Commodity issuing amount prediction method and device, electronic equipment and storage medium

Abstract

The invention provides a commodity issuing amount prediction method and device, electronic equipment and a storage medium. The commodity issuing amount prediction method comprises the steps of obtaining multi-dimensional data related to a to-be-predicted commodity issuing amount; wherein the multi-dimensional data at least comprises the sales data and the influence information of the sales volume; the trained target model is utilized to predict the issuing amount of the commodities based on the multi-dimensional data, and a target prediction result is generated; wherein the target model is a model formed by combining prediction models of at least two algorithm types, the at least two algorithm types at least comprise a linear type and a non-linear type, the prediction model applies an algorithm of the linear type or an algorithm of the non-linear type, and each algorithm type corresponds to at least one prediction model; the prediction model is a trained model. According to the invention, the prediction accuracy can be improved.

Description

Technical Field

[0001] This application relates to the field of forecasting technology, and in particular to a method, apparatus, electronic device, and storage medium for forecasting the distribution volume of a commodity. Background Technology

[0002] The 21st century is an era of highly developed information technology, in which the management and analysis of business information play a crucial role. Enterprises process vast amounts of sales data daily, and this data analysis is more direct and effective than traditional qualitative analysis methods. With the continuous advancement of data processing technology, enterprises have increasingly more ways to obtain dynamic market information, and the requirements for the accuracy and timeliness of data analysis are also rising. To meet the growing needs of enterprises for understanding future market trends, accurate forecasting of product issuance is becoming increasingly important; therefore, forecasting product issuance has significant practical implications for the stable future development of enterprises.

[0003] In related technologies, an intelligent algorithm is often used to predict the quantity of goods to be issued after a period of time by collecting sales data, or a machine learning model is trained and then used to predict the quantity of goods to be issued after a period of time by collecting sales data to obtain the product issuance quantity prediction result. However, the accuracy of the product issuance quantity prediction result obtained by using only an intelligent algorithm or a machine learning model is not high. Summary of the Invention

[0004] This application provides a method, apparatus, electronic device, and storage medium for predicting the quantity of goods issued, which can improve the accuracy of the prediction results.

[0005] The technical solution of this application embodiment is implemented as follows:

[0006] This application provides a method for predicting the issuance volume of a commodity, including:

[0007] Acquire multi-dimensional data related to the issuance volume of the product to be predicted; wherein, the multi-dimensional data includes at least sales data and information on the impact of sales volume;

[0008] Using the trained target model, the issuance volume of the product is predicted based on the multi-dimensional data, and a target prediction result is generated.

[0009] The target model is a model composed of prediction models combining at least two algorithm types, including at least linear and nonlinear types. The prediction model applies either the linear or nonlinear algorithm, with each algorithm type corresponding to at least one prediction model. The linear prediction model is used to predict the distribution volume of the product from a numerical calculation perspective, while the nonlinear prediction model is used to predict the distribution volume of the product from the perspective of nonlinear interaction of data in multi-dimensional data. The prediction model is a model trained using at least multi-dimensional historical data and corresponding actual distribution volumes as training samples. The multi-dimensional historical data includes at least historical sales data and information on the impact of historical sales volumes.

[0010] This application provides a device for predicting the quantity of goods issued, including:

[0011] The acquisition module is used to acquire multi-dimensional data related to the issuance volume of the product to be predicted; wherein, the multi-dimensional data includes at least sales data and information on the impact of sales volume;

[0012] The first generation module is used to predict the issuance volume of goods based on the multi-dimensional data using the trained target model, and generate the target prediction result.

[0013] The target model is a model composed of prediction models combining at least two algorithm types, including at least linear and nonlinear types. The prediction model applies either the linear or nonlinear algorithm, with each algorithm type corresponding to at least one prediction model. The linear prediction model is used to predict the distribution volume of the product from a numerical calculation perspective, while the nonlinear prediction model is used to predict the distribution volume of the product from the perspective of nonlinear interaction of data in multi-dimensional data. The prediction model is a model trained using at least multi-dimensional historical data and corresponding actual distribution volumes as training samples. The multi-dimensional historical data includes at least historical sales data and information on the impact of historical sales volumes.

[0014] In some possible implementations, the first generation module is used to predict the issuance volume of the commodity based on the multi-dimensional data using at least one prediction model corresponding to the linear type and at least one prediction model corresponding to the nonlinear type, and generate a first prediction result corresponding to each prediction model.

[0015] The target prediction result is generated based on the first prediction result corresponding to the prediction models of the at least two algorithm types, using the weight parameters of the target model; wherein the weight parameters of the target model include at least the weights corresponding to each prediction model.

[0016] In some possible implementations, the first generation module is used to perform normalization processing on the multi-dimensional data to generate first feature data;

[0017] Based on the input first feature data, each prediction model generates a first prediction result corresponding to each prediction model.

[0018] In some possible implementations, the first generation module is used to standardize the multi-dimensional data to generate first standardized data;

[0019] The first standardized data is subjected to feature extraction using a feature extraction function, and the first feature data is generated based on the extracted features.

[0020] In some possible implementations, the second generation module is used to train the initial prediction model of the at least two algorithm types using multi-dimensional historical data and the corresponding actual issuance volume as training samples, and generate the trained prediction model of the at least two algorithm types.

[0021] Based on the trained prediction models of the at least two algorithm types, the trained target model is generated.

[0022] In some possible implementations, the second generation module is used to generate second feature data based on the multi-dimensional historical data; wherein the second feature data includes at least the features of the multi-dimensional historical data;

[0023] Using a first algorithm, at least one initial prediction model corresponding to the linear type is trained based on the second feature data and the actual issuance volume to generate at least one prediction model corresponding to the linear type after training; wherein, the first algorithm is at least used to fit and optimize the parameters of the prediction model of the linear type through linear analysis;

[0024] Using the second algorithm, at least one initial prediction model corresponding to the nonlinear type is trained based on the second feature data and the actual issuance volume to generate at least one prediction model corresponding to the nonlinear type after training; wherein, the second algorithm is at least used to optimize the parameters of the prediction model of the nonlinear type through nonlinear analysis.

[0025] In some possible implementations, the second generation module is used to use at least one prediction model corresponding to the trained linear type to predict the issuance volume of the product based on the multi-dimensional historical data, and generate a second prediction result corresponding to the prediction model of each linear type.

[0026] Using at least one prediction model corresponding to the trained nonlinear type, the issuance volume of the commodity is predicted based on the multi-dimensional historical data, and a third prediction result corresponding to each of the nonlinear types is generated.

[0027] Based on the actual issuance volume corresponding to the multi-dimensional historical data, the second prediction results corresponding to each of the linear prediction models, and the third prediction results corresponding to each of the non-linear prediction models, the weight parameters of the target model are determined.

[0028] The target model is generated based on the weight parameters.

[0029] This application provides an electronic device, including:

[0030] Memory, used to store executable instructions;

[0031] A processor, when executing executable instructions stored in the memory, implements the method provided in the embodiments of this application.

[0032] This application provides a computer-readable storage medium storing a computer program or executable instructions thereon, which, when executed by a processor, implements the method provided in this application.

