Sales prediction method and system based on adaptive multi-kernel convolution space-time attention

Abstract

The invention discloses a sales volume prediction method and system based on adaptive multi-kernel convolution space-time attention, relates to the technical field of sales volume prediction of e-commerce platforms, and improves the perception ability of a model to commodity sales volume position features by enhancing a position coding mechanism and combining timestamp embedding with relative position coding. A self-adaptive multi-kernel convolution-cross attention module is put forward, a multi-scale convolution kernel group design is utilized, a relationship of spatial features in commodity sales volume and a local dependency relationship of single spatial features are captured at the same time, and time-space dependency fusion is realized. An e-commerce platform sales prediction model is provided based on adaptive multi-kernel convolution-cross attention. The model can capture a time sequence local dependency relationship, a long-term dependency relationship and a spatial dimension relationship in data, and performs fusion output; by capturing the time sequence local dependency relationship, the long-term dependency relationship and the spatial dimension in the data, the method has the advantages of being adaptive to multiple data sets and accurate in prediction.

Description

Technical Field

[0001] This invention relates to the field of sales prediction technology for e-commerce platforms, specifically a sales prediction method and system based on adaptive multi-kernel convolutional spatiotemporal attention. Background Technology

[0002] Product sales forecasting holds immense value across all industries, helping businesses better understand future trends, make more informed decisions, and positively impact societal development. The core characteristics of time-series data lie in its temporal dependence and non-stationarity. Traditional statistical models demonstrate stability in short-term forecasting, but their linear assumptions and fixed-parameter mechanisms struggle to capture multi-scale spatiotemporal dependencies when faced with complex and ever-changing long-term forecasting tasks. Given the crucial role of product sales data in real-world applications, and the challenges of complex data characteristics and model design, long-term product sales forecasting remains a challenging research area.

[0003] However, several issues remain to be addressed in current research. First, previous studies have focused primarily on modifying model components or architecture; a simple self-attention component cannot fully describe the characteristics of product sales. Second, there is a spatiotemporal coupling defect; traditional self-attention treats temporal and spatial dependencies as a one-dimensional sequence, leading to confusion between local fine-grained features and global features. Third, multi-scale perception is insufficient; existing work often relies on fixed windows or posterior decomposition (such as seasonal-trend separation), making it difficult to adapt to differences in data features and time periods arising from different datasets. This results in lower accuracy in predicting product sales. Summary of the Invention

[0004] To address the shortcomings mentioned in the background section, the present invention aims to provide a sales prediction method and system based on adaptive multi-kernel convolutional spatiotemporal attention.

[0005] Firstly, the objective of this invention can be achieved through the following technical solution: a sales prediction method based on adaptive multi-kernel convolutional spatiotemporal attention, the method comprising the following steps: Obtain product sales data within a historical time period, embed and encode the product sales data within the historical time period to obtain product historical sales feature parameters, and generate multiple one-dimensional convolutional kernels based on the number of product historical sales feature parameters. Pooling and multi-kernel one-dimensional convolution operations are performed on the historical sales feature parameters of the product using multiple one-dimensional convolution kernels to obtain time feature data at multiple scales. The time feature data at multiple scales are weighted to obtain time attention weights. The time attention weights are then applied to the historical sales feature parameters of the product to obtain historical sales data with time features. Based on a pre-defined two-dimensional spatial convolutional layer, max pooling and average pooling are performed on historical sales data with time characteristics to obtain two-dimensional spatial dimension data; convolution calculation is performed on the two-dimensional spatial dimension data to obtain the weights of spatial features, and the weights of spatial features are applied to historical sales data with time characteristics to obtain historical feature data with both time and spatial dimensions. An attention mechanism is used to obtain global product sales data from historical feature data with time and space dimensions. This global product sales data is then input into a pre-built feedforward neural network model, and the output is the product sales prediction result.

[0006] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of embedding and encoding the sales data of goods within the historical time period, as follows: By fusing the features of historical product data with date features, and adding location encoding with a time feature dimension, computation can be performed by mapping from a low-dimensional space to a high-dimensional space.

[0007] Set the product sales input data size to ,in For product sales volume sequence, For batch size, For the number of time features, Given the sequence length, convert the date attribute in the product sales data to... With each feature dimension, the product sales sequence after date-scale expansion is: , The expanded time feature dimension, after further high-dimensional mapping, results in the following product sales sequence: As a parameter representing the historical sales characteristics of a product, The dimension of the high-dimensional mapping; .

