Time sequence prediction method based on multi-dimensional feature fusion and computer program product

Abstract

The invention discloses a time sequence prediction method based on multi-dimensional feature fusion and a computer program product. The prediction method is constructed based on a lightweight architecture, and complex module stacking is avoided. The method comprises the following steps: firstly, designing a multi-dimensional convolution extraction module, respectively capturing short-term fluctuation, long-term trend and timestamp association by utilizing three-dimensional convolution operations, and realizing comprehensive coverage of local details, global trend and time dynamic; secondly, a mask mechanism is designed, and masks are used on the dimension based on periodic modeling so as to improve the prediction effect and the model stability. And finally, the features of the three dimensions are sent to a prediction module and weighted fusion is carried out to obtain a final result. Therefore, the features of different time dimensions can be effectively captured, the prediction efficiency is improved, and the computing power demand is reduced. In the face of different service data in the time sequence field, the model can actively capture the core law and potential association of different time sequence data without complex adjustment, so that accurate prediction is realized.

Description

Technical Field

[0001] This invention pertains to a prediction method, specifically a time series prediction method and computer program product based on multi-dimensional feature fusion. Background Technology

[0002] Time series forecasting plays a crucial role in numerous fields, such as traffic flow management, product sales strategy adjustments, and energy consumption planning. The high-precision forecasts it generates can help decision-makers plan ahead, enhancing the scientific rigor and foresight of their decisions. To achieve this, in the past, powerful and complex deep learning models, such as recurrent neural networks, convolutional neural networks, and attention-based Transformer models, have been commonly relied upon. In recent years, with the increasing forecasting requirements across industries, long-term series forecasting has gradually become a research hotspot. This field demands models capable of generating forecasts over longer time spans to meet the needs of advanced planning and strategic decision-making. For example, in the energy sector, long-term energy consumption forecasts can help relevant departments formulate power generation plans and energy allocation schemes in advance; in the transportation sector, long-term traffic flow forecasts help plan road construction and optimize traffic management measures; and in the business sector, long-term sales forecasts provide important information for companies' production, inventory, and marketing strategies.

[0003] In time series forecasting, while a longer forecast period provides decision-makers with more comprehensive forward-looking information, it also significantly increases the uncertainty of the forecast results. To generate reliable long-term forecasts, models must mine and extract complex and extensive time dependencies from longer historical data. This undoubtedly increases the difficulty and complexity of modeling, placing higher demands on model architecture design and algorithm optimization to accurately capture long-term time dependencies. Against this backdrop, attention-based models like the Transformer, while performing excellently, face numerous limitations in practical applications due to their massive parameter count, typically reaching millions or even tens of millions.

[0004] Furthermore, while existing feature extraction models have made progress, they still have significant limitations. Taking one-dimensional models as an example, LSTM, with its gating mechanism, attempts to capture short-term and long-term dependencies, but is limited by the temporal order, making it inadequate in distinguishing patterns at different scales. This highlights the inherent flaws of one-dimensional models: they focus only on linear or local dependencies in a single time dimension, failing to analyze periodic patterns in the frequency domain or capture collaborative relationships in multi-scale dynamic changes. While two-dimensional models have made breakthroughs in dimensional expansion, they still haven't achieved comprehensive coverage. TimesNet, using a time-frequency two-dimensional decomposition, can perform time-dynamic analysis of univariate periods, accurately capturing temporal changes in periodic intensity, but its application is limited to univariate data and cannot handle complex relationships in multivariate data. This directly leads to the failure of periodic pattern modeling in multivariate coupled scenarios, exposing the inadequacy of two-dimensional models in multivariate and multi-dimensional analysis. These limitations indicate that current feature extraction models face pressing problems in handling complex time series data, both in terms of dimensional coverage and the completeness of feature capture. Summary of the Invention

[0005] This invention addresses the problems of high computational complexity of Transformer-type models and the tendency of traditional denoising and anomaly removal methods to disrupt temporal correlations. It provides a time series prediction method based on multi-dimensional feature fusion, which achieves a precise balance between computational cost and model performance while ensuring the performance of time series prediction.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] In a first aspect, the present invention proposes a time series forecasting method, comprising:

[0008] The target to be predicted is input into the time series prediction model to obtain the prediction result;

[0009] The training method for the time series prediction model includes:

[0010] Establish a time series prediction dataset; the time series prediction dataset includes historical data of the target to be predicted and has several attributes;

[0011] The time series prediction model is trained using the time series prediction dataset; the time series prediction model includes a masking module, a multi-dimensional feature extraction module, and a fusion prediction module.

