CN-122022672-A - Unmanned warehouse output prediction method, system and storage medium thereof

CN122022672ACN 122022672 ACN122022672 ACN 122022672ACN-122022672-A

Abstract

The invention discloses a method, a system and a storage medium for predicting the warehouse output of an unmanned warehouse, wherein the method comprises the steps of collecting multi-source data with aligned time sequences; forming a time sequence data set and a multi-dimensional characteristic data set based on multi-source data, constructing a dynamic time sequence model, inputting the time sequence data set into the dynamic time sequence model to output a reference predicted value, constructing a machine learning model, inputting the multi-dimensional characteristic data set into the machine learning model to output a residual predicted value, fusing the reference predicted value and the residual predicted value to obtain a delivery predicted value, and outputting the delivery predicted value to a downstream system of a warehouse. The invention uses the dynamic ARIMA model as a reference time sequence predictor, uses XGBoost residual error correction model as an external influence catcher, realizes a self-optimized hybrid prediction framework through an online learning feedback loop, has high prediction precision and strong self-adaptability, and can make full use of multi-source data of an unmanned warehouse to make a warehouse quantity prediction.

Inventors

ZHAO ENJUN
FU LICAI
CHEN MAOJIN
WANG YUSHAN
WANG XIAOHUA
LIU JIAN
MENG XI
Ren Mengxiang

Assignees

四川泸天化股份有限公司

Dates

Publication Date: 20260512
Application Date: 20260108

Claims (7)

1. The unmanned warehouse output prediction method is characterized by comprising the following steps of Collecting time-aligned multi-source data including historical ex-warehouse data, warehouse operations data, campaign data, and external environmental data; Forming a time series data set and a multi-dimensional feature data set based on the multi-source data; Constructing a dynamic time sequence model, and inputting a time sequence data set into the dynamic time sequence model to output a reference predicted value; constructing a machine learning model, inputting a multidimensional characteristic data set into the machine learning model, and outputting a residual error predicted value; and fusing the reference predicted value and the residual predicted value to obtain a delivery predicted value, and outputting the delivery predicted value to a downstream system of the warehouse.
2. The unmanned warehouse outgoing amount prediction method according to claim 1, wherein the constructing the dynamic time sequence model comprises: using ARIMA model as dynamic time sequence model; Dynamically evolving parameters of the dynamic time sequence model by adopting an updating strategy; and predicting the future target period by using the updated dynamic time sequence model.
3. A method of unmanned warehouse outgoing volume prediction according to claim 2, wherein the update policy is a sliding window update or/and a performance triggered update.
4. The unmanned warehouse outgoing amount prediction method of claim 1, wherein constructing the machine learning model comprises: using XGBoost model as machine learning model; Calculating intra-sample prediction residual errors of the dynamic time sequence model on historical data; Constructing a multi-dimensional characteristic data set aligned with the residual sequence time sequence, wherein the characteristics comprise time sequence characteristics, warehouse operation characteristics, commercial activity characteristics and external environment characteristics; And training XGBoost a model regression model G by taking the historical residual as a target variable and the corresponding multidimensional characteristic as input, wherein the target function of the XGBoost model is a mean square error loss function with L2 regularization, and performing super-parameter optimization through time sequence cross validation.
5. The unmanned warehouse outgoing amount prediction method as claimed in claim 1, further comprising on-line learning and feedback, the on-line learning and feedback comprising: collecting actual ex-warehouse data, and calculating final prediction error data based on an ex-warehouse predicted value; updating sliding window data according to the final prediction error data, and inputting a dynamic time sequence model for dynamic updating; New data pairs are added to the training set of the machine learning model periodically, and the machine learning model is subjected to incremental learning or retraining.
6. An unmanned warehouse delivery prediction system is characterized by comprising The data acquisition module is used for acquiring time sequence aligned multi-source data, wherein the multi-source data comprises historical ex-warehouse data, warehouse operation data, business activity data and external environment data; A data processing module for forming a time series data set and a multi-dimensional feature data set based on the multi-source data; the dynamic time sequence model module is used for constructing a dynamic time sequence model, inputting a time sequence data set into the dynamic time sequence model and outputting a reference predicted value; the machine learning model module is used for constructing a machine learning model, inputting the multidimensional characteristic data set into the machine learning model and outputting residual error predicted values; and the fusion module is used for fusing the reference predicted value and the residual predicted value to obtain a delivery predicted value, and outputting the delivery predicted value to a downstream warehouse system.
7. A computer readable storage medium having stored thereon a computer program, the program being executable by a processor for implementing an unmanned warehouse outgoing volume prediction method as claimed in any one of claims 1 to 6.

