Supply chain demand prediction method and system based on artificial intelligence

Abstract

The invention relates to the technical field of supply chain intelligent prediction, and discloses a supply chain demand prediction method and system based on artificial intelligence. The method comprises the steps of obtaining a complete transaction record in a target time period, forming a material flow time sequence, and screening flow subsequences according to a predicted target material variety, a time span and a geographical range. And generating a material flow change map based on the subsequences, and visually describing a flow dynamic fluctuation track. Scanning the atlas by using a preset atlas feature recognition rule, recognizing a specific traffic change form, marking the specific traffic change form as a to-be-analyzed feature, and extracting a corresponding original transaction record subset; constructing a deep neural network, taking the to-be-analyzed features and the transaction record subsets as input, and training the network to learn a mapping relation between the to-be-analyzed features and the transaction record subsets and subsequent material demands; according to the method, complex non-stationary demand fluctuation is captured through combination of map form identification and transaction detail multi-modal learning, and prediction accuracy and business interpretability are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent supply chain forecasting technology, specifically to a supply chain demand forecasting method and system based on artificial intelligence. Background Technology

[0002] Existing supply chain demand forecasting technologies primarily rely on numerical time series analysis of historical transaction data, employing statistical models or machine learning algorithms to directly fit numerical patterns in the sequences for extrapolation and prediction. These methods are relatively mature in handling conventional trends and cyclical patterns. However, these approaches have inherent limitations. The modeling process heavily depends on the statistical properties of the numerical data itself, making it difficult to effectively identify and distinguish complex, non-stationary dynamic fluctuation patterns triggered by different business events. These patterns may only manifest as transient anomalies in the numerical sequences, and the underlying causal logic and structural characteristics cannot be fully captured through conventional numerical feature extraction. Therefore, existing methods struggle to achieve interpretability of the forecasting models, preventing decision-makers from correlating predicted demand changes with specific, past clusters of transaction events.

[0003] To overcome the aforementioned shortcomings, a new technological approach is needed. This method must be able to directly identify visualized macro-dynamic patterns representing different business scenarios from time-series data, rather than relying solely on numerical statistics. It requires precisely associating and binding each identified macro-pattern with the micro-level original transaction records that constitute that pattern. This provides the predictive model with a fusion input that simultaneously includes high-order pattern semantics and fine-grained factual evidence, enabling it to learn the mapping relationship from specific combinations of business events to future needs, achieving more accurate and traceable predictions. Summary of the Invention

[0004] The purpose of this invention is to provide an artificial intelligence-based supply chain demand forecasting method and system to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides an artificial intelligence-based supply chain demand forecasting method, the method comprising: Obtain the complete set of transaction records of all participants within the target time period, and use time-series processing to form a material flow time series with time as the dimension; Obtain the boundary conditions of the current forecasting task, including the type of target material, the time span of the forecast, and the geographical scope of the forecast; Filter out the flow subsequences from the material flow time series that match the predicted target material type, predicted time span, and predicted geographical range; A material flow rate change map is generated based on the flow rate subsequence, and the material flow rate change map is used to visually depict the dynamic fluctuation trajectory of the material flow rate; Obtain preset spectral feature recognition rules, and use the spectral feature recognition rules to scan the material flow rate change spectral and identify specific flow rate change patterns appearing in the spectral; The identified specific traffic change patterns are marked as features to be analyzed, and a subset of the original transaction records corresponding to each feature to be analyzed is extracted; A deep neural network is constructed, and the feature to be analyzed and its corresponding subset of original transaction records are used as input. The deep neural network is trained to learn the mapping relationship between the feature to be analyzed and subsequent material requirements.

[0006] Preferably, the step of obtaining a complete set of transaction records for all participants within the target time period and forming a time-series material flow data using time-series processing includes: The supply chain transaction database is formed by separating the supply chain-related transaction records from the complete set of transaction records. Locate the key nodes of each transaction in the supply chain transaction database and record the material flow information generated at each key node; The material flow information is processed in a time-series manner to form a material flow time series with time as the dimension; The process of locating the key nodes of each transaction in the supply chain transaction database includes: Extract the time and location associated with changes in the physical displacement state of the material from the records of each transaction; From the times and locations associated with the changes in the physical displacement state, select the times when the material enters the storage facility, when the material leaves the storage facility, when the material is sorted, when the material is loaded onto the transport vehicle, and when the material is unloaded from the transport vehicle; The selected time and location combination is defined as the key node of each transaction.

[0007] Preferably, when performing time-series processing on material flow information to form a material flow time series with time as the dimension, the following are included: Using each natural day as a unit time window, summarize the total quantity of the same type of material that passes through all key nodes within the unit time window; The total quantity of materials obtained from each unit time window is arranged into an ordered list of values ​​according to the chronological order. This ordered list of values ​​is the material flow time series.

[0008] Preferably, when selecting flow subsequences from the material flow time series that match the predicted target material type, predicted time span, and predicted geographical range, the process includes: From the material flow time series, identify entries whose material type matches the predicted target material type; From the identified entries, further filter out entries whose geographical coordinates of the transaction occurred within the predicted geographical range; From the re-selected entries, extract the data segments corresponding to consecutive time periods whose timestamps fall within the predicted time span; these data segments are the flow subsequences.

[0009] Preferably, when generating a material flow rate change map based on the flow rate subsequence, the process includes: Establish a two-dimensional coordinate system, where the horizontal axis represents time and the vertical axis represents material flow rate; In the two-dimensional coordinate system, each data point in the flow subsequence is located and marked according to its time and material flow value; By sequentially connecting the markers at adjacent time points using continuous line segments, a broken line reflecting the change in material flow over time is formed. The broken line and its coordinate system together constitute the material flow change map.

[0010] Preferably, when scanning the material flow rate change map using the aforementioned map feature recognition rules to identify specific flow rate change patterns appearing in the map, the process includes: Define a set of flow change pattern templates, including a rapid rise plateau pattern, a step-down pattern, and a periodic pulse pattern; Compare the flow change pattern template with the local broken line segments in the material flow change graph point by point; When the shape of a local broken line segment in the graph matches the shape of any flow change pattern template more than a preset pattern matching threshold, the local broken line segment is determined to be the specific flow change pattern.

