Multivariable time sequence prediction method based on GRU and computer program product

Abstract

The invention discloses a multivariable time sequence prediction method based on GRU and a computer program product. The method comprises the following steps: firstly, introducing the dynamic characteristics of self-adaptive different time sequences of self-defined time sequence decomposition, and splitting an original sequence into trend, seasonal and residual components so as to reduce the data complexity and improve the interpretability; afterwards, a value embedding module is used for unifying feature representation so as to ensure that the model fully captures the time dependency relationship; in the modeling stage, the model adopts a multi-channel recurrent neural network to independently model the three types of components so as to reduce the interference between modes and improve the learning ability. In the prediction stage, a feature splicing strategy is adopted, information of each component is integrated, and richer time sequence representation is provided. In addition, a segmented prediction strategy is designed for the model, the prediction process is divided into multiple time periods, prediction information is combined, error accumulation is reduced, and the stability and robustness of long-sequence prediction are improved. Experimental results show that the prediction precision can be improved on different data sets.

Description

Technical Field

[0001] The present invention belongs to a prediction method, and particularly relates to a multivariate time series prediction method based on GRU and a computer program product. Background Art

[0002] Time series data is observational data recorded in chronological order and having time dependence, and is widely applied in various fields. Time series prediction is an important research direction in the field of machine learning, and has wide applications in multiple practical scenarios such as energy consumption prediction, weather forecasting, traffic flow prediction, financial market analysis, and industrial control.

[0003] Traditional time series prediction methods often rely on linear modeling of historical data. However, many time series data in the real world have complex non-linear relationships. To address this challenge, many studies have begun to attempt to combine deep learning methods, especially architectures such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and gated recurrent units (GRUs), which improve prediction accuracy by capturing the time dependence relationships in time series. Although these methods have improved the prediction performance to a certain extent, they still face some challenges, including how to handle the complex change patterns of different components (such as trends, seasonality, and residuals) in time series. In addition, in recent years, models based on the attention mechanism have been widely applied in sequence data modeling. The self-attention mechanism can capture long-term dependence relationships and model the correlations between different time steps, especially when dealing with long sequences, showing better performance than traditional RNNs and LSTMs. Time series prediction models based on Transformer, leveraging their efficient computing power and ability to capture long-term dependence relationships, have become a research hotspot in the current field of time series prediction.

[0004] A multi-dimensional time series prediction method and system based on learning dynamic dependence relationships between variables are disclosed in the patent (CN118551189A), mainly including: dividing the original data into a training sample set and a test sample set; performing feature extraction and fusion on the time series data of the obtained sample set; extracting key information and global information from the fused information and adaptively adjusting weights; performing information transmission through node embedding of a dynamic graph structure and node feature update of a dynamic graph convolutional layer; and evaluating the effectiveness of the model. This invention utilizes the intrinsic attributes that vary with time between variables. At the same time, multi-level hidden features are effectively utilized, improving the correlation analysis ability, and can be used in fields such as macroeconomic analysis, environmental monitoring, and epidemiological analysis.

[0005] A long - term time - series prediction method based on normalizing flow is disclosed in the patent (CN118536054A), which includes the following steps: The data is a large amount of long - term time - series data of multiple types and inconsistent units, and data normalization is performed on it; Multivariate and multi - scale information of the long - term time - series data is extracted; Based on the multivariate and multi - scale information, long - term trend information, seasonal trend information, and dispersion information are respectively extracted; Different feature data are fused; Based on the fused features, data prediction based on normalizing flow is performed. This invention can improve the prediction accuracy of long - term time - series data, and is especially suitable for long - term meteorological prediction.

[0006] A time - series prediction method, device, and medium based on an optimized LSTM model are disclosed in the patent (CN117272051A), belonging to the field of data - processing systems or methods specifically applicable to prediction purposes. It includes the following steps: Initialize the hyperparameters of the LSTM model and pre - process the data samples to obtain training samples and test samples; Optimize the hyperparameters of the LSTM model through a multi - layer LSTM model based on the training samples, and calculate the initial fitness of the multi - layer LSTM model based on the test samples; When it is determined that the multi - layer LSTM model has not reached the preset iteration number threshold based on the initial fitness, perform population iteration on the hyperparameters of the LSTM model based on the population algorithm and determine the average score of the hyperparameters of the LSTM model after population iteration and parameter optimization; Calculate the current fitness of the multi - layer LSTM model based on the average score and the training samples, and obtain the global optimal parameters when the current fitness is greater than the current optimal solution; Obtain the optimized LSTM model based on the global optimal parameters and predict the time - series through the optimized LSTM model.

