can be used to dynamically adjust the feature fusion coefficients of each feature vector based on the error between the model's predicted results and the actual results on the scale coefficient optimization sample set. After multiple iterations and adjustments, an optimal set of feature fusion coefficients can be obtained. The iterative hierarchical frequency model can then be optimized using these coefficients to predict new multivariate time series data. That is, when new multivariate time series data arrives, the model can fuse the feature vectors of the low-frequency and high-frequency signals based on the optimized feature fusion coefficients, resulting in more accurate predictions.

[0077] Similarly, when predicting each new multivariate time series data, the above steps S3041~S3042 will be used to implement the optimization iteration of the feature fusion proportional coefficients of each feature vector corresponding to the feature fusion module in the hierarchical frequency model. That is, after each prediction of new multivariate time series data, the proportional coefficient optimization sample set will be reconstructed (the actual results of the new multivariate time series data will be included in the data set), and the low-frequency signal and high-frequency signal will be re-divided. Then, the model will be optimized and trained based on the new sample set to obtain a new optimal feature fusion proportional coefficient. Finally, the new feature fusion proportional coefficient will be used to predict the new round of multivariate time series data. In this way, the low-frequency and high-frequency signal demarcation points and proportional coefficients can be dynamically determined throughout the entire process to achieve adaptive adjustment of the model to changes in the characteristics of the multivariate time series data and effectively integrate key information.

[0078] The following is a detailed description of the training method of the above-mentioned hierarchical frequency model, see Figure 5 FIG. 5 is a flowchart of a method for training a hierarchical frequency model according to an embodiment of the present disclosure. The method for training a hierarchical frequency model includes the following steps S501 to S503: S501, obtaining a multivariate time series training dataset.

[0079] Specifically, the multivariate time series training dataset includes multiple groups of multivariate time series training data. Each group of multivariate time series training data consists of features of multiple different dimensions under the same time dimension. For example, assuming that a weather prediction model needs to be trained, the multivariate time series data may include multiple features such as temperature, humidity, wind speed, etc. that change over time within a day or a week.

[0080] S502: Determine a training sample set based on the multivariate time series training data set.

[0081] Here, the training sample set includes multiple training sample subsets, each of which corresponds to different multivariate time series training data and includes low-frequency signal training samples and high-frequency signal training samples corresponding to the multivariate time series training data. Specifically, the specific method for determining the training sample set based on the multivariate time series training data set can refer to steps S101-S102 above and is not further described here.

[0082] S503, obtaining a hierarchical frequency model to be trained; and training the hierarchical frequency model to be trained based on the training sample set until the training result meets the preset requirements, thereby obtaining the trained hierarchical frequency model.

[0083] Here, when training the to-be-trained hierarchical frequency model according to the training sample set, the following (a) to (b) may be included: (a) for each training sample subset, inputting the low-frequency signal training samples corresponding to the training sample subset into a low-frequency signal processing module to obtain a low-frequency signal training feature vector; and inputting the high-frequency signal training samples corresponding to the training sample subset into a high-frequency signal processing module to obtain a high-frequency signal training feature vector; and, based on a feature fusion module, the low-frequency signal training feature vector, and the high-frequency signal training feature vector, determining a fused training feature vector corresponding to the multivariate time series training data; and, based on a result prediction module and the fused training feature vector, determining a prediction result sample corresponding to the multivariate time series training data; (b) determining a target loss between the prediction result sample and the sample label corresponding to the multivariate time series training data based on a preset loss function, and adjusting model parameters of the hierarchical frequency model to be trained based on the target loss.

[0084] It is understood that the model operation process is similar to the principles of steps S301-S304 above. Please refer to steps S301-S304 for details and will not be repeated here. After obtaining the prediction result samples corresponding to the multivariate time series training data, the model performance is evaluated by calculating the target loss between the prediction result samples and the true sample labels. The target loss can be measured using a loss function to quantify the model's prediction error. The model parameters are then adjusted through an optimization algorithm to improve prediction accuracy.

