CN-121998333-A - Clothing demand prediction method and system based on hybrid modeling framework

CN121998333ACN 121998333 ACN121998333 ACN 121998333ACN-121998333-A

Abstract

The invention discloses a clothing demand prediction method and system based on a hybrid modeling framework, and relates to the technical field of demand prediction. The method comprises the steps of carrying out time sequence feature decomposition on a target data set for clothing demand pre-judging under a specific time sequence data monitoring scene, obtaining co-occurrence relation data for reflecting the association tightness degree between different data in the time sequence feature decomposition process, importing the obtained co-occurrence relation data into a mixed frame based on a fusion LightGBM model and a GRU model, carrying out differential demand prediction, outputting demand predicted values corresponding to different time periods and different data dimensions and reflecting clothing demands, carrying out scene adaptation effect evaluation according to the demand predicted values, and judging whether to output final demand predicted values or not by monitoring dynamic adaptation states of Poisson distribution parameters of demand sequences corresponding to the different data dimensions and the different time periods in the scene adaptation effect evaluation process so as to accurately capture real demand rules of different types of clothing.

Inventors

ZHANG DAI
CHENG LIANG
WANG HAILONG
YANG ZEHAO
Pang Zhongjie

Assignees

广州未来一手网络科技有限公司

Dates

Publication Date: 20260508
Application Date: 20260122

Claims (10)