[0033] The embodiments of this application have the following beneficial effects: Multi-dimensional data involves multiple dimensions of data related to sales, obtaining multi-dimensional data related to the quantity of goods to be predicted to provide a data foundation for subsequent predictions; the target model is a hybrid model composed of prediction models of at least two algorithm types, with at least two algorithm types including linear and non-linear types, thus enabling the use of at least one linear prediction model and at least one non-linear prediction model to jointly obtain the target prediction result; while focusing on the linear relationship between multi-dimensional data and the quantity of goods issued, it can also simultaneously focus on the non-linear relationship between multi-dimensional data and the quantity of goods issued, thereby using the linear relationship to predict conventional changes and using the non-linear relationship to focus on complex and non-linear market dynamics; the target prediction result obtained by combining linear and non-linear algorithms has a smaller gap with the actual quantity issued, improving the accuracy and robustness of prediction, while also improving prediction precision and system response speed, demonstrating extremely high adaptability and prediction efficiency in a volatile market environment, and providing prediction support for different industries and application scenarios. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application;

[0035] Figure 2 This is a flowchart illustrating a method for predicting the issuance volume of a commodity, as provided in an embodiment of this application.

[0036] Figure 3 This is a schematic diagram of a process for predicting issuance volume provided in an embodiment of this application. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0038] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0039] If the application documents contain similar descriptions such as "first / second", the following explanation shall be added: In the following description, the terms "first / second / third" are used only to distinguish similar objects and do not represent a specific order of objects. It is understood that "first / second / third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0040] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0041] In the field of commodity issuance volume prediction, the more mature methods for predicting commodity issuance volume currently mainly focus on traditional statistical analysis and modern intelligent algorithms. However, these methods still have certain limitations in application.

[0042] Traditional statistical methods rely on presupposed linear assumptions and lack the ability to handle nonlinear and complex market phenomena, making them ineffective in efficiently and accurately responding to rapidly changing market environments. Furthermore, these methods often rely solely on historical data for forecasting, neglecting dynamic responses to macroeconomic changes, emerging market trends, and unusual events, thus resulting in insufficient predictive accuracy.

[0043] While modern intelligent algorithms offer more flexible and accurate prediction methods, they also face challenges. First, most machine learning models use intelligent algorithms to predict data, but each model typically employs only one algorithm, leading to low accuracy in the predicted values. Even when using multiple machine learning models to predict the same value, the diversity of these models often results in inconsistent and non-universal predictions, making it difficult to establish a unified prediction framework, consuming significant resources, and maintaining low prediction accuracy.

[0044] To address the aforementioned problems, embodiments of this application provide a method, apparatus, electronic device, and storage medium for predicting the distribution volume of goods, which can improve prediction accuracy. The following describes exemplary applications of the electronic device provided in this application. The electronic device provided in this application can be implemented as various types of user terminals such as laptops, tablets, desktop computers, set-top boxes, and mobile devices (e.g., mobile phones, portable music players, personal digital assistants, dedicated messaging devices, portable gaming devices), or as a server. The following will describe exemplary applications when the device is implemented as a server.

[0045] See Figure 1 , Figure 1 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application, such as... Figure 1 As shown, Figure 1 The illustrated electronic device 100 includes at least one processor 110, a memory 150, at least one network interface 120, and a user interface 130. The various components of the electronic device 100 are coupled together via a bus system 140. It is understood that the bus system 140 is used to implement communication between these components. In addition to a data bus, the bus system 140 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 1 The general labeled all buses as Bus System 140.

[0046] The processor 110 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0047] User interface 130 includes one or more output devices 131 that enable the presentation of media content, including one or more speakers and / or one or more visual displays. User interface 130 also includes one or more input devices 132, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

[0048] The memory 150 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state storage, hard disk drives, optical disk drives, etc. The memory 150 may optionally include one or more storage devices physically located away from the processor 110.

[0049] Memory 150 may include volatile memory or non-volatile memory, or both. Non-volatile memory may be read-only memory (ROM), and volatile memory may be random access memory (RAM). The memory 150 described in this application embodiment is intended to include any suitable type of memory.

[0050] In some embodiments, memory 150 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0051] Operating system 151 includes system programs for handling various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic business functions and handling hardware-based tasks;

[0052] The network communication module 152 is used to reach other computing devices via one or more (wired or wireless) network interfaces 120, exemplary network interfaces 120 including Bluetooth, Wi-Fi, and Universal Serial Bus (USB), etc.

[0053] Presentation module 153 is configured to enable the presentation of information (e.g., a user interface for operating peripheral devices and displaying content and information) via one or more output devices 131 (e.g., a display screen, a speaker, etc.) associated with user interface 130;

[0054] The input processing module 154 is used to detect and translate one or more user inputs or interactions from one or more input devices 132.

[0055] In some embodiments, the apparatus provided in this application can be implemented in software. Figure 1 A predictive device 155 for the quantity of goods issued, stored in memory 150, is shown. This device can be software in the form of programs and plug-ins, and includes the following software modules: an acquisition module 1551 and a first generation module 1552. These modules are logically linked and can therefore be arbitrarily combined or further separated according to their implemented functions. The functions of each module will be described below.

[0056] In other embodiments, the apparatus provided in this application can be implemented in hardware. As an example, the apparatus provided in this application can be a processor in the form of a hardware decoding processor, which is programmed to execute the product distribution prediction method provided in this application. For example, the processor in the form of a hardware decoding processor can adopt one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0057] The present application will be used as a reference to the exemplary application and implementation of the electronic device provided in the embodiments of the present application to illustrate a method for predicting the issuance volume of goods.

[0058] Figure 2 This is a flowchart illustrating a method for predicting the issuance volume of goods provided in an embodiment of this application, which will be combined with... Figure 2 The steps shown illustrate the method for predicting the quantity of goods to be distributed as disclosed in this application, such as... Figure 2 As shown, the method includes:

[0059] Step 201: Obtain multi-dimensional data related to the product issuance volume to be predicted; wherein, the multi-dimensional data includes at least sales data and information on the impact of sales volume.

[0060] In this embodiment of the disclosure, the issuance volume is used to represent the expected quantity of goods to be issued, while the sales volume directly reflects the actual consumer demand for the goods. For example, both the issuance volume and the sales volume represent the quantity of goods sold; the issuance volume represents the expected quantity of goods to be sold, and the sales volume represents the actual quantity of goods sold. The issuance volume may include the sales volume and the quantity of goods not sold, or it may also represent the expected production quantity of the goods, etc. In order to predict the issuance volume of the goods to be predicted, the electronic device can acquire multi-dimensional data in real time.

[0061] Multi-dimensional data can also include the time and data source corresponding to sales data, as well as the time and data source corresponding to influencing information. Sales data should include at least one of the following: actual sales volume, sales revenue, and sales frequency of the product in the past. Influencing information should at least indicate the impact of at least one of the following factors on the sales volume of the product: marketing, unusual events, seasons, holidays, consumers, policies, emerging market trends, and macroeconomic changes. Influencing information includes, but is not limited to, at least one of the following: images, audio, news, videos, and promotional copy that affect sales volume.

[0062] For example, when predicting the sales volume of a product over a period of time, electronic devices acquire sales-related data from multiple dimensions. These multi-dimensional data include, but are not limited to, at least one of the following dimensions: sales, marketing, seasonal and holiday, and user. Specifically, sales-dimensional data includes at least one of the following: historical sales volume, historical sales revenue, and historical sales frequency. Marketing-dimensional data includes at least one of the following: promotional activities, advertising, and price discounts. Seasonal and holiday-dimensional data includes at least one of the following: the impact of seasonal changes, holidays, and special events on product sales; macroeconomic changes; emerging market trends; and dynamic response capabilities to abnormal events. User-dimensional data includes at least one of the following: user browsing history, shopping preferences, and feedback.