[0085] Here, when adjusting model parameters, the optimization algorithm can update the gradient information of the model parameters based on the loss function. The gradient information indicates the direction and rate at which the loss function value changes with the model parameters. By updating the model parameters in the opposite direction of the gradient, the loss function value can be reduced as quickly as possible. Furthermore, to accelerate convergence and avoid falling into local optimal solutions, the optimization algorithm can also incorporate methods such as learning rate decay, momentum terms, and adaptive learning rate adjustment. Common optimization algorithms include gradient descent and stochastic gradient descent (SGD).

[0086] In some possible embodiments, during the iterative model training process, a backpropagation algorithm is used for iterative training to adjust model parameters. Specifically, the model parameters are adjusted based on the target loss, and the parameters are updated using an optimizer (such as the Adam optimizer). The Adam optimizer can automatically adjust the learning rate based on data changes, thereby improving training efficiency and convergence speed. In addition, a dynamic learning rate strategy can be used. If the target loss does not decrease for multiple consecutive rounds (for example, if the target loss does not decrease for 10 consecutive rounds), the learning rate is reduced (for example, by half the original value) to accelerate convergence and avoid falling into local optimal solutions.

[0087] In this disclosure, the default loss function uses the mean square error (MSE), which measures the model performance by calculating the mean squared difference between the predicted result sample and the sample label corresponding to the multivariate time series training data as the target loss. The formula is: ; in, is the sample label corresponding to the multivariate time series training data, is the prediction result sample, and N is the number of samples.

[0088] In some other embodiments, the loss function may also include but is not limited to a cross entropy loss function and a logarithmic loss function, etc., which are not specifically limited here.

[0089] It can be understood that after calculating the target loss and adjusting the model parameters of the hierarchical frequency model to be trained based on the target loss, the above steps (a) to (b) are repeated until the training results meet the preset requirements and a trained hierarchical frequency model is obtained.

[0090] Specifically, after executing steps (a) and (b), the current model's training performance needs to be evaluated to determine whether the current training results meet the preset requirements. This can be achieved based on preset performance indicators such as accuracy, recall, and F1 score. If the current training results do not meet the preset requirements (i.e., the performance indicator does not reach the preset threshold or the loss function value during training does not converge to a sufficiently small range), it is necessary to return to step (a) and continue the next round of model training. During this process, the model parameters are updated based on the gradient information calculated by the previous round of backpropagation, in order to gradually optimize the model performance in subsequent iterations.

[0091] In some possible embodiments, after obtaining the trained hierarchical frequency model, further performance evaluation can be performed, that is, the model performance can be comprehensively measured by multiple indicators, including mean square error (MSE), mean absolute error (MAE) and determination coefficient ( ).

[0092] Here, MAE reflects the absolute level of prediction error, and the formula is: ; here, Evaluate the model's ability to explain data fluctuations. The formula is: .

[0093] For example, when determining the feature fusion ratio function for each feature vector, a learnable weight parameter (such as fusion_weights) can be introduced. These learnable weight parameters can be continuously learned and adjusted during model training to achieve dynamic weighted fusion of the outputs of the Bi-LSTM module, the CNN module, and the high-frequency module. Here, the learnable weight parameters are automatically optimized during model training using a backpropagation algorithm to adapt to the spectral characteristics and temporal dynamics of different input data. Specifically, each feature module is assigned an independent learnable weight parameter, which is initialized to a random value and updated using gradient descent during training. To ensure that the fused feature representation has a reasonable value range and that the weight parameters meet the constraints of non-negativity and summing to 1, the Softmax function can be used to normalize the weight parameters. The output features of each module are then weighted summed based on the normalized weights to obtain the fused feature representation.

[0094] For example, in the trend-dominant phase (i.e., when low-frequency vectors dominate), the weights of the low-frequency modules (Bi-LSTM and CNN) can be increased, allowing the fused feature representation to focus more on the global temporal characteristics of the signal. In the noise-significant phase (i.e., when high-frequency noise or rapidly changing details are present), the weights of the high-frequency modules can be increased, allowing the fused feature representation to better capture the signal's details. This dynamic weight fusion strategy significantly improves the model's adaptability to dynamic changes in temporal features.