1. A garment demand prediction method based on a hybrid modeling framework, the method comprising: Under a specific time sequence data monitoring scene, performing time sequence feature decomposition on a target data set for garment demand pre-judgment, and acquiring co-occurrence relation data for reflecting the association tightness degree between different data in the time sequence feature decomposition process; importing the acquired co-occurrence relation data into a mixed framework based on a fusion LightGBM model and a GRU model, carrying out differential demand prediction, and outputting demand prediction values corresponding to different time periods and different data dimensions and reflecting clothing demands; and (3) carrying out scene adaptation effect evaluation according to the demand predicted value, and judging whether to output the final demand predicted value by monitoring dynamic adaptation states of the demand sequence poisson distribution parameters corresponding to different data dimensions and different time periods in the scene adaptation effect evaluation process.
2. The garment demand prediction method based on a hybrid modeling framework according to claim 1, wherein the time series feature decomposition is used for splitting periodic fluctuation, trend components and random interference items of data in a target data set; the specific acquisition process of the co-occurrence relation data comprises the following steps: smoothing the acquired clothing product data by utilizing a sliding window filtering algorithm to obtain a stabilized time sequence data set, and carrying out frequency domain decomposition on the stabilized time sequence data set based on Fourier transformation to obtain periodic characteristic data; performing trend fitting and component separation on the periodic characteristic data by combining a least square method on a time domain to obtain trend characteristic data; Based on a time window division result of the stabilized time sequence data set, screening associated occurrence records corresponding to different trend characteristic data in the same time window, and further splitting to obtain discrete co-occurrence data and non-discrete co-occurrence data; Calculating the time stamp coincidence ratio of the discrete co-occurrence data and the non-discrete co-occurrence data on the same time axis, and summarizing the co-occurrence frequency of the discrete co-occurrence data and the non-discrete co-occurrence data in the same sliding time window, so as to quantify the association tightness degree between the data in different dimensions; Collecting the co-occurrence data corresponding to the co-occurrence frequency which is larger than the reference co-occurrence frequency, and judging that the co-occurrence data has obviously associated co-occurrence data pairs, classifying according to a preset time granularity and a data dimension level to obtain co-occurrence relation data of the same time period and different data dimensions; And summarizing co-occurrence data with the co-occurrence frequency not larger than the reference co-occurrence frequency, summarizing co-occurrence data pairs, judging that the co-occurrence data pairs do not have obvious association, and marking abnormal co-occurrence relation.
3. The garment demand prediction method based on the hybrid modeling framework of claim 2, wherein the specific acquisition process of the co-occurrence relationship data further comprises: If the historical co-occurrence frequency is not greater than the co-occurrence data corresponding to the reference co-occurrence frequency, calculating the time stamp coincidence ratio of the discrete co-occurrence data and the non-discrete co-occurrence data on the same time axis, taking the time stamp coincidence ratio as the input of a GRU model, carrying out short-term time sequence association trend prediction through a time sequence trend extrapolation algorithm, and outputting the co-occurrence frequency of the corresponding co-occurrence data in the next time period; if the co-occurrence frequency does not meet the preset association significance expected requirement, maintaining a judging result that the co-occurrence frequency does not have significant association, otherwise, judging that the co-occurrence frequency has significant association, classifying according to the preset time granularity and the data dimension level, and obtaining co-occurrence relation data of different time periods and different data dimensions.
4. A garment demand prediction method based on a hybrid modeling framework as claimed in claim 2 or 3, wherein said differentiated demand prediction comprises: Performing differential classification on the acquired co-occurrence relation data to obtain high-frequency co-occurrence data and low-frequency co-occurrence data, and performing corresponding demand prediction; The high-frequency co-occurrence data represent co-occurrence relation data with the frequency of co-occurrence of different data dimensions higher than a set threshold value in a preset sliding time window; And the low-frequency co-occurrence data represent co-occurrence relation data with the frequency of co-occurrence of different data dimensions lower than a set threshold value in a preset sliding time window.
5. The method for predicting clothing demand based on a hybrid modeling framework according to claim 4, wherein when the result of the differential classification is that only high-frequency co-occurrence data exists, the specific process of demand prediction is as follows: acquiring the time stamp coincidence degree of the high-frequency co-occurrence data in a sliding time window and a corresponding first-order differential slope, wherein the first-order differential slope represents a first-order differential calculation result of time stamp coincidence degree of the high-frequency co-occurrence data changing along with a time sequence; Taking the time stamp coincidence degree as a horizontal axis and the first-order differential slope as a vertical axis, introducing a pearson correlation coefficient corresponding to the time stamp coincidence degree and the first-order differential slope, performing linear mapping and normalization processing to obtain a dynamic weight coefficient, and generating a growth-time correlation coupling feature matrix for visualizing the high-frequency co-occurrence data growth trend and the time correlation feature; And taking the vector in the generated growth-time correlation coupling feature matrix as input of high-frequency co-scene demand prediction, carrying out demand prediction by relying on LightGBM and the collaborative fitting capacity of the GRU model, outputting demand prediction values of different time periods and different data dimensions under the high-frequency co-scene, and carrying out real-time scene adaptation effect evaluation.
6. The method for predicting clothing demand based on a hybrid modeling framework according to claim 4, wherein when the differential classification result is that only low-frequency co-occurrence data exists, the specific process of demand prediction is as follows: Acquiring the attenuation rate of a required sequence of the low-frequency co-occurrence data in a sliding time window and the corresponding time stamp coincidence ratio; Taking the attenuation rate of the required sequence as a vertical axis and the coincidence degree of the time stamp as a horizontal axis, introducing a pearson correlation coefficient corresponding to the attenuation rate of the required sequence and the coincidence degree of the time stamp, performing linear mapping and normalization processing to obtain a dynamic correction coefficient, and generating an attenuation-time correlation coupling characteristic matrix for visualizing the attenuation trend and the time correlation characteristic of the low-frequency co-occurrence data; and taking the vector in the obtained attenuation-time associated coupling feature matrix as input of low-frequency co-scene demand prediction, carrying out demand prediction through a prediction logic of LightGBM and GRU model hierarchical fitting, outputting demand prediction values of different time periods and different data dimensions under the low-frequency co-occurrence scene, and carrying out real-time scene adaptation effect evaluation.
7. The method for predicting clothing demand based on a hybrid modeling framework as claimed in claim 4, wherein when the result of the differential classification is that high-frequency co-occurrence data and low-frequency co-occurrence data coexist, the specific process of demand prediction is as follows: Respectively obtaining vectors of the associated coupling feature matrixes corresponding to the high-frequency co-occurrence data and the low-frequency co-occurrence data, and forming a double-branch parallel input channel of the high-frequency branch and the low-frequency branch after carrying out standardized processing according to a feature dimension alignment rule; and respectively calculating dynamic weight coefficients corresponding to the characteristic contribution degrees of the high-frequency co-occurrence data and the low-frequency co-occurrence data in the double-branch parallel input channel through LightGBM models, respectively capturing time sequence linear evolution rules of the two types of data through GRU models, outputting a demand predicted value, and carrying out real-time scene adaptation effect evaluation.
8. The clothing demand prediction method based on the hybrid modeling framework according to any one of claims 5 to 7, wherein the specific process of real-time scene adaptation effect evaluation is as follows: if the acquired demand predicted value is larger than the preset demand predicted value, judging that the demand prediction in the corresponding scene is unqualified, and sending a model prediction precision unqualified prompt; if the obtained demand predicted value is not greater than the preset demand predicted value, judging that the demand prediction in the corresponding scene is qualified, obtaining the absolute value of the poisson fitting residual error and the time sequence dimension characteristic ratio, and comprehensively judging whether to output the final clothing demand predicted result.
9. The method for predicting clothing demand based on the hybrid modeling framework according to claim 8, wherein the determining whether to output the final demand predicted value is specifically: if the absolute value of the poisson fitting residual error and the time sequence dimension characteristic ratio are both in the corresponding allowable interval, judging that the prediction precision of the current model and the data adaptation reach the standard, and directly outputting a final demand predicted value; If the single parameter exists in the absolute value of the poisson fitting residual error and the time sequence dimension characteristic ratio and is not in the corresponding allowable interval, the time sequence window length of the corresponding model is finely adjusted based on the offset degree of the corresponding parameter, after the differential demand prediction flow is re-executed, and if the acquired absolute value of the poisson fitting residual error and the acquired time sequence dimension characteristic ratio are both in the corresponding allowable interval, the final demand predicted value is output; If the absolute value of the poisson fitting residual error and the time sequence dimension characteristic ratio are not in the corresponding allowable interval, judging that the prediction precision of the current model and the data adaptation degree are not up to the standard, and sending a demand prediction abnormal prompt.
10. A garment demand prediction system based on a hybrid modeling framework, the system comprising: The co-occurrence relation data acquisition module is used for carrying out time sequence feature decomposition on a target data set for garment demand pre-judgment under a specific time sequence data monitoring scene, and acquiring co-occurrence relation data for reflecting the association tightness degree between different data in the time sequence feature decomposition process; The differentiated demand prediction module is used for importing the obtained co-occurrence relation data into a mixed framework based on a fusion LightGBM model and a GRU model, performing differentiated demand prediction, and outputting demand prediction values corresponding to different time periods and different data dimensions and reflecting clothing demands; The adaptive effect evaluation and judgment module is used for evaluating the scene adaptive effect according to the demand predicted value, and judging whether the final demand predicted value is output or not by monitoring the dynamic adaptive states of the demand sequence poisson distribution parameters corresponding to different data dimensions and different time periods in the scene adaptive effect evaluation process.