[0063] While acquiring data from multiple dimensions, we also need to understand the time frame and data source of this data. Data sources include, but are not limited to, historical sales data, marketing data, seasonal and holiday factors, and consumer behavior data. Let 'd' represent the data from multiple dimensions, 'm' represent the time frame of the data, and 'n' represent the data source. After acquiring the multi-dimensional sales-related data, we integrate this data to obtain multi-dimensional data, which includes 'd'. 11 d 12 ... d mn In this context, the time aspect of multi-dimensional data can be represented by time identifiers, where m represents the time identifier, and each time period or point in time corresponds to a time identifier. The data source of multi-dimensional data can be represented by source identifiers, where n represents the source identifier, and each data source corresponds to a source identifier. For example, in the above example of multi-dimensional data D, d... 23 In the text, "2" is a time marker, indicating the second point in time; "3" is a source marker, indicating seasonal and holiday factors; d 23This indicates data obtained from seasonal and holiday factors at the second point in time.

[0064] Multidimensional data can be represented in matrix form. For example, multidimensional data can be constructed in matrix form according to the time and data source of multiple dimensions. Multidimensional data in matrix form can be represented as follows:

[0065] Step 202: Using the trained target model, predict the distribution volume of the product based on multi-dimensional data to generate a target prediction result; wherein, the target model is a model composed of prediction models of at least two algorithm types, including at least linear and non-linear types, and the prediction model applies either linear or non-linear algorithms, with each algorithm type corresponding to at least one prediction model. The linear prediction model is used to predict the distribution volume of the product from the perspective of numerical calculation, and the non-linear prediction model is used to predict the distribution volume of the product from the perspective of non-linear interaction of data in multi-dimensional data; the prediction model is a model trained with at least multi-dimensional historical data and corresponding actual distribution volume as training samples, and the multi-dimensional historical data includes at least historical sales data and historical sales volume impact information.

[0066] In this embodiment of the disclosure, linear type algorithm refers to linear algorithm, nonlinear type algorithm refers to nonlinear algorithm, and the target model applies linear algorithm and nonlinear algorithm to obtain target prediction results.

[0067] Multidimensional historical data can be used to show the historical multidimensional data of the product to be predicted, or it can be used to show the historical multidimensional data of related products associated with the product to be predicted. For example, if the product to be predicted is a first model of mobile phone that has not been sold, another second model of mobile phone that has been sold can be regarded as a related product associated with the product to be predicted. The multidimensional historical data includes the multidimensional historical data of the second model of mobile phone, and the actual issuance volume includes the actual issuance volume of the second model of mobile phone.

[0068] The actual issuance volume corresponds to multi-dimensional historical data, and the actual issuance volume is used to at least indicate the actual issuance volume of the product based on multi-dimensional historical data.

[0069] Algorithm types can be categorized from different perspectives. Based on whether they are linear or nonlinear algorithms, they can be classified as linear or nonlinear. For example, linear algorithms correspond to linear types, and nonlinear algorithms correspond to nonlinear types; prediction models using linear algorithms correspond to linear types, and prediction models using nonlinear algorithms correspond to nonlinear types. Based on the domain of the target prediction result, algorithms can be classified as those applied to different domains; based on the type of target prediction result, algorithms can be classified as those predicting different types; and based on the function of the algorithm, algorithms can be classified as those with different functions, and so on.

[0070] Linear forecasting models focus more on the relationships between data and are suitable for situations where a clear linear relationship exists. When there is a clear linear relationship between the quantity of goods issued and the multi-dimensional data affecting the quantity of goods issued, linear models can fit the data well. For example, a linear forecasting model may show a linear relationship between price and quantity issued; when the price decreases within a certain range, the quantity issued will increase.

[0071] Linear prediction models, supported by mature mathematical theories and algorithms, consider the linear relationship between multi-dimensional data and product issuance volume, enabling rapid prediction of product issuance volume with high computational efficiency. Furthermore, the prediction results are more accurate when there is a clear linear relationship between product issuance volume and the multi-dimensional data affecting product issuance volume.

[0072] Non-linear predictive models comprehensively consider the impact of multi-dimensional data on product issuance volume, as well as the interactions between these multi-dimensional data. The relationships between these multi-dimensional data, and between each dimension of data and issuance volume, are often complex and non-linear. For example, there may be a non-linear relationship between price and issuance volume. When the price decreases within a certain range, the issuance volume increases, but when the price decreases to a certain extent, consumers may doubt the product quality, leading to a decrease in issuance volume.

[0073] Non-linear forecasting models can better capture this complexity, thus adapting to various types of data and relationships, making them highly adaptable. Furthermore, the more dimensions there are in multi-dimensional data, the better non-linear forecasting models can consider the non-linear relationships between factors such as actual sales volume, sales revenue, sales frequency, marketing, abnormal events, seasons, holidays, consumers, policies, emerging market trends, macroeconomic changes, and the distribution volume of goods, making the forecast results more accurate.

[0074] The target model is a hybrid model composed of prediction models of at least two algorithm types, including at least linear and nonlinear types. It can obtain the target prediction result by using at least one linear prediction model and at least one nonlinear prediction model. While focusing on the linear relationship between multi-dimensional data and the distribution volume of goods, it can also simultaneously focus on the nonlinear relationship between multi-dimensional data and the distribution volume of goods. Thus, the linear relationship can be used to predict conventional changes, while the nonlinear relationship can be used to focus on complex and nonlinear market dynamics. The target prediction result obtained by the combination of linear and nonlinear algorithms has a smaller gap with the actual distribution volume, which improves the accuracy and robustness of the prediction. At the same time, it also improves the prediction accuracy and system response speed. It can exhibit extremely high adaptability and prediction efficiency in a volatile market environment, providing comprehensive support for the prediction of the distribution volume of goods for different industries and application scenarios, thereby improving the competitiveness of enterprises under complex market conditions.

[0075] In some embodiments, the trained target model is used to predict the issuance volume of goods based on multi-dimensional data, generating a target prediction result, including steps 2021 and 2022, wherein:

[0076] Step 2021: Based on multi-dimensional data, predict the issuance volume of goods using at least one prediction model corresponding to the linear type and at least one prediction model corresponding to the non-linear type, and generate the first prediction result corresponding to each prediction model.

[0077] Step 2022: Using the weight parameters of the target model, generate the target prediction result based on the first prediction results corresponding to the prediction models of at least two algorithm types; wherein, the weight parameters of the target model include at least the weights corresponding to each prediction model.

[0078] In this embodiment of the disclosure, a trained linear prediction model is used to predict the issuance volume of goods over a period of time based on multi-dimensional data, generating a first prediction result corresponding to the linear prediction model; a trained non-linear prediction model is also used to predict the issuance volume of goods over a period of time based on multi-dimensional data, generating a first prediction result corresponding to the non-linear prediction model. The weight parameters of the target model include at least the parameters corresponding to the prediction models of different algorithm types in the target model. For example, the weight parameters of the target model include the weight coefficients corresponding to the prediction models of different algorithm types. The electronic device performs a weighted summation based on the weight coefficients corresponding to the prediction models of different algorithm types and the first prediction results corresponding to the prediction models of different algorithm types to obtain the target prediction result.

[0079] In some embodiments, the issuance volume of goods is predicted based on multi-dimensional data using at least one prediction model corresponding to linear type and at least one prediction model corresponding to non-linear type, generating a first prediction result corresponding to each prediction model, including: normalizing the multi-dimensional data to generate first feature data; and generating a first prediction result corresponding to each prediction model based on the input first feature data.

[0080] In this embodiment of the disclosure, the normalization process includes feature extraction processing, and also includes at least one of the following processing methods: standardization processing, normalization processing, and interpolation processing. The electronic device performs normalization processing on multi-dimensional data to extract features from the normalized multi-dimensional data according to multiple dimensions, and then constructs first feature data based on the extracted features from multiple dimensions. This enables the extraction of features corresponding to each dimension from multi-dimensional data, reducing the amount of data processing required for the subsequent target model.