[0095] In some other embodiments, the feature fusion ratio coefficient of each feature vector can also be obtained by iterative optimization through other optimization algorithms, such as the dung beetle algorithm, etc., which is not specifically limited here. A suitable optimization algorithm can be selected according to actual needs to determine the feature fusion ratio coefficient of each feature vector.

[0096] The multivariate time series data prediction method and device based on the hierarchical frequency model provided in the embodiments of the present disclosure adaptively determine the low-frequency signals and high-frequency signals in the multivariate time series data based on the dynamic division parameters of the frequency bands, and predict them using the trained hierarchical frequency model, thereby capturing the low-frequency features and high-frequency details in the data. This can improve the adaptability to the complex characteristics of the multivariate time series data, and further enhance the prediction accuracy and robustness of the model when processing complex time series data, thereby providing more accurate and reliable data support for practical application scenarios such as meteorological weather trend analysis and energy industry load forecasting.

[0097] Those skilled in the art will understand that in the above-mentioned method of the specific implementation method, the writing order of each step does not mean a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0098] Based on the same inventive concept, the embodiments of the present disclosure also provide a multivariate time series data prediction device based on a hierarchical frequency model corresponding to the multivariate time series data prediction method based on a hierarchical frequency model. Since the principle of solving the problem by the device in the embodiments of the present disclosure is similar to the above-mentioned multivariate time series data prediction method based on a hierarchical frequency model in the embodiments of the present disclosure, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be repeated.

[0099] Reference Figure 6 FIG. 1 is a schematic diagram of a multivariate time series data prediction device 600 based on a hierarchical frequency model provided by an embodiment of the present disclosure, wherein the device includes: A demarcation point determination module 601 is configured to obtain multivariate time series data; determine a frequency band dynamic division parameter based on the multivariate time series data; and determine a low-frequency and high-frequency signal demarcation point based on the frequency band dynamic division parameter and the multivariate time series data. A signal partitioning module 602 is configured to determine a signal set corresponding to the multivariate time series data based on the multivariate time series data and the low- and high-frequency signal demarcation points; wherein the signal set includes a low-frequency signal and a high-frequency signal; The data prediction module 603 is used to determine the prediction result corresponding to the multivariate time series data based on the trained hierarchical frequency model, the low-frequency signal and the high-frequency signal.

[0100] In some possible embodiments, the signal division module 602 is specifically configured to: Determine a frequency domain signal corresponding to the multivariate time series data based on a first time-frequency transformation method and the multivariate time series data; Determine a low-frequency spectrum and a high-frequency spectrum based on the frequency domain signal and the low- and high-frequency signal dividing points respectively; The low-frequency signal is determined based on a second time-frequency transform method and the low-frequency spectrum; and the high-frequency signal is determined based on a second time-frequency transform method and the high-frequency spectrum.

[0101] In some possible embodiments, the hierarchical frequency model includes a low-frequency signal processing module, a high-frequency signal processing module, a feature fusion module, and a result prediction module; the data prediction module 603 is specifically used to: Inputting the low-frequency signal into the low-frequency signal processing module to obtain a low-frequency signal feature vector; Inputting the high-frequency signal into the high-frequency signal processing module to obtain a high-frequency signal feature vector; Determine a fused feature vector corresponding to the multivariate time series data based on the feature fusion module, the feature fusion ratio coefficient of each feature vector, the low-frequency signal feature vector, and the high-frequency signal feature vector; A prediction result corresponding to the multivariate time series data is determined based on the result prediction module and the fused feature vector.

[0102] In some possible embodiments, the data prediction module 603 is further configured to: Obtaining a true result corresponding to the multivariate time series data with the same time length as the predicted result, and constructing a proportional coefficient optimization sample set based on the low-frequency signal and the high-frequency signal corresponding to the multivariate time series data and the true result; The hierarchical frequency model is optimized and trained based on the proportional coefficient optimization sample set, and the feature fusion proportional coefficients of each feature vector corresponding to the feature fusion module in the hierarchical frequency model are optimized and iterated, so as to realize the prediction of new multivariate time series data by using the hierarchical frequency model after the feature fusion proportional coefficient optimization iteration.

