Methods, devices, and equipment for predicting the quantity of business objects required

Abstract

This application relates to a method, apparatus, and device for predicting the quantity of business objects in the field of computer intelligent analysis technology, aiming to improve the accuracy of predicting the quantity of business objects in the field of computer intelligent analysis technology. The method includes: establishing a hierarchical structure of various processing flows in a target scenario based on the impact of the results of executing a specific processing flow on the results of other different processing flows; selecting a target processing flow from multiple processing flows in response to a user input selection command; collecting multi-dimensional related features generated by the target processing flow in processing the business objects before the current time node to form a historical time node sequence; predicting the predicted quantity of business objects processed by the target processing flow at least at a future time node after the current time using a single-flow prediction model; inputting the predicted quantity into a harmonic model pre-trained based on the hierarchical structure, and adjusting the predicted quantity using the harmonic model with the inferred quantity to obtain the harmonic prediction vector.

Description

Technical Field

[0001] This application relates to the field of computer intelligent analysis technology, specifically to a method, apparatus, and device for predicting the demand quantity of business objects. Background Technology

[0002] Predicting the future consumption of business entities is one of the most important tools in modern business activities. It effectively helps companies formulate short-term and long-term plans and decisions, such as determining consumption patterns, rationally managing inventory, and optimizing product structure to reduce operating costs and increase profits. The accuracy of forecasting directly impacts the rationality and effectiveness of corporate decisions.

[0003] In deep learning, features are one of the most important factors affecting the performance of deep learning methods. Features that influence the prediction results of business objects include: historical processing data of the business object, date, holidays, weather, etc. Current learning methods directly input the above multi-dimensional features into the model, ignoring the differences between different features in the time dimension, as well as the correlation between various features, resulting in poor model performance.

[0004] Furthermore, the actual number of business objects consumed during the execution of a certain process is related to the scenario in which the process is executed. Existing technologies that predict the number of objects consumed in the future often ignore the correlation constraints between the scenarios in which a certain process is executed to process business objects, resulting in poor model prediction performance. Summary of the Invention

[0005] The technical problem to be solved by this application is to provide a method, apparatus and equipment for predicting the demand quantity of business objects, so as to improve the accuracy of predicting the demand quantity of business objects.

[0006] In one aspect, an embodiment provides a method for predicting the demand quantity of business objects, applied to an electronic device. The method includes: establishing a hierarchical structure of processing processes in a target scenario based on the impact of the result of executing a specific processing process on the results of other different processing processes; the processing process is the process of running a predetermined program to process business objects; the target scenario includes multiple processing processes; responding to a selection instruction input by a user, selecting a target processing process from the multiple processing processes; collecting multi-dimensional related features generated by the target processing process in processing the business objects before the current time node to form a historical time node sequence; the multi-dimensional related features include, but are not limited to, the number of business objects, the resources consumed in processing the business objects, time information, and scene environment; inputting the historical time node sequence into a pre-established single-process prediction model, and using the single-process prediction model to predict the predicted number of business objects processed by the target processing process at least at a future time node after the current time, as basic prediction information; inputting the predicted number into a harmonic model pre-trained based on the hierarchical structure, and using the harmonic model to adjust the predicted number with the inferred number to obtain the harmonic prediction vector; the inferred number is the number of objects consumed by other processing processes in the hierarchical structure at future time nodes.

[0007] In one embodiment, the method further includes: training the harmonic model using multiple samples based on the hierarchical structure; the process of training the harmonic model using a single sample includes:

[0008] Based on the characteristics generated by the target processing flow in processing the consumed object at the first time node, predict the first value of the target processing flow in processing the consumed object at the second time node, and collect the second value of the target processing flow in processing the consumed object at the second time node.

[0009] Based on the characteristics of the consumed object processed by at least one other processing flow at the first time node, predict the third value of the consumed object processed by the other processing flow at the second time node, and collect the fourth value of the consumed object processed by the other processing flow at the second time node.

[0010] According to the hierarchical structure, the third value and the first value are combined to form a prediction vector;

[0011] Initial weights are set for different processing flows and different time points in the hierarchical structure, and a relationship matrix is ​​constructed.

[0012] The prediction vector is reset using the relation matrix to obtain a reset prediction vector; the reset prediction vector includes a reset third value and a reset first value.

[0013] Combining the second and fourth values ​​yields the marker sequence;

[0014] The loss function is calculated based on the reset prediction vector and the label sequence, and the weight parameters in the relation matrix are adjusted according to the loss function until the loss function converges.

[0015] In one embodiment, calculating the loss function based on the reset prediction vector and the label sequence includes:

[0016] (1)

[0017] in, It is the reset prediction vector. It is a labeled sequence; It is the sequence of the first value after reset corresponding to the second time node. It is the sequence of the third value after reset, corresponding to the second time node. Represents the loss function. This indicates the sequence of the currently used sample. It represents the total number of samples.

[0018] In one embodiment, the harmonic model is trained based on the hierarchical structure; the process of training the harmonic model includes:

[0019] Initial weights are set for different processing flows and different time points in the hierarchical structure, and a relationship matrix is ​​constructed.

[0020] Using the relation matrix as an encoder and the transformation matrix as a decoder, a deep learning framework is obtained.

[0021] The harmonic model is obtained by training based on the deep learning framework.

[0022] In one embodiment, the historical time node sequence is input into a pre-established single-process prediction model, and the single-process prediction model is used to predict the predicted number of objects to be processed and consumed by the target processing process at least at a future time node after the current time, including:

[0023] Responding to user settings, determine the predicted future time points;

[0024] Set up a basic prediction network, set the prediction output structure of the basic prediction network according to the future time nodes, and set the retrospective output structure of the basic prediction network according to the node information in the historical time node sequence.

[0025] Input the historical time node sequence into any basic prediction network, output the first prediction information according to the prediction output structure, and output the first review information according to the review output structure;

[0026] The first review information is input into the basic prediction network corresponding to the next dimension feature, and the second prediction information and the second review information are output. The second review information is input into the next basic prediction network until the outputs of all basic prediction networks are obtained.

[0027] By combining the prediction information output by each basic prediction network, the predicted number of business objects to be processed by the target processing flow at least at a future time node after the current time is obtained.

[0028] In one embodiment, the method further includes:

[0029] For each basic prediction network, the difference between the output of the retrospective information and the input is calculated to obtain the residual information filtered by that basic prediction network.

[0030] The prediction information, combined with the output of each base prediction network, includes:

[0031] When the number of basic prediction networks is less than a preset number, the predicted number of objects to be processed by the target processing flow at least at a future time node after the current time is obtained by combining the prediction information output by each basic prediction network and the residual information.