[0081] In some embodiments, normalizing multi-dimensional data to generate first feature data includes: normalizing the multi-dimensional data to generate first normalized data; extracting features from the first normalized data using a feature extraction function; and generating the first feature data based on the extracted features.

[0082] In this embodiment of the disclosure, the value range or unit of different data may be different. In order to avoid the problem of inaccurate prediction caused by the value range and unit of data, it is necessary to standardize the data so that the standardized data adopts the same value range and reduces the influence of the unit on the data.

[0083] For numerical values ​​in multidimensional data, the electronic device calculates the mean μ and standard deviation σ based on the multidimensional data; using a standardization formula, it calculates the difference between any data point in the multidimensional data and the mean μ, and generates standardized multidimensional data d based on the ratio of the difference to the standard deviation σ. ′ The standardized formula can be expressed as: Standardized multidimensional data d ′ It follows a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0084] For example, an electronic device processes the first data in the same dimension with the same value range and the same unit in multi-dimensional data, calculates the mean μ and standard deviation σ based on multiple first data, and calculates the standardized first data through a standardization formula.

[0085] Electronic devices generate first standardized data S(D) based on standardized multi-dimensional data. The values ​​in the first standardized data S(D) have the same range of values; for example, the range of standardized values ​​is greater than or equal to 0 and less than or equal to 1, thus ensuring data consistency and comparability. The first standardized data can be represented in matrix form.

[0086] The feature extraction function is used to extract a first feature from the first standardized data. For example, feature extraction can be performed based on the data source or dimensions of multi-dimensional data to extract the first features of the product from different data sources or different dimensions.

[0087] In some embodiments, the multi-dimensional data includes at least one of sales data, marketing data, and seasonal and holiday factor data. When the multi-dimensional data includes sales data, marketing data, and seasonal and holiday factor data, feature extraction functions are used to extract features from the first standardized data, and first feature data is generated based on the extracted features. This includes: extracting sales trend features from sales data using a first feature extraction function; extracting market pattern features from marketing data using a second feature extraction function; extracting social influence features from seasonal and holiday factor data using a third feature extraction function; and generating the first feature data based on the sales trend features, market pattern features, and social influence features.

[0088] In this embodiment of the disclosure, sales data is used at least to indicate the issuance volume of goods in different time periods or points in time, and sales trend features are used at least to indicate the trend and / or fluctuation of sales data. The sales trend features include at least trend features and / or fluctuation features.

[0089] The first feature extraction function is used at least to indicate the extraction of trends and / or fluctuations in sales data. This first function includes at least the functions applied by time series analysis methods. The electronic device uses these functions to analyze whether the sales trend of a product over a period of time shows an increasing or decreasing trend, and uses this increasing or decreasing trend as a trend feature. For example, it can analyze the long-term trend of a product's distribution volume since its launch to determine if it shows an increasing or decreasing trend. The electronic device can also use the functions applied by time series analysis methods to analyze whether the sales trend of a product over a period of time shows a fluctuating trend, and use this fluctuating trend as a fluctuation feature. For example, the distribution volume of a product at the beginning and end of the month is always lower than the distribution volume in the middle of the month, and the monthly distribution volume always shows an increasing trend first, then a decreasing trend. The electronic device extracts the fluctuation trend of the distribution volume according to a periodic pattern from the sales trend.

[0090] Marketing data must include at least multi-source marketing data such as marketing time, marketing platform, and marketing content. Market pattern features must at least be used to indicate the marketing activity pattern. A second feature extraction function must at least be used to extract the marketing activity pattern from the marketing data. This second feature extraction function must at least include functions applied by Principal Component Analysis (PCA). Electronic devices, through the functions applied by PCA, reduce the dimensionality of multi-source marketing data, extract the main marketing activity patterns, and use the extracted main marketing activity patterns as market pattern features.

[0091] Seasonal and holiday factor data must include at least social media data, and social influence features must be used to indicate consumer sentiment index. A third feature extraction function must be used to extract the consumer sentiment index from the seasonal and holiday factor data. This third feature extraction function must include at least the function applied by the natural language processing algorithm. Electronic devices analyze social media data using the function applied by the natural language processing algorithm to extract consumers' willingness to purchase goods, and the extracted consumer willingness to purchase goods is used as the social influence feature.

[0092] The feature extraction function includes at least one of a first feature extraction function, a second feature extraction function, and a third feature extraction function. The first feature includes at least one of a sales trend feature, a market pattern feature, and a social influence feature. The feature extraction function enables comprehensive analysis of multi-dimensional data, thereby deeply mining the first feature influencing product distribution volume. When performing feature extraction, the first feature can be extracted from multi-dimensional data, or it can be extracted from first standardized data. The process of extracting features from first standardized data will not be elaborated here.

[0093] In some embodiments, the electronic device performs Figure 2 Before the method shown, the target model needs to be trained. This method also includes steps 301 and 302, wherein:

[0094] Step 301: Using multi-dimensional historical data and the corresponding actual issuance volume as training samples, train the initial prediction model of at least two algorithm types to generate the trained prediction model of at least two algorithm types.

[0095] Step 302: Generate the trained target model based on the prediction models of at least two algorithm types that have been trained.

[0096] In this embodiment of the disclosure, before using the target model for prediction, it is necessary to train the target model with training samples so that the target model learns the relationship between multi-dimensional historical data and actual issuance volume.

[0097] Multi-dimensional historical data should include at least historical sales data and information on the impact of historical sales volume. Multi-dimensional historical data may also include the time and data source corresponding to the historical sales data, as well as the time and data source corresponding to the impact information. Historical sales data should include at least one of the following: actual sales volume, sales revenue, and sales frequency. Information on the impact of historical sales volume should at least indicate the influence of at least one of the following factors on historical sales volume: marketing, unusual events, seasons, holidays, consumers, policies, emerging market trends, and macroeconomic changes. Information on the impact of historical sales volume includes, but is not limited to, at least one of the following: images, audio, news, videos, and promotional copy that influence sales volume.

[0098] The target model consists of at least one linear prediction model and at least one nonlinear prediction model. Multidimensional historical data and actual issuance volume are used as training samples for training the model. Electronic devices can train at least one initial linear prediction model and at least one initial nonlinear prediction model separately or simultaneously using multidimensional historical data and actual issuance volume to obtain at least two algorithmic prediction models. The target model is then generated from the at least two algorithmic prediction models.

[0099] For example, at least one initial prediction model of linear type and at least one initial prediction model of nonlinear type can be trained separately, and then the trained at least one prediction model of linear type and at least one prediction model of nonlinear type can be mixed into the target model. Alternatively, at least one initial prediction model of linear type and at least one initial prediction model of nonlinear type can be concatenated first, and then the concatenated at least one initial prediction model and at least one initial prediction model of nonlinear type can be trained simultaneously, and the trained at least one prediction model of linear type and at least one prediction model of nonlinear type can be mixed into the target model.

[0100] When training the initial prediction model, two types of initial prediction models are first constructed: linear and non-linear. The linear initial prediction model includes models using linear algorithms. These models are used to capture the linear relationship between multi-dimensional historical data and actual data. Linear initial prediction models include generalized linear models, such as the General Language Model (GLM). The non-linear initial prediction model includes models using non-linear algorithms. These models are used to capture the non-linear relationship between multi-dimensional historical data and actual data. Examples of non-linear initial prediction models include Random Forest or Support Vector Machines (SVM).

[0101] After the initial predictive model is built, the parameters required for it are not yet determined. Bayesian optimization is a strategy for global optimization of functions, particularly suitable for computationally expensive black-box functions, such as hyperparameter tuning of machine learning models. Bayesian optimization can be used to adjust model hyperparameters, thereby quickly and effectively finding the optimal solution for the parameters required for the initial predictive model, greatly improving the efficiency of model training and optimization, and making it suitable for application scenarios that require rapid response to market changes.