[0012] The masking module is used to perform periodic-level masking on the dimension based on periodic modeling to obtain masked time series data;

[0013] The multi-dimensional feature extraction module is used to extract multi-dimensional features from the time series prediction dataset to obtain data features under different dimensions.

[0014] The fusion prediction module is used to predict multi-dimensional features separately and introduces a learnable dynamic weight mechanism to weight and fuse the results of each dimension to obtain the final prediction result.

[0015] Furthermore, after establishing the time series prediction dataset, the method further includes:

[0016] The time series prediction dataset is filled with imputation by replacing missing values ​​with the mean of the two values ​​before and after the missing value.

[0017] The time series prediction dataset after completion is standardized so that the mean is 0 and the variance is 1.

[0018] Furthermore, the time series prediction dataset meets the following requirements:

[0019] The time difference between adjacent data in the time series prediction dataset is constant.

[0020] The feature type and number of features are consistent at each time step in the time series prediction dataset;

[0021] The time series prediction dataset includes the date and timestamp of each data collection, and the date and timestamp format of each data is consistent.

[0022] Furthermore, the specific process of the mask module includes:

[0023] In the dimension of periodic modeling, the data from the past L time steps are first divided into fixed period lengths. Then, the first 70% of the periods are processed, and 30% of the complete periods are randomly selected for masking. After obtaining the mask matrix, multiplying it with the original sequence yields the masked sequence. The masking formula for the i-th feature is as follows:

[0024]

[0025] in, This represents the observations over the past L time steps, and mask(·) indicates that a mask is used for the historical sequence.

[0026] Furthermore, the specific process of the multi-dimensional feature extraction module includes:

[0027] (1) Design a periodic-aware one-dimensional convolutional module

[0028] The convolutional kernel is designed according to the specified period length, while padding is used to maintain the sequence length and avoid information loss. Then, the designed one-dimensional convolutional kernel is used to perform convolution operations on the temporal data to extract one-dimensional features. Finally, a residual connection mechanism is used to add the extracted features to the original temporal data to alleviate the gradient decay problem of deep networks. The formula for the i-th feature is as follows:

[0029]

[0030] in, This represents the observations over the past L time steps, and 1DConv(·) indicates that a one-dimensional convolution kernel is applied for feature extraction. This represents the final one-dimensional feature obtained.

[0031] (2) Design two-dimensional convolution to capture the correlation between different periods.

[0032] First, the time-series data is transformed into two-dimensional data according to a specified period. Then, intra-period convolutions are designed to capture local temporal patterns, and inter-period convolutions are designed to capture dependencies between different periods. Finally, a residual connection mechanism is used to alleviate gradient vanishing and improve prediction performance. The formula for the i-th feature is as follows:

[0033]

[0034] Where 2DConv(·) represents the application of a two-dimensional convolution kernel for feature extraction. This represents the final two-dimensional feature.

[0035] (3) Design a three-dimensional convolution to directly learn the three-level time structure of "long-term-medium-short-term".

[0036] First, the original time series is reshaped according to different sampling times in the dataset to meet the requirements of 3D convolution. Then, a depthwise separable 3D convolution is designed, first mapping the input channels to an intermediate dimension, and then mapping the features to the output channels using the same structure, while expanding the receptive field to capture the correlation between discontinuous time points. Finally, residual connections are used to alleviate the gradient vanishing problem. The formula for realizing the i-th feature is as follows:

[0037]

[0038] Where 3DConv(·) represents the application of a three-dimensional convolution kernel for feature extraction. This represents the final three-dimensional feature obtained.