[0032] In one embodiment, the single-process prediction model includes multiple basic prediction networks; the basic prediction network includes a time-related feature extraction network layer and a non-time-related feature extraction network layer; the time-related feature extraction network layer is constructed based on LSTM results, and the non-time-related feature extraction network layer is constructed based on MLP.

[0033] In one embodiment, the historical time node sequence is input into any basic prediction network, and first prediction information is output according to the prediction output structure, and first review information is output according to the review output structure, including:

[0034] The historical time node sequence is calculated by the time-related feature extraction network layer, and the first latent variable features associated with time in the historical time node sequence are filtered out.

[0035] When time-independent features exist in the historical time node sequence, a non-time-related feature extraction network layer is activated to filter out second latent variable features in the historical time node sequence that are not related to time; time-independent features refer to features that do not have a linear relationship with time.

[0036] The first latent variable feature is enhanced to obtain a first enhanced feature, and the second latent variable feature is enhanced to obtain a second enhanced feature;

[0037] The features obtained by fusing the first enhancement feature and the second enhancement feature are used to calculate the first prediction information and the first review information.

[0038] Secondly, in one embodiment, a business object demand quantity prediction device is provided, disposed in an electronic device, the device comprising:

[0039] The hierarchical structure building module is used to establish a hierarchical structure of each processing flow in the target scenario based on the impact of the result of executing a specific processing flow on the results of other different processing flows; the processing flow is the process of running a predetermined program to process a business object; the target scenario contains multiple processing flows;

[0040] The process selection module is used to select the target processing flow from the multiple processing flows in response to the user's input selection command.

[0041] The feature acquisition module is used to collect multi-dimensional related features generated by the target processing flow in processing the business object before the current time node, forming a historical time node sequence; the multi-dimensional related features include, but are not limited to, the number of business objects, the resources consumed in processing the business objects, time information, and scene environment;

[0042] The sequence input module is used to input the historical time node sequence into a pre-established single-process prediction model, and predict the number of business objects to be processed by the target processing process at least at a future time node after the current time through the single-process prediction model, as the basic prediction information.

[0043] The adjustment module is used to input the predicted quantity into a harmonic model pre-trained based on the hierarchical structure, and adjust the predicted quantity using the inferred quantity through the harmonic model to obtain the harmonic prediction vector; the inferred quantity is the number of objects consumed by other processing flows in the hierarchical structure at future time nodes.

[0044] The beneficial effects of this application are:

[0045] The business object demand quantity prediction method proposed in this application reveals that the results of different processing flows can affect the results of other processing flows. In a sales scenario, this manifests as the sales quantity of one product category influencing the sales quantity of another, or the sales quantity of a certain product in one region being affected by the sales quantity of the same product in other regions. Based on this finding, and according to the current application scenario of processing business objects, the method mines the hierarchical structure of the business object processing flow from aspects such as region, city, store, date, and business object category. This hierarchical structure is used to optimize the predicted business object demand quantity for different processing flows, improving the accuracy of business object demand quantity prediction. Simultaneously, this application also extracts the dependency relationship between features and time, extracting time-related features and time-independent features respectively, and strengthening both time-related and time-independent features to amplify the differences between features. This clarifies the feature learning path of the deep learning network, thereby improving the accuracy of the model's output prediction of the business object quantity. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the steps of the business object demand quantity prediction method proposed in the embodiments of this application;

[0047] Figure 2 This is a schematic diagram of a hierarchical structure established as an example in this application;

[0048] Figure 3 This is a schematic diagram of the hierarchical structure established by another example in this application;

[0049] Figure 4 This is a schematic diagram of the structure of an example single-process prediction model of this application;

[0050] Figure 5 This is a schematic diagram of the structure of a basic prediction network constructed as an example in this application;

[0051] Figure 6 This is a schematic diagram of the geographical extent in one example of this application;

[0052] Figure 7 This is a schematic diagram of the learning framework of a harmonic model in this application;

[0053] Figure 8 This is a functional block diagram of the business object demand quantity prediction device proposed in the embodiments of this application;

[0054] Figure 9 This is a schematic diagram of the structure of a computing device provided in an embodiment of this application. Detailed Implementation

[0055] The present application will now be described in further detail with reference to the accompanying drawings and specific embodiments. Similar elements in different embodiments are referred to by related similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of the present application. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to the present application are not shown or described in the specification. This is to avoid obscuring the core parts of the present application with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0056] Furthermore, the features, operations, or characteristics described in the specification can be combined in any suitable manner to form various embodiments. At the same time, the steps or actions in the method description can be rearranged or adjusted in a manner obvious to those skilled in the art. Therefore, the various orders in the specification and drawings are only for the clear description of a particular embodiment and do not imply a necessary order, unless otherwise stated that a particular order must be followed.

[0057] The serial numbers assigned to components in this document, such as "first" and "second," are used only to distinguish the described objects and have no sequential or technical meaning. The terms "connection" and "linkage" used in this application, unless otherwise specified, include both direct and indirect connections (linkages).

[0058] Figure 1 This is a flowchart of the steps for predicting the number of business object requirements proposed in the embodiments of this application. The executing entity can be an electronic device, which can be a server, computer terminal, virtual computing system, etc.

[0059] The steps include:

[0060] S101: Based on the impact of the results achieved by executing a specific processing flow on the results of other different processing flows, establish a hierarchical structure formed by each processing flow in the target scenario.

[0061] The processing flow is the process of running a predetermined program to process business objects; the target scenario includes multiple processing flows.

[0062] The applicant discovered that the number of different business objects consumed in processing specific processes for business objects under different scenarios is mutually constrained.

[0063] In one example, the target scenario is product processing. The target scenario may include multiple intelligent devices that execute processing procedures and mechanical devices that perform operations on the product according to instructions from the intelligent devices. Each mechanical device's operation on the product according to instructions from the intelligent devices constitutes a processing procedure.

[0064] In this example, business object 'a' is a part to be processed, and the specific processing flow is the process of the intelligent device running a predetermined program to process the part. Business object 'b' is a complete assembly to be assembled, and the specific processing flow is the process of the intelligent device running a predetermined program to assemble the complete assembly. Business object 'c' is also a part to be processed, and the specific processing flow is the process of the intelligent device running a predetermined program to process the part. Business object 'b' is assembled from business objects 'a' and 'c'. Therefore, there is a constraint relationship between the number of business objects consumed in executing the specific processing flow: + =N. It represents the consumption quantity of business object 'a'. N is the consumption quantity of business object c, and N is the consumption quantity of business object b.