[0007] In summary, existing deep - learning - based time - series prediction methods have achieved excellent performance in many applications, but they still face some challenges. For example, how to efficiently capture various complex change patterns in time - series, how to maintain computational efficiency in long - sequence prediction, how to solve the problems of model generalization ability and interpretability, etc. are still key issues in current time - series prediction research. Therefore, there is an urgent need for a prediction method that can be faster and have a higher prediction accuracy to achieve better prediction effects. Summary of the Invention

[0008] The purpose of the present invention is to provide a multivariate time - series prediction method based on GRU, which can effectively improve the accuracy of the prediction effect.

[0009] To achieve the above - mentioned purpose, the present invention provides the following solutions:

[0010] A multivariate time - series prediction method includes:

[0011] Obtain a multivariate data - sequence data set and perform pre - processing to meet the model input requirements;

[0012] Custom time series decomposition decomposes the original time series into trend, seasonal, and residual components to reduce data complexity and improve interpretability;

[0013] Multi-channel recurrent neural network modeling uses a multi-channel recurrent neural network to independently model the trend, seasonal, and residual components respectively to reduce interference between different patterns and improve the ability to learn time series features;

[0014] Feature fusion and prediction concatenate the modeling results of different components in the feature dimension and input them into the prediction layer;

[0015] Segmented prediction divides the prediction task into multiple consecutive time periods during the prediction stage to reduce error accumulation and improve the prediction efficiency of long time series;

[0016] Adding non-linear factors adds activation functions during the prediction stage, enabling the model to learn complex time series patterns.

[0017] Optionally, obtaining the multi-variable data sequence dataset and performing preprocessing to meet the model input requirements specifically includes:

[0018] Normalizing, detrending, and segmenting the multi-variable time series dataset.

[0019] Optionally, the custom time series decomposition decomposes the original time series into trend, seasonal, and residual components to reduce data complexity and improve interpretability, specifically including:

[0020] Adopt the moving average method of a sliding window, set a fixed window size W for local mean calculation to dynamically adjust the trend, reduce errors and avoid information loss. The seasonal component describes periodic fluctuations. After extracting the trend, the time series is detrended, and a Butterworth low-pass filter is used to extract the seasonal component. This method can effectively suppress high-frequency noise, retain the main periodic features, and avoid overfitting. Finally, by removing the trend and seasonal components, the obtained residual component mainly reflects short-term perturbations, providing accurate input for subsequent modeling.

[0021] Optionally, the multi-channel recurrent neural network modeling uses a multi-channel recurrent neural network to independently model the trend, seasonal, and residual components respectively to reduce interference between different patterns and improve the ability to learn time series features, specifically including:

[0022] For the three independent components of trend, seasonality, and residuals, independent GRU networks are constructed respectively. Each GRU network independently processes its own time series features to avoid information mixing between different components. During the update process of the hidden state of the GRU, a channel attention mechanism is introduced to calculate the importance weights of different time series components, and the contributions of each channel are adaptively adjusted according to these weights to enhance the expression ability of key time features. The multi-head self-attention mechanism is applied to the hidden state sequence processed by the GRU to capture the global time dependencies between different time steps, thereby enhancing the model's ability to model long-term dependence information, and extracting multi-scale time features through multiple attention heads to further improve the modeling ability for different channel information and long-term time dependencies.

[0023] Optionally, for the feature fusion and prediction, the modeling results of different components are concatenated in the feature dimension and input into the prediction layer, specifically including:

[0024] The final hidden states are extracted from the multi-channel recurrent neural networks of the trend, seasonality, and residuals channels respectively. Each hidden state represents the feature extraction result of the channel for the time series data and contains short-term and long-term time dependencies. The hidden states of the trend, seasonality, and residuals channels are concatenated according to the feature dimension to form a comprehensive vector. This vector contains the feature information of different time series components and enhances the information sharing and interaction between components through the concatenation operation. Thus, a comprehensive vector is concatenated, and this comprehensive vector will be used as the input for each prediction segment.

