Disclosure of Invention
Aiming at the defects in the prior art, the invention provides a supply chain intelligent task optimization method based on a language model, which comprises the following steps:
Acquiring task data of a supply chain, and identifying and extracting to give key information;
analyzing and generating product demand prediction data of different task addresses based on the identified and extracted key information;
based on the product demand prediction data, giving a product inventory distribution adjustment scheme containing all task addresses;
and confirming or modifying the product inventory distribution adjustment scheme, and giving a final product inventory distribution adjustment scheme to realize intelligent task optimization of the supply chain.
The recognition and extraction processing comprises the step of recognizing and extracting key information from task data by using a GPT algorithm of a natural language processing NLP technology.
The task data comprises unstructured data and structured data, and the identifying and extracting processes comprise:
The GPT model carries out structuring treatment on unstructured data;
the GPT model predicts the category to which the task data belongs based on all the structured data;
based on the category to which the task data belongs, the GPT model identifies and extracts key information from all structured data.
The key information at least comprises product names, quantity, task addresses and time.
Wherein analyzing and generating product demand forecast data for different task addresses includes:
obtaining key features from the identification and extraction processing results, wherein the key features comprise product names, quantity units, task addresses, time and state labels;
Constructing a corresponding feature matrix based on the key features, wherein the construction of the feature matrix specifically comprises:
let t be the original time data, convert t to days d from the reference time point, convert to binary vector using one-hot encoding for product name p and task address lAndDetermining the number of product demands as ys;
Combining feature matrices, combining all features into one feature matrix X, wherein
Wherein after constructing the feature matrix, the feature matrix is processed by firstly performing the following steps ofOrdering and then within each product category, ordering according to task addressAnd sorting, namely sorting out historical sales data for each product name p and each task address l.
The process of analyzing and generating the product demand prediction data of different task addresses comprises the steps of determining global predicted sales and predicted sales of the task addresses, and giving the product demand prediction data of the different task addresses based on an objective function and combining the global predicted sales and the predicted sales of the task addresses;
Determining a global predicted sales amount G defined as the sum of all task addresses and predicted sales of the product g= Σl L (p, L, t), where L (p, L, t) represents the predicted sales amount of the product name p at the task address L at time t;
Determining a predicted sales volume of the task addresses, wherein the predicted sales volume L (p, L, t) should reflect the latest sales volume for each task address L;
The objective function is defined as:
minimize∑l[L(p,l,t)·α·local_recent(p,l,t)·D(p,l,t)/(1-α)·local_avg(p,l)]2;
Wherein, the
Alpha is a weight parameter between 0 and 1, which controls the weight of the latest sales and the historical sales average in prediction;
p, product name identification;
A task address or an identification of a region;
time stamp, generally representing the current time or a particular predicted point in time;
local_Current (p, l, t) sales of product name p at task address l at time t-1, i.e., sales data of the last phase;
local_avg (p, l) average demand of product name p over the number of historic days of consideration for task address l;
d (p, l, t) decay factor reflecting the purchase saturation and trend of the task address l at time t.
Wherein, based on the product demand forecast data, a product inventory distribution adjustment scheme containing all task addresses is provided, including obtaining the product inventory distribution adjustment scheme based on an optimization process or obtaining the product inventory distribution adjustment scheme of all task addresses based on a prioritization.
Wherein obtaining a product inventory distribution adjustment scheme based on an optimization process includes:
determining a product inventory distribution adjustment scheme using optimization objectives to minimize the sum of product inventory gap and shipping costs for each task address, comprising:
minimize∑l[gap(p,l,t)+K·dist(w,l)·transfer(w,p,l,t)]
gap(p,l,t)=L(p,l,t)-stock(w,p)
Wherein K is a weight parameter, the importance of the control transportation cost in the optimization target, dist (w, l) is the distance between the product warehouse w and the task address l, gap (p, l, t) is the stock gap of the product name p of the task address l at the time t, transfer (w, p, l, t) is the number of the product names p allocated from the product warehouse w to the task address l at the time t, and stock (w, p) is the stock number of the product names p existing in the product warehouse w;
obtaining a product inventory distribution adjustment scheme for all task addresses based on prioritization, comprising:
the method comprises the steps of calculating priorities of all task addresses, sorting all task addresses according to descending order of the priorities, and adjusting product inventory according to sorting results by using an allocation rule until the product inventory is exhausted or the requirements of all task addresses are met, wherein the allocation rule meets the following relation:
transfer(w,p,l,t)=min(stock(w,p),gap(p,l,t))
In the formula, gap (p, l, t) is an inventory gap of a product name p of a task address l at time t, dist (w, l) is a distance from a warehouse w to the task address l, and priority (p, l, t) is a priority.