[0039] Furthermore, the specific process of the fusion prediction module includes:

[0040] Design an adaptive multi-dimensional fusion network for time series prediction, capturing temporal features at different granularities through three branches. Each branch is followed by a multilayer perceptron for feature transformation. This structure maps the features of each time window to a higher-dimensional space for feature combination, and then projects them onto the prediction length. Two learnable dynamic weight parameters are introduced, enabling the model to automatically allocate different scales based on data characteristics. The calculation of the i-th feature is as follows:

[0041]

[0042] Where α and β represent learnable parameters, and MLP(·) represents the transformation of features across each dimension. This represents the value predicted for the next H time steps.

[0043] Furthermore, the loss function used when training the time series prediction model includes:

[0044] The optimization objective of the model is the mean squared error (MSE), the most commonly used metric in time series forecasting. This index quantifies the H-step predicted sequence given by the model using the form of squared differences. Compared with actual observation The overall deviation between them is specifically implemented as follows:

[0045]

[0046] Where L is the loss value obtained from the MSE loss function.

[0047] Secondly, the present invention provides a computer program product, which includes a computer program or instructions that, when executed by a communication device, cause the above-described method to be performed.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] This invention proposes a time series forecasting method. First, a masking strategy is implemented on the model data based on periodic modeling, randomly masking 30% of the first 70% of the time series data. This forces the model to learn anti-interference capabilities without losing core periodic information, thus improving model robustness. Next, a multi-dimensional feature extraction module is constructed. The data is first reshaped in different dimensions, and then features are extracted from the processed data in different dimensions. Compared to existing single-dimensional feature extraction, this method covers more comprehensive time series information. Finally, the obtained three-dimensional features are input into a multilayer perceptron for prediction. The prediction results are de-standardized and then weighted and fused to highlight the differences in importance of features across different dimensions. This method outperforms existing simple averaging fusion, and the prediction results better reflect the core patterns of the time series data.

[0050] The present invention also proposes a computer program product that possesses all the advantages of the above-mentioned prediction methods. Attached Figure Description

[0051] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the drawings described below are only some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart illustrating the time series prediction model training method of the present invention.

[0053] Figure 2 This is a flowchart illustrating the time series prediction model of the present invention.

[0054] Figure 3 This is a flowchart illustrating the multi-dimensional feature extraction module of the present invention.

[0055] Figure 4 This is a flowchart illustrating the masking mechanism of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0057] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0058] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0059] In the description of the embodiments of the present invention, it should be noted that if terms such as "upper," "lower," "horizontal," or "inner" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of the invention is in use, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as limiting the present invention. Furthermore, terms such as "first" and "second" are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0060] Furthermore, the use of the term "horizontal" does not imply that the component must be absolutely horizontal, but rather that it can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal than "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0061] In the description of the embodiments of the present invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention according to the specific circumstances.

[0062] Time series forecasting, based on historical observation data of multivariate time series, aims to predict the future values ​​of a target variable by uncovering the inherent patterns within the sequence. Specifically, each observation sample at any given time point contains quantitative information across multiple feature dimensions, and all observation points are arranged in strict chronological order to form a continuous sequence. The key to this research lies in scientifically predicting the values ​​of the target sequence at subsequent time points by mining the historical evolution patterns of each feature variable.

[0063] Time series forecasting has a wide range of applications. For example, in the healthcare field, epidemic transmission can be predicted based on multi-source time series data such as historical incidence rates, population movement trajectories, and climate conditions. By predicting regional infection peaks and the distribution of key populations, precise prevention and control and efficient services can be achieved. In the industrial field, real-time time series data such as equipment vibration frequency, temperature, and current can be collected by sensors to predict future equipment failures. As a result, companies can arrange shutdowns for maintenance in advance to avoid losses caused by unplanned downtime.

