CN110135647A - A control method and control device for realizing trend prediction based on feature modeling - Google Patents

Abstract

The present invention provides a kind of control methods that trend prediction is realized based on feature modeling, it is based on dynamic acquisition data, summarizes rule modeling, and be applied to external data to realize trend prediction in modeling, include the following steps: that the date determines multiple contextual informations at one or more dates to a. based on one or more；B. each contextual information situational model corresponding with the contextual information is matched one by one, one or more prediction results under each situational model state are obtained, the prediction result includes at least a core-prediction result and multiple auxiliary prediction results；C. final prediction result is determined according to a core-prediction result and multiple auxiliary prediction results.The present invention is easy to use, high-efficient, forecasting accuracy is high, has high commercial value.

Description

A control method and control device for realizing trend prediction based on feature modeling

technical field

The invention belongs to the field of software technology architecture, and relates to a trend prediction model developed based on machine learning technology, in particular to a control method and a control device for realizing trend prediction based on feature modeling.

Background technique

Most of the operation and maintenance work is done manually by operation and maintenance personnel. With the rapid expansion of Internet business and high labor costs, it is difficult for human operation and maintenance to maintain daily operation. Therefore, by creating scripts that can be automatically triggered and predefined rules to perform common and repetitive operation and maintenance tasks to achieve automated operation and maintenance, it is considered an expert system based on industry domain knowledge and operation and maintenance scenario domain knowledge . Automated operation and maintenance builds an operation and maintenance automation platform with the core of orchestration operations, which can provide customers with functions such as configuration changes, task inspections, script execution control, and custom workflows; covering inspections, file distribution, backup recovery, SQL operations, etc. Operation and maintenance scenarios, and provide scalability; the automated operation and maintenance system configures batch jobs on the server side, schedules scripts according to policies, and realizes automated operations of business flows; the console is the operating platform for each manager, realizing roles, objects, and operations hierarchical and decentralized management. Build a unified device model on the basis of CMDB to realize automatic delivery of resources and services.

From its birth to the present, IT operation and maintenance has experienced manual operation and maintenance to the current automatic operation and maintenance, and has been able to gradually adapt to complex businesses, diverse user needs, and continuously expanding IT applications to ensure that IT services are flexible, convenient, safe, stable, and automated. Operation and maintenance, as a starting point to replace manual operations, is being widely researched and applied.

O&M automation is a set of strategies that convert static equipment structures into dynamic and elastic responses based on IT service requirements. The purpose is to achieve the quality of IT O&M and reduce costs. Automated operation and maintenance is mainly to create scripts that can be automatically triggered and pre-defined rules to perform common and repetitive operation and maintenance tasks to achieve automated operation and maintenance. But once a problem that is not predefined occurs, it needs to be perceived, checked, and located manually. This kind of operation and maintenance service requires a lot of manpower. Due to lack of manpower or due to investigation and positioning, the service will be suspended for a long time, and the stable operation of the system cannot be achieved. This deficiency has become increasingly prominent. For example, in business activities or special time points, resource allocation problems are caused due to insufficient estimation of business volume and untimely response of personnel, resulting in abnormal business operation

However, at present, there is no specific method on the market that can effectively solve the above problems, especially a control method and a control device for realizing trend prediction based on feature modeling.

Contents of the invention

Aiming at the technical defects existing in the prior art, the object of the present invention is to provide a control method and control device for realizing trend prediction based on feature modeling, which is based on dynamically acquiring data, summarizing rule modeling, and applying external data to modeling To achieve trend forecasting, including the following steps:

a. Determining a plurality of context information on one or more dates based on the one or more dates;

b. Match each of the situation information with the situation model corresponding to the situation information one by one, and obtain one or more prediction results in each situation model state, and the prediction results include at least one core prediction result and multiple Auxiliary prediction results;

c. Determine the final prediction result according to one core prediction result and multiple auxiliary prediction results.

Preferably, after the step b, a step d is further included: dynamically displaying the final prediction result.

Preferably, in the step a, the context information at least includes:

XGBoost basic model information;

Holiday model information;

week model information; and

Marketing campaign information.

Preferably, said step b includes the following steps:

b1. Determine the core prediction result matching the XGBoost basic model information based on the XGBoost basic model information;

b2. Determining a first auxiliary prediction result that matches the holiday model information based on the holiday model information;

b3. Determining a second auxiliary prediction result that matches the week model information based on the week model information;

b4. Determine a third auxiliary prediction result that matches the marketing activity model information based on the marketing activity information.