[0008] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of generating multiple one-dimensional convolutional kernels based on the number of historical sales feature parameters of the goods, as follows: Calculate multiple one-dimensional convolution kernels to form a set of convolution kernels and initialize a multi-kernel one-dimensional convolutional layer. The formula for calculating the one-dimensional convolution kernel set is as follows: in For the first The size of a one-dimensional convolution kernel For the expanded time feature dimension, Use the nearest odd number obtained by rounding up the square root.

[0009] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of performing pooling and multi-kernel one-dimensional convolution operations on the historical sales feature parameters of the goods based on multiple one-dimensional convolution kernels, as follows: To obtain data in the time dimension of historical sales data, an average pooling operation is performed on the historical sales data. The calculation formula is as follows: in, The first characteristic representing the sales volume of the input product 1 eigenvector For batch size, For the dimension of the high-dimensional mapping, The pooled tensor is then subjected to a one-dimensional global average pooling operation to obtain... ; right Multi-scale temporal feature data are obtained by performing convolution calculations with different kernels. The calculation formula for multi-kernel one-dimensional convolution is as follows: in, It is a one-dimensional convolution function with a kernel size of . , For batch size, Dimensions of the high-dimensional mapping In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the calculation process of the historical sales data with time characteristics is as follows: The maximum value of multiple time feature data is calculated along the time feature dimension, the weight of time features is calculated along the time feature dimension, and historical sales data with time features are calculated. .

[0010] in, For function, To find the maximum value function, To obtain temporal feature weights for the activation function, For multi-kernel one-dimensional convolutional feature vectors, For a multi-kernel one-dimensional convolution feature vector to take the maximum value in the time feature dimension, For product sales volume sequence, This is a sequence of product sales based on learning time characteristics.

[0011] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of performing max pooling and average pooling on historical sales data with time characteristics based on a preset two-dimensional spatial convolutional layer, as follows: Initialize a two-dimensional spatial convolutional layer based on the size of the convolutional kernel in the spatial dimension; The calculation formula is: in, The first characteristic representing the sales volume of the input product The feature vector at time 1 For batch size, For the expanded time feature dimension, The pooled tensor is then subjected to a one-dimensional global average pooling operation to obtain... ; Average pooling is performed on the time-dimensional data to obtain... Max pooling yields And concatenate them to obtain two-dimensional time dimension data, the calculation formula is: in, The first characteristic representing the sales volume of the input product The feature vector at time 1 For batch size, for , The concatenated tensor, where max is the maximum value function. This is a join function.

[0012] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the processing procedure for the historical feature data having time and spatial dimensions, as follows: Spatial convolution and gating calculations are performed on two-dimensional time-dimensional data to obtain spatial feature weights. These spatial feature weights are then applied to historical sales data with temporal features to obtain historical feature data with both time and spatial dimensions. The calculation is as follows: in, It is a convolution function. The activation function is used to obtain the weights of the time features.

[0013] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of using an attention mechanism on historical feature data with time and spatial dimensions to obtain commodity sales data with global features, as follows: Three feature tensors are obtained by performing three fully connected operations on historical feature data with time and spatial dimensions. , , Self-attention is calculated for the three feature tensors using the following formula: in, The activation function is used to obtain global feature weights. This is the value of the dimension length.

[0014] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the pre-established feedforward neural network model as follows: in, It is a fully connected layer. Y represents the predicted sales volume of goods, derived from historical feature data learned through self-attention learning.

[0015] Secondly, in order to achieve the above objectives, this invention discloses a sales prediction system based on adaptive multi-kernel convolutional spatiotemporal attention, comprising: The feature encoding module is used to acquire product sales data within a historical time period, embed and encode the product sales data within the historical time period to obtain product historical sales feature parameters, and generate multiple one-dimensional convolutional kernels based on the number of product historical sales feature parameters. The feature processing module is used to perform pooling and multi-kernel one-dimensional convolution operations on the historical sales feature parameters of the product based on multiple one-dimensional convolution kernels to obtain time feature data at multiple scales. The time feature data at multiple scales are weighted to obtain time attention weights. The time attention weights are applied to the historical sales feature parameters of the product to obtain historical sales data with time features. The spatiotemporal fusion module is used to perform max pooling and average pooling on historical sales data with time features based on a preset two-dimensional spatial convolutional layer to obtain two-dimensional spatial dimension data; convolution calculation is performed on the two-dimensional spatial dimension data to obtain the weights of spatial features, and the weights of spatial features are applied to historical sales data with time features to obtain historical feature data with both time and spatial dimensions. The sales forecasting module uses an attention mechanism to obtain global product sales data from historical feature data with time and spatial dimensions. The global product sales data is then input into a pre-built feedforward neural network model, and the product sales forecast results are output.