Wherein the confirmation or modification of the product inventory distribution adjustment scheme is used to provide a final product inventory distribution adjustment scheme, comprising:
after confirming that the user has the authority of managing the user, displaying the product demand prediction data and the product inventory distribution adjustment scheme, and explaining natural language;
And the user receives the given product inventory distribution adjustment scheme, or the user modifies the product inventory distribution adjustment scheme after inputting feedback information.
The invention also discloses a supply chain intelligent task optimizing device based on the language model, which comprises:
the data processing module is used for acquiring task data of the supply chain, identifying and extracting the task data and giving out key information;
The analysis generation module is used for analyzing and generating product demand prediction data of different task addresses based on the identified and extracted key information;
an initial scheme generation module for giving a product inventory distribution adjustment scheme containing all task addresses based on product demand prediction data;
And the final scheme determining module is used for confirming or modifying the product inventory distribution adjusting scheme and giving a final product inventory distribution adjusting scheme.
The data processing module comprises a GPT algorithm which uses natural language processing NLP technology to identify and extract key information from task data.
Wherein the data processing module comprises:
The GPT model carries out structuring treatment on unstructured data;
the GPT model predicts the category to which the task data belongs based on all the structured data;
based on the category to which the task data belongs, the GPT model identifies and extracts key information from all structured data.
Wherein, analysis generation module includes:
Analyzing and generating product demand forecast data for different task addresses, including:
obtaining key features from the identification and extraction processing results, wherein the key features comprise product names, quantity units, task addresses, time and state labels;
Constructing a corresponding feature matrix based on the key features, wherein the construction of the feature matrix specifically comprises:
let t be the original time data, convert t to days d from the reference time point, convert to binary vector using one-hot encoding for product name p and task address lAndDetermining the number of product demands as ys;
Combining feature matrices, combining all features into one feature matrix X, wherein
Wherein the analysis generation module comprises the steps of firstly, matching the product name withOrdering and then within each product category, ordering according to task addressAnd sorting, namely sorting out historical sales data for each product name p and each task address l.
The analysis generation module further comprises a step of determining the global predicted sales volume and the predicted sales volume of the task address, and providing product demand prediction data of different task addresses based on an objective function and combining the global predicted sales volume and the predicted sales volume of the task address;
Wherein determining the global predicted sales amount G is defined as the sum of all task addresses and predicted sales of the product g= Σl L (p, L, t), where L (p, L, t) represents the predicted sales amount of the product name p at the task address L at time t;
determining a predicted sales volume of the task addresses, wherein the predicted sales volume L (p, L, t) reflects the latest sales volume for each task address L;
The objective function is defined as:
minimize∑l[L(p,l,t)·α·local_recent(p,l,t)·D(p,l,t)/(1-α)·local_avg(p,l)]2;
In the formula,
Alpha is a weight parameter between 0 and 1, and controls the weight of the latest sales and the historical sales average in prediction;
p, product name identification;
A task address or an identification of a region;
t is a time mark which represents the current time or a specific predicted time point;
local_Current (p, l, t) sales of product name p at task address l at time t-1;
local_avg (p, l) average demand of product name p over the number of historic days of consideration for task address l;
d (p, l, t) decay factor reflecting the purchase saturation and trend of the task address l at time t.
The initial scheme generation module comprises a product inventory distribution adjustment scheme obtained based on an optimization process or a product inventory distribution adjustment scheme obtained based on priority ordering for all task addresses.