Preferably, the step c includes c1: adjusting the core prediction result based on the first auxiliary prediction result, the second auxiliary prediction result and the third auxiliary prediction result to determine a final prediction result.

Preferably, the establishment of the situation model will be obtained through the following methods:

i: Obtain one or more original external information on one or more data collection platforms based on one or more data collection methods;

ii: Preprocessing the one or more original external information, and determining one or more preprocessed external information;

iii: performing feature extraction on the one or more preprocessed external information, and determining one or more context feature information;

iv: Based on the one or more context information models, determine one or more context models that match the context feature information.

Preferably, the data collection methods include at least any one or more of the following methods:

SQL export;

third-party API export; or

Reptile collection.

Preferably, the data collection platform includes at least any one or more of the following platforms:

Operation and maintenance dispatch system;

configuration management database;

application performance monitoring;

Kanban system; or

Unified logging platform.

Preferably, the original external information includes at least any one or more of the following information:

history information;

Kilometer, lunar calendar information;

holiday information; or

Marketing campaign data.

Preferably, said preprocessing includes attachment collection, data screening and data encoding.

According to another aspect of the present invention, a control device for realizing trend prediction based on feature modeling is provided, including:

The first determining means 1: based on one or more dates, determine a plurality of context information in which the one or more dates are located;

The first acquiring means 2: matching each of the context information with the context model corresponding to the context information one by one, and acquiring one or more prediction results under the state of each context model;

The second determining means 3: determine the final prediction result according to a core prediction result and multiple auxiliary prediction results.

The first processing device 4: dynamically display the final prediction result.

Preferably, the first acquiring device 2 includes:

The third determining means 21: determine the core prediction result matching the XGBoost basic model information based on the XGBoost basic model information;

The fourth determination means 22: determine a first auxiliary prediction result matching the holiday model information based on the holiday model information;

Fifth determining means 23: determining a second auxiliary prediction result matching the week model information based on the week model information;

The sixth determining means 24: determine a third auxiliary prediction result that matches the marketing activity model information based on the marketing activity information.

Preferably, it also includes:

The second obtaining means 5: obtain one or more original external information on one or more data collection platforms based on one or more data collection methods;

The seventh determining means 6: perform preprocessing on the one or more original external information, and determine one or more preprocessed external information;

The eighth determining means 7: performing feature extraction on the one or more preprocessed external information, and determining one or more context feature information;

Ninth determining means 8: Based on the one or more situational information models, determine one or more situational models that match the situational feature information.

The present invention provides a method that does not rely on human-specified rules, and uses machine learning algorithms to automatically learn from massive business data, continuously refine and summarize rules, and form trend predictions. That is, on the basis of automated operation and maintenance, a brain based on machine learning is added, and the command monitoring system collects the data required for brain decision-making, makes business data predictions, and directs the container platform to complete the decision-making actions of elastic scaling, so as to achieve The overall goal of the operation and maintenance system. Combining big data visualization technology to display business operation status to business and technical parties in real time, provide future forecast operation data, enable the operation and maintenance team to do capacity management and performance management in advance, and as the central brain of the container cloud platform, direct the relevant container management platform to complete In the decision-making action of elastic scaling, the present invention determines multiple context information on one or more dates based on one or more dates, and performs each context model corresponding to the context information one by one. Matching, obtaining one or more prediction results in each situation model state, the prediction results include at least one core prediction result and multiple auxiliary prediction results, and determining the final prediction result according to one core prediction result and multiple auxiliary prediction results, The invention is convenient to use, high in efficiency, high in prediction accuracy, and has extremely high commercial value.

Description of drawings

Other characteristics, objects and advantages of the present invention will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings:

FIG. 1 shows a specific structural schematic diagram of a control method for realizing trend prediction based on feature modeling according to a specific embodiment of the present invention;

Fig. 2 shows the first embodiment of the present invention, each of the context information is matched with the context model corresponding to the context information one by one, and one or more prediction results in each context model state are obtained Schematic diagram of the specific process;

Fig. 3 shows the specific flowchart of establishing the situation model according to the second embodiment of the present invention;

Fig. 4 shows another specific embodiment of the present invention, a schematic diagram of module connection of a control device for realizing trend prediction based on feature modeling; and

Fig. 5 shows a schematic diagram of module connection of a control device for realizing trend prediction based on feature modeling according to the third embodiment of the present invention.