Description

Clothing demand prediction method and system based on hybrid modeling framework Technical Field The invention relates to the technical field of demand prediction, in particular to a clothing demand prediction method and system based on a hybrid modeling framework. Background In the clothing industry, demand prediction is a core support link of enterprise production planning, inventory optimization and supply chain coordination, and the prediction accuracy directly determines the fund turnover efficiency and market competitiveness of an enterprise. A method for predicting the demand of clothes includes such steps as obtaining the original data set, generating the countering network model, expanding the original data set to obtain the synthetic data set, merging the original data set with the synthetic data set to become new data set, or directly using the original data set as new data set, building the predicting model of clothes demand, initial training, optimizing the model parameters by black-fin iris optimizing algorithm, reconstructing the demand predicting model, and finally obtaining the mixed predicting model, and predicting clothes demand. In the prior art, more prediction logic leading by a single model is adopted, namely a data processing system for clothing demand prediction is firstly built, through information such as historical sales data, user behavior data, trend data and the like, core data dimensions and association relations such as demand influence factors, clothing characteristics, consumption scenes and the like are built, then multi-source data acquired in real time are accessed, the multi-source data are input into a prediction model after being processed through cleaning, normalization and the like, then a time sequence model such as a gate control circulation unit (GRU, gated Recurrent Unit) or a deep learning model such as a lightweight gradient elevator model (LightGBM, light Gradient Boosting Machine Model) is utilized for demand prediction, a prediction result is adjusted by combining with manual experience, finally a demand prediction conclusion is output, decision basis is provided for production stock preparation and inventory allocation, and the whole flow of demand prediction is completed. However, when facing complex data scenes of various products in the clothing industry, the requirements of which are influenced by multiple factors such as seasons, tides and the like, a single time sequence method represented by GRU generally depends on linear time sequence characteristics of historical data, so that prediction precision of different products is uneven, layering adaptation and dynamic iteration capability of different products in product data management are insufficient, meanwhile, as the existing product data management lacks an effective association mechanism of non-time sequence data and basic sales data, the multi-dimensional characteristic extraction coverage of the single model on the product data is not high, if discrete data such as size distribution, color preference and the like are encountered, prediction errors of core demand data are aggravated, and real demand rules of different products are difficult to accurately capture. Disclosure of Invention In order to solve the technical problems in the prior art, the embodiment of the invention provides a clothing demand prediction method and a clothing demand prediction system based on a hybrid modeling framework. The technical scheme is as follows: On the one hand, the clothing demand prediction method based on the hybrid modeling framework comprises the steps of carrying out time sequence feature decomposition on a target data set for clothing demand prediction under a specific time sequence data monitoring scene, obtaining co-occurrence relation data for reflecting the association tightness degree between different data in the time sequence feature decomposition process, importing the obtained co-occurrence relation data into the hybrid framework based on a fusion LightGBM model and a GRU model to carry out differential demand prediction, outputting demand prediction values corresponding to different time periods and different data dimensions to reflect clothing demands, carrying out scene adaptation effect assessment according to the demand prediction values, and judging whether to output final demand prediction values by monitoring dynamic adaptation states of demand sequence Poisson distribution parameters corresponding to different data dimensions and different time periods in the scene adaptation effect assessment process. On the other hand, the clothing demand prediction system based on the mixed modeling framework comprises a co-occurrence relation data acquisition module, a differentiation demand prediction module and an adaptation effect evaluation and judgment module, wherein the co-occurrence relation data acquisition module is used for carrying out time sequence feature decomposition on a target da