Wherein the final scheme determination module comprises:
after confirming that the user has the authority of managing the user, displaying the product demand prediction data and the product inventory distribution adjustment scheme, and explaining natural language;
And the user receives the given product inventory distribution adjustment scheme, or the user modifies the product inventory distribution adjustment scheme after inputting feedback information.
The method provided by the invention provides efficient and accurate demand prediction and inventory optimization through integrating and intelligently processing the multi-source data, enhances user interaction and decision support, and greatly improves the overall performance and efficiency of supply chain management.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail below with reference to the accompanying drawings, and it is apparent that the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
The terminology used in the embodiments of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. As used in this application and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise, the "plurality" generally includes at least two.
It should be understood that although the terms first, second, third, etc. may be used in embodiments of the present invention to describe the terms. These terms are only used to distinguish. For example, the first..once again may be referred to as the second..once again, and similarly, the second..once again may be referred to as the first..once again without departing from the scope of embodiments of the present invention.
It should be understood that the term "and/or" as used herein is merely an association relationship describing the associated object, and means that there may be three relationships, e.g., a and/or B, and that there may be three cases where a exists alone, while a and B exist together, and B exists alone. In addition, the character "/" herein generally indicates that the front and rear associated objects are an "or" relationship.
The words "if", as used herein, may be interpreted as "at" or "when" or "in response to a determination" or "in response to a detection", depending on the context. Similarly, the phrase "if determined" or "if detected (stated condition or event)" may be interpreted as "when determined" or "in response to determination" or "when detected (stated condition or event)" or "in response to detection (stated condition or event), depending on the context.
It should also be noted that the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a product or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such product or apparatus. Without further limitation, an element defined by the phrase "comprising one does not exclude the presence of additional like elements in a commodity or device comprising the element.
Supply chain management typically involves a large amount of data processing including, but not limited to, mail, orders, accounting vouchers, and the like. Traditional methods rely on manual processing of such data, which is time consuming and error prone. To solve this problem, the present invention proposes an automated solution that can efficiently process and analyze supply chain data and support real-time decisions.
Traditional demand forecasting methods often rely on historical sales data, lacking a quick response to sudden events or market changes. The method utilizes the extracted data, and improves the accuracy and response speed of prediction through a demand prediction and inventory reset module. In the context of multitasking address distribution, the complexity of inventory management increases and traditional approaches have difficulty achieving efficient inventory optimization. The multitask address inventory optimization module of the method can automatically adjust the inventory distribution of products, solves the problem of excessive inventory or shortage, and improves the utilization rate of resources. In complex supply chain management, a decision maker needs to understand the complex data analysis results and the output of the predictive model.
As shown in fig. 1, the method for optimizing the supply chain intelligent task based on the language model comprises the following steps:
acquiring task data with a plurality of sources in a supply chain, and identifying and extracting to give key information;
analyzing and generating product demand prediction data of different task addresses based on the identified and extracted key information;
based on the product demand prediction data, giving a product inventory distribution adjustment scheme containing all task addresses;
and confirming or modifying the product inventory distribution adjustment scheme, and giving a final product inventory distribution adjustment scheme to realize intelligent task optimization of the supply chain.
The system comprises a data integration and information extraction module, a multi-task address inventory optimization module, a user interaction module, a product inventory distribution adjustment scheme and a product inventory distribution information analysis module, wherein the data integration and information extraction module is used for processing task data from a plurality of sources, the extracted data are transmitted to a demand prediction and inventory reset module for analysis and generation of demand prediction data, the multi-task address inventory optimization module is used for receiving the prediction data and automatically giving a product inventory distribution adjustment scheme containing all task addresses, and the user interaction module is used for carrying out interaction to obtain advice and interpretation information of the product inventory distribution adjustment scheme. The method provided by the invention provides efficient and accurate product demand prediction data and inventory optimization through integrating and intelligently processing the multi-source data, enhances user interaction and decision support, and greatly improves the overall performance and efficiency of supply chain management.
The task data comprises unstructured data and structured data, and the data integration and information extraction module automatically identifies and extracts key information from unstructured data sources (mails and reports) of the task data (text data) by using a Natural Language Processing (NLP) technology GPT algorithm. Text classification and named entity recognition techniques may be included to extract data points such as date, task address, and number. Wherein the task address may be a location where the product is received.