[0011] Preferably, when identifying specific traffic change patterns as features to be analyzed and extracting a subset of original transaction records corresponding to each feature to be analyzed, the process includes: In the material flow rate change map, each identified specific flow rate change pattern is labeled with an independent feature identifier; Based on the start and end times of each specific flow change pattern on the time axis, search the supply chain transaction database for all transaction records within the time period between the start and end times. All the found transaction records are combined to form the original transaction record subset corresponding to the specific flow change pattern.

[0012] Preferably, when using the feature to be analyzed and its corresponding subset of original transaction records as input to train the deep neural network to learn the mapping relationship between the feature to be analyzed and subsequent material requirements, the process includes: For each feature to be analyzed, its morphological description parameters are calculated, including the duration of the morphology, the magnitude of flow rate change, the flow rate value at the beginning of the morphology, and the flow rate value at the end of the morphology. For each subset of original transaction records, calculate its transaction attribute parameters, which include the average number of transaction participants, the average geographical distance involved in the transaction, and the distribution density of transaction amount; The morphological description parameters are concatenated with the corresponding transaction attribute parameters to form a comprehensive feature vector; The comprehensive feature vector and the actual material demand change trend within a set time period after the occurrence of the feature to be analyzed are used together as a set of training samples and input into a deep neural network for iterative training until the neural network can output a prediction of subsequent demand changes based on the input comprehensive feature vector.

[0013] Preferably, after the deep neural network has been trained, when the neural network is used for prediction, it includes: Obtain the latest material flow data and generate the latest material flow change map according to the method described above; Identify the latest specific flow change pattern in the latest material flow change map and generate the corresponding latest comprehensive feature vector; The latest comprehensive feature vector is input into a trained deep neural network, which outputs a predicted material demand value for a future period following the latest specific flow change pattern.

[0014] Preferably, the present invention also includes an artificial intelligence-based supply chain demand forecasting system, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the artificial intelligence-based supply chain demand forecasting method described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are: By transforming material flow data into visual graphs and scanning them using pre-defined graph feature recognition rules, the overall dynamic trajectory and local morphological features of flow changes can be directly depicted and captured. This method overcomes the limitations of traditional statistical features, enabling the identification of specific spatial-temporal patterns that may be masked or simplified in pure numerical analysis, such as periodic clusters, trend breaks, and irregular pulses. The identified patterns are directly linked to their corresponding subsets of original transaction records, allowing each macroscopic flow fluctuation pattern to be traced back to the specific set of transaction events that constitute it. This achieves the transformation from "numerical anomalies" to "interpretable business patterns," providing a direct, evidence-based tracing path for understanding the composition of demand fluctuations.

[0016] By training a deep neural network using the identified graph patterns and their corresponding subsets of original transaction records as input, a multimodal learning architecture integrating high-order semantic features and fine-grained factual data was constructed. The model not only learns the statistical regularities of the numerical sequences themselves, but also the specific visual patterns and the complex mapping relationship between the specific transaction combinations supporting these patterns and future demand. This enables the model to distinguish between superficially similar but fundamentally different fluctuations, thus making more refined inferences. When encountering new data, the trained model can simultaneously analyze its potential graph patterns and refer to similar micro-transactional contexts for prediction, improving the accuracy and robustness of predicting demand triggered by specific events or transaction patterns. Furthermore, the model's decision-making basis becomes more transparent due to its association with specific transaction data. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating the working principle of the AI-based supply chain demand forecasting method described in this invention. Figure 2 A flowchart for filtering flow subsequences; Figure 3 A flowchart for generating a material flow rate change graph; Figure 4 This is a time-series fluctuation chart of material flow for antibiotic drug A (East China region). Figure 5 A comparison chart of demand forecasts based on material flow characteristics in the medical device supply chain. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 This invention provides an artificial intelligence-based supply chain demand forecasting method. The method includes: acquiring a complete set of transaction records for all participants within a target time period, and forming a material flow time series with time as the dimension using time-series processing; acquiring the boundary conditions of the current forecasting task, including the target material type, the forecast time span, and the forecast geographical range; filtering out flow subsequences from the material flow time series that match the target material type, the forecast time span, and the forecast geographical range; generating a material flow change map based on the flow subsequences to visually depict the dynamic fluctuation trajectory of material flow; acquiring preset map feature recognition rules and scanning the material flow change map using the map feature recognition rules to identify specific flow change patterns appearing in the map; marking the identified specific flow change patterns as features to be analyzed and extracting the original transaction record subset corresponding to each feature to be analyzed; constructing a deep neural network and using the features to be analyzed and their corresponding original transaction record subsets as input to train the deep neural network to learn the mapping relationship between the features to be analyzed and subsequent material demand.

[0020] Example 1: Transaction records related to the supply chain are separated from the complete transaction record set to form a supply chain transaction database. Key nodes of each transaction are located in the supply chain transaction database, and material flow information generated at each key node is recorded. The material flow information is processed temporally to form a material flow time series. When locating key nodes of each transaction in the supply chain transaction database, the time and location associated with changes in the physical displacement state of the materials are extracted from the records of each transaction. From the times and locations associated with changes in the physical displacement state, the times when materials enter the warehousing facility, leave the warehousing facility, are sorted, are loaded onto transport vehicles, and are unloaded from transport vehicles are selected. The selected times and locations are combined to define the key nodes of each transaction.

[0021] In practice, the scenario involves a large-scale cross-regional distribution network for consumer electronics products. This network generates a complete transaction record set containing 15 million records within a quarter. These records cover orders, payments, logistics tracking, warehouse receipts, delivery orders, and customs declarations. In practice, supply chain-related transaction records are separated from the complete transaction record set to form a supply chain transaction database. This operation is accomplished through a pre-defined supply chain business rule filter. The filter rules explicitly identify the physical flow, storage, and transportation of materials as supply chain-related. For example, a record containing a logistics waybill number, material SKU code, warehouse number, and actual inbound / outbound timestamps is identified and extracted, while a record containing only the end customer's payment amount and payment method is filtered out. After separation, the original complete transaction record set is reduced to 8 million records, which together constitute the supply chain transaction database upon which subsequent analysis relies.

[0022] In some embodiments, key nodes of each transaction are located in the supply chain transaction database and material flow information generated at each key node is recorded. This is achieved by parsing the detailed event logs recorded in each supply chain transaction record. In specific implementation, the time and location associated with the change in the physical displacement state of the material are extracted from the records of each transaction. This information is usually stored in the logistics status update field or the operation log of the warehouse management system. The parsing process identifies descriptive statuses such as "arrived at Shanghai central warehouse", "sent out from Beijing sorting center", and "container loaded at Shenzhen port" and their corresponding timestamps and geographical coordinates.