[0102] When forecasting issuance volume, multi-dimensional historical data should be used to represent historical data across at least several dimensions related to sales, and actual issuance volume should be used to represent data related to the actual issuance volume. Multi-dimensional historical data corresponds to actual issuance volume.

[0103] By using multi-dimensional historical data and actual issuance volume as training samples, initial prediction models of different algorithm types learn the relationship between these two data points. Linear initial prediction models learn the linear relationship through linear algorithms, while non-linear initial prediction models learn the non-linear relationship through non-linear algorithms. The target model integrates linear and non-linear algorithms, considering the relationship between multi-dimensional historical data and actual issuance volume from both linear and non-linear perspectives, thus increasing the model's robustness. Furthermore, linear algorithms can handle predictable and routine changes, while non-linear algorithms focus on complex and non-linear market dynamics. The target model comprehensively considers both routine changes and complex market dynamics, resulting in target predictions that better match actual market issuance volume, thus improving prediction accuracy.

[0104] Alternatively, multi-dimensional data can be used as multi-dimensional historical data, while simultaneously obtaining the actual issuance volume corresponding to the target prediction result. The above steps can be performed to train initial prediction models for at least two algorithm types based on multi-dimensional historical data and actual issuance volume, generating trained prediction models for at least two algorithm types. Based on the trained prediction models for at least two algorithm types, a trained target model can be generated. This allows for obtaining the actual issuance volume after each prediction, retraining the target model, and timely optimization. Alternatively, before optimizing the target model, a difference value can be generated based on the target prediction result and the actual issuance volume. If the difference value is within a difference threshold range, the target model is not optimized; if the difference value is outside the difference threshold range, the target model is optimized. This allows for determining whether the target model's prediction result closely matches reality based on actual data, and timely optimization of the target model when it does not.

[0105] In some embodiments, multi-dimensional historical data and corresponding actual issuance volume are used as training samples to train initial prediction models of at least two algorithm types, generating trained prediction models of at least two algorithm types, including steps 3011 to 3014, wherein:

[0106] Step 3011: Train the initial prediction model of at least two algorithm types based on the first multi-dimensional historical data, the first actual issuance volume and the preparatory weight parameters of the target model, and generate the prediction model of at least two algorithm types.

[0107] Step 3012: Using at least two prediction models of different algorithm types, based on the second multi-dimensional historical data and the acquired target model data, generate the fourth prediction result corresponding to each prediction model;

[0108] Step 3013: Generate the first target prediction result based on the target model data and the fourth prediction results corresponding to each prediction model;

[0109] Step 3014: Based on the first target prediction result and the second actual issuance volume, determine the preliminary weight parameters of the target model as the weight parameters of the target model, and use the generated prediction models of at least two algorithm types as the prediction models of at least two algorithm types after training; or, based on the first target prediction result and the second actual issuance volume, update the preliminary weight parameters of the target model, and perform the step of training the initial prediction models of at least two algorithm types based on the first multi-dimensional historical data, the first actual issuance volume and the obtained preliminary weight parameters of the target model to generate the prediction models of at least two algorithm types after training.

[0110] In this embodiment of the disclosure, prediction models of at least two algorithm types can be concatenated first, and then the initial prediction models of the concatenated at least two algorithm types can be trained simultaneously. The preliminary weight parameters of the target model can be enumerated using enumeration or grid search, and these preliminary weight parameters at least include the weight coefficients corresponding to each prediction model in the target model.

[0111] First, at least one prediction model of linear type and at least one prediction model of nonlinear type can be concatenated. Then, at least one initial prediction model after concatenation and at least one initial prediction model of nonlinear type can be trained simultaneously. Based on the trained at least one prediction model of linear type and at least one prediction model of nonlinear type, the target model can be formed by mixing them.

[0112] The training set consists of the target model's pre-set weight parameters, the first multi-dimensional historical data, and the first actual issuance volume. The test set consists of the target model's pre-set weight parameters, the second multi-dimensional historical data, and the second actual issuance volume. First, using a fusion formula, the proportion of the first actual issuance volume is calculated based on the target model's pre-set weight parameters to determine the third actual issuance volume corresponding to each prediction model.

[0113] The fusion formula can be expressed as: Among them, Y final Y represents the actual issuance volume, α represents the weighting coefficient corresponding to the linear prediction model, and Y represents the actual issuance volume. linear Y represents the fourth prediction result corresponding to the linear prediction model. nonlinear This represents the fourth prediction result corresponding to the non-linear prediction model. For example, the weight coefficient for the linear prediction model is 0.4, the weight coefficient for the non-linear prediction model is 0.6, and the first actual issuance is 5. Using the above fusion formula, αY is obtained. linear Equals 4, (1-α)×Y nonlinear Equals 6, thus yielding Y linear Equals 10, Y nonlinear The value is 10. 10 can be used as the third actual issuance amount corresponding to the linear prediction model and the third actual issuance amount corresponding to the nonlinear prediction model. When the effect of using the pre-required weight parameters of the target model to predict the issuance amount of the test set is poor, the pre-required weight parameters of the next target model can be used as the training set to train the prediction model.

[0114] Then, based on the first multi-dimensional historical data and the third actual issuance volume corresponding to each prediction model, the prediction models are trained to generate prediction models.

[0115] Then, based on the prediction model trained this time, the distribution volume of the test set is predicted to obtain the fourth prediction result. Then, based on the preliminary weight parameters of the target model, the fourth prediction result is weighted and summed using the above fusion formula to generate the first target prediction result. The difference between the first target prediction result and the second actual distribution volume is calculated. If the difference value is within the difference threshold range, the preliminary weight parameters of the target model can be used as the weight parameters of the target model. Step 302 includes: obtaining the target model based on the weight parameters of the target model, at least one prediction model corresponding to the linear type after training, and at least one prediction model corresponding to the nonlinear type after training. If the difference value is not within the difference threshold range, the next preliminary weight parameter can be obtained by enumeration or grid search, and retraining can be performed.

[0116] When training prediction models of at least two algorithm types based on first multi-dimensional historical data, first actual issuance data, and the pre-set weight parameters of the target model, and generating prediction models of at least two algorithm types, one part of the prediction models of the at least two algorithm types can be the trained models, and the other part of the prediction models can be the untrained models. The untrained prediction models and the trained prediction models are trained using the first multi-dimensional historical data, the first actual issuance data, and the pre-set weight parameters.

[0117] In some embodiments, initial prediction models for at least two algorithm types are trained based on multi-dimensional historical data and actual issuance volume to generate trained prediction models for at least two algorithm types, including steps 401 and 402, wherein:

[0118] Step 401: Standardize the multi-dimensional historical data to generate second feature data;

[0119] Step 402: Based on the second feature data and the actual issuance volume, train the initial prediction model for at least two algorithm types to generate the trained prediction model for at least two algorithm types.

[0120] In this embodiment of the disclosure, the normalization process includes feature extraction processing, and also includes at least one of the following processing methods: standardization processing, normalization processing, and interpolation processing. The electronic device performs normalization processing on multi-dimensional historical data to extract features from the normalized multi-dimensional historical data according to multiple dimensions, and then uses the extracted features from multiple dimensions to form second feature data. This enables the extraction of features corresponding to each dimension from the multi-dimensional historical data, reducing the amount of data required for subsequent target model training.

[0121] In some embodiments, the multi-dimensional historical data is normalized to generate second feature data, including steps 4011 and 4012, wherein:

[0122] Step 4011: Standardize the multi-dimensional historical data to generate second standardized data;

[0123] Step 4012: Use the feature extraction function to extract features from the second standardized data, and generate second feature data based on the extracted features.