[0065] Based on the above scenario, the calculation process by which the electronic device "establishes a hierarchical structure of each processing flow in the target scenario based on the impact of the result of executing a specific processing flow on the results of other different processing flows" can be as follows: Based on the impact of the number of parts obtained from executing the parts processing process on the workpiece assembly process, establish a hierarchical structure such as... Figure 2 As shown, Figure 2 This is a schematic diagram of a hierarchical structure established as an example in this application. Figure 2 In the diagram, business object b corresponds to processing flow b as the root node, and business object a corresponds to processing flow a and business object c corresponds to processing flow c as leaf nodes of the root node.

[0066] In one example, the target scenario is the sale of business objects. The target scenario may include the process of multiple vending machines each executing their local reservation program. The process of each vending machine executing its local reservation program to sell goods is a processing flow. The process of the server receiving instructions from users through the mini-program to control the vending machines to dispense goods can also be considered a processing flow. Predicting the demand for a business object at a certain time can be understood as the sales quantity of the business object at a certain time. In practical applications, accurately predicting the sales quantity of a business object at a certain time provides a basis for providing the demand for the business object in the future time period, ensuring that the business object will not be in short supply or become unsaleable.

[0067] In this example, the business object is product A to be sold. The computing device set up in vending machine 1 is used to receive user instructions and execute the program to deliver product A to be sold. The program to deliver product A to be sold is a pre-defined program for this scenario. The computing device controls the vending machine's dispensing mechanism to deliver product A to be sold. The program to deliver product B to be sold is also a pre-defined program for this scenario. The computing device controls the vending machine's dispensing mechanism to deliver product B to be sold. Product A and product B to be sold are located in different areas. During the execution of the program to deliver the product to be sold, the vending machine's computing device can also upload user requests to the server. The server executes the process of supplementing business objects for each vending machine based on the program execution results of the vending machines located in different locations.

[0068] Based on the business objects consumed by different vending machines in a certain region during the process of delivering goods according to a pre-order program, and considering the impact of the server in that region executing the pre-order program to control the robot's ability to replenish goods in the vending machines, a hierarchical structure is established as follows: Figure 3 As shown, Figure 3 This is a schematic diagram of the hierarchical structure established by another example of this application. Figure 3 In the diagram, the root node is the processing flow C corresponding to product C to be sold, and the leaf nodes are the processing flow B corresponding to product B to be sold and the processing flow A corresponding to product A to be sold.

[0069] Based on the application of step S101 in the scenario shown in the example above, we get: In the sales scenario, "establishing a hierarchical structure formed by each processing flow in the target scenario based on the impact of the result achieved by executing a specific processing flow on the results of other different processing flows" means: establish a hierarchical structure for each sales scenario based on the impact of the sales quantity of a business object in a certain sales scenario on the sales quantity of business objects in other sales scenarios.

[0070] Generally, factors influencing hierarchical structures may include region, city, store, and date. Therefore, electronic devices can establish hierarchical structures by analyzing the constraints of factors such as region, city, store, date, and business object classification on the consumption quantity of business objects, based on different scenarios where specific processing flows are executed.

[0071] S102: In response to the user's input selection instruction, select the target processing flow from the plurality of processing flows.

[0072] The objective conditions for executing different processing flows are different, so it is necessary to predict the number of business objects consumed for each processing flow separately. Specifically, each processing flow can be taken as the target processing flow for the current processing flow, and then the number of business objects consumed for the target processing flow can be predicted.

[0073] In this application embodiment, a single-process prediction model is set for each processing flow. The consumption quantity of business objects in each processing flow is predicted based on the single-process prediction model. Before predicting the consumption quantity of business objects in the target processing flow based on the single-process prediction model, features related to business objects are collected during the execution of the target processing flow and the consumption of business objects.

[0074] S103: The target processing flow processes the multi-dimensional related features generated by the business object before the current time node to form a historical time node sequence.

[0075] The multidimensional related features include, but are not limited to, the number of business objects, the resources consumed in processing business objects, time information, and scene environment; each historical time node in the multiple historical time node sequences has a corresponding specific dimension related feature.

[0076] Taking the business object sales scenario as an example, the number of business objects can be the number of goods sold in a certain period of time, the resources consumed in processing business objects can include promotional activities and advertising investment, and the scenario environment can include holidays, temperature, weather, etc.

[0077] The multidimensional related features generated from processing the business object before the current time point can represent: historical data on the consumption quantity of the business object. For example, if the current time is T month TT day, historical time points before the current time include: T month TT-1 day, T month TT-2 day, and T month TT-3 day, and the multidimensional related features generated from processing the business object on T month TT-1 day are collected. Collect multidimensional related features generated from business objects processed on day T (T-2). Collect multidimensional related features generated from business objects processed on day T of month TT-3. The historical timeline is obtained as follows: { Month T, Day T-1 - ... ,T month TT-2nd- ,T month TT-3rd- }

[0078] In real-world scenarios involving the consumption of business objects, such as the sale of business objects, the characteristics affecting the sales volume of these objects include endogenous and exogenous features. Endogenous features include lagging historical sales data and features added by the merchant to sell the business object, such as promotional activities and advertising investment. Exogenous features reflect external factors such as date, holidays, and weather. All of these factors affect the sales volume of the business object, and each factor has a different impact on the sales volume. Therefore, features can be collected separately based on each dimension.

[0079] For example, suppose the multidimensional features include: quantity, Advertising investment, temperature, Weather, so T month TT-1- 、T month TT-1- 、T month TT-1- 、T month TT-1- Each can be used as a specific dimension-related feature of the historical time node T month TT-1 day.

[0080] S104: Input the historical time node sequence into a pre-established single-process prediction model, and predict the number of business objects to be processed by the target processing process at least at a future time node after the current time through the single-process prediction model.

[0081] Taking the business object sales scenario as an example, the target processing flow is the sales of business objects at venue A in a certain street. The historical time node sequence generated by venue A is input into the single-process prediction model to predict the number of business objects sold at venue A at a certain point in the future.

[0082] In one example, if a business entity sells at location A for the first time, multi-dimensional features of similar products can be collected to obtain a historical timeline sequence.

[0083] Figure 4 This is a schematic diagram of the structure of an example single-process prediction model of this application, as shown below. Figure 4 As shown, the single-process prediction model includes multiple basic prediction networks, consisting of... Figure 4 As can be seen, the output of the basic prediction network includes both review information and prediction information. The review information output by basic prediction network 1 is input into basic prediction network 2. To enhance feature extraction capabilities, multiple basic prediction networks are stacked to obtain a single-process prediction model. At the same time, skip connections are used to prevent model degradation due to the increase in the number of layers.

[0084] The execution step S104 of this application embodiment includes the following sub-steps:

[0085] S1041: Respond to the user's setting command and determine the predicted future time node.