[0023] In some embodiments, the time when materials enter the storage facility, the time when materials leave the storage facility, the time when materials are sorted, the time when materials are loaded onto the transport vehicle, and the time when materials are unloaded from the transport vehicle are selected from the times and locations associated with the change in the physical displacement state. In specific implementation, the system matches and selects based on predefined key event keywords. For example, keywords such as "warehousing completed", "outbound confirmed", "sorting completed", "loading", and "unloading" trigger the capture of the corresponding time and location. The selected time and location combination is defined as the key node of each transaction. A material transaction from Shanghai to Chengdu is located with five key nodes, including the outbound time T1 and location L1 at the Shanghai warehouse, the loading time T2 and location L2 at the Shanghai logistics center, the sorting time T3 and location L3 at the transit station, the unloading time T4 and location L4 at the Chengdu logistics center, and the inbound time T5 and location L5 at the Chengdu warehouse.

[0024] It is understood that performing time-series processing on the material flow information to form a material flow time series is the foundation for subsequent steps. In specific implementation, each located key node generates a material flow information record. This record includes at least the material type code, the key node timestamp, the key node location, and the quantity of materials flowing through this node. For the aforementioned example transaction, a material flow information record for "outbound" will be generated at time T1 and location L1, with the quantity being the number of material items outbound in this transaction. The system globally sorts the material flow information records generated by all key nodes of all transactions by timestamp, preparing for the construction of a unified material flow time series. In specific implementation, the scattered material flow information records are integrated into time series data using the following aggregation formula: ; Where: symbol Indicates the first The total quantity of a specific type of material passing through all key nodes within a given natural day, symbolized by [symbol]. Indicates the first The total number of all key events occurring within a calendar day, symbol Indicates the first The first day of the calendar day The quantity of materials flowing through secondary key node events is recorded. All material flow information records are categorized into the corresponding natural day unit time window based on their key node timestamps, and the quantities of the same material type are accumulated. For example, on March 15th, the system accumulates the quantities of all materials marked as SKU "A-Phone-128G" at all "outbound" key nodes on that day, obtaining the total outbound flow of "A-Phone-128G" for that day. This process is repeated daily, ultimately generating an ordered list of values ​​arranged by date that reflects the total daily flow of a specific material, i.e., the material flow time series.

[0025] Example 2: See Figure 2 The total quantity of the same type of material passing through all key nodes within each natural day is summarized in a unit time window. The total quantity of materials summarized in each unit time window is arranged into an ordered numerical list according to the time sequence to form a material flow time series. From the material flow time series, entries whose material type is consistent with the predicted target material type are identified. From the identified entries, entries whose geographical coordinates of the transaction are located within the predicted geographical range are further filtered. From the further filtered entries, data segments corresponding to continuous time periods whose timestamps fall within the predicted time span are extracted as flow subsequences.

[0026] In practice, this involves a nationwide food and fresh produce supply chain network. This network processed over 30 million time-series material flow information records in the past year, forming an initial material flow time series. In practice, each natural day is used as a unit time window to summarize the total quantity of the same type of material passing through all key nodes within that unit time window. This operation is performed independently for each material type recorded in the material flow time series. The system reads the date, material type identifier, and flow value contained in each record in the series. For each natural day and each material type, the system performs an aggregation calculation, retrieving all records belonging to that date and material type. The flow rate values ​​are accumulated. For example, for a product labeled "Fresh Milk - 1L," the system accumulates the total quantity of all "Fresh Milk - 1L" items recorded on May 10th at key nodes in warehouse entry / exit, sorting, and loading / unloading. This total value is recorded as the total flow rate of "Fresh Milk - 1L" on May 10th. The total quantity of materials obtained from each unit time window is arranged into an ordered list of values ​​according to chronological order. This ordered list of values ​​constitutes the material flow rate time series. Thus, the daily total material flow rate of "Fresh Milk - 1L" from January 1st to December 31st is arranged sequentially, forming a material flow rate time series containing 365 ordered values. In specific implementation, the daily total material flow rate is calculated using the following accumulation formula: ; Where: symbol This indicates that for a specific material type, on the [number]th [year]... Total quantity of materials for a calendar day, symbol Indicates the first The total number of all key events involving this material type that occurred within a calendar day, symbolized by Indicates the first The first natural day involving this material type The material flow rate value corresponding to the secondary critical node is calculated by iterating through all dates and all material types in the material flow rate time series.

[0027] In some embodiments, flow subsequences that match the predicted target material type, predicted time span, and predicted geographical range are filtered from the material flow time series. The boundary conditions of the current prediction task are set as follows: the predicted target material type is "fresh milk - 1L", the predicted time span is the next thirty days, and the predicted geographical range is North China. In specific implementation, entries whose material type matches the predicted target material type are identified from the material flow time series. The system traverses each entry in the material flow time series and performs an exact match between the material type identifier field in the entry and the string "fresh milk - 1L". Successfully matched entries are retained. For example, an entry containing the material type identifier "yogurt - 200g" will be filtered out, while entries identified as "fresh milk - 1L" will be collected. After this step, the system filters out tens of thousands of time-sorted entries specifically for "fresh milk - 1L" from a material flow time series containing tens of thousands of materials and hundreds of millions of records.

[0028] In some embodiments, entries are further filtered from the identified entries to include those whose geographical coordinates of the transaction occurred within the predicted geographical range. Each entry in the material flow time series is associated with the geographical coordinates of its key generating node. The system loads the boundary polygon definition of the predicted geographical range and uses a spatial coordinate judgment algorithm to check whether the geographical coordinates associated with each entry are located inside the boundary polygon of North China. The coordinates (116.40, 39.90) in the entry "2023-08-15, Fresh Milk - 1L, 5000, 116.40, 39.90" are determined to be located within Beijing, so this entry is retained. However, the coordinates (121.47, 31.23) in the entry "2023-08-15, Fresh Milk - 1L, 3000, 121.47, 31.23" are determined to be located in Shanghai, which is outside the North China range, so this entry is filtered out. After geographical range filtering, tens of thousands of entries specifically for "Fresh Milk - 1L" are reduced to thousands of entries containing only flow data from North China.