[0025] Optionally, for the segmented prediction, in the prediction stage, the prediction task is divided into multiple consecutive time periods to reduce error accumulation and improve the prediction efficiency of long time series, specifically including:

[0026] First, the concatenated hidden state is input into the fully connected layer. According to the target prediction length and the dynamic characteristics of the historical data, the overall prediction task is divided into multiple consecutive small time periods, and each time period is predicted independently. Thus, prediction results for multiple time periods are generated. When the prediction results of all small time periods are completed through independent modeling, the outputs of each prediction segment are concatenated in chronological order. This concatenation method ensures that the prediction results of each time period can be connected in an orderly manner into a complete time series prediction and effectively reduces the impact of the error within each segment on the global prediction.

[0027] Optionally, for adding non-linear factors, an activation function is added in the prediction stage, so that the model can learn complex time series patterns, specifically including:

[0028] During the prediction stage of the model, an activation function is added to introduce non-linear transformation, enhancing the model's fitting ability for complex time-series data. By introducing appropriate activation functions between the prediction layers of the network, the model can not only capture the linear relationships in the time series but also learn more complex and variable non-linear time-series patterns, improving the model's expressive ability for data.

[0029] A computer device, comprising: a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement a multivariate time series prediction method according to any one of claims 1-7.

[0030] According to the specific embodiments provided by the present invention, the following technical effects are disclosed:

[0031] The present invention provides a multivariate time series prediction method and a computer program product based on GRU. Multivariate data sequence datasets are obtained and preprocessed to meet the model input requirements; custom time series decomposition is performed to decompose the original time series into trend, seasonal, and residual components to reduce data complexity and improve interpretability; multi-channel recurrent neural network modeling is carried out, using a multi-channel recurrent neural network to independently model the trend, seasonal, and residual components respectively to reduce interference between different patterns and improve the time series feature learning ability; feature fusion and prediction are performed by concatenating the modeling results of different components in the feature dimension and inputting them into the prediction layer; segmented prediction is carried out, during the prediction stage, dividing the prediction task into multiple consecutive time periods to reduce error accumulation and improve the long-time series prediction efficiency; adding non-linear factors, adding an activation function during the prediction stage, so that the model can learn complex time series patterns. A hybrid model combining a custom decomposition method, a multi-channel recurrent neural network modeling mechanism, and a segmented prediction strategy is used for multivariate time series prediction. First, the original time series is split into trend, seasonal, and residual components through custom decomposition, and specialized modeling strategies are adopted for different components. A model based on a multi-channel recurrent neural network is used to independently model each component, effectively reducing interference between components and enhancing the time series feature learning ability. Secondly, a segmented prediction strategy is introduced, dividing the prediction task into multiple consecutive time periods to reduce error accumulation and improve the prediction stability and accuracy of long-time series. Through custom decomposition and multi-module collaborative processing, this model can effectively capture complex dynamic patterns in the time series, significantly improving the prediction accuracy and robustness, and providing more accurate and stable prediction results in variable time series data.

[0032] The present invention also proposes a computer program product with all the advantages of the above prediction method. Description of the Drawings

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required in the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0034] Figure 1 It is a flowchart of the multi-variable time series prediction method based on GRU provided by this application;

[0035] Figure 2 It is a data decomposition flowchart in the multi-variable time series prediction method based on GRU provided by this application;

[0036] Figure 3 It is a specific flowchart of the multi-variable time series prediction method based on GRU provided by this application. Detailed implementation manners

[0037] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0038] The present invention proposes a multi-variable time series prediction method and a computer program product based on GRU. First, obtain a multi-variable data sequence data set and perform preprocessing to meet the model input requirements; secondly, customize time series decomposition, decompose the original time series, and split it into trend, seasonal, and residual components to reduce data complexity and improve interpretability; then, perform multi-channel recurrent neural network modeling, use a multi-channel recurrent neural network to independently model the trend, seasonal, and residual components respectively to reduce interference between different patterns and improve the ability to learn time series features; after that, perform feature fusion and prediction, splice the modeling results of different components in the feature dimension and input them into the prediction layer; thus, perform segmented prediction. In the prediction stage, divide the prediction task into multiple consecutive time periods to reduce error accumulation and improve the long-time series prediction efficiency; finally, add non-linear factors, add an activation function in the prediction stage, so that the model can learn complex time series patterns. The present invention realizes multi-variable time series prediction and solves the problems of low prediction efficiency and inaccurate prediction accuracy of existing models.