[0086] If the current time is day T, the user selects to predict the number of business objects consumed in the next Q days, and the future time nodes include {T+1, T+2, ..., T+Q}.

[0087] The single-process prediction model includes a base prediction network for each dimension of features.

[0088] S1042: Set up a basic prediction network, set the prediction output structure of the basic prediction network according to the future time nodes, and set the retrospective output structure of the basic prediction network according to the node information in the historical time node sequence.

[0089] For example, if the historical timeline sequence Let {T+1, T+2, ..., T+0} be the current time, and the future time nodes be {T+1, T+2, ..., T+Q}. Then the predicted output structure can be a 0-bit output structure, and the retrospective output structure can be a Q-bit output structure.

[0090] S1043: Input the historical time node sequence into any basic prediction network, output the first prediction information according to the prediction output structure, and output the first review information according to the review output structure.

[0091] For example, assuming future time points are {T+1, T+2, ..., T+Q}, the predicted information output based on the predicted output structure is represented as follows: ,in, It is the number of business objects required at time point (T+1) predicted by the basic prediction network. This is the number of business objects required at time point (T+2) predicted by the basic prediction network. It is the number of business objects required at the (T+Q) time point predicted by the basic prediction network.

[0092] The applicant discovered that among the features across different dimensions, there were both time-dependent and time-independent features. For example, in a sales scenario, features such as advertising complaints and temperature are time-independent, while features such as date and sales volume are time-dependent. If a unified network is used to calculate all features, the potential risk is overlooked: interference from time information during feature mixing extraction may lead to performance degradation.

[0093] In view of the above problems, the basic prediction network set in the embodiments of this application includes a time-related feature extraction network layer and a non-time-related feature extraction network layer; the time-related feature extraction network layer is constructed based on LSTM results, and the non-time-related feature extraction network layer is constructed based on MLP.

[0094] Figure 5 This is a schematic diagram of the structure of a basic prediction network constructed as an example in this application, such as... Figure 5As shown, the time-related feature extraction network layer calculates the first latent variable feature by analyzing the historical time node sequence. Then, the first latent variable feature is enhanced to obtain the enhanced time-related feature sequence. Combined with the time-independent feature sequence, the factors affecting the historical consumption quantity of business objects are first mined from two aspects, and then integrated. This avoids noise when extracting time-related features and time-independent features separately, and solves the problem that interference from time information in feature mixing extraction may lead to performance degradation.

[0095] based on Figure 5 The network shown in the diagram executes step S1043 by running sub-steps S201-S204:

[0096] S201: The historical time node sequence is calculated through the time-related feature extraction network layer, and the first latent variable feature associated with time in the historical time node sequence is selected.

[0097] The temporal correlation feature extraction network layer is built on a Long Short-Term Memory (LSTM) network.

[0098] S202: When time-independent features exist in the historical time node sequence, the non-time-related feature extraction network layer is activated to filter out second latent variable features in the historical time node sequence that are not related to time. Time-independent features refer to features that do not have a linear relationship with time.

[0099] The non-temporally related feature extraction network layer is built based on the multi-layer perceptron (MLP) mechanism.

[0100] For example, in a sales scenario, features such as promotional activities, advertising investment, and weather can be discrete features. In the actual implementation of methods for predicting the demand quantity of business objects, a feature library can be pre-established, or curves showing the changes of features over historical time points can be created to determine whether a feature has a linear relationship with time, thereby determining whether to enable the non-temporally related feature extraction network layer.

[0101] One example of this application involves feature segmentation. First, the historical time node sequence T is divided into time-independent features unrelated to time, and time-dependent features (TDFs) related to time. The input historical time node sequence is a matrix of length L with r-dimensional features. A sliding window is used to sample the historical time node sequence to obtain m-dimensional time-related features (TDFs), denoted as follows: Time-Independent Features (TiDFs), which are not related to time in the rm dimension, are represented as follows: When obtained through sliding window sampling This indicates that there are time-independent features in the historical time node sequence.

[0102] The LSTM network layer for extracting time-related features can be used to calculate historical time node sequences using formula (2);

[0103] (2)

[0104] This represents the first latent variable feature.

[0105] The non-temporally correlated feature extraction network layer MLP can calculate the historical time node sequence through formula (3);

[0106] (3)

[0107] This represents the characteristics of the second latent variable.

[0108] S203: Strengthen the first latent variable feature to obtain a first strengthened feature, and strengthen the second latent variable feature to obtain a second strengthened feature.

[0109] The enhancement of the first latent variable feature can be achieved through formula (4);

[0110] (4)

[0111] This indicates the use of the sigmoid activation function on... Perform calculations. This indicates the first strengthening feature.

[0112] Calculate using the sigmoid activation function This yields the representation of the importance of time-dependent features. And then Element-wise multiplication and filtering yield the most prominent results. Time-dependent features .

[0113] The enhancement of the second latent variable features can be achieved through formula (5);

[0114] (5)

[0115] This represents the second enhancement feature. It is calculated using the sigmoid activation function. This yields the representation of the importance of time-independent features. And then Element-wise multiplication and filtering yield the most prominent results. Time-independent characteristics This study aims to clarify the connections between time-related features in historical time sequence and to enhance the network's learning and memorization of time-independent features.

[0116] S204: The features obtained by fusing the first enhanced feature and the second enhanced feature are used to calculate the first prediction information and the first review information.

[0117] The fusion of the first and second enhancement features can be achieved through formula (6);

[0118] (6)

[0119] This represents element-wise matrix multiplication. This represents the element-wise addition of matrices, yielding... Then, the fused information is calculated using a fully connected layer FC1, and the predicted information is output. The fused information is calculated using a fully connected layer FC2, and the resulting information is a review. .

[0120] In one possible implementation, a single-process prediction model is obtained by stacking basic prediction networks. The basic prediction network currently performing the computation outputs first review information and first prediction information. The first review information serves as the input to the basic prediction network that performs the next computation task. The basic prediction network that performs the next computation task outputs second review information and second prediction information, and so on.

[0121] S1044: Input the first review information into the basic prediction network corresponding to the next dimension feature, output the second prediction information and the second review information, input the second review information into the next basic prediction network, until the output of all basic prediction networks is obtained.

[0122] The implementation method of inputting the second review information into the basic prediction network corresponding to the next dimension feature includes: performing residual calculation on the second review information and the input of the previous basic prediction network, and using the parameters as all inputs to the next basic prediction network.

[0123] Continue to refer to Figure 4 In one example of this application, a historical time node sequence Input base prediction network 1, base prediction network 1 outputs review information and predictive information Calculate the residuals of the basic prediction network 1. = - residual As input to the base prediction network 2, the output of the base prediction network 2 is... and Calculate the residuals of the basic prediction network 2. = - residual As the input to the basic prediction network 3, the residuals of the output of the basic prediction network k-1 are also used. As input to the base prediction network k, the output of the base prediction network k is... and .