[0029] Optionally, data segments corresponding to continuous time periods whose timestamps fall within the predicted time span can be extracted from the further filtered entries. The prediction task requires predicting future demand based on historical data, so the selected historical data time period needs to have a certain continuity and length. In specific implementation, the historical data time span used for model training is set to the past 180 days. The system finds all entries with datestamps within the past 180 days from the entries filtered by material type and geographical range, and extracts and arranges the total material flow values ​​of these entries according to date order. The data segment is the flow subsequence. For example, if the current date is October 1, 2023, the system extracts all daily flow data of "fresh milk - 1L" in North China from April 4, 2023 to September 30, 2023, forming a flow subsequence containing 180 continuous values.

[0030] It is understandable that through the above-mentioned step-by-step screening and extraction process, the final flow subsequence is strictly consistent with the boundary conditions of the prediction task in terms of material type, geographic space, and time. In specific implementation, the initial material flow time series contains a total of about 365 days × hundreds of nodes nationwide for all data of "fresh milk - 1L" in the past year. The flow subsequence formed after screening precisely corresponds to the daily total flow data of "fresh milk - 1L in North China in the past 180 days". The data scale is reduced from hundreds of millions of records to a sequence containing only 180 values.

[0031] Optionally, when extracting data segments to form traffic subsequences, if there are individual days without data, the system will use the average of the data from adjacent days to fill in the missing data to ensure the continuity of the time series. However, this filling operation is only performed when the proportion of missing days to the total time period is less than a preset threshold. For example, if there are three days without data in the past 180 days, the average traffic of the two days before and after will be used to generate the filling value. If the number of missing days exceeds five days, the system will prompt that the data is incomplete and suggest adjusting the prediction time span.

[0032] Example 3: See Figure 3 A two-dimensional coordinate system is established, where the horizontal axis represents time and the vertical axis represents material flow rate. In the two-dimensional coordinate system, each data point in the flow rate subsequence is located and marked according to its time and material flow rate value. Continuous line segments are used to connect the marked points of adjacent time points in sequence to form a broken line reflecting the change of material flow rate over time. The broken line and the coordinate system it is located in together constitute the material flow rate change map.

[0033] In the specific implementation, the target material type is "automotive engine ECU", the prediction time span is the next four weeks, and the prediction geographical range is the Yangtze River Delta region. Based on this, a flow subsequence containing 28 consecutive days of historical material flow data has been obtained. Each data point in the flow subsequence contains a date stamp and a corresponding total flow value for that day. For example, the data point sequence is [(Day1, 1500), (Day2, 1620), (Day3, 1480), ..., (Day28, 1710)]. In the specific implementation, a two-dimensional coordinate system is established, where the horizontal axis represents time and the vertical axis represents material flow. This coordinate system has a definite mathematical ratio. The time scale of the horizontal axis is evenly divided according to the time range covered by the flow subsequence and marked in natural days. The material flow scale of the vertical axis is set according to the maximum and minimum value range of all flow values ​​in the flow subsequence and marked in "pieces". The origin of the coordinate system is usually set as the intersection of the starting date of the horizontal axis and the zero flow of the vertical axis.

[0034] In some embodiments, each data point in the flow subsequence is located and marked in the two-dimensional coordinate system according to its time and material flow value. In a specific implementation, the system reads the first data point (Day1, 1500) in the flow subsequence, finds the position of the scale line representing "Day1" on the horizontal axis of the established two-dimensional coordinate system, and finds the position representing the flow value 1500 in the vertical direction of this scale line according to the vertical axis scale. The two positions determine a unique coordinate point, and the system draws a mark symbol at this coordinate point, such as a solid circle or a cross. This process is performed sequentially for each data point in the flow subsequence until the twenty-eighth data point (Day28, 1710) is also accurately located and marked in the coordinate system, thus forming twenty-eight discrete, time-sequentially distributed mark points in the coordinate system. The distribution of these mark points intuitively reflects the daily flow changes of "automobile engine ECU" in the Yangtze River Delta region over the past twenty-eight days.

[0035] In some embodiments, continuous line segments are used to sequentially connect adjacent time points to form a broken line reflecting the change of material flow over time. In specific implementations, the system processes these points in chronological order. First, a straight line segment is drawn between the point representing Day 1 and the point representing Day 2. Then, the next straight line segment is drawn between the point representing Day 2 and the point representing Day 3, and so on, until the last straight line segment is drawn between the point representing Day 27 and the point representing Day 28. These sequentially connected straight line segments together constitute a continuously fluctuating broken line that runs through the entire time range. The broken line and its coordinate system together constitute the material flow change map. For example, if the flow on Day 1 is 1500 units and the flow on Day 2 is 1620 units, the line segment connecting these two points shows an upward trend; if the flow on Day 3 is 1480 units, the line segment connecting Day 2 and Day 3 shows a downward trend. The fluctuation shape of the entire broken line visually demonstrates the dynamic trajectory of the material flow within the observation period.

[0036] Optionally, when locating and marking data points, the system uses the following mapping function to convert data values ​​into graphical coordinates: ; Where: symbol and This represents the pixel position of a data point in a two-dimensional coordinate system, indicated by the x and y coordinates. and The x and y coordinates of the origin of the drawing area in a two-dimensional coordinate system are represented by the pixel positions. Represents the pixel length corresponding to each unit of time on the horizontal axis, with the symbol... Indicates the date sequence number corresponding to the current data point, symbol Indicates the starting date sequence of the flow subsequence, symbol This represents the pixel height corresponding to each unit of flow on the vertical axis, with the symbol... This indicates the material flow rate value corresponding to the current data point.