[0016] The beneficial effects of this invention are: This invention enhances the location encoding mechanism by combining timestamp embedding with relative location encoding, thereby improving the model's ability to perceive the location features of product sales. It proposes an adaptive multi-kernel convolution-cross-attention module, utilizing a multi-scale convolutional kernel design to simultaneously capture the spatial relationships within product sales data, as well as the local dependencies of single spatial features, achieving spatiotemporal dependency fusion. Based on this adaptive multi-kernel convolution-cross-attention approach, a sales prediction model for e-commerce platforms is proposed. This model can capture temporal local dependencies, long-term dependencies, and spatial dimensional relationships in the data, and fuse them into a single output, thus improving the accuracy of product sales prediction. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the system structure of the present invention; Figure 3 This is a schematic diagram of the model construction of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: like Figure 1 As shown, a sales prediction method based on adaptive multi-kernel convolutional spatiotemporal attention includes the following steps: S101: Obtain product sales data within a historical time period, embed and encode the product sales data within the historical time period to obtain product historical sales feature parameters; generate multiple one-dimensional convolutional kernels based on the number of product historical sales feature parameters. Specifically, the data loader retrieves historical data from product sales data within a set historical time period; The process of embedding and encoding product sales data within a historical time period is as follows: By fusing the features of historical product data with date features, and adding location encoding with a time feature dimension, computation can be performed by mapping from a low-dimensional space to a high-dimensional space.

[0020] Set the product sales input data size to ,in For product sales volume sequence, For batch size, For the number of time features, Given the sequence length, convert the date attribute in the product sales data to... With each feature dimension, the product sales sequence after date-scale expansion is: , The expanded time feature dimension, after further high-dimensional mapping, results in the following product sales sequence: As a parameter representing the historical sales characteristics of a product, The dimension of the high-dimensional mapping; The process of generating multiple one-dimensional convolutional kernels based on the number of historical sales feature parameters of a product is as follows: Calculate multiple one-dimensional convolution kernels to form a set of convolution kernels and initialize a multi-kernel one-dimensional convolutional layer. The formula for calculating the one-dimensional convolution kernel set is as follows: in For the first The size of a one-dimensional convolution kernel For the expanded time feature dimension, The nearest odd number is obtained by rounding up the square root; the minimum value is 3.

[0021] S102: Based on multiple one-dimensional convolution kernels, pooling and multi-kernel one-dimensional convolution operations are performed on the historical sales feature parameters of the product to obtain time feature data at multiple scales. The time feature data at multiple scales are weighted to obtain time attention weights. The time attention weights are applied to the historical sales feature parameters of the product to obtain historical sales data with time features. The process of pooling and multi-kernel one-dimensional convolution operations on the historical sales feature parameters of a product using multiple one-dimensional convolution kernels is as follows: To obtain data in the time dimension of historical sales data, an average pooling operation is performed on the historical sales data. The calculation formula is as follows: in, The first characteristic representing the sales volume of the input product 1 eigenvector For batch size, For the dimension of the high-dimensional mapping, The pooled tensor is then subjected to a one-dimensional global average pooling operation to obtain... ; right Multi-scale temporal feature data are obtained by performing convolution calculations with different kernels. The calculation formula for multi-kernel one-dimensional convolution is as follows: in, It is a one-dimensional convolution function with a kernel size of . , For batch size, The dimension of the high-dimensional mapping; The calculation process for historical sales data with time characteristics is as follows: The maximum value of multiple time feature data is calculated along the time feature dimension, the weight of time features is calculated along the time feature dimension, and historical sales data with time features are calculated. .

[0022] in, For function, To find the maximum value function, To obtain temporal feature weights for the activation function, For multi-kernel one-dimensional convolutional feature vectors, For a multi-kernel one-dimensional convolution feature vector to take the maximum value in the time feature dimension, For product sales volume sequence, This is a sequence of product sales based on learning time characteristics.