[0124] S1045: Combining the prediction information output by each basic prediction network, obtain the predicted number of business objects to be processed by the target processing flow at least at a future time node after the current time.

[0125] For each basic prediction network, the difference between the output of the review information and the input is calculated to obtain the residual information filtered by that basic prediction network. When the number of basic prediction networks is less than the preset number, the prediction information and residual information output by each basic prediction network are combined to obtain the predicted number of objects to be processed and consumed by the target processing flow at least at a future time node after the current time.

[0126] When the number of basic prediction networks stacked is small, the last basic prediction network may still contain a lot of information. Therefore, when the number of basic prediction networks is less than the preset number, the prediction information output by each basic prediction network and the residual information corresponding to the last basic prediction network are fused by formula (7) to obtain the prediction number of objects to be processed at least one future time node after the current time.

[0127] (7)

[0128] This indicates the predicted number of objects to be processed at least one future time point after the current time, i.e., basic forecast information. In a sales scenario, it can be the predicted number of sales business objects at least one future time point after the current time. Accurately predicting sales volume can provide data support for goods preparation, ensuring that the prepared goods are within a reasonable range and avoiding waste and insufficient goods.

[0129] This indicates that a fully connected computation is performed on the residual information corresponding to the last basic prediction network in the arrangement.

[0130] Indicates the first The predicted information output by each basic prediction network is accumulated, and then combined with the calculation results of the residual information by the fully connected network to obtain the predicted number of objects to be processed and consumed by the target processing flow at least at a future time node after the current time.

[0131] Each processing flow can determine the predicted number of business objects to be processed at least at a future time node after the current time through the above embodiments of this application. By mining the hierarchical structure between processing flows through the harmonic model set in the embodiments of this application, the predicted number of business object demand quantities corresponding to each processing flow can be further optimized.

[0132] For example, in a sales scenario, business data sales region The number of predicted business objects is Business data sales region The number of predicted business objects is Business data sales region The number of predicted business objects is The features generated by a single processing flow may not be comprehensive enough, or special factors may cause prediction errors. In view of this, the embodiments of this application set up a harmonic model to mine the hierarchical constraints between the prediction results of different processing flows and adjust the prediction results to further optimize the processing results of each processing flow.

[0133] Because the prediction information output by the basic prediction network according to the prediction output structure is represented as The prediction information from the output of each base prediction network is expressed in the same way.

[0134] S105: Input the predicted quantity into the harmonic model pre-trained based on the hierarchical structure, and adjust the predicted quantity with the inferred quantity through the harmonic model to obtain the harmonic prediction vector.

[0135] The estimated quantity is the number of objects that other processing flows in the hierarchy will process at a future time point.

[0136] For example, the business data shows the first sales region. The number of predicted business objects is The second sales region corresponding to the business data The number of predicted business objects is Business data for the third sales region The geographic region includes the number of forecasts for the forecast processing business objects. ,analyze , and The map's extent data, such as Figure 6 As shown, Figure 6 This is a schematic diagram of the geographical extent in one example of this application. Sales volume of business objects Objectively speaking and the sum of Figure 6 This is a diagram of the location area corresponding to the business object processing flow in one example of this application, and the hierarchical structure is as follows: = + .

[0137] If based on Angle, and As a presumed quantity, for Optimize to make the optimized The number approaches and The sum. If based on Angle, and As a presumed quantity, for Optimize to make the optimized The number approaches and difference.

[0138] The business object demand quantity prediction method proposed in this application reveals that the results of different processing flows can affect the results of other processing flows. In a sales scenario, this manifests as the sales quantity of one product category influencing the sales quantity of another, or the sales quantity of a certain product in one region being affected by the sales quantity of the same product in other regions. Based on this finding, and according to the current application scenario of processing business objects, the method mines the hierarchical structure of the business object processing flow from aspects such as region, city, store, date, and business object category. This hierarchical structure is used to optimize the predicted business object demand quantity for different processing flows, improving the accuracy of business object demand quantity prediction. Simultaneously, this application also extracts the dependency relationship between features and time, extracting time-related features and time-independent features respectively, and strengthening both time-related and time-independent features to amplify the differences between features. This clarifies the feature learning path of the deep learning network, thereby improving the accuracy of the model's output prediction of the business object quantity.

[0139] This application also provides a method for training a harmonic model. Figure 7 This is a schematic diagram of the learning framework of a harmonic model proposed in this application, as shown below. Figure 7 As shown, the harmonic model includes an encoder and a decoder. In this embodiment, the relation matrix S is used as the encoder and the transformation matrix P is used as the decoder.

[0140] Initial weights are set for different processing flows and time points in the hierarchical structure to construct a relation matrix; the relation matrix is ​​used as an encoder and the transformation matrix as a decoder to obtain a deep learning framework; the harmonic model is trained based on the deep learning framework.

[0141] The transformation matrix P is an M*N matrix used for scaling matrices. Relationship matrix. Different initial weight parameters are set for different levels. The initial weight parameters are adjusted during training until the reconciliation model adjusts the loss function of the number of predictions and labels to converge.

[0142] Training the learning framework to obtain the harmonic model involves the following steps:

[0143] S301: Based on the characteristics generated by the target processing flow processing the consumed object at the first time node, predict the first value of the target processing flow processing the consumed object at the second time node, and collect the second value of the target processing flow processing the consumed object at the second time node.

[0144] The first value can be predicted using the methods proposed in other embodiments of this application. The characteristics generated by the processed object include date, weather, temperature, quantity, etc.

[0145] For example, the first time point includes , , The second time node corresponds to ( ), ( ), ( ); Collected features , Collected features , Collected features Using the single-process prediction model proposed in other embodiments of this application, based on the above features { , , The target prediction processing flow is in ( ), ( ), ( The number of business objects consumed are respectively , , , as the first value; and the target processing flow is in ( ), ( ), ( The actual number of business objects consumed in the time was respectively , , , as the second value.

[0146] S302: Based on the characteristics of the consumed object processed by at least one other processing flow at the first time node, predict the third value of the consumed object processed by the other processing flow at the second time node, and collect the fourth value of the consumed object processed by the other processing flow at the second time node.

[0147] The characteristics of data acquisition and processing flow 1 include: Collected features , Collected features , Collected features The features of data acquisition and processing flow 2 include: Collected features , Collected features , Collected features Based on the above characteristics { , , Predictive processing flow 1 in ( ), ( ), ( The number of business objects consumed is , , As the third value corresponding to processing flow 1, and the data collected for processing flow 1 in ( ), ( ), ( The actual number of business objects consumed in the time is , , This serves as the fourth value corresponding to processing flow 1.