The data integration and information extraction module adopts a natural language processing algorithm based on GPT, and the GPT model converts unstructured data of multiple sources into structured data which can be understood by the GPT model in a preprocessing (structuring processing) stage, for example, the text is divided into tokens (tokens). The data integration and information extraction module collects data from various unstructured data sources (such as e-mail, PDF report, electronic certificate), and is implemented by calling through an API or directly exporting the data from an enterprise internal system, and then extraneous characters such as HTML tags, special symbols, redundant blank spaces and the like are removed.
A suitable pre-trained GPT model needs to be selected. OpenAI provide various versions of a GPT model, such as GPT-3, that have been pre-trained on large-scale text data. The pre-trained GPT model is selected and fine-tuned on a specific data set to accommodate specific text classification and named entity recognition tasks. The GPT model carries out structuring processing on unstructured data, predicts the category (such as order information, accounting documents and the like) to which the task data belong based on all structured data, and then uses the GPT model which is specially adjusted and adapted to the task data to identify and extract key information from all structured data based on the category to which the task data belong.
And (3) performing fine tuning on the GPT model by using the labeled classification data set, wherein the fine tuning comprises setting a proper learning rate, and selecting a proper batch processing size and iteration times. During the fine-tuning process, the last layer of the model needs to be customized to output a particular class prediction. During the fine tuning process, the model's performance on the independent validation data set is continually validated to monitor the overfitting and adjust the training parameters.
And classifying the input task data by using a GPT algorithm, and identifying the subject or class of the task data. Classification helps the system understand the primary purpose and content of the document, thereby more accurately targeting information extraction.
After identifying the subject or category, the task data is sent to a GPT model extracted by corresponding key information, and the GPT model outputs structured data containing the key information, wherein the GPT model is fine-tuned corresponding to the text category.
Using the trimmed GPT model in combination with named entity recognition NER techniques, extracting specific information from text, including:
the training data is marked with entities (key information) including product name, quantity, task address, time and status tags.
And further training and fine-tuning the GPT model by using the marked data set, so that the GPT model can identify and extract specific information in the document. And inputting the classified text into a fine-tuned model, and identifying and extracting key information by the model.
The identified key information is converted to a structured format, such as JSON or XML format, for further processing and analysis.
For example, after analyzing a certain order document, the NER-based GPT model outputs the following results:
[
{
"Product name": "sample Product",
"Number" means "1000",
"Task address": "Shanghai",
"Time": "2024-01-15",
"State tags": "shipped"
},
{
"Product name": "Another Product",
"Number" means "500",
"Task address": beijing ",
"Time": "2024-01-15",
"State label": "to be delivered"
}
]
LSTM models are constructed using Keras or other deep learning frameworks. The model is trained using historical data, which refers to previous supply chain data, arranged in chronological order. The sales records of the product are continuously tracked and trained using these data as inputs to the model.
The data processed by the data integration and information extraction module is subjected to cleaning and standardization processes so as to ensure the quality and consistency of the data. For example, a presentation of a uniform date format, a naming of a task address, and a number.
The cleaned and standardized data is integrated into a downstream module and fed into a demand forecast and inventory reset module for further analysis and decision support. And enabling other systems or modules to send requests to the model by developing corresponding APIs or service interfaces to obtain results of text classification or entity recognition.
The data integration and information extraction module analyzes and generates product demand prediction data of different task addresses, wherein the data integration and information extraction module extracts key features including product names, quantity units, task addresses, time and state labels from identification and extraction processing results (each JSON or XML record), constructs a feature matrix corresponding to each record, and the feature matrix construction specifically comprises the following steps:
1. And (3) time characteristic processing:
let t be the original time data, e.g. "2024-01-15".
-Converting t into a number of days d starting from a certain reference point in time.
2. Category feature coding:
conversion into binary vectors using one-hot encoding for product name p and task address lAnd。
3. Quantity:
-determining the number of product demands as ys.