[0023] S103: Based on a preset two-dimensional spatial convolutional layer, max pooling and average pooling are performed on historical sales data with time characteristics to obtain two-dimensional spatial dimension data; convolution calculation is performed on the two-dimensional spatial dimension data to obtain the weights of spatial features, and the weights of spatial features are applied to historical sales data with time characteristics to obtain historical feature data with time and spatial dimensions. The process of performing max pooling and average pooling on historical sales data with time characteristics based on a pre-defined two-dimensional spatial convolutional layer is as follows: Initialize a two-dimensional spatial convolutional layer based on the size of the convolutional kernel in the spatial dimension; The calculation formula is: in, The first characteristic representing the sales volume of the input product The feature vector at time 1 For batch size, For the expanded time feature dimension, The pooled tensor is then subjected to a one-dimensional global average pooling operation to obtain... ; Average pooling is performed on the time-dimensional data to obtain... Max pooling yields And concatenate them to obtain two-dimensional time dimension data, the calculation formula is: in, The first characteristic representing the sales volume of the input product The feature vector at time 1 For batch size, for , The concatenated tensor, where max is the maximum function. This is a join function.

[0024] The processing procedure for historical data with both temporal and spatial dimensions is as follows: Spatial convolution and gating calculations are performed on two-dimensional time-dimensional data to obtain spatial feature weights. These spatial feature weights are then applied to historical sales data with temporal features to obtain historical feature data with both time and spatial dimensions. The calculation is as follows: in, It is a convolution function. The activation function is used to obtain the weights of the time features.

[0025] S104: Use an attention mechanism to obtain global product sales data from historical feature data with time and space dimensions. Input the global product sales data into a pre-established feedforward neural network model and output the product sales prediction result.

[0026] The process of using an attention mechanism to obtain product sales data with global features from historical feature data that has time and spatial dimensions is as follows: Three feature tensors are obtained by performing three fully connected operations on historical feature data with time and spatial dimensions. , , Self-attention is calculated for the three feature tensors using the following formula: in, The activation function is used to obtain global feature weights. This is the value of the dimension length.

[0027] The pre-established feedforward neural network model is as follows: in, It is a fully connected layer. Y represents the predicted sales volume of goods, derived from historical feature data learned through self-attention learning.

[0028] Specifically, the present invention will be further illustrated below through embodiments: To achieve the above objectives, such as Figure 3 As shown in the figure, the model construction of this method is illustrated. When predicting product sales, this method can simultaneously focus on the spatiotemporal features of the product sales time series and adaptively adjust the size of the convolutional kernel according to the number of features, enabling it to be applied to different product sales prediction scenarios. The specific training implementation steps are as follows: Obtain the product sales data set, divide the product sales data into a training dataset, a validation set, and a test set, with a division ratio of 0.7:0.1:0.2, and then standardize the data.

[0029] Set the hyperparameters required for training: batch size of 32, training feature of M (multivariate prediction of multivariate), feature dimension of 512, input sequence length of 90, window size of 28, and prediction length of 28.

[0030] The training objective is to predict the sales volume of a product over the next 28 days using 90 days of historical sales data. The objective loss function used in this method is MSE.

[0031] The training phase of this method involves inputting the training dataset into the neural network parameters of the model to be learned. The validation phase fine-tunes the neural network parameters of the model. The testing phase evaluates the predictive performance of the method.

[0032] In practice, this method was used to predict product sales volume on a product dataset (M5). In the sales feature encoding stage, historical sales data of a certain product (Hobbies_1_002) are selected for feature encoding. The encoded data is then input into the adaptive spatiotemporal module for temporal and spatial feature extraction. After feature extraction, the data is input into the attention module for attention learning, and finally into the feedforward neural network for prediction of product sales results.

[0033] To verify the scalability of the proposed scheme, weather forecasting was performed using the Weather dataset and electricity consumption forecasting was performed using the ECL dataset. The specific test set results are shown in Table 1.