[0064] To achieve the prediction goals in the aforementioned complex scenarios, time series forecasting tasks typically rely on machine learning or deep learning algorithms to build predictive models. These models can effectively characterize the inherent patterns in time series data: on the one hand, they can capture linear dependencies in the data; on the other hand, they can reveal nonlinear coupling mechanisms. By integrating the co-evolutionary information of multiple feature variables from historical observations, the model can transform the discovered time-series patterns into predictive capabilities, ultimately achieving quantitative inferences about the future values ​​of the target sequence.

[0065] Currently, there are three main problems with using mainstream time series forecasting methods for long series forecasting:

[0066] (1) Regarding model complexity. In time series forecasting, while a longer forecast range can provide decision-makers with more comprehensive forward-looking information, it also significantly increases the uncertainty of the forecast results. To support reliable long-series forecasting, the model needs to mine complex and extensive temporal dependencies from long-term historical data. In this context, although deep stacked network models have powerful time series modeling capabilities, their number of parameters grows exponentially with network depth, leading to a surge in computational resource consumption. This high complexity characteristic significantly limits their practical application and makes them difficult to implement in real-world business scenarios.

[0067] (2) Regarding feature extraction. While existing feature extraction models have made progress, they still have significant limitations. Taking one-dimensional models as an example, although LSTM captures long-term and short-term dependencies through gating mechanisms, it is limited by the unidirectional nature of time and has difficulty distinguishing different scale patterns. This makes it unable to analyze periodic patterns or capture the collaborative relationships of multi-scale dynamics in prediction. Although two-dimensional models have made breakthroughs in dimensional expansion, they still have not achieved comprehensive coverage. Although TimesNet achieves accurate analysis of single-variable periodic dynamics through "time-frequency" two-dimensional decomposition, it is limited to single-variable scenarios: when facing multi-variable coupled prediction tasks, it cannot model the interaction between variables and has difficulty capturing the joint "time-frequency" features.

[0068] (3) Regarding masking. Although there are many masking strategies available, they all share some common problems. For example, they are not well adapted to the essential characteristics of time series data, failing to fully consider the temporal dependencies, periodicity, trends, and multi-scale characteristics of time series data, making it difficult for the model to capture core patterns. Secondly, the granularity and location of the masking lack specificity, and using fixed granularity or simple random locations will prevent the model from covering key information such as short-term fluctuations, long-term trends, and abrupt change points. Finally, the multivariate interaction relationships are ignored. When dealing with multivariate time series, most strategies independently handle single-variable or random cross-variable masks without considering complex dependencies such as causality between variables, making it difficult for the model to establish joint features.

[0069] For the reasons mentioned above, this invention proposes a time series prediction method based on multi-dimensional feature fusion. The invention will be described in detail below with reference to embodiments and accompanying drawings.

[0070] As a first embodiment of the time series forecasting method of the present invention, it may include: inputting the target to be predicted into the time series forecasting model to obtain the forecast result.

[0071] This invention uses a time series forecasting model to directly obtain the forecast results. The specific structure of the time series forecasting model can be partially adjusted to meet the subsequent basic structure requirements.

[0072] like Figure 1 The diagram shown is a flowchart illustrating an embodiment of a time series forecasting method according to the present invention, which may include:

[0073] S1, Establish a time series prediction dataset.

[0074] By aggregating historical data of the target to be predicted, a multidimensional time-series data sample containing several attributes is collected. The multidimensional time-series data sample must meet the following requirements:

[0075] (1) Uniform spacing. The collected time-series data are sorted in chronological order, and the time difference between adjacent data is required to be constant. If some timestamps are missing, their timestamps are added and the empty positions are filled with the character NULL.

[0076] (2) Constant number of features. For each time step in the time series prediction dataset, the feature type and number must be consistent. If features are missing at some time steps, the empty positions should be filled with the character NULL.

[0077] (3) Include timestamp information. For each value in the time series prediction dataset, in addition to recording the value itself, the date and timestamp when the value was collected should also be recorded. The timestamp format should be uniform and should not contain Chinese characters.