4. Combining the feature matrix:
combining all features into a feature matrix X, wherein
Processing the feature matrix according to the product nameOrdering and then within each product category, according to the task addressOrdering is performed, and for each product p and each task address l, historical purchase data is sorted out, wherein the cycle time is days or weeks.
The attenuation factor is introduced with the purpose of adjusting the predicted value to take into account the historical trend and volatility of the product demand. Based on the overall historical average sales and recent sales, and the purchase history for the particular task address. The attenuation factor calculation process comprises the following steps:
Global historical average sales global avg is calculated as follows, where Y (p, i) represents the total demand of product p at all task addresses, and n is the number of historical days considered:
historical average sales for a particular task address:
The historical average sales localavg (p, l) for a particular task address l is calculated as follows, where Y (p, l, i) is the latest demand for product p at task address l at time i;
Average demand:
Recent sales of task address l:
The latest sales local_fraction (p, l, t) of the task address l is defined as the sales of the product p in the latest phase (i.e., time t-1) of the task address l: localjfraction (p, l, t) =y (p, l, t-1).
The decay factor D (p, l, t) reflects the purchase saturation and trend of the task address l at time t, defined as follows:
where β and γ are adjustable weight parameters for controlling the extent of influence of different factors in the attenuation factor.
The process of analyzing and generating the product demand prediction data of different task addresses comprises the steps of determining global predicted sales and predicted sales of the task addresses, and giving the product demand prediction data of the different task addresses based on an objective function and combining the global predicted sales and the predicted sales of the task addresses;
Wherein the global predicted sales amount G is defined as the sum of all task addresses and predicted sales of the product g= Σl L (p, L, t), where L (p, L, t) represents the predicted sales amount of the product p at the task address L at time t.
The predicted sales amount of the task address is determined, and for each task address L, the predicted sales amount L (p, L, t) should reflect the influence of the latest sales amount local_direct (p, L, t) and the attenuation factor D (p, L, t), while also reflecting the influence of the historical sales average value local_avg (p, L). Optionally, the objective function is defined as:
minimize∑l[L(p,l,t)·α·local_recent(p,l,t)·D(p,l,t)/(1-α)·local_avg(p,l)]2;
Where α is a weight parameter between 0 and 1, controlling the weight of the most recent sales and the historical sales averages in the forecast.
Alternatively, in order to approximate the global predicted sales amount G to the global average sales amount global_avg, the objective function may be :minimize∑l[L(p,l,t)·α·local_recent(p,l,t)·D(p,l,t)/(1-α)·local_avg(p,l)]2+λ·(G-globa_avg)2.
Wherein, the parameters in the above process are defined as follows:
-p product identification.
-I: identification of the task address or region.
Time identification, generally representing the current time or a specific predicted point in time.
Historical days, i.e. days considered in calculating the average sales.
Y (p, i) the total amount of demand of product p at all task addresses at time i.
Y (p, l, t) the demand of the product p at a specific task address l at time i.
Local_direct (p, l, t) sales of product p at task address l at time t-1, i.e. last-phase sales data.
Global avg, the average total demand for all task addresses and products over the number of historic days considered.
Localavg (p, l) average demand of product p over the number of historic days considered for the task address l.
- Β, a weight parameter for adjusting the degree of influence of the latest sales on the attenuation factor.
-Gamma weight parameter for adjusting the degree of influence of the historical average sales on the attenuation factor.
D (p, l, t) decay factor reflecting the purchase saturation and trend of the task address l at time t.
L (p, L, t) at time t, the predicted sales of product p at task address L.
-Alpha, a weight parameter between 0 and 1, controlling the relative importance of the latest sales and the historical sales averages in the predictions.
A weight parameter for adjusting the influence of the deviation between the global predicted sales and the global average sales.
And G, predicting the sum of sales of all task addresses and products, namely globally predicting sales.
Wherein obtaining a product inventory distribution adjustment scheme based on an optimization process includes:
The obtained data includes a predicted demand for each product for each task address over a future period of time, and a predicted demand L (p, L, t) for each task address L for the product at a future time t is determined.
The distance between the repository and each task address is determined, and dist (w, l) is defined as the distance between the repository w and the task address l.