[0034] Table 1. Model performance compared to existing benchmark methods Example 2: To achieve the above objective, such as Figure 2 As shown, based on Embodiment 1, this invention discloses a sales prediction system based on adaptive multi-kernel convolutional spatiotemporal attention, comprising: Feature encoding module 11 is used to acquire product sales data within a historical time period, embed and encode the product sales data within the historical time period to obtain product historical sales feature parameters, and generate multiple one-dimensional convolution kernels based on the number of product historical sales feature parameters. Feature processing module 12 is used to perform pooling and multi-kernel one-dimensional convolution operations on the historical sales feature parameters of the product based on multiple one-dimensional convolution kernels to obtain time feature data at multiple scales. The time feature data at multiple scales are weighted to obtain time attention weights. The time attention weights are applied to the historical sales feature parameters of the product to obtain historical sales data with time features. The spatiotemporal fusion module 13 is used to perform max pooling and average pooling on historical sales data with time characteristics based on a preset two-dimensional spatial convolutional layer to obtain two-dimensional spatial dimension data; to perform convolution calculation on the two-dimensional spatial dimension data to obtain the weights of spatial features; and to apply the weights of spatial features to historical sales data with time characteristics to obtain historical feature data with time and spatial dimensions. The sales forecast module 14 is used to obtain global product sales data by using an attention mechanism on historical feature data with time and space dimensions. The global product sales data is then input into a pre-established feedforward neural network model, and the product sales forecast result is output.

[0035] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0036] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0037] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0038] The foregoing has shown and described the basic principles, main features, and advantages of this disclosure. Those skilled in the art should understand that this disclosure is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of this disclosure. Various changes and modifications can be made to this disclosure without departing from its spirit and scope, and all such changes and modifications fall within the scope of this disclosure as claimed.

Claims

1. A sales prediction method based on adaptive multi-kernel convolutional spatiotemporal attention, characterized in that, The method includes the following steps: Obtain product sales data within a historical time period, embed and encode the product sales data within the historical time period, and obtain product historical sales feature parameters. Multiple one-dimensional convolutional kernels are generated based on the number of historical sales feature parameters of the product. Pooling and multi-kernel one-dimensional convolution operations are performed on the historical sales feature parameters of the product using multiple one-dimensional convolution kernels to obtain time feature data at multiple scales. The time feature data at multiple scales are weighted to obtain time attention weights. The time attention weights are then applied to the historical sales feature parameters of the product to obtain historical sales data with time features. Based on a pre-defined two-dimensional spatial convolutional layer, max pooling and average pooling are performed on historical sales data with time characteristics to obtain two-dimensional spatial dimension data; convolution calculation is performed on the two-dimensional spatial dimension data to obtain the weights of spatial features, and the weights of spatial features are applied to historical sales data with time characteristics to obtain historical feature data with both time and spatial dimensions. An attention mechanism is used to obtain global product sales data from historical feature data with time and space dimensions. This global product sales data is then input into a pre-built feedforward neural network model, and the output is the product sales prediction result.

2. The sales prediction method based on adaptive multi-kernel convolutional spatiotemporal attention as described in claim 1, characterized in that, The process of embedding and encoding product sales data within a historical time period is as follows: By fusing the features of historical product data with date features, and adding location encoding with a time feature dimension, computation can be performed by mapping from a low-dimensional space to a high-dimensional space. Set the product sales input data size to ,in For product sales volume sequence, For batch size, For the number of time features, Given the sequence length, convert the date attribute in the product sales data to... With each feature dimension, the product sales sequence after date-scale expansion is: , The expanded time feature dimension, after further high-dimensional mapping, results in the following product sales sequence: As a parameter representing the historical sales characteristics of a product, The dimension of the high-dimensional mapping; 。 3. The sales prediction method based on adaptive multi-kernel convolutional spatiotemporal attention as described in claim 1, characterized in that, The process of generating multiple one-dimensional convolutional kernels based on the number of historical sales feature parameters of the product is as follows: Calculate multiple one-dimensional convolution kernels to form a set of convolution kernels and initialize a multi-kernel one-dimensional convolutional layer. The formula for calculating the one-dimensional convolution kernel set is as follows: in For the first The size of a one-dimensional convolution kernel For the expanded time feature dimension, Use the nearest odd number obtained by rounding up the square root.

4. The sales prediction method based on adaptive multi-kernel convolutional spatiotemporal attention according to claim 1, characterized in that, The process of pooling and multi-kernel one-dimensional convolution operations on the historical sales feature parameters of the product based on multiple one-dimensional convolution kernels is as follows: Product sales sequence Let X be the historical sales data. Perform average pooling on the time dimension of X to obtain the data in the time feature dimension. The calculation formula is as follows: in, The first characteristic representing the sales volume of the input product 1 eigenvector For batch size, For the dimension of the high-dimensional mapping, The pooled tensor is then subjected to a one-dimensional global average pooling operation to obtain... ; right Multi-scale temporal feature data are obtained by performing convolution calculations with different kernels. The calculation formula for multi-kernel one-dimensional convolution is as follows: in, It is a one-dimensional convolution function with a kernel size of . , For batch size, The dimension of the high-dimensional mapping.