[0103] In some possible embodiments, the low-frequency signal processing module includes a Bi-LSTM module and a CNN module, and the low-frequency signal feature vector includes a first low-frequency signal feature vector and a second low-frequency signal feature vector; The data prediction module 603 is specifically used to: Using the Bi-LSTM module to capture the global temporal dependency of the low-frequency signal, and generate the first low-frequency signal feature vector; The CNN module is used to extract local spatial features of the low-frequency signal to generate the second low-frequency signal feature vector.

[0104] In some possible embodiments, the data prediction module 603 is specifically configured to: The high-frequency signal is subjected to noise suppression and feature compression by the lightweight linear transformation layer of the high-frequency signal processing module to generate the high-frequency signal feature vector.

[0105] In some possible embodiments, reference Figure 7 Said device further comprises: The data acquisition module 604 is used to acquire a multivariate time series training data set; wherein the multivariate time series training data set includes multiple sets of multivariate time series training data; A sample determination module 605 is configured to determine a training sample set based on the multivariate time series training data set; wherein the training sample set includes a plurality of training sample subsets, each training sample subset corresponding to different multivariate time series training data and including low-frequency signal training samples and high-frequency signal training samples corresponding to the multivariate time series training data; The model training module 606 is configured to obtain a hierarchical frequency model to be trained and train the hierarchical frequency model to be trained based on the training sample set, including: For each training sample subset, inputting the low-frequency signal training sample corresponding to the training sample subset into a low-frequency signal processing module to obtain a low-frequency signal training feature vector; and inputting the high-frequency signal training sample corresponding to the training sample subset into a high-frequency signal processing module to obtain a high-frequency signal training feature vector; and, based on the feature fusion module, the low-frequency signal training feature vector, and the high-frequency signal training feature vector, determining a fused training feature vector corresponding to the multivariate time series training data; and, based on the result prediction module and the fused training feature vector, determining a prediction result sample corresponding to the multivariate time series training data; Determining a target loss between the prediction result sample and the sample label corresponding to the multivariate time series training data based on a preset loss function, and adjusting the model parameters of the hierarchical frequency model to be trained based on the target loss; Repeat the above steps until the training results meet the preset requirements, and obtain the trained hierarchical frequency model; The step of determining the fused training feature vector corresponding to the multivariate time series training data based on the feature fusion module, the low-frequency signal training feature vector, and the high-frequency signal training feature vector includes: Determining a feature fusion ratio coefficient for each feature vector based on the multivariate time series training data and the learnable weight parameter; Based on the feature fusion module, the feature fusion ratio coefficient of each feature vector, the low-frequency signal training feature vector and the high-frequency signal training feature vector, a fusion training feature vector corresponding to the multivariate time series training data is determined.

[0106] Based on the same technical concept, the embodiment of the present disclosure also provides a computer device. Figure 8 800 according to an embodiment of the present disclosure, including a processor 801, a memory 802, and a bus 803. The memory 802 is used to store execution instructions and includes a memory 8021 and an external memory 8022. The memory 8021 is also referred to as internal memory and is used to temporarily store calculation data in the processor 801 and data exchanged with an external memory 8022, such as a hard disk. The processor 801 exchanges data with the external memory 8022 through the memory 8021.

[0107] In the embodiment of the present application, the memory 802 is specifically used to store application code for executing the solution of the present application, and the execution is controlled by the processor 801. That is, when the computer device 800 is running, the processor 801 communicates with the memory 802 via the bus 803, so that the processor 801 executes the application code stored in the memory 802, thereby performing the method described in any of the aforementioned embodiments.

[0108] The memory 802 may be, but is not limited to, a random access memory (RAM), a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), etc.

[0109] Processor 801 may be an integrated circuit chip with signal processing capabilities. Such processors may be general-purpose processors, including central processing units (CPUs) and network processors (NPs). They may also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. These processors may implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of the present invention. A general-purpose processor may be a microprocessor or any conventional processor.