The stock existing in the warehouse is determined, and the stock (w, p) is defined as the stock number of the product p existing in the warehouse w.
The stock gap of each task address is calculated, and gap (p, L, t) is defined as the stock gap of the product p of the task address L at time t, wherein gap (p, L, t) =l (p, L, t) -stock (w, p).
Determining an inventory allocation strategy:
transfer (w, p, l, t) is defined as the number of products p allocated to task address l at time t from warehouse w. This number needs to be determined based on inventory gap (p, l, t) and shipping costs. The optimization objective employed is to minimize the sum of inventory gap and shipping costs for each task address:
minimize∑l[gap(p,l,t)+K·dist(w,l)·transfer(w,p,l,t)],
wherein K is a weight parameter, which controls the importance of the transportation cost in the optimization objective.
The product inventory distribution adjustment scheme for obtaining all task addresses based on the priority ordering comprises the following steps:
Defining priorities, for each task address l, product p, inventory allocation priority (p, l, t) at time t is calculated as follows:
Wherein, the
Gap (p, l, t) is the stock gap of product p at time t for task address l.
Dist (w, l) is the distance of the warehouse w to the task address l.
Inventory allocation is determined, wherein the inventory allocation rules are as follows according to the priority from high to low:
transfer(w,p,l,t)=min(stock(w,p),gap(p,l,t)),
For the task address l with the highest priority, the stock is distributed according to the rule, and the task address with the large gap and close distance of the priority stock is indicated.
The specific steps include 1, calculating priority (p, l, t) of all task addresses. 2. All task addresses are sorted in descending order of priority (p, l, t). 3. And (3) according to the sorting result, performing inventory allocation by using the transfer formula until the warehouse inventory is exhausted or the requirements of all task addresses are met.
The optimization problem may be solved using various algorithms, such as gradient descent, simulated annealing, genetic algorithm, linear programming, or mixed integer linear programming. For example, the present invention uses a nonlinear optimization algorithm, such as a gradient descent algorithm, because it involves continuous variables and objective functions that are nonlinear.
The following is a solution using gradient descent method, which includes that first, all variables need to be initialized, which can be set to 0 or a random value. For each variable, the gradient of the optimization objective function to it, i.e. the partial derivative of the optimization objective function, is calculated.
The variables are updated according to the gradient and the learning rate (a parameter that needs to be set in advance).
Checking for convergence, including repeating the gradient calculation and updating steps until a predetermined number of iterations is reached or the change in the optimization objective function is less than a certain set threshold.
The method comprises the steps of confirming or modifying a product inventory distribution adjustment scheme, and giving out a final product inventory distribution adjustment scheme, wherein the final product inventory distribution adjustment scheme comprises the steps of displaying product demand prediction data and the product inventory distribution adjustment scheme and performing natural language explanation after confirming that a user has the authority of managing the user, and the final product inventory distribution adjustment scheme is modified after the user receives the given product inventory distribution adjustment scheme or the user inputs feedback information.
Specifically, the user interacts with the system through an interaction module to obtain advice and interpretation information of the system, wherein the module is a part of the supply chain intelligent task optimization system, and the main function is to enable the user to interact with the system to obtain advice and interpretation information of the system and provide feedback.
The user interaction interface is the medium for the user to interact with the system, and can be realized through a webpage, a mobile application or a desktop application. On the interface, after the system confirms that the user has the authority to manage the user, the system can display the product demand prediction data and the product inventory distribution adjustment scheme, explain the result in natural language, and the user can interact on the natural language interaction interface with respect to a specific parameter meaning and adjustment reason on the interface. And the interactive interface also provides an input box, wherein the input box can comprise selectable items of final adjustment results of the inventory, feedback information can be input by a user, and the user can accept a product inventory distribution adjustment scheme provided by the system or manually modify the product inventory distribution adjustment scheme. This may be achieved by front-end development techniques.
The method provides an interactive module, and enables a decision maker to better understand and utilize the output product inventory distribution adjustment scheme through intelligent interpretation and suggestion.