[0148] Using the same method, obtain the third value corresponding to processing flow 2: , , Obtain the fourth value corresponding to processing flow 2: , , .

[0149] S303: According to the hierarchical structure, combine the third value and the first value to form a prediction vector.

[0150] Based on the method of outputting prediction information by the single-process prediction model described in other embodiments of this application, the output of the single-process prediction model corresponding to processing flow 1 can be expressed as follows: , ={( )- 、( )- 、( )- The output of the single-process prediction model corresponding to processing flow 2 can be expressed as follows: , ={( )- 、( )- 、( )- The output of the single-process prediction model corresponding to the target processing flow can be expressed as: , ={( )- 、( )- 、( )- }

[0151] For example, suppose the target process is as follows: Figure 3 The processing flow C in the illustrated case hierarchy is the target flow, which is the root node in the hierarchy. Processing flow 1 corresponds to... Figure 3 The processing flow A and processing flow 2 shown are as follows: Figure 3 The processing flow shown is B, refer to Figure 3 In the hierarchical structure shown, processing flow 1 and processing flow 2 are leaf nodes of the target processing flow. Therefore, according to the hierarchical structure, combining the third values ​​predicted by the other processing flows with the first value predicted by the target processing flow to form a prediction vector can be represented as follows: = .

[0152] S304: Set initial weights for different processing flows and different time nodes in the hierarchical structure, and construct a relationship matrix.

[0153] Using the relation matrix S as an encoder, features are extracted based on the predicted information corresponding to the positions of different hierarchical structures, with attention weights of different weights.

[0154] For example, by constructing initial weights, the relationship matrix is ​​controlled during the encoding of the predicted vector. The computational strategy is determined by the position of the target process within the hierarchical structure, thereby adjusting the weight information and the relation matrix during model training. The computational strategy matches the hierarchical constraint relationship among the three.

[0155] S305: The prediction vector is reset using the relation matrix to obtain a reset prediction vector; the reset prediction vector includes a reset third value and a reset first value.

[0156] For reference Figure 7 The learning framework diagram shown illustrates that the process of resetting the prediction vector using the relation matrix can be represented by formula (8):

[0157] (8)

[0158] in, The prediction vector is formed by the third value and the first value. It is the reset prediction vector. It is a relation matrix. It is a transformation matrix.

[0159] S306: Combine the second value and the fourth value to obtain the marker sequence.

[0160] The target acquisition and processing flow is in ( ), ( ), ( The actual number of business objects consumed in the time was respectively , , As the second value, the actual target processing flow is in ( ), ( ), ( The actual sequence information of the business data consumed at each time point is represented as follows: Processing flow 1 in ( ), ( ), ( The actual sequence information of the business data consumed at each time point is represented as follows: Processing flow 1 in ( ), ( ), ( The actual sequence information of the business data consumed at each time point is represented as follows: .

[0161] Tag sequence for .

[0162] S307: Calculate the loss function based on the reset prediction vector and the label sequence, and adjust the weight parameters in the relation matrix according to the loss function until the loss function converges.

[0163] This application provides an example of how the loss function is calculated:

[0164] (1)

[0165] in, It is the reset prediction vector. It is a labeled sequence; It is the sequence of the first value after reset corresponding to the second time node. It is the sequence of the third value after reset, corresponding to the second time node. Represents the loss function. This indicates the sequence of the currently used sample. It represents the total number of samples.

[0166] This can represent the difference between the root node prediction information and the leaf node prediction information. The root node prediction information is the prediction information corresponding to the root node in the hierarchical structure, and the leaf node prediction information is the prediction information corresponding to the leaf node connected to the root node.

[0167] Other examples given in this application, such as Figure 3 In the hierarchical structure of the case shown, the target processing flow, processing flow 1, and processing flow 2 are adjusted using a harmonic model to optimize the output prediction information of the single-flow prediction model corresponding to each of the target processing flow, processing flow 1, and processing flow 2. This is done by using the constraint of the number of business objects consumed by processing flow 1 and processing flow 2 in the hierarchical structure on the number of business objects consumed by the target flow, and optimizing the output prediction information of the single-flow prediction model corresponding to the target processing flow.

[0168] Substitute the predicted vector and label sequence formed by the third and first values ​​in the above example into the loss function. ,like This is the first value after the reset. It is the third value after the reset, that is yes ,So yes and .

[0169] Figure 8 This is a functional block diagram of the business object demand quantity prediction device proposed in the embodiments of this application, such as... Figure 8 As shown, the business object demand quantity prediction device includes:

[0170] The hierarchical structure establishment module 81 is used to establish a hierarchical structure of each processing flow in the target scenario based on the impact of the result achieved by executing a specific processing flow on the results of other different processing flows; the processing flow is the process of running a predetermined program to process a business object; the target scenario contains multiple processing flows.

[0171] The process selection module 82 is used to select the target processing flow from the plurality of processing flows in response to the user's input selection command.

[0172] The feature acquisition module 83 is used to acquire multi-dimensional related features generated by the target processing flow in processing the business object before the current time node, forming a historical time node sequence; the multi-dimensional related features include, but are not limited to, the number of business objects, the resources consumed in processing the business objects, time information, and scene environment;

[0173] The sequence input module 84 is used to input the historical time node sequence into a pre-established single-process prediction model, and predict the number of business objects to be processed by the target processing process at least at a future time node after the current time through the single-process prediction model, as the basic prediction information.

[0174] The adjustment module 85 is used to input the predicted quantity into the harmonic model pre-trained based on the hierarchical structure, and adjust the predicted quantity by the inferred quantity through the harmonic model to obtain the harmonic prediction vector; the inferred quantity is the number of objects consumed by other processing flows in the hierarchical structure at future time nodes.

[0175] Figure 8 The business object demand quantity prediction provided in the illustrated embodiment can be used to perform the work described in this specification. Figures 1 to 7 The technical solution of the method embodiment shown can be further described in the relevant descriptions in the method embodiment for its implementation principle and technical effect.

[0176] Optionally, the device further includes a training module, which is used to train the harmonic model using multiple samples based on the hierarchical structure. Specifically, it is used to predict the first value of the target processing flow processing the consumed object at the second time node based on the features generated by the target processing flow processing the consumed object at the first time node, and to collect the second value of the target processing flow processing the consumed object at the second time node.

[0177] Based on the characteristics of the consumed object processed by at least one other processing flow at the first time node, predict the third value of the consumed object processed by the other processing flow at the second time node, and collect the fourth value of the consumed object processed by the other processing flow at the second time node.