Detailed ways

In order to better clearly express the technical solution of the present invention, the present invention will be further described below in conjunction with the accompanying drawings.

Fig. 1 shows a specific structural diagram of a control method based on feature modeling to realize trend prediction according to a specific embodiment of the present invention. Specifically, it is based on dynamically acquiring data, summarizing rule modeling, and applying external data to Realize trend prediction in modeling, including the following steps:

First, enter step S101, determine a plurality of context information on one or more dates based on one or more dates, in such an embodiment, the context information includes at least XGBoost basic model information, holiday model information, Weekly model information as well as marketing campaign information. Those skilled in the art understand that the present invention judges according to historical business data when selecting feature data, such as business volume has obvious periodicity, including week, month and year cycle; There are obvious differences in the degree of influence; the overall fluctuation is relatively stable, and the peaks and valleys caused by business activities are not obvious. Based on the above, we have selected the following features: 1. Gregorian month and day; 2. Week 3. Whether it is a certain day during the Spring Festival holiday 4. Whether it is a certain day during the National Day holiday 5. Whether it is a certain day during the small long holiday, that is, here, we can choose The date of a single day is used as the input to determine the situational information, and a time period can also be selected as the input. At the same time, multiple time periods and multiple moments can also be selected as the input, and we can determine the date corresponding to the date based on the above-mentioned dates situational information.

Then, enter step S102, match each of the situation information with the situation model corresponding to the situation information one by one, and obtain one or more prediction results under the state of each situation model, and the prediction results include at least one core A forecast result and a plurality of auxiliary forecast results. In such an embodiment, each context information corresponding to the date is matched with the context model to obtain one or more forecast results in each context model state . The prediction results include at least one or more of the prediction results of XGBoost basic model information, the prediction results of holiday model information, the prediction results of week model information, and the prediction results of marketing activity information, and the prediction results include at least one core prediction results and multiple auxiliary prediction results, and in this embodiment, the prediction results of the XGBoost basic model information are the core prediction results, and the prediction results of holiday model information, week model information and marketing activity information The prediction result of is the auxiliary prediction result.

Then, enter step S103, and determine the final prediction result according to a core prediction result and multiple auxiliary prediction results. In such an embodiment, the final prediction result determined by the multiple auxiliary prediction results can be converted into a certain coefficient for the The calibration of the core forecast results can also be calculated on the basis of the core forecast results according to the relationship of the weight ratio, and finally the final forecast result is obtained.

Finally, enter step S104, and dynamically display the final prediction result. Those skilled in the art understand that the final prediction result will be displayed in the form of a table, an image, etc., that is, in such an embodiment, we can set Enter the specified date or time period in the intelligent operation terminal of the computer, and display the final forecast results after calculation, and the displayed results may change according to the continuously input data or continuously added historical data, so that the final forecast results In a dynamic display screen. Intelligent operation and maintenance does not rely on human-specified rules, and advocates that machine learning algorithms automatically learn from massive operation and maintenance data (including the event itself and manual processing logs of operation and maintenance personnel), and continuously refine and summarize rules. That is, on the basis of automated operation and maintenance, a brain based on machine learning is added, and the command and monitoring system collects the data required for brain decision-making, makes analysis and decision-making, and directs automated scripts to execute brain decisions, so as to achieve operation and maintenance. overall goal of the system.

Fig. 2 shows the first embodiment of the present invention, each of the context information is matched with the context model corresponding to the context information one by one, and one or more prediction results in each context model state are obtained The specific flow diagram, as understood by those skilled in the art, the above-mentioned Figure 2 is the expansion step of the step S102, more specifically, includes the following steps:

First, enter step S1021, and determine the core prediction result matching the XGBoost basic model information based on the XGBoost basic model information. In such an embodiment, the establishment of the XGBoost basic model will be described later in the specific implementation It is further described in , and the information of the XGBoost basic model is substituted into the XGBoost basic model, and a matching operation is performed to finally determine the core prediction result.

Then, enter step S1022, determine the first auxiliary prediction result matching the holiday model information based on the holiday model information, the first auxiliary prediction result is the result obtained based on the holiday model, and the The establishment of the above-mentioned holiday model will be further described in the specific implementation manner described later.