After obtaining the product demand forecast data and the product inventory distribution adjustment scheme, easily understood text information is generated based on the GPT model, explaining why such adjustments are to be made. This may be accomplished by applying a pre-trained language model such as GPT. The language model generates a piece of text that explains why such an adjustment is to be made, using the prediction result and the adjustment result as inputs.
In generating the explanatory text, product demand forecast data and a product inventory distribution adjustment scheme are used as inputs to generate a piece of text explaining why such adjustments are to be made by applying a pre-trained GPT model. The following specific steps are realized:
First, the product demand forecast data and the product inventory distribution adjustment scheme need to be converted into a form understandable by the GPT model, for example, the GPT model may be input in a tabular form. The preprocessed input data is then fed to the GPT model. The model will generate a piece of text from the input data that explains why inventory adjustments are to be made. In particular implementations, a generation mode (also referred to as a decoding mode) of the GPT model may be used, such as greedy decoding, beam search decoding, or sample decoding. The GPT model can be a specially trimmed model, which comprises trimming the GPT model by using a marked table and a corresponding explanatory data set, and comprises setting a proper learning rate and selecting a proper batch size and iteration times.
Finally, post-processing of the generated text is required, e.g. it may be necessary to check whether the generated text contains any unsuitable content, or whether certain length restrictions are met, etc.
When the user inputs feedback information, including that the user does not accept such modification, the user can modify the product inventory distribution adjustment scheme in the feedback information, the system receives and processes the feedback information of the user, and after the system receives the feedback information, the management user ID and the feedback content thereof are stored in a database for subsequent viewing and rewinding.
The interactive system needs to interact with the database, and comprises a prediction result, an inventory adjustment result and user feedback, wherein the interactive system internally comprises a special GPT module for user interpretation of the prediction result and the inventory adjustment result.
Therefore, the method of the invention provides efficient and accurate product demand prediction and inventory optimization by integrating and intelligently processing the multi-source data, enhances user interaction and decision support, and greatly improves the overall performance and efficiency of supply chain management.
As shown in fig. 2, the present invention further provides a supply chain intelligent task optimizing device based on a language model, including:
the data processing module is used for acquiring task data of the supply chain, identifying and extracting the task data and giving out key information;
The analysis generation module is used for analyzing and generating product demand prediction data of different task addresses based on the identified and extracted key information;
an initial scheme generation module for giving a product inventory distribution adjustment scheme containing all task addresses based on product demand prediction data;
And the final scheme determining module is used for confirming or modifying the product inventory distribution adjusting scheme and giving a final product inventory distribution adjusting scheme.
It should be noted that the computer readable medium described in the present disclosure may be a computer readable signal medium or a computer readable storage medium, or any combination of the two. The computer readable storage medium can be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or a combination of any of the foregoing. More specific examples of a computer-readable storage medium may include, but are not limited to, an electrical connection having one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the context of this disclosure, a computer-readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. In the present disclosure, however, the computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, with the computer-readable program code embodied therein. Such a propagated data signal may take any of a variety of forms, including, but not limited to, electro-magnetic, optical, or any suitable combination of the foregoing. A computer readable signal medium may also be any computer readable medium that is not a computer readable storage medium and that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. Program code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to electrical wiring, fiber optic cable, RF (radio frequency), and the like, or any suitable combination of the foregoing.
The computer readable medium may be included in the electronic device or may exist alone without being incorporated into the electronic device.
Computer program code for carrying out operations of the present disclosure may be written in one or more programming languages, including an object oriented programming language such as Java, smalltalk, C ++ and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any kind of network, including a Local Area Network (LAN) or a Wide Area Network (WAN), or may be connected to an external computer (for example, through the Internet using an Internet service provider).
The flowcharts and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems which perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.
The units involved in the embodiments of the present disclosure may be implemented by means of software, or may be implemented by means of hardware. Wherein the names of the units do not constitute a limitation of the units themselves in some cases.
The foregoing description of the preferred embodiments of the present invention has been presented for purposes of clarity and understanding, and is not intended to limit the invention to the particular embodiments disclosed, but is intended to cover all modifications, alternatives, and improvements within the spirit and scope of the invention as outlined by the appended claims.