[0178] According to the hierarchical structure, the third value and the first value are combined to form a prediction vector;

[0179] Initial weights are set for different processing flows and different time points in the hierarchical structure, and a relationship matrix is ​​constructed.

[0180] The prediction vector is reset using the relation matrix to obtain a reset prediction vector; the reset prediction vector includes a reset third value and a reset first value.

[0181] Combining the second and fourth values ​​yields the marker sequence;

[0182] The loss function is calculated based on the reset prediction vector and the label sequence, and the weight parameters in the relation matrix are adjusted according to the loss function until the loss function converges.

[0183] Optionally, the training module is also used to set initial weights for different processing flows and different time nodes in the hierarchical structure, and to construct a relationship matrix;

[0184] Using the relation matrix as an encoder and the transformation matrix as a decoder, a deep learning framework is obtained.

[0185] The harmonic model is obtained by training based on the deep learning framework.

[0186] Optionally, the sequence input module includes:

[0187] The node determination submodule is used to determine the predicted future time nodes in response to user setting instructions;

[0188] The network setup submodule is used to set up the basic prediction network, set the prediction output structure of the basic prediction network according to the future time nodes, and set the retrospective output structure of the basic prediction network according to the node information in the historical time node sequence.

[0189] The first output submodule is used to input the historical time node sequence into any basic prediction network, output first prediction information according to the prediction output structure, and output first review information according to the review output structure.

[0190] The second output submodule is used to input the first review information into the basic prediction network corresponding to the next dimension feature, output the second prediction information and the second review information, input the second review information into the next basic prediction network, and so on until the outputs of all basic prediction networks are obtained.

[0191] The fusion submodule is used to combine the prediction information output by each basic prediction network to obtain the predicted number of business objects to be processed by the target processing flow at least at a future time node after the current time.

[0192] Optionally, the device further includes:

[0193] The residual calculation submodule is used to calculate the difference between the output review information and the input for each basic prediction network, and obtain the residual information filtered by the basic prediction network.

[0194] The sub-module is specifically used to, when the number of basic prediction networks is less than a preset number, combine the prediction information output by each basic prediction network and the residual information to obtain the predicted number of objects to be processed and consumed by the target processing flow at least at a future time node after the current time.

[0195] Optionally, the single-process prediction model includes multiple basic prediction networks; the basic prediction network includes a time-related feature extraction network layer and a non-time-related feature extraction network layer; the time-related feature extraction network layer is constructed based on LSTM results, and the non-time-related feature extraction network layer is constructed based on MLP.

[0196] Optionally, the first output submodule includes:

[0197] The calculation submodule is used to calculate the historical time node sequence through the time-related feature extraction network layer and filter the first latent variable features related to time in the historical time node sequence;

[0198] The feature extraction submodule is used to enable the non-temporally related feature extraction network layer when there are time-independent features in the historical time node sequence, and to filter the second latent variable features in the historical time node sequence that are not related to time; time-independent features refer to features that do not have a linear relationship with time.

[0199] The enhancement submodule is used to enhance the first latent variable feature to obtain the first enhanced feature, and to enhance the second latent variable feature to obtain the second enhanced feature;

[0200] The fusion submodule is used to fuse the features obtained by fusing the first enhanced feature and the second enhanced feature, and to calculate the first prediction information and the first recall information.

[0201] Regarding the modules / units included in the various devices described in the above embodiments, they can be software modules / units, hardware modules / units, or a combination of both. For example, for devices applied to or integrated into a chip, all modules / units can be implemented using hardware methods such as circuits, or at least some modules / units can be implemented using software programs running on a processor integrated within the chip, while the remaining modules / units can be implemented using hardware methods such as circuits. For devices applied to or integrated into a chip module, all modules / units can be implemented using hardware methods such as circuits. Different modules / units can be located in the same component (e.g., chip, circuit module, etc.) or different components of the chip module, or at least some modules / units can be implemented using software programs. The software program runs on the processor integrated inside the chip module, and the remaining modules / units can be implemented using hardware methods such as circuits. For each device applied to or integrated into an electronic terminal device, each of its modules / units can be implemented using hardware methods such as circuits. Different modules / units can be located in the same component (e.g., chip, circuit module, etc.) or different components within the electronic terminal device. Alternatively, at least some modules / units can be implemented using software programs that run on the processor integrated inside the electronic terminal device, and the remaining (if any) modules / units can be implemented using hardware methods such as circuits.

[0202] Figure 9 This is a schematic diagram of a computing device 900 provided in an embodiment of this application. The computing device 900 includes a processor 910, a memory 911, and a computer program stored in the memory 911 and executable on the processor 910. When the processor 910 executes the program, it implements the steps in the aforementioned method embodiments. The computing device provided in this embodiment can be used to execute the technical solutions of the above-described method embodiments, and its implementation principles and technical effects can be further referenced to the method.

[0203] One embodiment of this application provides a computer storage medium storing a program, the stored program including methods that can be loaded by a processor and processed in any of the above embodiments.

[0204] One embodiment of this application provides a computer storage medium, which is a readable storage medium storing a program. The stored program includes methods that can be loaded by a processor and processed in any of the above embodiments.

[0205] Those skilled in the art will understand that all or part of the functions of the various methods in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to achieve the above functions. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be achieved. In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the program can also be stored in a server, another computer, disk, optical disk, flash drive, or external hard drive, etc., and can be downloaded or copied to the memory of a local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be achieved.

[0206] The above examples illustrate this application only to aid understanding and are not intended to limit its scope. Those skilled in the art to which this application pertains can make various simple deductions, modifications, or substitutions based on the ideas presented.