Then, enter step S1023, determine the second auxiliary prediction result matching the week model information based on the week model information, the second auxiliary prediction result is the result obtained based on the week model, and The establishment of the week model will be further described in the specific implementation manner described later.

Finally, enter step S1024, determine the third auxiliary prediction result matching the marketing activity model information based on the marketing activity information, the third auxiliary prediction result is the result obtained based on the marketing activity model, The establishment of the marketing activity model will be further described in the specific implementation manner described later.

Preferably, the core prediction result is adjusted based on the first auxiliary prediction result, the second auxiliary prediction result and the third auxiliary prediction result to determine a final prediction result. Those skilled in the art understand that industry policies and business activities of insurance companies have a great impact on business volume, and it is difficult to collect corresponding data. However, property insurance is much less affected by such effects than life insurance. The auto insurance business of Pacific Insurance has been relatively stable, with few marketing activities, and the reporting and closing of cases are mainly affected by customer activities, and have little to do with enterprise-side factors. When modeling the amount of closed property insurance cases, factors such as policy and marketing are not considered for the time being. In parameter tuning, that is, how to adjust the core prediction result based on the first auxiliary prediction result, the second auxiliary prediction result and the third auxiliary prediction result, and determine the final prediction result, this study uses The parameter tuning of the Xgboost model mainly refers to booster parameters, including:

First: eta [default 0.3], similar to the learning rate parameter in GBM, by reducing the weight of each step, the robustness of the model can be improved. Wherein the typical value is 0.01-0.2.

Second: min_child_weight[default 1], which determines the minimum leaf node sample weight sum. Similar to GBM's min_child_leaf parameter, but not exactly the same. This parameter of XGBoost is the sum of the minimum sample weights, while the GBM parameter is the minimum total number of samples. This parameter is used to avoid overfitting. When its value is large, it can prevent the model from learning local special samples. But if this value is too high, it will lead to underfitting. This parameter needs to be tuned using CV.

Third: max_depth [default 6], the same as the parameters in GBM, this value is the maximum depth of the tree. This value is also used to avoid overfitting. The larger max_depth, the model will learn more specific and local samples. The CV function needs to be used for tuning. Among them, the typical value: 3-10.

Fourth: max_leaf_nodes, the largest number of nodes or leaves on the tree. Can replace the role of max_depth. Because if a binary tree is generated, a tree with a depth of n can generate at most n2 leaves. If this parameter is defined, GBM ignores the max_depth parameter.

Fifth: gamma [default 0] When a node is split, the node will only be split if the value of the loss function decreases after the split. Gamma specifies the minimum loss function drop value required for node splits. The larger the value of this parameter, the more conservative the algorithm. The value of this parameter is closely related to the loss function, so it needs to be adjusted.

Sixth: max_delta_step[default 0], this parameter limits the maximum step size of each tree weight change. If the value of this parameter is 0, it means that there is no constraint. If it is given some positive value, it makes the algorithm more conservative. Normally, this parameter does not need to be set. But it is very helpful for logistic regression when the samples of each category are very unbalanced. This parameter is generally not used, but you can find more uses for it.

Seventh: subsample [default 1], which is exactly the same as the subsample parameter in GBM. This parameter controls the proportion of random sampling for each tree. Decreasing the value of this parameter will make the algorithm more conservative and avoid overfitting. However, if this value is set too small, it may cause underfitting. Typical value: 0.5-1.

Eighth: colsample_bytree [default 1], similar to the max_features parameter in GBM. Used to control the ratio of the number of columns randomly sampled per tree (each column is a feature). Typical value: 0.5-1.

Ninth: colsample_bylevel [default 1], used to control each split of each level of the tree, the proportion of the sampling of the number of columns. The subsample parameter and the colsample_bytree parameter can play the same role.

Tenth: lambda [default 1], the L2 regularization term of the weight. (Similar to Ridge regression). This parameter is used to control the regularization part of XGBoost. Although most data scientists rarely use this parameter, this parameter can still find more uses in reducing overfitting.

Eleventh: alpha [default 1], the L1 regularization term of the weight. (Similar to Lasso regression). It can be applied in very high-dimensional situations, making the algorithm faster.

Twelfth: scale_pos_weight [default 1], when the samples of each category are very unbalanced, setting this parameter to a positive value can make the algorithm converge faster.