[0039] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific implementation manners.

[0040] In an exemplary embodiment, the present application provides a multivariate time series prediction method based on GRU. As Figure 1 、 Figure 2 and Figure 3 shown, the multivariate time series prediction method based on GRU provided by the embodiments of the present invention includes the following steps 1 to 6.

[0041] Step 1: Obtain a multivariate data sequence dataset and perform preprocessing to meet the model input requirements.

[0042] Obtain a multivariate time series dataset and perform preprocessing on the time series dataset. First, an optional RevIN normalization mechanism is introduced in the input stage, and the steps are as follows:

[0043] If RevIN is enabled, the model first performs instance-level normalization on each time series variable to generate standardized data x norm :

[0044]

[0045] where is the mean calculated by channel, and is the standard deviation calculated by channel (to prevent division by zero, ε is a very small number).

[0046] If RevIN is not enabled, the model subtracts the value of the last time step of each time series, X -1 , from the input data to remove the static offset term and generate a new input sequence:

[0047] X' = X - X -1 (2)

[0048] The dimension of the input sequence is adjusted to X' = ∈R B×C×T , where: B represents the batch size, C represents the number of variables in the time series (multivariate time series scenario), and T represents the time step of the input sequence. Through the above preprocessing, it is made compatible with the subsequent decomposition and modeling modules.

[0049] Step 2: Customize time series decomposition, decompose the original time series, and split it into trend, seasonal, and residual components to reduce data complexity and improve interpretability.

[0050] The long-term trend reflects the continuous upward or downward change of the time series on a macroscopic scale and is usually used to depict the long-term evolution pattern of the data. To extract this trend component, a moving average method based on a sliding window is adopted in this paper. Specifically, a fixed window size W is set, and the time series is convolved through the sliding window to calculate the local mean corresponding to each time step t, thereby obtaining the smoothed trend component.

[0051]

[0052] Among them, T t is the trend part corresponding to the time step t, and W is the size of the sliding window. This method can effectively weaken the influence of short-term fluctuations, highlight the overall change trend of the time series, and at the same time suppress periodic and random perturbations, making the subsequent modeling more stable and robust.

[0053] The seasonal component describes the periodic fluctuation pattern presented in the time series, usually reflecting the repeated change trend of the data within a specific period, such as the annual periodic change of temperature or the monthly periodic fluctuation of sales volume. After extracting the trend component, a detrended residual sequence is obtained by removing the trend component, and a Butterworth Low-pass Filter is further used to extract the seasonal component.

[0054] Specifically, this paper designs an adaptive filtering scheme, in which the cut-off frequency of the filter is dynamically adjusted according to the characteristics of the input sequence to ensure the effective separation of periodic components and aperiodic high-frequency noise. Through this method, the periodic pattern in the time series can be effectively retained, while suppressing aperiodic interference and high-frequency noise, improving the stability and accuracy of the decomposition.

[0055] S t = Butterworth_Filter(x t - T t ) (4)

[0056] Among them, S t is the seasonal component corresponding to the time step t, x t is the original sequence, and T t is the trend component extracted from the original sequence.

[0057] By subtracting the trend and seasonal components from the original sequence, the residual component is obtained:

[0058] R = X' - T - S (5)

[0059] After passing through this decomposition module, the time series is transformed into trend T, season S, and residual R, which are respectively input into the subsequent feature extraction module.

[0060] Step 3: Multi-channel recurrent neural network modeling. A multi-channel recurrent neural network is used to independently model the trend, seasonal, and residual components respectively to reduce the interference between different patterns and improve the time series feature learning ability.

[0061] First, extract the temporal dependencies in the time series to generate the hidden state at each time step. Assume that at time step t, for the t-th channel, the output of the GRU is denoted as This hidden state can be calculated using the standard formula of the GRU:

[0062]

[0063] Then, calculate the channel attention weights at each time step These weights represent the importance of each channel at a given time step. Assuming there are C channels, calculate the attention weight for each channel to adjust the importance of each channel at each time step. Specifically, given the GRU output at each time step use a neural network to calculate the attention weights:

[0064]

[0065] where f(·) is an activation function, W α and b α are the trained parameters. Through this process, the model can dynamically adjust the importance of each channel at each time step and weight the contributions of different channels according to the task requirements.