Description

Unmanned warehouse output prediction method, system and storage medium thereof Technical Field The invention belongs to the technical intersection field of industrial Internet of things and intelligent supply chain prediction, and particularly relates to a method and a system for predicting warehouse output of an unmanned warehouse and a storage medium thereof. Background In the operation of modern unmanned warehouses, accurate prediction of the amount of delivery (i.e., demand) is a core premise for realizing inventory optimization, efficient scheduling of automated equipment, and reasonable planning of transportation resources. The accuracy of the prediction results directly determines the operation cost of the warehouse and the customer satisfaction. Currently, the main prediction technology in the industry has the following limitations: 1. The static and univariate limitation of classical time series models (such as ARIMA) is that the ARIMA model can effectively capture the inherent trend and rule of time series, but the model parameters are fixed once determined, and cannot adapt to the rapid service growth or the change of market environment. More importantly, the model is a univariate model, and abundant internal and external characteristics (such as sales promotion activities, social media heat and warehouse equipment operation efficiency) cannot be utilized, so that the demand fluctuation prediction capability of sudden and event-driven type is extremely poor. 2. The pure machine learning model has the disadvantage of capturing time sequence dependency, namely that models such as XGBoost can process multidimensional features, but generally have difficulty in capturing long-term dependency and periodicity rules of time sequences in a deep way, which are natural like ARIMA, and complex feature engineering is needed to construct time sequence features such as lag terms. 3. Roughness of existing hybrid models some simple hybrid methods only weight average the prediction results of different models and cannot realize depth complementation between models. In particular, prediction residuals of conventional ARIMA models are often discarded as noise, however, these residuals contain exactly the systematic information driven by external factors that the model does not capture due to insufficient information. Therefore, there is an urgent need in the art for an innovative predictive solution that can adaptively evolve and deeply fuse timing rules with multi-source information. Disclosure of Invention Therefore, in order to solve the above-mentioned shortcomings, the invention provides a method, a system and a storage medium for predicting the warehouse-out quantity of the unmanned warehouse, wherein the method, the system and the storage medium take a dynamic ARIMA model as a reference time sequence predictor, take a XGBoost residual error correction model as an external influence catcher, realize a self-optimized mixed prediction framework through an online learning feedback loop, have high prediction precision and strong self-adaptability, and can fully utilize multi-source data of the unmanned warehouse to make warehouse-out quantity prediction. In a first aspect, the present invention provides a method for predicting the delivery of an unmanned warehouse, including: collecting time-aligned multi-source data including historical ex-warehouse data, warehouse operations data, campaign data, and external environmental data; Forming a time series data set and a multi-dimensional feature data set based on the multi-source data; Constructing a dynamic time sequence model, and inputting a time sequence data set into the dynamic time sequence model to output a reference predicted value; constructing a machine learning model, inputting a multidimensional characteristic data set into the machine learning model, and outputting a residual error predicted value; and fusing the reference predicted value and the residual predicted value to obtain a delivery predicted value, and outputting the delivery predicted value to a downstream system of the warehouse. Optionally, the building the dynamic time sequence model includes: using ARIMA model as dynamic time sequence model; Dynamically evolving parameters of the dynamic time sequence model by adopting an updating strategy; and predicting the future target period by using the updated dynamic time sequence model. Optionally, the update policy is a sliding window update or/and a performance triggered update. Optionally, building the machine learning model includes: using XGBoost model as machine learning model; Calculating intra-sample prediction residual errors of the dynamic time sequence model on historical data; Constructing a multi-dimensional characteristic data set aligned with the residual sequence time sequence, wherein the characteristics comprise time sequence characteristics, warehouse operation characteristics, commercial activity characteristics and external en