Claims

1. A method for predicting the quantity of business object demand, characterized in that, The method includes: Based on the impact of the results achieved by executing a specific processing flow on the results of other different processing flows, a hierarchical structure is established for each processing flow in the target scenario; the processing flow is the process of running a predetermined program to process a business object; the target scenario contains multiple processing flows. In response to a user's selection command, select the target processing flow from the plurality of processing flows; The target processing flow processes the multi-dimensional related features generated by the business object before the current time node to form a historical time node sequence; the multi-dimensional related features include, but are not limited to, the number of business objects, the resources consumed in processing the business objects, time information, and scene environment; The historical time node sequence is input into a pre-established single-process prediction model. The single-process prediction model is used to predict the number of business objects that the target processing process will process at least at a future time node after the current time, which serves as the basic prediction information. The predicted quantity is input into a harmonic model pre-trained based on the hierarchical structure. The harmonic model is then used to adjust the predicted quantity with the inferred quantity to obtain the harmonic prediction vector. The inferred quantity is the number of objects that other processing flows in the hierarchical structure will process and consume at future time nodes. The method further includes: training the harmonic model using multiple samples based on the hierarchical structure; the process of training the harmonic model using a single sample includes: Based on the characteristics generated by the target processing flow in processing the consumed object at the first time node, predict the first value of the target processing flow in processing the consumed object at the second time node, and collect the second value of the target processing flow in processing the consumed object at the second time node. Based on the characteristics of the consumed object processed by at least one other processing flow at the first time node, predict the third value of the consumed object processed by the other processing flow at the second time node, and collect the fourth value of the consumed object processed by the other processing flow at the second time node. According to the hierarchical structure, the third value and the first value are combined to form a prediction vector; Initial weights are set for different processing flows and different time points in the hierarchical structure, and a relationship matrix is ​​constructed. The prediction vector is reset using the relation matrix to obtain a reset prediction vector; the reset prediction vector includes a reset third value and a reset first value. Combining the second and fourth values ​​yields the marker sequence; The loss function is calculated based on the reset prediction vector and the label sequence, and the weight parameters in the relation matrix are adjusted according to the loss function until the loss function converges.

2. The method according to claim 1, characterized in that, The loss function is calculated based on the reset prediction vector and the labeled sequence, including: （1） in, It is the reset prediction vector. It is a labeled sequence; It is the sequence of the first value after reset corresponding to the second time node. It is the sequence of the third value after reset, corresponding to the second time node. Represents the loss function. This indicates the sequence of the currently used sample. It represents the total number of samples.

3. The method according to claim 1, characterized in that, The harmonic model is obtained by training based on the hierarchical structure; The process of training the harmonic model includes: Initial weights are set for different processing flows and different time points in the hierarchical structure, and a relationship matrix is ​​constructed. Using the relation matrix as an encoder and the transformation matrix as a decoder, a deep learning framework is obtained. The harmonic model is obtained by training based on the deep learning framework.

4. The method according to claim 1, characterized in that, The historical timeline is input into a pre-established single-process prediction model. The single-process prediction model then predicts the number of objects that the target processing flow will process at least at a future time after the current time, including: Responding to user settings, determine the predicted future time points; Set up a basic prediction network, set the prediction output structure of the basic prediction network according to the future time nodes, and set the retrospective output structure of the basic prediction network according to the node information in the historical time node sequence. Input the historical time node sequence into any basic prediction network, output the first prediction information according to the prediction output structure, and output the first review information according to the review output structure; The first review information is input into the basic prediction network corresponding to the next dimension feature, and the second prediction information and the second review information are output. The second review information is input into the next basic prediction network until the outputs of all basic prediction networks are obtained. By combining the prediction information output by each basic prediction network, the predicted number of business objects to be processed by the target processing flow at least at a future time node after the current time is obtained.

5. The method according to claim 4, characterized in that, The method further includes: For each basic prediction network, the difference between the output of the retrospective information and the input is calculated to obtain the residual information filtered by that basic prediction network. The prediction information, combined with the output of each base prediction network, includes: When the number of basic prediction networks is less than a preset number, the predicted number of objects to be processed by the target processing flow at least at a future time node after the current time is obtained by combining the prediction information output by each basic prediction network and the residual information.

6. The method according to claim 4, characterized in that, The single-process prediction model includes multiple basic prediction networks; the basic prediction network includes a time-related feature extraction network layer and a non-time-related feature extraction network layer; the time-related feature extraction network layer is constructed based on LSTM results, and the non-time-related feature extraction network layer is constructed based on MLP.

7. The method according to claim 6, characterized in that, Input the historical time node sequence into any basic prediction network, output first prediction information according to the prediction output structure, and output first review information according to the review output structure, including: The historical time node sequence is calculated by the time-related feature extraction network layer, and the first latent variable features associated with time in the historical time node sequence are filtered out. When time-independent features exist in the historical time node sequence, a non-time-related feature extraction network layer is activated to filter out second latent variable features in the historical time node sequence that are not related to time; time-independent features refer to features that do not have a linear relationship with time. The first latent variable feature is enhanced to obtain a first enhanced feature, and the second latent variable feature is enhanced to obtain a second enhanced feature; The features obtained by fusing the first enhancement feature and the second enhancement feature are used to calculate the first prediction information and the first review information.

8. A device for predicting the quantity of business object demand, characterized in that, The device is disposed in an electronic device, the device comprising: The hierarchical structure building module is used to establish a hierarchical structure of each processing flow in the target scenario based on the impact of the result of executing a specific processing flow on the results of other different processing flows; the processing flow is the process of running a predetermined program to process a business object; the target scenario contains multiple processing flows; The process selection module is used to select the target processing flow from the multiple processing flows in response to the user's input selection command. The feature acquisition module is used to collect multi-dimensional related features generated by the target processing flow in processing the business object before the current time node, forming a historical time node sequence; the multi-dimensional related features include, but are not limited to, the number of business objects, the resources consumed in processing the business objects, time information, and scene environment; The sequence input module is used to input the historical time node sequence into a pre-established single-process prediction model, and predict the number of business objects to be processed by the target processing process at least at a future time node after the current time through the single-process prediction model, as the basic prediction information. An adjustment module is used to input the predicted quantity into a harmonic model pre-trained based on the hierarchical structure, and adjust the predicted quantity using the inferred quantity through the harmonic model to obtain the harmonic prediction vector; the inferred quantity is the number of objects consumed by other processing flows in the hierarchical structure at future time nodes. The adjustment module is further configured to: train the harmonic model using multiple samples based on the hierarchical structure; the process of training the harmonic model using a single sample includes: Based on the characteristics generated by the target processing flow in processing the consumed object at the first time node, predict the first value of the target processing flow in processing the consumed object at the second time node, and collect the second value of the target processing flow in processing the consumed object at the second time node. Based on the characteristics of the consumed object processed by at least one other processing flow at the first time node, predict the third value of the consumed object processed by the other processing flow at the second time node, and collect the fourth value of the consumed object processed by the other processing flow at the second time node. According to the hierarchical structure, the third value and the first value are combined to form a prediction vector; Initial weights are set for different processing flows and different time points in the hierarchical structure, and a relationship matrix is ​​constructed. The prediction vector is reset using the relation matrix to obtain a reset prediction vector; the reset prediction vector includes a reset third value and a reset first value. Combining the second and fourth values ​​yields the marker sequence; The loss function is calculated based on the reset prediction vector and the label sequence, and the weight parameters in the relation matrix are adjusted according to the loss function until the loss function converges.

9. A computing device, characterized in that, It includes at least one processor and at least one memory connected to the processor; the memory stores program instructions for execution by the processor, which, when invoked, can perform the method as described in any one of claims 1 to 7.

Images