[0066] Finally, perform channel attention weighting on the hidden state at each time step. Assume the hidden state of each channel is (where c represents the channel index), and weight the importance of each channel through the channel attention weights.

[0067]

[0068] where is the channel attention weight, representing the importance of channel c at time step t, and C is the number of channels.

[0069] Use linear transformation on the combined hidden state h t to calculate the query (Q), key (K), and value (V) matrices.

[0070] Q t = W Q h t , K t = W K h t , V t = W V h t (11)

[0071] where W Q , W K , WV is a linear transformation matrix representing the transformations of queries, keys, and values.

[0072] The attention weights at each time step t are calculated using the self-attention mechanism. Specifically, the dot product of the query Q t and the key K t is computed and normalized, and then the result is applied to the value V t .

[0073]

[0074] where d k is the dimension of the key.

[0075] For each attention head i, an independent attention output is calculated separately, and then these outputs are concatenated and passed through a linear transformation to obtain the final multi-head attention output.

[0076] MultiHeadAttention(h t ) = Concat(head1, head2, ···, head H )W O (13)

[0077] where head i = Attention(Q i , K i , V i ), and each attention head i has independent queries, keys, and values. The outputs of all heads are concatenated and passed through the linear transformation W O to obtain the final multi-head attention output.

[0078] Step 4: Feature fusion and prediction. The modeling results of different components are concatenated in the feature dimension and input into the prediction layer.

[0079] Predictions are made using the hidden states obtained from each component (trend, seasonality, residual). The hidden states output by the module are concatenated into a comprehensive vector, which will serve as the input for each prediction segment. The hidden states of the three components are concatenated as the comprehensive representation:

[0080] H combined = Concat(H trend , H seasonal , H residual ) (14)

[0081] Step 5: Segment prediction. In the prediction stage, the prediction task is divided into multiple consecutive time periods to reduce error accumulation and improve the efficiency of long time series prediction.

[0082] The concatenated hidden state is input into a fully connected layer to generate the prediction results for each segment. At this stage, the model generates an output for a predicted segment based on the hidden state:

[0083] Y segment = Linear(H combined )(15)

[0084] This output Y segment represents the predicted value for the next time period. At this point, the length of the predicted value should be L, that is, the time length of each prediction.

[0085] Step 6: Incorporate non - linear factors. Add an activation function during the prediction stage so that the model can learn complex temporal patterns.

[0086] Add an activation function after the hidden state output of each predicted segment. The role of the activation function is to introduce non - linear characteristics, enabling the model to learn complex temporal patterns.

[0087] In the calculation of the output of each segment, add an activation function (GELU). This activation function helps the model perform non - linear mapping before making predictions, thus better fitting the non - linear features in the data.

[0088] The application method of the activation function is as follows:

[0089] H combined～activated = Activation Funcation(H combined )(16)

[0090] This activated hidden state vector then passes through a fully connected layer to generate the predicted value:

[0091] Y segment～activated = Linear(H combined～activated )(17)

[0092] The addition of the activation function effectively helps the model capture complex temporal patterns, improving the model's expressive ability and prediction accuracy. Finally, the output of each segment is processed by the activation function and finally concatenated into a complete prediction result.

[0093] The time - series dataset is predicted using the present invention and other methods respectively. For the comparison of the prediction accuracy results, see Table 1.

[0094] Table 1 Comparison table of prediction accuracy of the present invention and other methods on the time - series dataset

[0095]

[0096] After processing the time series dataset and inputting it into the constructed model, the present invention has obtained the highest prediction accuracy compared to other models. The present invention proposes the following three innovations in time series prediction: 1. A decomposition method combining a low-pass filter and moving average is proposed to decompose the time series into three parts: trend, seasonality, and residual, thereby reducing the mutual interference between components and facilitating independent modeling. 2. To capture the temporal dependencies and interaction information of different components, a recurrent neural network combining self-attention mechanism and multi-head attention mechanism is designed. The self-attention mechanism is introduced in time series modeling to enhance the long-term and short-term dependence modeling ability between different time steps, and the dynamic interaction relationships of different components such as trend, seasonality, and residual are captured through the multi-head attention mechanism, thereby optimizing the overall prediction performance. 3. To solve the prediction problem of long time spans, the model adopts a segmented prediction strategy, decomposing the prediction task of a long time series into multiple sub-tasks of short time periods, thereby improving the prediction accuracy.