Further, the model is integrated. After the basic model is generated according to the xgboost algorithm, the prediction of the model for the Spring Festival, National Day, small holiday, week and other patterns, the trend is basically correct, and then based on historical data, the week model, the Spring Festival model, the National Day model, Small holiday model, used to make further adjustments to the output of the base model.

Fig. 3 shows the specific flowchart of establishing the situation model according to the second embodiment of the present invention, and the establishment of the situation model will be obtained by the following methods:

First, enter step S201, and obtain one or more original external information on one or more data collection platforms based on one or more data collection methods, the data collection methods include at least SQL export, third-party API export or crawler collection, The data collection platform includes at least an operation and maintenance dispatch system, a configuration management database, application performance monitoring, a Kanban system or a unified log platform. The original external information includes at least historical information, kilometers, lunar calendar information, holiday information or marketing activity data.

Then, enter step S202, perform preprocessing on the one or more original external information, and determine one or more preprocessed external information, preferably, the preprocessing includes attachment collection, data screening and data encoding. Part of the data is placed in the attachment. Although these data are not large in number, they are more important. This system downloads attachment information from the system in the form of a crawler and adds it to the relevant program.

Then, enter step S203, perform feature extraction on the one or more pre-processed external information, and determine one or more context feature information, and the present invention selects features by combining business knowledge and data mining. First, collect sample cases and possible related features by visiting business personnel, and then calculate the correlation between each feature and the corresponding variable, and select suitable candidate features from them. In the present invention, the Pearson correlation coefficient and the mutual information coefficient are mainly used to measure the single-variable correlation. After obtaining the candidate features, train the model based on these features, and evaluate the significance of the features according to the accuracy and recall of the model.

Finally, enter step S204, and determine one or more situational models matching the situational feature information based on the one or more situational information models. Those skilled in the art understand that the situation model includes XGBoost basic model, holiday model information, week model information and marketing activity model information. Based on the foregoing steps S201 to S203, one or more situation models matching the situation characteristic information are determined.

Fig. 4 shows another specific embodiment of the present invention, a schematic diagram of module connection of a control device for trend prediction based on feature modeling, including a first determination device 1, which determines one or more dates based on one or more dates For the multiple context information of the date, the working principle of the first determining device 1 can refer to the aforementioned step S101, which will not be repeated here.

Further, it also includes a first obtaining means 2: matching each of the situation information with the situation model corresponding to the situation information one by one, and obtaining one or more prediction results under the state of each situation model, the first For the working principle of the acquiring device 2, reference may be made to the aforementioned step S102, which will not be repeated here.

Further, a second determination means 3 is also included: determining the final prediction result according to a core prediction result and multiple auxiliary prediction results. The working principle of the second determination means 3 can refer to the aforementioned step S103, which will not be repeated here.

Further, it also includes a first processing device 4: dynamically displaying the final prediction result, the working principle of the first processing device 4 can refer to the aforementioned step S104, which will not be repeated here.

Further, the first acquiring device 2 includes a third determining device 21, which determines the core prediction result matching the XGBoost basic model information based on the XGBoost basic model information, and the working principle of the third determining device 21 can refer to the aforementioned Step S1021 will not be described in detail here.

Further, the first obtaining device 2 also includes a fourth determining device 22, which determines the first auxiliary prediction result matching the holiday model information based on the holiday model information, and the working principle of the fourth determining device 22 can be referred to The aforementioned step S1022 will not be described in detail here.

Further, the first acquiring device 2 also includes a fifth determining device 23, which determines a second auxiliary prediction result that matches the week model information based on the week model information. The working principle of the fifth determining device 23 is Reference may be made to the aforementioned step S1023, which will not be described in detail here.

Further, the first obtaining means 2 also includes a sixth determining means 24, which determines a third auxiliary prediction result matching the marketing activity information based on the marketing activity information, and the working principle of the sixth determining means 24 Reference may be made to the aforementioned step S1024, which will not be described in detail here.

Fig. 5 shows a schematic diagram of module connection of a control device for realizing trend prediction based on feature modeling according to the third embodiment of the present invention, and also includes a second acquisition device 5: based on one or more data acquisition methods in one One or more data collection platforms obtain one or more original external information, and the working principle of the second obtaining device 5 can refer to the aforementioned step S201, which will not be repeated here.

Further, it also includes a seventh determining means 6: preprocessing the one or more original external information, and determining one or more preprocessed external information, the working principle of the seventh determining means 6 can refer to the aforementioned step S202, in This will not be repeated.