[0124] In this embodiment of the disclosure, the steps for normalizing multi-dimensional historical data can refer to the steps for normalizing multi-dimensional data described above, and will not be repeated here.

[0125] In some embodiments, multi-dimensional historical data and corresponding actual issuance volume are used as training samples to train initial prediction models of at least two algorithm types, generating trained prediction models of at least two algorithm types, including steps 301a to 301c:

[0126] Step 301a: Generate second feature data based on multi-dimensional historical data; wherein the second feature data includes at least the features of the multi-dimensional historical data;

[0127] Step 301b: Using the first algorithm, at least one initial prediction model corresponding to the linear type is trained based on the second feature data and the actual issuance volume to generate at least one prediction model corresponding to the linear type after training; wherein, the first algorithm is at least used to fit and optimize the parameters of the prediction model of the linear type through linear analysis;

[0128] Step 301c: Using the second algorithm, at least one initial prediction model corresponding to the nonlinear type is trained based on the second feature data and the actual issuance volume to generate at least one prediction model corresponding to the nonlinear type after training; wherein, the second algorithm is at least used to optimize the parameters of the prediction model of the nonlinear type through nonlinear analysis.

[0129] In this embodiment of the disclosure, each trained prediction model is obtained by training the initial prediction model separately.

[0130] Secondary features that can influence issuance volume are extracted from multi-dimensional historical data. These secondary features must at least indicate characteristics related to sales or influencing sales, and must include at least one of the following: sales trend features, market pattern features, or social influence features. The secondary feature data is then composed of these secondary features. This secondary feature data, along with the actual issuance volume, serves as the training set for training both linear and non-linear initial prediction models.

[0131] Linear prediction models are based on a model formula that uses a linear algorithm to learn the linear relationship between the second feature data and the actual issuance volume. The corresponding model formula for a linear prediction model is: Among them, Y linear This represents a linear prediction model, or a second prediction result corresponding to a linear prediction model; β0 represents the intercept, β... i The characteristic coefficients of the second characteristic; f iLet represent the second feature, where 'i' represents the i-th second feature in the second feature data, and 'k' represents the number of second features in the second feature data. The model formula for a linear prediction model represents the model continuously learning the linear relationship between multi-dimensional historical data and actual data.

[0132] The first algorithm can be used to find approximate solutions for the linear model parameters of a linear prediction model through linear analysis, in order to obtain the best-fit parameters. For example, the first algorithm includes the least squares method, which is commonly used in linear analysis to find approximate solutions and obtain the best-fit parameters; the linear model parameters include the feature coefficients β of the second feature. i When training a linear prediction model using multiple second feature data, the electronic device uses the least squares method to fit the feature coefficients β of the second features. i The characteristic coefficients β of the fitted second feature are obtained. i .

[0133] Non-linear prediction models are based on a model formula and learn the non-linear relationship between the second feature data and the actual issuance volume through a non-linear algorithm. The model formula corresponding to the non-linear prediction model is Y. nonlinear =Model nonlinear (F); where Y nonlinear It can also represent a non-linear prediction model, or a third prediction result corresponding to a non-linear prediction model. nonlinear (F) indicates the nonlinear algorithm used in the prediction model of the nonlinear type.

[0134] The second algorithm is used at least to optimize the nonlinear model parameters of the nonlinear prediction model. Depending on the characteristics of the nonlinear prediction model, different second algorithms can be applied. For example, when the nonlinear prediction model is a random forest, the second algorithm includes the algorithms used in decision tree construction and node splitting methods.

[0135] In some embodiments, a trained target model is generated based on prediction models of at least two algorithm types after training, including steps 3021 to 3024, wherein:

[0136] Step 3021: Using at least one prediction model corresponding to the trained linear type, predict the issuance volume of the product based on multi-dimensional historical data, and generate the second prediction result corresponding to the prediction model of each linear type.

[0137] Step 3022: Using at least one prediction model corresponding to the trained nonlinear type, predict the issuance volume of the product based on multi-dimensional historical data, and generate the third prediction result corresponding to the prediction model of each nonlinear type.

[0138] Step 3023: Determine the weight parameters of the target model based on the actual issuance volume corresponding to the multi-dimensional historical data, the second prediction results corresponding to each linear prediction model, and the third prediction results corresponding to each non-linear prediction model.

[0139] Step 3024: Generate the target model based on the weight parameters.

[0140] In this embodiment, multi-dimensional historical data is used as a test set to provide data for testing the linear prediction model. The multi-dimensional historical data used as the test set may be different from or the same as the multi-dimensional historical data used as the training set. The weight parameters of the target model can be determined using a third algorithm, which is used at least to optimize the weight parameters of the target model. The third algorithm includes at least one of the following algorithms: cross-validation algorithm, grid search algorithm, etc. The cross-validation algorithm involves constructing multiple test sets to obtain multiple preliminary model parameters, and then determining the weight parameters of the target model based on these preliminary model parameters. The cross-validation algorithm can be a median algorithm, a mean algorithm, etc., and this disclosure does not limit the algorithm used for cross-validation. The grid search algorithm is used at least to enumerate multiple preliminary model parameters, obtain the difference between the target prediction result and the actual issuance volume corresponding to each preliminary model parameter, and generate the weight parameters of the target model based on the preliminary model parameter corresponding to the smallest difference value. The weight parameters of the target model include at least the weight coefficients corresponding to the linear prediction model. In determining the weight parameters of the target model, multiple determinations and evaluations are usually performed to determine the optimal solution for the weight coefficients corresponding to the linear prediction model using the third algorithm.

[0141] For example, the fusion formula corresponding to the target model can be expressed as: Among them, Y final Y represents the actual data, α represents the weighting coefficients corresponding to the linear prediction model, and Y represents the actual data. linear Y represents the second prediction result corresponding to the linear prediction model. nonlinear This represents the third prediction result corresponding to the nonlinear prediction model. Using the aforementioned fusion formula, a weighted average is calculated between the second prediction result corresponding to the linear prediction model and the third prediction result corresponding to the nonlinear prediction model. The weight coefficients corresponding to the linear prediction model are obtained using known actual data, thereby determining the weight parameters of the target model.

[0142] When generating the target model based on the weight parameters of the target model, the process involves using the target model and multi-dimensional historical data to generate a second target prediction result, generating an error value based on the second target prediction result and the actual data, generating the target model if the error value is within the error range, or, if the error value is outside the error range, executing a second prediction result corresponding to each linear prediction model by using a trained linear prediction model and multi-dimensional historical data to predict the distribution volume of the goods. The error value is used to at least indicate the difference between the second target prediction result and the actual data; for example, the error value is the mean squared error (MSE). The multi-dimensional historical data in this section may differ from the multi-dimensional historical data used as the test set and the multi-dimensional historical data used as the training set mentioned above.

[0143] Alternatively, when generating the target model based on the weight parameters of the target model, the target model is used to generate a second target prediction result and a corresponding confidence level based on multi-dimensional historical data; the target model is then generated based on the confidence level falling within the confidence range. The confidence level is used to at least indicate the credibility or accuracy of the second target prediction result. Generating the confidence level includes: using a linear prediction model and multi-dimensional historical data to generate a second prediction result and a first confidence level corresponding to the linear prediction model; using a non-linear prediction model and multi-dimensional historical data to generate a third prediction result and a second confidence level corresponding to the non-linear prediction model; generating a second target prediction result based on the weight parameters of the target model, the second prediction result corresponding to the linear prediction model, and the third prediction result corresponding to the non-linear prediction model; and generating a confidence level based on the weight parameters of the target model, the first confidence level, and the second confidence level.