[0097] Compared with the prior art, the present invention has the following advantages: 1. Higher prediction accuracy. The present invention combines custom time series decomposition, multi-channel recurrent neural network modeling, and segmented prediction strategy, enabling the model to more accurately capture the trend, seasonality, and residual components in the time series, thereby effectively improving the prediction accuracy, especially performing excellently in long time span prediction tasks. 2. Stronger generalization ability: Compared with traditional fixed decomposition methods, the custom decomposition method adopted by the present invention can adaptively adjust the decomposition strategy according to the dynamic characteristics of different datasets, enabling the model to maintain good prediction performance on different types of time series data, improving the adaptability and generalization ability of the model. 3. Effectively reduce interference between components: By separately modeling the trend, seasonality, and residual components through a multi-channel recurrent neural network and combining self-attention and multi-head attention mechanisms, the model can effectively reduce the mutual influence between different time series components, ensuring that each component can independently learn its temporal characteristics, thereby improving the modeling effect. 4. Stronger long-term and short-term dependence relationship modeling ability: Based on traditional GRU / LSTM, the present invention introduces self-attention and multi-head attention mechanisms, enabling the model to better learn the long-term and short-term dependence relationships between different time steps, optimize information transmission, and enhance the learning ability of complex time patterns, especially showing significant advantages in long time series prediction tasks. 5. Alleviate the problem of error accumulation: The present invention adopts a segmented prediction strategy, dividing the prediction task of a long time span into multiple short time periods, and using the information of the previous prediction segment to guide the generation of the subsequent prediction segment, thereby effectively reducing the problem of gradual accumulation of errors and improving the stability and consistency of prediction. 6. Higher computational efficiency: Compared with traditional end-to-end long sequence prediction methods, the segmented prediction strategy of the present invention reduces the complexity of a single prediction task and reduces computational redundancy through independent modeling after decomposition, enabling the model to improve computational efficiency and reduce inference time while ensuring prediction accuracy.

[0098] In an exemplary embodiment, the present invention provides a computer device, which includes a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the relevant steps in the above method embodiments can be implemented.

[0099] In addition, in another exemplary embodiment, the present invention further provides a computer-readable storage medium, in which a computer program is stored. When the processor executes the computer program, the corresponding operations in the above method embodiments can be completed.

[0100] Furthermore, the present invention also provides a computer program product, which includes a computer program. When the processor executes the computer program, the specific steps in the above method embodiments can be implemented.

[0101] It should be emphasized that the user information (such as user device information, user personal information, etc.) and data (including but not limited to data for analysis, storage, and display) involved in the present invention can only be used after being authorized by the user or fully authorized by the relevant party, and the collection, use, and processing of all data must comply with the requirements of applicable laws and regulations.

[0102] Those skilled in the art can understand that to implement part or all of the processes in the above method embodiments, it can be instructed by a computer program to complete the relevant hardware. The computer program can be stored in a non-volatile computer-readable storage medium and implement the functions of the above method embodiments when executed. It should be noted that the memory, database, or other storage media involved in the present invention can include two types: non-volatile memory and volatile memory. Examples of non-volatile memory include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random-access memory (ReRAM), magnetoresistive random-access memory (MRAM), ferroelectric random-access memory (FRAM), phase change memory (PCM), and graphene memory, etc.; while volatile memory includes random access memory (RAM) and external cache memory, such as static random access memory (SRAM) and dynamic random access memory (DRAM).

[0103] It is worth noting that the technical features of the above embodiments can be combined in different ways. Although this specification does not describe all possible combinations of technical features one by one, as long as these combinations do not have logical contradictions, they should all be regarded as the scope of application of the present invention.

[0104] This text elaborates on the basic principles and implementation modes of the present invention through specific examples. These examples are only used to assist in understanding the core idea of the present invention and do not limit the present invention. Those skilled in the art can make appropriate adjustments and optimizations based on the present invention to make it applicable to different application scenarios. Therefore, the content of this specification should not be regarded as a limitation on the scope of the present invention, but rather the scope of the appended claims shall prevail.