[0037] It is understandable that the generated material flow rate change graph is a standard technical output that converts numerical sequences into visual graphics. In specific implementation, after the above steps, the original flow rate subsequence containing twenty-eight values ​​is transformed into a visual chart containing horizontal and vertical axes, scale labels, twenty-eight marker points, and twenty-seven line segments connecting these points. This visual chart is the material flow rate change graph. This graph allows the trend, fluctuation, and anomaly patterns of time series data to be directly observed by the human eye or processed by subsequent image recognition algorithms. For example, a periodic pulse pattern or a sudden flow rate drop plateau will appear as a geometric shape on the line graph. The specific implementation of the image recognition algorithm is based on preset graph feature recognition rules. First, a set of flow rate change pattern templates are defined, such as a rapid rise plateau pattern, a stepped drop pattern, and a periodic pulse pattern. These templates describe typical flow rate fluctuation patterns with standardized geometric shapes. The image recognition algorithm discretizes the continuous line in the material flow rate change graph into a sequence of coordinate points arranged in chronological order, and extracts local line segments through a sliding time window. Specifically, the discretization operation is based on the generation method of the material flow rate change map, which is a continuous polyline formed by connecting data points in the flow rate subsequence. Each data point corresponds to a specific timestamp and material flow rate value, and has a unique coordinate position in a two-dimensional coordinate system. The algorithm extracts the coordinate values ​​of all data points on the polyline, arranges them in chronological order to form a coordinate point sequence, thereby converting the continuous graphical representation into a digital point sequence, which is convenient for subsequent calculation and processing. This discretization process essentially restores the visualized polyline to its original set of data points, ensuring that each point retains the order in the time dimension and the accuracy of the flow rate value. Next, the operation of sliding the time window to extract local polyline segments starts from the starting point of the coordinate point sequence and slides gradually along the time axis with a preset window size. The window size is usually set according to the duration of the flow rate change pattern template, such as the minimum matching length required by the template, to ensure that the extracted local segment can cover the complete pattern features. In each slide, the algorithm extracts the coordinate point subsequence corresponding to consecutive time points within the window as the local polyline segment to be analyzed; the window sliding step size can be set to one time unit to achieve a comprehensive scan of the entire sequence. This sliding window mechanism allows the system to examine the polyline shape segment by segment and compare it with a preset template, thereby identifying specific flow change patterns and providing structured input for subsequent feature extraction and neural network training. Each local segment is compared point by point with the coordinate sequence of the shape template to calculate shape similarity; when the similarity exceeds a preset shape matching threshold, the local segment is determined to be a specific flow change pattern.Specifically, the system first represents both the local line segments and the preset flow change pattern templates in the material flow change map as normalized flow value sequences arranged in chronological order, ensuring consistent data scale. Each data point in the local line segment corresponds to a timestamp and a normalized flow value, while the pattern template is stored as a standardized point sequence, representing typical flow fluctuation patterns such as a sharp rise or a step-down. During point-by-point comparison, the system sequentially compares the normalized flow values ​​of the local segment sequence and the template sequence at the same time position, calculating the absolute difference between each corresponding point pair. These differences are accumulated to measure the degree of deviation of the overall shape. By comparing the accumulated differences with the total value of the template sequence, the system evaluates the matching degree between the local segment and the template. If the matching degree exceeds a preset pattern matching threshold, the local segment is determined to exhibit a specific flow change pattern. After recognition, the algorithm labels each pattern with an independent feature identifier and extracts the corresponding subset of original transaction records from the supply chain transaction database based on its time range, thereby associating the visualized pattern with micro-transaction data and providing structured input for deep neural network training. Optionally, when establishing a two-dimensional coordinate system, the scale range of the material flow rate on the vertical axis can be dynamically calculated based on the minimum and maximum values ​​of the flow rate subsequence data and expanded by a certain proportion to accommodate subsequent data point markers. The drawing of the polyline can use an anti-aliasing algorithm to make the lines smooth. The style of the marker points can be set to use red markers when the data point values ​​exceed the historical average and blue markers when they are below the average to enhance the visual contrast effect.

[0038] Example 4: Define a set of flow change pattern templates. The flow change pattern templates include a rapid rise plateau pattern, a step-down pattern, and a periodic pulse pattern. The flow change pattern templates are compared point by point with local line segments in the material flow change graph. When the shape of a local line segment in the graph matches the shape of any flow change pattern template more than a preset pattern matching threshold, the local line segment is determined to be the specific flow change pattern. Each identified specific flow change pattern is labeled with an independent feature identifier in the material flow change graph. Based on the start and end times covered by each specific flow change pattern on the time axis, all transaction records within the time period between the start and end times of the transaction timestamp are searched in the supply chain transaction database. All the found transaction records are combined to form the original transaction record subset corresponding to the specific flow change pattern.

[0039] In practical implementation, an example scenario involving a pharmaceutical supply chain is presented. The target material is predicted to be "antibiotic drug A," and the predicted geographical area is East China. The system has generated a corresponding material flow change map based on historical data from the past ninety days. This map is a continuous broken line that fluctuates over time in a two-dimensional coordinate system. In practical implementation, a set of flow change pattern templates are defined, including a rapid rise plateau pattern, a stepped decline pattern, and a periodic pulse pattern. The template data for the rapid rise plateau pattern shows that the flow value rises rapidly by more than a threshold percentage within three consecutive time units and then remains at a high level within the next five time units, with fluctuations not exceeding a specific percentage. The template data for the stepped decline pattern shows that the flow value drops significantly at several discrete time points and then remains stable for a period of time, forming a step-like shape. The template data for the periodic pulse pattern shows that the flow value rises and falls in a peak shape at approximately fixed time intervals.

[0040] In some embodiments, the flow change pattern template is compared point by point with the local broken line segments in the material flow change map. In a specific implementation, the system discretizes the continuous broken line of the material flow change map into a series of coordinate points arranged in chronological order. For each flow change pattern template, the system also stores it as a standardized coordinate point sequence. The comparison operation starts from the starting point of the broken line of the material flow change map, and a subsequence of coordinate points corresponding to a local broken line segment is extracted with a sliding time window. The similarity of this subsequence with the standard coordinate point sequence of each flow change pattern template is calculated. The similarity calculation involves comparing the relative position and change trend of the flow values ​​of the corresponding points.

[0041] In some embodiments, when the shape of a local polyline segment in the graph matches the shape of any flow change pattern template more than a preset shape matching threshold, the local polyline segment is determined to be the specific flow change pattern. In a specific implementation, the shape matching threshold is set to 85%, and the matching degree is calculated using the following formula: ; Where: symbol Represents the calculated shape matching percentage, symbol The symbol represents the length of the local polyline segment being compared with the morphological template sequence. This indicates the first segment of a local broken line in the material flow rate change graph. Normalized flow values ​​at each point, sign The template representing the pattern of flow change The normalized flow values ​​at each point, if calculated If the value is greater than 85%, it is determined that the currently scanned local line segment matches the template pattern. For example, if the system scans a line segment and the normalized value sequence matches the "rapid rise platform pattern" template sequence by 88%, then this line segment is identified as a specific flow change pattern of a rapid rise platform pattern.