[0144] When applying the target model, Y final Y represents the target prediction result, α represents the weight coefficient corresponding to the linear prediction model, and Y represents the target prediction result. linear Y represents the first prediction result corresponding to the linear prediction model. nonlinear This represents the first prediction result corresponding to the nonlinear prediction model. Therefore, when applying the target model, the first prediction result corresponding to the linear prediction model and the first prediction result corresponding to the nonlinear prediction model can be mixed to make the target prediction result obtained after back-mixing more accurate and improve the prediction accuracy.

[0145] This application provides only an example of a fusion formula. The fusion formula can also be other formulas that can obtain the weight coefficients corresponding to linear prediction models and / or the weight coefficients corresponding to nonlinear prediction models. This disclosure does not limit the fusion formula.

[0146] The following describes an exemplary application of the embodiments of this application in a real-world sales scenario. When selling a product, it is necessary to first purchase goods from a supplier before proceeding with the sale. Purchasing goods from a supplier requires determining the quantity to be purchased. To avoid stockpiling or insufficient supply, merchants often predict the quantity of goods to be issued and then purchase goods based on that prediction. However, accurately predicting the quantity to be issued has become a pressing problem.

[0147] In some embodiments, Figure 3 This is a schematic diagram of a process for predicting issuance volume provided in an embodiment of this application, such as... Figure 3 As shown, steps 31 to 34 are included, wherein:

[0148] Step 31, Data Collection: Collect multi-source historical data, including sales data, social media data, etc., and represent the multi-source historical data in matrix form;

[0149] Step 32, Data Processing: Through data normalization and feature extraction, based on multi-source historical data, the second features of the multi-source historical data are obtained, and the second features are composed of matrix-represented second feature data.

[0150] Step 33, Model Training: Based on the second feature data and the actual issuance volume, train the target model consisting of a linear prediction model and a non-linear prediction model to obtain the weight parameters of the optimized target model, and obtain the target model based on the weight parameters of the target model.

[0151] Step 34, Model Application: Using the target model, predict the issuance volume of goods based on multi-source data to obtain the target prediction result of the issuance volume of goods.

[0152] In this embodiment, steps 31 to 34 can be executed cyclically. After obtaining the target prediction result through multi-source data, the actual issuance volume corresponding to the target prediction result can be obtained; the multi-source data is used as multi-source historical data; the model is trained again based on the issuance volume corresponding to the target prediction result and the multi-source data, so that prediction can be made in real time based on the target model, and the target model can be optimized again after prediction, thereby improving the accuracy of prediction.

[0153] Electronic devices collect multi-source data related to sales. To respond efficiently and accurately to the rapidly changing market environment, multi-source data can include macroeconomic changes, emerging market trends, anomalies, real-time market changes, historical sales data, etc. Multi-source data is composed of the data sources, the collection time of the data, and the sources of the data.

[0154] Standardizing multi-source historical data unifies the value range and format, reducing data clutter caused by differing value ranges and units. Then, from the standardized multi-source historical data, secondary features are extracted from multiple dimensions. These extracted secondary features are then combined to form secondary feature data, enabling analysis of data relationships and identification of sales-related patterns or characteristics. Based on these patterns or characteristics, sales-related features can be derived, such as summarizing peak sales times each month, day, and year, product distribution trends, user preferences, and current consumption trends. By extracting secondary features for each dimension from multi-dimensional multi-source historical data, the amount of data required for training the subsequent target model is reduced.

[0155] The second feature data and the actual issuance volume are used, either separately or simultaneously, to train both linear and nonlinear prediction models. The second feature is used to indicate at least the characteristics related to sales and influencing sales, and includes at least sales trend characteristics, market pattern characteristics, or social influence characteristics. The second feature data includes at least one of these three characteristics. A linear prediction model (such as GLM) is used to capture the linear relationship between the feature and the product issuance volume, while a nonlinear prediction model is used to capture the nonlinear relationship between the feature and the target variable. This yields the prediction results for both the linear and nonlinear prediction models. Based on these two prediction results and the actual issuance volume, the weight coefficients for the linear prediction model are obtained, resulting in the trained linear and nonlinear prediction models. The weight coefficients are weight parameters of the target model, indicating the proportion of the linear prediction model's prediction result in the target prediction result. Meanwhile, since the target model is a hybrid of linear and nonlinear prediction models, it considers both the linear and nonlinear relationships between multi-source data and issuance volume. This reduces the number of training samples required to train the target model compared to training a single prediction model, thus reducing the amount of data needed for training and resource consumption. This allows for rapid training of the target model and improves prediction accuracy.

[0156] The trained target model is applied to predict the issuance volume of goods after a certain period of time based on multi-source data, thus obtaining the target prediction result. The multi-source data consists of the data required to predict the issuance volume of goods after a certain period. After predicting the target result, the actual issuance volume can be obtained, and the multi-source data can be used as multi-source historical data to execute step 31, thereby enabling timely updates to the target model.

[0157] The following description continues to illustrate the exemplary structure of the commodity issuance volume prediction device 155 provided in the embodiments of this application as a software module. In some embodiments, such as Figure 1 As shown, the software modules in the product issuance prediction device 155 stored in the memory 140 may include: an acquisition module 1551, used to acquire multi-dimensional data related to the product issuance volume to be predicted; wherein the multi-dimensional data includes at least sales data and sales volume impact information; a first generation module 1552, used to predict the product issuance volume based on the multi-dimensional data using a trained target model, and generate a target prediction result; wherein the target model is a model composed of prediction models of at least two algorithm types, the at least two algorithm types including at least linear and nonlinear types, the prediction model applies a linear algorithm or a nonlinear algorithm, each algorithm type corresponds to at least one prediction model, the linear prediction model is used to predict the product issuance volume from the perspective of numerical calculation, and the nonlinear prediction model is used to predict the product issuance volume from the perspective of nonlinear interaction of data in the multi-dimensional data; the prediction model is a model trained with at least multi-dimensional historical data and corresponding actual issuance volume as training samples, the multi-dimensional historical data including at least historical sales data and historical sales volume impact information.

[0158] In some possible implementations, the first generation module 1552 is used to predict the issuance volume of goods based on multi-dimensional data using at least one prediction model corresponding to the linear type and at least one prediction model corresponding to the nonlinear type, and generate the first prediction result corresponding to each prediction model.

[0159] Using the weight parameters of the target model, a target prediction result is generated based on the first prediction results corresponding to at least two prediction models of different algorithm types; wherein, the weight parameters of the target model include at least the weights corresponding to each prediction model.

[0160] In some possible implementations, the first generation module 1552 is used to perform normalization processing on multi-dimensional data to generate first feature data;

[0161] For each prediction model, based on the first feature data input, a first prediction result is generated for each prediction model.

[0162] In some possible implementations, the first generation module 1552 is used to standardize the multi-dimensional data to generate first standardized data;

[0163] The first standardized data is used to extract features using a feature extraction function, and the first feature data is generated based on the extracted features.

[0164] In some possible implementations, the second generation module is used to train the initial prediction model of at least two algorithm types using multi-dimensional historical data and the corresponding actual issuance volume as training samples, and generate the trained prediction model of at least two algorithm types.

[0165] Generate a trained target model based on prediction models of at least two algorithm types that have been trained.

[0166] In some possible implementations, the second generation module is used to generate second feature data based on multi-dimensional historical data; wherein the second feature data includes at least the features of the multi-dimensional historical data;

[0167] Using a first algorithm, at least one initial prediction model corresponding to the linear type is trained based on the second feature data and the actual issuance volume to generate at least one prediction model corresponding to the linear type after training; wherein, the first algorithm is at least used to fit and optimize the parameters of the prediction model of the linear type through linear analysis;

[0168] Using a second algorithm, at least one initial prediction model corresponding to the nonlinear type is trained based on the second feature data and the actual issuance volume, generating at least one prediction model corresponding to the nonlinear type after training; wherein, the second algorithm is used at least to optimize the parameters of the prediction model of the nonlinear type through nonlinear analysis.