Claims

1. A multi-variable time series prediction method, characterized in that, Including: Obtain a multi-variable data sequence dataset and perform preprocessing to meet the model input requirements; Customize time series decomposition, decompose the original time series, and split it into trend, seasonal, and residual components to reduce data complexity and improve interpretability; Multi-channel recurrent neural network modeling, adopt a multi-channel recurrent neural network, and independently model the trend, seasonal, and residual components respectively to reduce interference between different patterns and improve the time series feature learning ability; Feature fusion and prediction, splice the modeling results of different components in the feature dimension and input them into the prediction layer; Segmented prediction, in the prediction stage, divide the prediction task into multiple consecutive time periods to reduce error accumulation and improve the long-term time series prediction efficiency; Add non-linear factors, add an activation function in the prediction stage, so that the model can learn complex time series patterns.

2. The multivariate time series prediction method according to claim 1, characterized in that Obtain a multi-variable data sequence dataset and perform preprocessing to meet the model input requirements, specifically including: Perform normalization processing, de-offset processing, and sequence segmentation processing on the multi-variable time series dataset.

3. A multivariable time series prediction method according to claim 1, wherein Customize time series decomposition, decompose the original time series, and split it into trend, seasonal, and residual components to reduce data complexity and improve interpretability, specifically including: Adopt the moving average method of a sliding window, set a fixed window size for local mean calculation to dynamically adjust the trend, reduce errors and avoid information loss. The seasonal component describes periodic fluctuations. After extracting the trend, the time series is detrended, and a Butterworth low-pass filter is used to extract the seasonal component. This method can effectively suppress high-frequency noise, retain the main periodic features, and avoid overfitting. Finally, by removing the trend and seasonal components, the obtained residual component mainly reflects short-term perturbations, providing accurate input for subsequent modeling.

4. A multivariable time series prediction method according to claim 1, characterized in that Multi-channel recurrent neural network modeling, adopt a multi-channel recurrent neural network, and independently model the trend, seasonal, and residual components respectively to reduce interference between different patterns and improve the time series feature learning ability, specifically including: For the three independent components of trend, seasonality, and residuals, construct independent GRU networks respectively. Each GRU network independently processes its own time series features to avoid information mixing between different components. In the process of updating the hidden state of the GRU, introduce a channel attention mechanism, calculate the importance weights of different time series components, and adaptively adjust the contributions of each channel according to these weights to enhance the expression ability of key time features. Apply a multi-head self-attention mechanism to the hidden state sequence processed by the GRU to capture the global time dependencies between different time steps, thereby enhancing the model's ability to model long-term dependency information, and extracting multi-scale time features through multiple attention heads to further improve the modeling ability of different channel information and long-term time dependencies.

5. A multivariable time series prediction method according to claim 1, characterized in that Feature fusion and prediction, splice the modeling results of different components in the feature dimension and input them into the prediction layer, specifically including: Extract the final hidden states from the multi-channel recurrent neural network of the trend, seasonal, and residual channels respectively. Each hidden state represents the feature extraction result of the channel for the time series data and contains short-term and long-term time dependencies. Concatenate the hidden states of the trend, seasonal, and residual channels according to the feature dimension to form a comprehensive vector. This vector contains the feature information of different time series components and enhances the information sharing and interaction among components through the concatenation operation. Thus, a comprehensive vector is concatenated, and this comprehensive vector will be used as the input for each prediction segment.

6. A multivariable time series prediction method according to claim 1, wherein Segmented prediction. In the prediction stage, divide the prediction task into multiple consecutive time periods to reduce error accumulation and improve the prediction efficiency of long time series, specifically including: First, input the concatenated hidden states into the fully connected layer. According to the target prediction length and the dynamic characteristics of the historical data, divide the overall prediction task into multiple consecutive small time periods, and each time period is predicted independently. Thus, prediction results for multiple time periods are generated. When the prediction results for all small time periods are completed through independent modeling, concatenate the outputs of each prediction segment in chronological order. This concatenation method ensures that the prediction results for each time period can be connected in an orderly manner into a complete time series prediction and effectively reduces the impact of the error within each segment on the global prediction.

7. A multivariate time series prediction method according to claim 1, characterized in that, Add non-linear factors. Add activation functions in the prediction stage so that the model can learn complex time series patterns, specifically including: In the prediction stage of the model, add activation functions to introduce non-linear transformations and enhance the model's fitting ability for complex time series data. By introducing appropriate activation functions between the prediction layers of the network, the model can learn more complex and variable non-linear time series patterns in addition to capturing the linear relationships in the time series, improving the model's ability to represent data.

8. A computer device, comprising: A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement a multi-variable time series prediction method according to any one of claims 1-7.

Images