[0042] Table 1: Template Parameter Table for Flow Change Patterns Shape Template Name Key geometric feature description Normalized numerical sequence (first 5 points) Minimum match length (time unit) Rapid rise platform pattern After a rapid and steep rise, it remained at a high level. [0.2,0.5,0.9,0.88,0.91] 8 Step descent pattern Decline in stages and levels [1.0,0.95,0.6,0.58,0.2] 6 Periodic pulse pattern Regularly spaced peaks and troughs [0.3,0.8,0.35,0.85,0.33] 10

[0043] It is understood that each specific flow change pattern identified in the material flow change map is marked with an independent feature identifier. In specific implementation, the system assigns a globally unique string as a feature identifier to each successfully identified local polyline segment. The encoding rules of the feature identifier include the abbreviation of the pattern type and the identification timestamp. For example, for a rapid rise plateau pattern identified in the time range of "Day45 to Day52", the system marks it with the feature identifier "RSP-20230415-001", where "RSP" represents the rapid rise plateau pattern, "20230415" represents the identification date, and "001" represents the identification sequence number of the day. This feature identifier serves as a unique index for the pattern in the map and database.

[0044] Optionally, based on the start and end times of each specific flow change pattern covered on the time axis, the system searches the supply chain transaction database for all transaction records within the time period between the start and end times. In a specific implementation, the feature identifier "RSP-20230415-001" is associated with a start time of "2023-04-01 00:00:00" and an end time of "2023-04-08 23:59:59". The system submits a structured query statement to the supply chain transaction database. The query conditions are that the transaction timestamp field is greater than or equal to the start time and less than or equal to the end time, the material type is "antibiotic drug A", and the geographical scope is located in East China. The database executes the query and returns all transaction records that meet the conditions within this seven-day time window. The returned records include drug outbound orders from different batches, waybills from different logistics companies, and receipt confirmations from different hospitals.

[0045] Optionally, all the found transaction records are combined to form a subset of the original transaction records corresponding to the specific flow change pattern. In a specific implementation, the database query returned 872 transaction records. The system extracts these 872 records from the original storage format of the database and encapsulates them into an independent data set. This data set is marked and bound to the feature identifier "RSP-20230415-001", becoming the subset of the original transaction records corresponding to this rapid rise platform pattern. Each record in the subset of original transaction records contains complete transaction details such as transaction ID, buyer and seller information, material quantity, amount, precise timestamp, and geographical location.

[0046] In some embodiments, the identification and feature extraction process is iterative. In a specific implementation, after the system completes the identification of the first specific flow change pattern and the extraction of the original transaction record subset, the sliding time window moves forward by one time unit and continues to scan the next local polyline segment of the material flow change graph. The process of template comparison, matching degree calculation, identifier labeling and record subset extraction is repeated until the entire material flow change graph polyline is traversed. The final output is a list, and each item in the list contains a feature identifier, a time range description of a specific flow change pattern, and a corresponding pointer to the original transaction record subset data.

[0047] See Figure 4 This chart presents a "Time-Series Fluctuation Chart of Material Flow for Antibiotic Drug A (East China Region)," which essentially displays the dynamic trajectory of the target material's flow over a 90-day period. The horizontal axis represents time (days), and the vertical axis represents the normalized material flow. The curves visually reflect the fluctuation characteristics of the flow over time, with different background color blocks corresponding to specific flow change patterns (such as rapid plateaus, periodic pulses, etc.) that can be identified.

[0048] From a professional perspective, this graph, as a visualization of material flow change patterns, allows for the identification of local line shapes using pre-defined flow change pattern templates (rapid rise plateau, stepped decline, periodic pulse). For example, the rapid increase in flow rate followed by a sustained high level during the 15-25 day period matches the template characteristics of a "rapid rise plateau" pattern; while the high-frequency fluctuations during the 55-65 day period correspond to the typical manifestation of a "periodic pulse" pattern. The value of such graphs lies in providing a correlation between the features to be analyzed and a subset of the original transaction records for subsequent deep neural network training. It serves as a crucial visualization support tool for the "graph feature recognition" stage in supply chain demand forecasting methods.

[0049] Example 5: Calculate the morphological description parameters for each feature to be analyzed. The morphological description parameters include the duration of the morphology, the magnitude of flow change, the flow value at the start of the morphology, and the flow value at the end of the morphology. Calculate the transaction attribute parameters for each subset of original transaction records. The transaction attribute parameters include the average number of transaction participants, the average geographical distance involved in the transaction, and the distribution density of transaction amount. Concatenate the morphological description parameters with the corresponding transaction attribute parameters to form a comprehensive feature vector. Use the comprehensive feature vector and the actual material demand change trend within a set time after the occurrence of the feature to be analyzed as a set of training samples and input them into a deep neural network for iterative training until the neural network can output a prediction of subsequent demand changes based on the input comprehensive feature vector. After the deep neural network completes training, when using the neural network for prediction, obtain the latest material flow data and generate the latest material flow change map according to the method. Identify the latest specific flow change pattern in the latest material flow change map and generate the corresponding latest comprehensive feature vector. Input the latest comprehensive feature vector into the trained deep neural network to output the predicted material demand value for a future period after the latest specific flow change pattern.

[0050] In a specific implementation example involving a medical device supply chain, multiple specific flow change patterns have been identified, and a corresponding subset of original transaction records has been extracted for each pattern. For example, a periodic pulse pattern labeled "Feature Identifier Pulse-2023Q3-005" and its corresponding subset of original transaction records containing 350 transactions are analyzed. In the implementation, the pattern description parameters are calculated for each feature to be analyzed. The pattern description parameters include the duration of the pattern, the magnitude of the flow change, the flow value at the beginning of the pattern, and the flow value at the end of the pattern. For the feature identifier Pulse-2023Q3-005, its duration is obtained by subtracting the beginning time from the end time of the pattern on the time axis, resulting in seven days. The magnitude of the flow change is obtained by calculating the absolute difference between the highest and lowest flow values ​​during the duration of the pattern, resulting in 850 units. The flow value at the beginning of the pattern and the flow value at the end of the pattern are directly obtained from the vertical coordinates corresponding to the beginning and end points of the pattern in the material flow change graph, which are 1,200 units and 1,150 units, respectively.