[0078] For ease of subsequent description, each data point in the time series prediction dataset will be referred to as a sequence, and each data point in the sequence will be referred to as a point.

[0079] S2 cleans and standardizes time-series data.

[0080] Perform completion and standardization processing on the time series prediction dataset after preprocessing in step S1:

[0081] (1) Fill in the missing values. Fill the missing values ​​(NULL) in the time series prediction dataset with the mean of the two values ​​before and after the missing interval.

[0082] (2) Standardization. The completed time series forecast dataset is standardized using the following formula (1) to adjust each attribute column in the entire time series forecast dataset to a mean of 0 and a variance of 1:

[0083]

[0084] Where seq is the original sequence, mean(seq) is the mean of the entire sequence, std(seq) is the standard deviation of the sequence, and x is the standardized sequence.

[0085] S3, Construct a time series forecasting model.

[0086] Design time series prediction models with multi-dimensional feature extraction and lightweight architecture, such as Figure 2 The diagram illustrates the process of a time series prediction model. First, the input time series data is standardized, and a masking strategy is implemented for the periodic modeling data: the first 70% of the time series data is randomly masked by 30% to improve model robustness. Then, the data undergoes reshaping operations in different dimensions to meet the needs of the multi-dimensional feature extraction module. Next, features are extracted from the processed data in different dimensions. Finally, the obtained three-dimensional features are restored to their initial dimensions through reshaping operations and input into a multilayer perceptron for prediction. The prediction results are then destandardized and weighted to obtain the final prediction output.

[0087] The specific process is as follows:

[0088] Sequence standardization.

[0089] In time series prediction tasks, statistical distribution bias often exists between the training and test sets. To mitigate this impact, recent studies have widely adopted a lightweight "sample normalization" strategy. Following this approach, before the sequence is fed into the model, its mean is zeroed out to eliminate global bias. After the model outputs the data, the mean is added back to ensure the prediction results are restored to their original scale. This process is as follows:

[0090]

[0091] in, This represents the observations over the past L time steps. This represents the mean of past observations. This indicates the prediction result.

[0092] Construct a multi-dimensional feature extraction module.

[0093] like Figure 3 The diagram shown is a flowchart of the multi-dimensional feature extraction module.

[0094] (1) Establish a one-dimensional convolution model

[0095] To capture potential periodic patterns in time series data, a period-aware one-dimensional convolutional module is designed. First, a convolutional kernel is designed based on a specified period length, while padding is used to maintain the sequence length. Then, the designed one-dimensional convolutional kernel is used to perform convolution operations on the time series data to extract one-dimensional features. Finally, the extracted features are added to the original data to obtain a one-dimensional output. The above implementation is shown in formula (4):

[0096]

[0097] in, This represents the observations over the past L time steps, and 1DConv(·) indicates that a one-dimensional convolution kernel is applied for feature extraction. This represents the final one-dimensional feature obtained.

[0098] By employing a residual connection mechanism, the periodic features extracted by convolution are added to the original sequence features. This approach not only preserves the basic trend information but also alleviates the gradient decay problem in deep networks and enhances the robustness of the features.

[0099] (2) Establish a two-dimensional convolution model

[0100] Time series data often exhibits significant periodicity, such as daily traffic flow and monthly electricity consumption. One-dimensional convolutional operations cannot effectively capture the correlations between different periods when processing this type of data. Therefore, we use two-dimensional convolution to improve the accuracy of time series prediction.

[0101] First, the time-series data is transformed into two-dimensional data according to a specified period. Then, a two-dimensional convolution kernel is designed to perform a convolution operation on the time-series data to extract two-dimensional features. Finally, the extracted features are added to the original data to obtain the two-dimensional output. The above implementation is shown below:

[0102]

[0103] Where 2DConv(·) represents the application of a two-dimensional convolution kernel for feature extraction. This represents the final two-dimensional feature.