[0110] It should be understood that the structure illustrated in the embodiments of the present application does not constitute a specific limitation on the computer device 800. In other embodiments of the present application, the computer device 800 may include more or fewer components than shown, or may combine or separate certain components, or arrange the components differently. The components shown in the illustrations may be implemented in hardware, software, or a combination of software and hardware.

[0111] The present disclosure also provides a computer-readable storage medium having a computer program stored thereon. When executed by a processor, the computer program executes the steps of the multivariate time series data prediction method based on a hierarchical frequency model described in the above method embodiment. The storage medium may be a volatile or non-volatile computer-readable storage medium.

[0112] An embodiment of the present disclosure also provides a computer program product, which carries a program code. The instructions included in the program code can be used to execute the steps of the multivariate time series data prediction method based on the hierarchical frequency model described in the above method embodiment. For details, please refer to the above method embodiment, which will not be repeated here.

[0113] The computer program product may be implemented in hardware, software, or a combination thereof. In one embodiment, the computer program product is implemented as a computer storage medium. In another embodiment, the computer program product is implemented as a software product, such as a software development kit (SDK).

[0114] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system and device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. In the several embodiments provided in the present disclosure, it should be understood that the disclosed system and method can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0115] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0116] In addition, each functional unit in each embodiment of the present disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0117] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present disclosure, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present disclosure. The aforementioned storage medium includes: various media that can store program code, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0118] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present disclosure, which are used to illustrate the technical solutions of the present disclosure, rather than to limit them. The scope of protection of the present disclosure is not limited thereto. Although the present disclosure has been described in detail with reference to the above-described embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above-described embodiments within the technical scope disclosed in the present disclosure, or replace some of the technical features therein with equivalents. Such modifications, changes, or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure, and should be included in the scope of protection of the present disclosure. Therefore, the scope of protection of the present disclosure shall be subject to the scope of protection of the claims.

Claims

1. A multivariate time series data prediction method based on a hierarchical frequency model, characterized in that: include: Obtain multivariate time series data; and determining frequency band dynamic division parameters based on the multivariate time series data; and determining a low- and high-frequency signal demarcation point based on the frequency band dynamic division parameter and the multivariate time series data; Determining a signal set corresponding to the multivariate time series data based on the multivariate time series data and the low- and high-frequency signal demarcation points; wherein the signal set includes a low-frequency signal and a high-frequency signal; A prediction result corresponding to the multivariate time series data is determined based on the trained hierarchical frequency model, the low-frequency signal and the high-frequency signal.

2. The method according to claim 1, characterized in that The determining of a signal set corresponding to the multivariate time series data based on the multivariate time series data and the low-high frequency signal demarcation point includes: Determine a frequency domain signal corresponding to the multivariate time series data based on a first time-frequency transformation method and the multivariate time series data; Determine a low-frequency spectrum and a high-frequency spectrum based on the frequency domain signal and the low- and high-frequency signal dividing points respectively; The low-frequency signal is determined based on a second time-frequency transform method and the low-frequency spectrum; and the high-frequency signal is determined based on a second time-frequency transform method and the high-frequency spectrum.

3. The method according to claim 2, characterized in that The hierarchical frequency model includes a low-frequency signal processing module, a high-frequency signal processing module, a feature fusion module, and a result prediction module; the prediction result corresponding to the multivariate time series data based on the trained hierarchical frequency model, the low-frequency signal, and the high-frequency signal includes: Inputting the low-frequency signal into the low-frequency signal processing module to obtain a low-frequency signal feature vector; Inputting the high-frequency signal into the high-frequency signal processing module to obtain a high-frequency signal feature vector; Determine a fused feature vector corresponding to the multivariate time series data based on the feature fusion module, the feature fusion ratio coefficient of each feature vector, the low-frequency signal feature vector, and the high-frequency signal feature vector; A prediction result corresponding to the multivariate time series data is determined based on the result prediction module and the fused feature vector.