Further, it also includes an eighth determining means 7: performing feature extraction on the one or more pre-processed external information, and determining one or more context feature information; the working principle of the eighth determining means 7 can refer to the aforementioned step S203, in This will not be repeated.

Further, it also includes a ninth determination means 8: based on the one or more context information models, determine one or more context models that match the context feature information, the working principle of the ninth determination means 8 Reference may be made to the aforementioned step S204, which will not be described in detail here.

Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention.

Claims (13)

1. A control method based on feature modeling to realize trend forecasting, which is based on dynamic acquisition of data, summary rule modeling, and applying external data in modeling to realize trend forecasting, characterized in that, comprising the steps: a. Determining a plurality of context information on one or more dates based on the one or more dates; b. Match each of the situation information with the situation model corresponding to the situation information one by one, and obtain one or more prediction results in each situation model state, and the prediction results include at least one core prediction result and multiple Auxiliary prediction results; c. Determine the final prediction result according to one core prediction result and multiple auxiliary prediction results. 2. The control method according to claim 1, characterized in that, after the step b, further comprising a step d: dynamically displaying the final prediction result. 3. The control method according to claim 1 or 2, characterized in that, in the step a, the context information at least includes: XGBoost basic model information; Holiday model information; week model information; and Campaign model information. 4. The control method according to claim 3, wherein said step b comprises the steps of: b1. Determine the core prediction result matching the XGBoost basic model information based on the XGBoost basic model information; b2. Determining a first auxiliary prediction result that matches the holiday model information based on the holiday model information; b3. Determining a second auxiliary prediction result that matches the week model information based on the week model information; b4. Determine a third auxiliary prediction result that matches the marketing activity model information based on the marketing activity information. 5. The control method according to claim 4, wherein the step c comprises c1: based on the first auxiliary prediction result, the second auxiliary prediction result and the third auxiliary prediction result, the The core forecast results are adjusted to determine the final forecast result. 6. The control method according to claim 5, characterized in that, the establishment of the situational model will be obtained in the following manner: i: Obtain one or more original external information on one or more data collection platforms based on one or more data collection methods; ii: Preprocessing the one or more original external information, and determining one or more preprocessed external information; iii: performing feature extraction on the one or more preprocessed external information, and determining one or more context feature information; iv: Based on the one or more context information models, determine one or more context models that match the context feature information. 7. The control method according to claim 6, characterized in that, the data collection methods at least include any one or more of the following methods: SQL export; third-party API export; or Reptile collection. 8. The control method according to claim 6, wherein the data collection platform includes at least any one or more of the following platforms: Operation and maintenance dispatch system; configuration management database; application performance monitoring; Kanban system; or Unified logging platform. 9. The control method according to claim 6, wherein the original external information at least includes any one or more of the following information: history information; Kilometer, lunar calendar information; holiday information; or Marketing campaign data. 10. The control method according to claim 6, characterized in that the preprocessing includes attachment collection, data screening and data encoding. 11. A control device for realizing trend prediction based on feature modeling, characterized in that it comprises: First determining means (1): based on one or more dates, determine a plurality of context information in which one or more dates are located; The first obtaining means (2): matching each of the situation information with the situation model corresponding to the situation information one by one, and obtaining one or more prediction results under the state of each situation model; The second determination means (3): determine the final prediction result according to a core prediction result and multiple auxiliary prediction results. The first processing device (4): dynamically display the final prediction result. 12. The control device according to claim 11, characterized in that, the first obtaining device (2) comprises: The third determining means (21): determine the core prediction result matching the XGBoost basic model information based on the XGBoost basic model information; A fourth determining means (22): determining a first auxiliary prediction result matching the holiday model information based on the holiday model information; Fifth determining means (23): determining a second auxiliary prediction result matching the week model information based on the week model information; Sixth determining means (24): determining a third auxiliary prediction result that matches the marketing activity model information based on the marketing activity information. 13. The control device according to claim 11, further comprising: The second obtaining device (5): obtain one or more original external information on one or more data collection platforms based on one or more data collection methods; The seventh determining means (6): perform preprocessing on the one or more original external information, and determine one or more preprocessed external information; The eighth determining device (7): performing feature extraction on the one or more pre-processed external information, and determining one or more context feature information; Ninth determining means (8): based on the one or more situational information models, determine one or more situational models that match the situational feature information.