[0104] Two-dimensional convolution primarily captures local temporal patterns through intra-period convolution and inter-period convolution to capture dependencies between different periods. Finally, residual connections are used to alleviate the vanishing gradient problem and improve prediction performance.

[0105] (3) Establish a three-dimensional convolution model

[0106] Inspired by human cognition of time (such as year-month-day), a 3D convolution module is designed to directly learn the three-level time structure of "long-term-medium-short-term". Combined with the channel attention mechanism, the model's sensitivity to key time features is enhanced while reducing the computational complexity of 3D convolution.

[0107] First, the original time series is reshaped according to different sampling times in the dataset to meet the requirements of 3D convolution. Then, a 3D convolution kernel is designed to perform convolution operations on the time series data to extract 3D features. Finally, the extracted features are added to the original data to obtain the 3D output. The above implementation is shown below:

[0108]

[0109] Where 3DConv(·) represents the application of a three-dimensional convolution kernel for feature extraction. This represents the final three-dimensional feature obtained.

[0110] The design employs a depthwise separable 3D convolution, first mapping the input channels to an intermediate dimension, and then mapping the features to the output channels using the same structure. This expands the receptive field, capturing correlations between discontinuous time points. Finally, residual connections are used to mitigate the vanishing gradient problem in deep networks.

[0111] Construct a masking mechanism.

[0112] like Figure 4 The diagram shown is a schematic of the masking mechanism.

[0113] To improve prediction performance and overall model stability, a data augmentation strategy is designed to address the periodicity of time series data. Specifically, in both one-dimensional and two-dimensional periodic modeling, the data from the past L time steps are first divided into segments with fixed period lengths. Then, the first 70% of the segments are processed, and 30% of the complete segments are randomly selected for masking. The resulting mask matrix is ​​then multiplied by the original sequence to obtain the masked sequence. This process is illustrated below:

[0114]

[0115] in, This represents the observations over the past L time steps, and mask(·) indicates that a mask is used for the historical sequence.

[0116] This strategy forces the model to learn more fundamental inter-cycle correlations and general periodic characteristics rather than relying on specific periodic segments, thereby reducing overfitting to local periods. In this way, the model can better adapt to missing periods, anomalous periods, or unseen periodic patterns, thus improving the overall robustness of predictions.

[0117] Construct a fusion prediction module.

[0118] An adaptive multi-dimensional fusion network for time series prediction is designed, capturing temporal features at different granularities through three branches: a one-dimensional convolutional branch focuses on short-term local dependencies; a two-dimensional convolutional branch explicitly models two-dimensional relationships within and between periods; and a three-dimensional convolutional branch utilizes a multi-level temporal structure to simultaneously model the complex interactions of long-term trends, medium-term cycles, and short-term fluctuations. Each branch is followed by a multilayer perceptron for feature transformation. This structure maps the features of each time window to a higher-dimensional space for feature combination, and then projects them onto the prediction length. Essentially, it learns a non-linear mapping from historical segments to future segments. The above process is shown below:

[0119]

[0120] Where α and β represent learnable parameters, and MLP(·) represents the transformation of features across each dimension. This represents the value predicted for the next H time steps.

[0121] In particular, a learnable dynamic weighting mechanism is introduced, enabling the model to automatically allocate weights across different time scales based on data characteristics. This design preserves complementary information across different time scales while avoiding the limitations of manual parameter tuning through a data-driven approach.

[0122] S4, training the time series prediction model.

[0123] Based on the time series prediction dataset obtained in step S2 and the time series prediction model constructed in step S3, the parameters in the time series prediction model are trained.

[0124] First, construct the MSE loss function as follows:

[0125]

[0126] Where L is the loss value obtained from the MSE loss function, and this index quantifies the H-step prediction sequence given by the model in the form of squared difference. Compared with actual observation The overall degree of deviation between them.

[0127] Secondly, the time series prediction dataset is divided into training, test, and validation sets in a 6:3:1 ratio. The loss is tested on the test set after each training round. Training stops when the loss increases for five consecutive rounds, and the weight file with the minimum loss on the test set is stored.