4. The method according to claim 3, characterized in that After determining the prediction result corresponding to the multivariate time series data based on the result prediction module and the fused feature vector, the method includes: Obtaining a true result corresponding to the multivariate time series data with the same time length as the predicted result, and constructing a proportional coefficient optimization sample set based on the low-frequency signal and the high-frequency signal corresponding to the multivariate time series data and the true result; The hierarchical frequency model is optimized and trained based on the proportional coefficient optimization sample set, and the feature fusion proportional coefficients of each feature vector corresponding to the feature fusion module in the hierarchical frequency model are optimized and iterated, so as to realize the prediction of new multivariate time series data by using the hierarchical frequency model after the feature fusion proportional coefficient optimization iteration.

5. The method according to claim 3, characterized in that The low-frequency signal processing module includes a Bi-LSTM module and a CNN module, and the low-frequency signal feature vector includes a first low-frequency signal feature vector and a second low-frequency signal feature vector; Inputting the low-frequency signal into the low-frequency signal processing module to obtain a low-frequency signal feature vector includes: Using the Bi-LSTM module to capture the global temporal dependency of the low-frequency signal, and generate the first low-frequency signal feature vector; The CNN module is used to extract local spatial features of the low-frequency signal to generate the second low-frequency signal feature vector.

6. The method according to claim 3, characterized in that Inputting the high-frequency signal into the high-frequency signal processing module to obtain a high-frequency signal feature vector includes: The high-frequency signal is subjected to noise suppression and feature compression by the lightweight linear transformation layer of the high-frequency signal processing module to generate the high-frequency signal feature vector.

7. The method according to any one of claims 1 to 6, characterized in that The hierarchical frequency model is trained by the following steps: Acquire a multivariate time series training data set; wherein the multivariate time series training data set includes multiple sets of multivariate time series training data; Determining a training sample set based on the multivariate time series training data set; wherein the training sample set includes a plurality of training sample subsets, each training sample subset corresponds to different multivariate time series training data, and includes low-frequency signal training samples and high-frequency signal training samples corresponding to the multivariate time series training data; Acquiring a hierarchical frequency model to be trained; and training the hierarchical frequency model to be trained based on the training sample set, including: For each training sample subset, inputting the low-frequency signal training sample corresponding to the training sample subset into a low-frequency signal processing module to obtain a low-frequency signal training feature vector; and inputting the high-frequency signal training sample corresponding to the training sample subset into a high-frequency signal processing module to obtain a high-frequency signal training feature vector; and, based on the feature fusion module, the low-frequency signal training feature vector, and the high-frequency signal training feature vector, determining a fused training feature vector corresponding to the multivariate time series training data; and, based on the result prediction module and the fused training feature vector, determining a prediction result sample corresponding to the multivariate time series training data; Determining a target loss between the prediction result sample and the sample label corresponding to the multivariate time series training data based on a preset loss function, and adjusting the model parameters of the hierarchical frequency model to be trained based on the target loss; Repeat the above steps until the training results meet the preset requirements, and obtain the trained hierarchical frequency model; The step of determining the fused training feature vector corresponding to the multivariate time series training data based on the feature fusion module, the low-frequency signal training feature vector, and the high-frequency signal training feature vector includes: Determining a feature fusion ratio coefficient for each feature vector based on the multivariate time series training data and the learnable weight parameter; Based on the feature fusion module, the feature fusion ratio coefficient of each feature vector, the low-frequency signal training feature vector and the high-frequency signal training feature vector, a fusion training feature vector corresponding to the multivariate time series training data is determined.

8. A multivariate time series data prediction device based on a hierarchical frequency model, characterized in that: include: A cut-off point determination module is used to obtain multivariate time series data; and determining frequency band dynamic division parameters based on the multivariate time series data; and determining a low- and high-frequency signal demarcation point based on the frequency band dynamic division parameter and the multivariate time series data; A signal partitioning module, configured to determine a signal set corresponding to the multivariate time series data based on the multivariate time series data and the low- and high-frequency signal demarcation points; wherein the signal set includes a low-frequency signal and a high-frequency signal; A data prediction module is used to determine a prediction result corresponding to the multivariate time series data based on the trained hierarchical frequency model, the low-frequency signal and the high-frequency signal.

9. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

10. A computer device comprising a storage medium, a processor, and a computer program stored in the storage medium and executable on the processor, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

Images