[0051] In some embodiments, transaction attribute parameters are calculated for each subset of original transaction records. These parameters include the average number of transaction participants, the average geographical distance involved in the transaction, and the distribution density of transaction amount. In a specific implementation, for the subset of original transaction records containing 350 transaction records corresponding to the feature identifier Pulse-2023Q3-005, the system reads the buyer and seller information recorded in each transaction record, calculates the number of participants in each transaction in the subset, and then calculates the arithmetic mean of the number of all transaction participants. The average geographical distance involved in the transaction is obtained by calculating the spherical distance between the shipping location and the receiving location of each transaction in the subset and then averaging all distance values. The distribution density of transaction amount is calculated using the following formula: ; Where: symbol The distribution density of transaction amounts, symbol Indicates the total number of transaction records in the original transaction record subset, symbol Represents the maximum value of all transaction amounts in the subset, with the symbol [symbol missing]. This represents the minimum transaction amount in the subset, and the transaction amount distribution density of that subset is calculated.

[0052] In specific implementation, the morphological description parameters are concatenated with the corresponding transaction attribute parameters to form a comprehensive feature vector. The system will arrange the four morphological description parameter values ​​and three transaction attribute parameter values ​​calculated for a feature to be analyzed in a predetermined order, for example, the order is [duration, flow change amplitude, flow value at the beginning of the morphological pattern, flow value at the end of the morphological pattern, average number of transaction participants, average geographical distance involved in the transaction, distribution density of transaction amount], and combine these seven values ​​into a one-dimensional array. This array is the comprehensive feature vector representing the specific flow change pattern and the underlying transaction pattern. The comprehensive feature vector of feature identifier Pulse-2023Q3-005 is [7,850,1200,1150,2,350.5,0.8].

[0053] It is understood that the comprehensive feature vector and the actual material demand change trend within a set time period after the occurrence of the feature to be analyzed are used as a set of training samples and input into a deep neural network for iterative training until the neural network can output a prediction of subsequent demand changes based on the input comprehensive feature vector. In specific implementation, the set time period is determined to be thirty days. For the feature identifier Pulse-2023Q3-005, the system searches for historical data of actual material demand within thirty days after the end of this form and uses this demand change trend data as the expected output label of the set of training samples. A complete set of training samples includes a seven-dimensional comprehensive feature vector input and an output label representing the daily demand sequence for the next thirty days. The system constructs a deep neural network with an input layer, multiple hidden layers and an output layer, inputs a large number of such training samples into the network, and iteratively adjusts the internal weight parameters of the network through the backpropagation algorithm so that the network output gradually approaches the real subsequent demand change trend.

[0054] In some embodiments, when the deep neural network is trained and then used for prediction, the following steps are included: obtaining the latest material flow data and generating the latest material flow change map according to the method. In a specific implementation, the system accesses a real-time data stream, obtains the latest material flow data from the past ninety days up to yesterday, and performs the same processing flow as the training phase, including filtering the flow subsequences that match the prediction target and generating the latest material flow change map reflecting the flow changes in the past ninety days.

[0055] Optionally, the system identifies the latest specific flow change pattern in the latest material flow change map and generates a corresponding latest comprehensive feature vector. The system scans the latest map using the same map feature recognition rules as in the training phase to identify the latest specific flow change pattern that has recently appeared and has not been used in training. For example, a new step-down pattern is identified in the last seven days. The system calculates the pattern description parameters for this pattern and calculates the transaction attribute parameters from the transaction records of the corresponding time period. These parameters are then concatenated to form a new latest comprehensive feature vector for prediction. Optionally, the latest comprehensive feature vector is input into a trained deep neural network. The deep neural network outputs a predicted value for material demand in the future period following the latest specific flow change pattern. In a specific implementation, the latest comprehensive feature vector representing the newly identified step-down pattern is input into the trained deep neural network. The forward propagation process of the neural network processes the input vector based on its internally learned mapping relationship and generates a prediction sequence. This sequence is the quantitative prediction value of material demand for the next thirty days, and the prediction value is output in the form of daily demand.

[0056] Furthermore, the internally learned mapping relationship is a non-linear correlation established by the deep neural network through the training process, from the comprehensive feature vector to the subsequent trend of material demand changes. This relationship is essentially a pattern encoded by the weight parameters within the neural network, dynamically linking the specific flow change pattern represented by the input comprehensive feature vector and its corresponding transaction attributes with future material demand. This mapping relationship captures the complex causal logic between visible morphological features and transaction contexts and demand fluctuations in historical data, enabling the neural network to infer potential demand changes based on newly input features. This mapping relationship is obtained through iterative training of the deep neural network using a large number of historical training samples. During training, each training sample consists of a comprehensive feature vector and the actual material demand change trend within a set time period after the occurrence of the feature to be analyzed. The neural network uses the backpropagation algorithm to repeatedly adjust the connection weights of its internal layers, gradually reducing the error between the network output value and the actual demand trend. After sufficient training, the internal parameters of the neural network stabilize, thus solidifying the learning results of the mapping relationship between input and output, forming a reusable predictive model. In the prediction phase, when the latest comprehensive feature vector is input into the trained neural network, the forward propagation process is initiated. The input vector passes sequentially through the input layer, hidden layer, and output layer of the neural network. Each layer performs linear combinations and nonlinear transformations on the data, progressively extracting and fusing feature information. Finally, the output layer generates an ordered numerical sequence, where each value corresponds to a demand forecast for the next day, with the entire sequence covering a thirty-day time span. This forecast sequence directly reflects the neural network's analysis of the input features based on learned mapping relationships, outputting daily demand as a quantitative basis for supply chain decisions.

[0057] See Figure 5 In this medical device supply chain demand forecasting scenario, this graph presents a time-series comparison of actual demand and predicted demand after identifying specific flow change characteristics. The horizontal axis represents the number of forecast days (unit: days), and the vertical axis represents the material demand (unit: units), showing the trajectories of change in "actual demand after Pulse feature" and "predicted demand after StepDown feature," respectively. The line representing "actual demand after the Pulse feature" exhibits high-frequency fluctuations, consistent with the flow dynamics corresponding to the periodic pulse pattern. Its value range is concentrated between 900 and 1500 units, reflecting the fluctuation range of actual demand after the occurrence of this feature. The line representing "predicted demand after StepDown feature" shows a gradually increasing trend, with the value range concentrated between 400 and 600 units. This corresponds to the model prediction results after the step-down pattern recognition, reflecting the demand change trend based on the comprehensive feature vector output of the deep neural network.