[0128] S5, Predicted Output.

[0129] Based on the network parameters updated in step S4, predictions are made about the data.

[0130] In practical applications, the dataset used for prediction should be independent and identically distributed from the dataset used for training, and should also be processed by steps S1 and S2.

[0131] This invention has been practically verified and demonstrated to be accurate and fast in real-world tasks. In algorithm testing, the performance of the time series prediction model was evaluated in prediction tasks across multiple domains, and the results are shown in Table 1.

[0132] Table 1. Prediction results of the time series forecasting model in multiple domain forecasting tasks.

[0133]

[0134] The top of the table shows different prediction models horizontally and different tasks vertically, with the best prediction results highlighted in bold. As can be seen from the results in Table 1, the method proposed in this invention achieves comprehensive advantages across six datasets in four domains, demonstrating the effectiveness of the proposed method.

[0135] This invention proposes a time series forecasting method and a computer program product for predicting time series data. Addressing the problems and challenges of traditional models in feature extraction and model complexity, it proposes a multi-dimensional feature fusion model based on a lightweight architecture. This model includes a masking mechanism, multi-dimensional feature extraction, and a fusion prediction module. Experimental results demonstrate that the proposed fusion prediction method, by introducing a learnable dynamic weight mechanism, preserves complementary information at different time scales while avoiding the limitations of manual parameter tuning. The prediction method is simple to train and deploy, achieving excellent performance in prediction tasks across various fields, providing a typical example for time series forecasting tasks.

[0136] Compared with existing time series forecasting methods, the advantages of this invention are mainly reflected in the following aspects:

[0137] 1. Based on a lightweight architecture design, and centered around the "feature extraction-masking-prediction" design pattern, the design focuses on simplifying the structure of each module. While retaining the ability to capture multi-dimensional features, it reduces the number of model parameters and calculation steps. The core objective is to improve prediction efficiency and reduce computing power requirements.

[0138] 2. Design a multi-dimensional convolutional architecture that utilizes three-dimensional convolutional operations to capture short-term fluctuations, long-term trends, and timestamp correlations, achieving comprehensive coverage of local details, global trends, and temporal dynamics. The complementary fusion of three-dimensional features enhances the ability to capture complex patterns in time-series data, providing robust feature support for prediction.

[0139] 3. A periodic-based masking mechanism is proposed. This mechanism operates by randomly masking entire periodic segments before one-dimensional and two-dimensional convolutions that model the input based on a predefined period. It forces the model to learn anomalous fluctuations in the data, reducing its sensitivity to noise, and as a dynamic masking strategy, effectively mitigates the sensitivity to distribution shifts caused by fixed-period modeling.

[0140] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0141] The above are merely preferred embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A time series prediction method based on multi-dimensional feature fusion, characterized in that, include: The target to be predicted is input into the time series prediction model to obtain the prediction result; The training method for the time series prediction model includes: Establish a time series prediction dataset; the time series prediction dataset includes historical data of the target to be predicted and has several attributes; The time series prediction model is trained using the time series prediction dataset; the time series prediction model includes a masking module, a multi-dimensional feature extraction module, and a fusion prediction module. The masking module is used to perform periodic-level masking on the dimension based on periodic modeling to obtain masked time series data; The multi-dimensional feature extraction module is used to extract multi-dimensional features from the time series prediction dataset to obtain data features under different dimensions. The fusion prediction module is used to predict multi-dimensional features separately and introduces a learnable dynamic weight mechanism to weight and fuse the results of each dimension to obtain the final prediction result.

2. The time series forecasting method according to claim 1, characterized in that, After establishing the time series prediction dataset, the following is also included: The time series prediction dataset is filled with imputation by replacing missing values ​​with the mean of the two values ​​before and after the missing value. The time series prediction dataset after completion is standardized so that the mean is 0 and the variance is 1.