5. The sales prediction method based on adaptive multi-kernel convolutional spatiotemporal attention according to claim 1, characterized in that, The calculation process for the historical sales data with time characteristics is as follows: The maximum value of multiple time feature data is calculated along the time feature dimension, the weight of time features is calculated along the time feature dimension, and historical sales data with time features are calculated. ; in, For function, To find the maximum value function, To obtain temporal feature weights for the activation function, For multi-kernel one-dimensional convolutional feature vectors, For a multi-kernel one-dimensional convolution feature vector to take the maximum value in the time feature dimension, For product sales volume sequence, This is a sequence of product sales based on learning time characteristics.

6. The sales prediction method based on adaptive multi-kernel convolutional spatiotemporal attention according to claim 1, characterized in that, The process of performing max pooling and average pooling on historical sales data with time characteristics based on a preset two-dimensional spatial convolutional layer is as follows: Initialize a two-dimensional spatial convolutional layer based on the size of the convolutional kernel in the spatial dimension; The calculation formula is: in, The first characteristic representing the sales volume of the input product The feature vector at time 1 For batch size, For the expanded time feature dimension, The pooled tensor is then subjected to a one-dimensional global average pooling operation to obtain... ; Average pooling is performed on the time-dimensional data to obtain... Max pooling yields And concatenate them to obtain two-dimensional time dimension data, the calculation formula is: in, The first characteristic representing the sales volume of the input product The feature vector at time 1 For batch size, for , The concatenated tensor, where max is the maximum value function. This is a join function.

7. The sales prediction method based on adaptive multi-kernel convolutional spatiotemporal attention according to claim 1, characterized in that, The processing procedure for the historical feature data with time and spatial dimensions is as follows: Spatial convolution and gating calculations are performed on two-dimensional time-dimensional data to obtain spatial feature weights. These spatial feature weights are then applied to historical sales data with temporal features to obtain historical feature data with both time and spatial dimensions. The calculation is as follows: in, It is a convolution function. The activation function is used to obtain the weights of the time features.

8. The sales prediction method based on adaptive multi-kernel convolutional spatiotemporal attention according to claim 1, characterized in that, The process of using an attention mechanism to obtain global product sales data from historical feature data with time and spatial dimensions is as follows: Three feature tensors are obtained by performing three fully connected operations on historical feature data with time and spatial dimensions. , , Self-attention calculation is performed on the three feature tensors to obtain The calculation formula is: in, The activation function is used to obtain global feature weights. This is the value of the dimension length.

9. The sales prediction method based on adaptive multi-kernel convolutional spatiotemporal attention according to claim 1, characterized in that, The pre-established feedforward neural network model is as follows: in, It is a fully connected layer. Y represents the predicted sales volume of goods, derived from historical feature data learned through self-attention learning.

10. A sales forecasting system based on adaptive multi-kernel convolutional spatiotemporal attention, employing the sales forecasting method based on adaptive multi-kernel convolutional spatiotemporal attention as described in any one of claims 1 to 9, characterized in that, include: The feature encoding module is used to acquire product sales data within a historical time period, embed and encode the product sales data within the historical time period, and obtain product historical sales feature parameters. Multiple one-dimensional convolutional kernels are generated based on the number of historical sales feature parameters of the product. The feature processing module is used to perform pooling and multi-kernel one-dimensional convolution operations on the historical sales feature parameters of the product based on multiple one-dimensional convolution kernels to obtain time feature data at multiple scales. The time feature data at multiple scales are weighted to obtain time attention weights. The time attention weights are applied to the historical sales feature parameters of the product to obtain historical sales data with time features. The spatiotemporal fusion module is used to perform max pooling and average pooling on historical sales data with time features based on a preset two-dimensional spatial convolutional layer to obtain two-dimensional spatial dimension data; convolution calculation is performed on the two-dimensional spatial dimension data to obtain the weights of spatial features, and the weights of spatial features are applied to historical sales data with time features to obtain historical feature data with both time and spatial dimensions. The sales forecasting module uses an attention mechanism to obtain global product sales data from historical feature data with time and spatial dimensions. The global product sales data is then input into a pre-built feedforward neural network model, and the product sales forecast results are output.

Images