[0058] This figure visually quantifies the temporal relationship between actual demand and model prediction results corresponding to different flow characteristics, providing a visual basis for verifying the effectiveness of the "specific flow pattern - subsequent demand" mapping relationship in supply chain demand forecasting methods.

[0059] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0060] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-based supply chain demand forecasting method, characterized in that, Includes the following steps: Obtain the complete set of transaction records of all participants within the target time period, and use time-series processing to form a material flow time series with time as the dimension; Obtain the boundary conditions of the current forecasting task, including the type of target material, the time span of the forecast, and the geographical scope of the forecast; Filter out the flow subsequences from the material flow time series that match the predicted target material type, predicted time span, and predicted geographical range; A material flow rate change map is generated based on the flow rate subsequence, and the material flow rate change map is used to visually depict the dynamic fluctuation trajectory of the material flow rate; Obtain preset spectral feature recognition rules, and use the spectral feature recognition rules to scan the material flow rate change spectral and identify specific flow rate change patterns appearing in the spectral; The identified specific traffic change patterns are marked as features to be analyzed, and a subset of the original transaction records corresponding to each feature to be analyzed is extracted; A deep neural network is constructed, and the feature to be analyzed and its corresponding subset of original transaction records are used as input. The deep neural network is trained to learn the mapping relationship between the feature to be analyzed and subsequent material requirements.

2. The supply chain demand forecasting method based on artificial intelligence according to claim 1, characterized in that, The process of obtaining a complete set of transaction records for all participants within the target time period and forming a time-series material flow data using time-series processing includes: The supply chain transaction database is formed by separating the supply chain-related transaction records from the complete set of transaction records. Locate the key nodes of each transaction in the supply chain transaction database and record the material flow information generated at each key node; The material flow information is processed in a time-series manner to form a material flow time series with time as the dimension; The process of locating the key nodes of each transaction in the supply chain transaction database includes: Extract the time and location associated with changes in the physical displacement state of the material from the records of each transaction; From the times and locations associated with the changes in the physical displacement state, select the times when the material enters the storage facility, when the material leaves the storage facility, when the material is sorted, when the material is loaded onto the transport vehicle, and when the material is unloaded from the transport vehicle; The selected time and location combination is defined as the key node of each transaction.

3. The supply chain demand forecasting method based on artificial intelligence according to claim 2, characterized in that, When processing material flow information to form a time series of material flow, the following steps are included: Using each natural day as a unit time window, summarize the total quantity of the same type of material that passes through all key nodes within the unit time window; The total quantity of materials obtained from each unit time window is arranged into an ordered list of values ​​according to the chronological order. This ordered list of values ​​is the material flow time series.

4. The supply chain demand forecasting method based on artificial intelligence according to claim 3, characterized in that, When selecting flow subsequences from the material flow time series that match the predicted target material type, predicted time span, and predicted geographical range, the process includes: From the material flow time series, identify entries whose material type matches the predicted target material type; From the identified entries, further filter out entries whose geographical coordinates of the transaction occurred within the predicted geographical range; From the re-selected entries, extract the data segments corresponding to consecutive time periods whose timestamps fall within the predicted time span; these data segments are the flow subsequences.

5. The supply chain demand forecasting method based on artificial intelligence according to claim 4, characterized in that, When generating a material flow rate change map based on the flow rate subsequence, the following steps are included: Establish a two-dimensional coordinate system, where the horizontal axis represents time and the vertical axis represents material flow rate; In the two-dimensional coordinate system, each data point in the flow subsequence is located and marked according to its time and material flow value; By sequentially connecting the markers at adjacent time points using continuous line segments, a broken line reflecting the change in material flow over time is formed. The broken line and its coordinate system together constitute the material flow change map.

6. The supply chain demand forecasting method based on artificial intelligence according to claim 5, characterized in that, When scanning a material flow rate change graph using the aforementioned graph feature recognition rules, and identifying specific flow rate change patterns appearing in the graph, the process includes: Define a set of flow change pattern templates, including a rapid rise plateau pattern, a step-down pattern, and a periodic pulse pattern; Compare the flow change pattern template with the local broken line segments in the material flow change graph point by point; When the shape of a local broken line segment in the graph matches the shape of any flow change pattern template more than a preset pattern matching threshold, the local broken line segment is determined to be the specific flow change pattern.

7. The supply chain demand forecasting method based on artificial intelligence according to claim 6, characterized in that, When identifying specific patterns of traffic changes as features to be analyzed, and extracting a subset of the original transaction records corresponding to each feature, the process includes: In the material flow rate change map, each identified specific flow rate change pattern is labeled with an independent feature identifier; Based on the start and end times of each specific flow change pattern on the time axis, search the supply chain transaction database for all transaction records within the time period between the start and end times. All the found transaction records are combined to form the original transaction record subset corresponding to the specific flow change pattern.

8. The supply chain demand forecasting method based on artificial intelligence according to claim 7, characterized in that, When using the feature to be analyzed and its corresponding subset of original transaction records as input, and training the deep neural network to learn the mapping relationship between the feature to be analyzed and subsequent material requirements, the process includes: For each feature to be analyzed, its morphological description parameters are calculated, including the duration of the morphology, the magnitude of flow rate change, the flow rate value at the beginning of the morphology, and the flow rate value at the end of the morphology. For each subset of original transaction records, calculate its transaction attribute parameters, which include the average number of transaction participants, the average geographical distance involved in the transaction, and the distribution density of transaction amount; The morphological description parameters are concatenated with the corresponding transaction attribute parameters to form a comprehensive feature vector; The comprehensive feature vector and the actual material demand change trend within a set time period after the occurrence of the feature to be analyzed are used together as a set of training samples and input into a deep neural network for iterative training until the neural network can output a prediction of subsequent demand changes based on the input comprehensive feature vector.

9. The supply chain demand forecasting method based on artificial intelligence according to claim 8, characterized in that, After the deep neural network completes training, when it is used for prediction, it includes: Obtain the latest material flow data and generate the latest material flow change map according to the method described above; Identify the latest specific flow change pattern in the latest material flow change map and generate the corresponding latest comprehensive feature vector; The latest comprehensive feature vector is input into a trained deep neural network, which outputs a predicted material demand value for a future period following the latest specific flow change pattern.

10. An artificial intelligence-based supply chain demand forecasting system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the artificial intelligence-based supply chain demand forecasting method according to any one of claims 1 to 9.

Images