3. The time series prediction method according to claim 1 or 2, characterized in that, The time series prediction dataset must meet the following requirements: The time difference between adjacent data in the time series prediction dataset is constant. The feature type and number of features are consistent at each time step in the time series prediction dataset; The time series prediction dataset includes the date and timestamp of each data collection, and the date and timestamp format of each data is consistent.

4. The time series forecasting method according to claim 1, characterized in that, The specific process of the mask module includes: In the dimension of periodic modeling, the data from the past L time steps are first divided into fixed period lengths. Then, the first 70% of the periods are processed, and 30% of the complete periods are randomly selected for masking (entire segments are set to zero). After obtaining the mask matrix, it is multiplied with the original sequence to obtain the masked sequence. The mask of the i-th feature is calculated as follows: Here, 'i' is used to identify the feature dimension. Mask represents the observations over the past L time steps. · ) indicates that a mask is used for the historical sequence.

5. The time series forecasting method according to claim 1, characterized in that, The specific process of the multi-dimensional feature extraction module includes: Design a periodic-aware one-dimensional convolutional module. First, design the convolutional kernel according to the specified period length, while using padding to maintain the sequence length and avoid information loss. Then, use the designed one-dimensional convolutional kernel to perform convolution operations on the temporal data to extract one-dimensional features. Finally, use a residual connection mechanism to add the extracted features to the original temporal data to alleviate the gradient decay problem of deep networks. The formula for the i-th feature is as follows: in, Represents the observations over the past L time steps, 1DConv( · This indicates that a one-dimensional convolutional kernel is used for feature extraction. This represents the final one-dimensional feature; Two-dimensional convolutions are designed to capture the correlations between different periods. First, time series data is transformed into two-dimensional data according to a specified period. Then, intra-period convolutions are designed to capture local time patterns, and inter-period convolutions are designed to capture the dependencies between different periods. Finally, a residual connection mechanism is used to alleviate gradient vanishing and improve prediction performance. The calculation of the i-th feature is as follows: Among them, 2DConv( · This indicates that a two-dimensional convolutional kernel is used for feature extraction. This represents the final two-dimensional feature; The design employs a 3D convolutional architecture to directly learn the "long-term, intermediate-term, and short-term" time structure. First, the original time series is reshaped according to different sampling times in the dataset to meet the requirements of 3D convolution. Then, a depthwise separable 3D convolution is designed, mapping the input channels to the intermediate dimension and then mapping the features to the output channels using the same structure, while simultaneously expanding the receptive field to capture the correlation between discontinuous time points. Finally, residual connections are designed to alleviate the gradient vanishing problem. The formula for the i-th feature is as follows: Among them, 3DConv( · This indicates that a 3D convolutional kernel is used for feature extraction. This represents the final three-dimensional feature obtained.

6. The time series forecasting method according to claim 1, characterized in that, The specific process of the fusion prediction module includes: Design an adaptive multi-dimensional fusion network for time series prediction. It captures temporal features at different granularities through three branches, and each branch is followed by a multilayer perceptron (MLP) for feature transformation. This structure maps the features of each time window to a higher-dimensional space for feature combination, and then projects them onto the prediction length. Furthermore, two learnable dynamic weight parameters are introduced, enabling the model to automatically allocate different scales based on data characteristics. The calculation of the i-th feature is as follows: Where α and β represent learnable parameters, and MLP(·) represents the transformation of features across each dimension. This represents the value predicted for the next H time steps.

7. The time series forecasting method according to claim 1, characterized in that, The loss function used when training the time series prediction model includes: The optimization objective of the model is the mean squared error (MSE), which is the most commonly used metric in the field of time series forecasting; this metric quantifies the H-step predicted sequence given by the model in the form of squared difference. Compared with actual observation The overall deviation between them is specifically implemented as follows: Where L is the loss value obtained from the MSE loss function.

8. A computer program product, comprising a memory, a processor, and a computer program executable on the processor, characterized in that, The memory stores a computer program executable by the processor, which executes the computer program to implement the time series prediction method based on multi-dimensional feature fusion as described in any one of claims 1-7.

Images