[0169] In some possible implementations, the second generation module is used to predict the issuance volume of goods based on multi-dimensional historical data using at least one prediction model corresponding to the trained linear type, and generate a second prediction result corresponding to the prediction model of each linear type.

[0170] Using at least one prediction model corresponding to the trained nonlinear type, the issuance volume of goods is predicted based on multi-dimensional historical data, and a third prediction result corresponding to each nonlinear type prediction model is generated.

[0171] Based on the actual issuance volume corresponding to multi-dimensional historical data, the second prediction results corresponding to each linear prediction model, and the third prediction results corresponding to each non-linear prediction model, the weight parameters of the target model are determined.

[0172] The target model is generated based on the weight parameters.

[0173] In some possible implementations, the second generation module is used to normalize the multi-dimensional historical data to generate second feature data;

[0174] Based on the second feature data and the actual issuance volume, train the initial prediction model for at least two algorithm types to generate the trained prediction model for at least two algorithm types.

[0175] In some possible implementations, the second generation module is used to standardize the multi-dimensional historical data to generate second standardized data;

[0176] The second standardized data is used to extract features using a feature extraction function, and the extracted features are used to generate the second feature data.

[0177] This application provides a computer program product comprising a computer program or executable instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer program or executable instructions from the computer-readable storage medium and executes the computer program or executable instructions, causing the electronic device to perform the product distribution prediction method described above in this application.

[0178] This application provides a computer-readable storage medium storing a computer program or executable instructions. When the computer program or executable instructions are executed by a processor, the processor will execute the product distribution prediction method provided in this application. For example, ... Figure 2 The method for predicting the volume of goods issued is shown.

[0179] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EP ROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or it may be a variety of devices including one or any combination of the above-mentioned memories.

[0180] In some embodiments, executable instructions may take the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0181] As an example, executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a Hyper Text Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple collaborating files (e.g., a file that stores one or more modules, subroutines, or code sections).

[0182] As an example, executable instructions can be deployed to execute on a single computing device, or on multiple computing devices located in one location, or on multiple computing devices distributed across multiple locations and interconnected via a communication network.

[0183] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. A method for predicting the issuance volume of a commodity, characterized in that, include: Acquire multi-dimensional data related to the issuance volume of the product to be predicted; wherein, the multi-dimensional data includes at least sales data and information on the impact of sales volume; Using the trained target model, the issuance volume of the product is predicted based on the multi-dimensional data, and a target prediction result is generated. The target model is a model composed of prediction models combining at least two algorithm types, including at least linear and nonlinear types. The prediction model applies either the linear or nonlinear algorithm, with each algorithm type corresponding to at least one prediction model. The linear prediction model is used to predict the distribution volume of the product from a numerical calculation perspective, while the nonlinear prediction model is used to predict the distribution volume of the product from the perspective of nonlinear interaction of data in multi-dimensional data. The prediction model is a model trained using at least multi-dimensional historical data and corresponding actual distribution volumes as training samples. The multi-dimensional historical data includes at least historical sales data and information on the impact of historical sales volumes.

2. The method according to claim 1, characterized in that, The step of using the trained target model to predict the distribution volume of goods based on the multi-dimensional data and generating a target prediction result includes: Based on the multi-dimensional data, the issuance volume of the commodity is predicted by at least one prediction model corresponding to the linear type and at least one prediction model corresponding to the nonlinear type, and the first prediction result corresponding to each prediction model is generated. The target prediction result is generated based on the first prediction results corresponding to the prediction models of the at least two algorithm types, using the weight parameters of the target model; wherein the weight parameters of the target model include at least the weights corresponding to each prediction model.

3. The method according to claim 2, characterized in that, The step of predicting the distribution volume of goods based on the multi-dimensional data using at least one prediction model corresponding to the linear type and at least one prediction model corresponding to the nonlinear type, and generating a first prediction result corresponding to each prediction model, includes: The multi-dimensional data is normalized to generate the first feature data; Based on the input first feature data, each prediction model generates a first prediction result corresponding to each prediction model.

4. The method according to claim 3, characterized in that, The step of normalizing the multi-dimensional data to generate first feature data includes: The multi-dimensional data is standardized to generate first standardized data; The first standardized data is subjected to feature extraction using a feature extraction function, and the first feature data is generated based on the extracted features.

5. The method according to claim 1, characterized in that, The method further includes: Using multi-dimensional historical data and corresponding actual issuance volume as training samples, the initial prediction models of the at least two algorithm types are trained to generate the trained prediction models of the at least two algorithm types. Based on the trained prediction models of the at least two algorithm types, the trained target model is generated.

6. The method according to claim 5, characterized in that, The step of training the initial prediction models of the at least two algorithm types using multi-dimensional historical data and corresponding actual issuance volume as training samples, and generating trained prediction models of the at least two algorithm types, includes: Second feature data is generated based on the multi-dimensional historical data; wherein, the second feature data includes at least the features of the multi-dimensional historical data; Using a first algorithm, at least one initial prediction model corresponding to the linear type is trained based on the second feature data and the actual issuance volume to generate at least one prediction model corresponding to the linear type after training; wherein, the first algorithm is at least used to fit and optimize the parameters of the prediction model of the linear type through linear analysis; Using the second algorithm, at least one initial prediction model corresponding to the nonlinear type is trained based on the second feature data and the actual issuance volume to generate at least one prediction model corresponding to the nonlinear type after training; wherein, the second algorithm is at least used to optimize the parameters of the prediction model of the nonlinear type through nonlinear analysis.

7. The method according to claim 5, characterized in that, The process of generating the target model based on the trained prediction model of at least two algorithm types includes: Using at least one prediction model corresponding to the trained linear type, the issuance volume of the product is predicted based on the multi-dimensional historical data, and a second prediction result corresponding to the prediction model of each linear type is generated. Using at least one prediction model corresponding to the trained nonlinear type, the issuance volume of the commodity is predicted based on the multi-dimensional historical data, and a third prediction result corresponding to each of the nonlinear types is generated. Based on the actual issuance volume corresponding to the multi-dimensional historical data, the second prediction results corresponding to each of the linear prediction models, and the third prediction results corresponding to each of the non-linear prediction models, the weight parameters of the target model are determined. The target model is generated based on the weight parameters.

8. A device for predicting the quantity of goods issued, characterized in that, The device includes: The acquisition module is used to acquire multi-dimensional data related to the issuance volume of the product to be predicted; wherein, the multi-dimensional data includes at least sales data and information on the impact of sales volume; The first generation module is used to predict the issuance volume of goods based on the multi-dimensional data using the trained target model, and generate the target prediction result. The target model is a model composed of prediction models combining at least two algorithm types, including at least linear and nonlinear types. The prediction model applies either the linear or nonlinear algorithm, with each algorithm type corresponding to at least one prediction model. The linear prediction model is used to predict the distribution volume of the product from a numerical calculation perspective, while the nonlinear prediction model is used to predict the distribution volume of the product from the perspective of nonlinear interaction of data in multi-dimensional data. The prediction model is a model trained using at least multi-dimensional historical data and corresponding actual distribution volumes as training samples. The multi-dimensional historical data includes at least historical sales data and information on the impact of historical sales volumes.

9. An electronic device, characterized in that, The electronic device includes: Memory, used to store executable instructions; A processor, when executing executable instructions stored in the memory, implements the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program or executable instructions stored thereon, characterized in that, When the computer program or the executable instructions are executed by the processor, they implement the method according to any one of claims 1 to 7.

Images