E-commerce intelligent prediction management method and system based on big data

Abstract

The invention relates to the technical field of e-commerce, in particular to an e-commerce intelligent prediction management method and system based on big data, and the method comprises the following steps: obtaining a real-time message in a buffer area, generating a throughput rate sequence based on a sliding window, carrying out the difference and mapping of throughput rates of adjacent periods, and constructing a gradient slope sequence; and extracting a transient abrupt change ratio by means of a moving average model to generate a pre-trigger level identifier, extracting a set boundary extreme value, carrying out quantitative evaluation to generate discrete parameters, correcting a transfer matrix, and executing steady-state iterative solution to obtain prediction confidence to construct an elastic control instruction. According to the method, the sudden change trend of the business time domain dimension is accurately captured and static reference deviation interference is eliminated, so that deep analysis and real-time correction of the dynamic transfer characteristics of the core transaction link are realized, and the dilemma of research and judgment distortion and scheduling lag existing in a traditional rule when facing burst traffic is thoroughly avoided; and the full-link prediction precision and the elastic resource management and control efficiency in a complex transaction scene are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of e-commerce technology, and in particular to an e-commerce intelligent prediction management method and system based on big data. Background Technology

[0002] The field of e-commerce technology mainly involves core aspects such as displaying product information, facilitating transactions, processing orders, making payments, settling logistics, and analyzing user behavior through internet platforms. It encompasses multiple stages, including data collection (such as user browsing history, click patterns, purchase history, and review information), data storage (such as distributed databases and data warehouses), data processing (such as cleaning, integration, and feature construction), and business applications (such as product recommendations, inventory management, price adjustments, and marketing strategy formulation). Overall, it constitutes a data-driven online transaction and operation management system.

[0003] Traditional e-commerce intelligent forecasting management methods refer to technical matters such as changes in product demand, user purchasing tendencies, inventory turnover, and sales trends. By summarizing and statistically analyzing historical order records in chronological order, combining holiday markings, promotional activity records, and product category tags, time series data is generated using a fixed period division method. Then, future sales volume is estimated through mean calculation, sliding window accumulation, or simple regression fitting. At the same time, inventory replenishment time and quantity are determined based on manually set thresholds, and user access frequency and purchase frequency are classified through a pre-set rule table to achieve prediction and management of product sales and operational rhythm.

[0004] Existing technologies typically aggregate historical order records in chronological order, generate time-series data using a fixed period division method, estimate future sales volume using mean calculation or simple regression fitting, determine replenishment quantity based on manually set thresholds, and classify user access and purchase frequency based on preset fixed rules. This operating mode is limited by a fixed time span and static evaluation rules, making it difficult to accurately capture sudden traffic characteristics and dynamic demand changes. This results in low traffic prediction accuracy and serious lag in resource scheduling, which can easily lead to business processing congestion or idle and wasted system resources. Summary of the Invention

[0005] To address the technical problems of existing technologies that typically aggregate historical order records chronologically, generate time-series data using a fixed period division method, estimate future sales volume using mean calculation or simple regression fitting, determine replenishment quantities based on manually set thresholds, and classify user access and purchase frequencies based on preset fixed rules, this operational mode is limited by a fixed time span and static evaluation rules. It is difficult to accurately capture sudden traffic characteristics and dynamic demand changes, resulting in low traffic prediction accuracy and serious lag in resource scheduling. This can easily lead to business processing congestion or idle and wasted system resources. The present invention provides an e-commerce intelligent prediction management method based on big data.

[0006] To achieve the above objectives, this invention employs a big data-based intelligent predictive management method for e-commerce, comprising the following steps: S1: Obtain real-time order messages from the streaming data buffer of the e-commerce platform, parse and extract the order timestamp and order type identifier, perform time span segmentation based on the sliding window algorithm, classify and statistically analyze the order type identifier, and generate an order throughput sequence. S2: Obtain the order throughput sequence, perform differential operation on the order throughput of adjacent sampling periods to extract the throughput difference and perform time-domain rate of change mapping to construct a throughput gradient slope sequence; S3: Obtain the throughput gradient slope sequence, input it into the weighted moving average model to perform baseline deviation analysis to extract the transient mutation ratio, perform matching and filtering based on the preset trigger threshold interval, and generate peak pre-trigger level identifier; S4: Obtain the peak pre-trigger level identifier to construct a level convergence set, extract the highest frequency identifier as the periodic dominant trigger level, extract the boundary extreme value of the level convergence set and perform state span quantitative evaluation to generate trigger fluctuation discrete parameters; S5: Use the triggering fluctuation discrete parameters to correct the initial state transition matrix constructed based on the preset full-link logs, and perform steady-state probability distribution iterative calculation to obtain the traffic prediction confidence, extract the corresponding elastic quota parameters, and construct e-commerce prediction and control instructions.

[0007] As a further aspect of the present invention, the order throughput sequence includes the throughput of home appliance orders, the throughput of apparel orders, and the throughput of digital product orders; the throughput gradient slope sequence includes the first-order difference value, the second-order derivative, and the rate of change; the peak pre-trigger level identifier includes the normal release identifier, the flow-limiting queuing identifier, and the circuit breaker degradation identifier; the trigger fluctuation discrete parameters include the overall variance, the sample standard deviation, and the interquartile range; and the e-commerce predictive control instructions include the elastic scaling instruction, the token bucket flow-limiting instruction, and the service degradation instruction.

[0008] As a further aspect of the present invention, the specific steps of S1 are as follows: S101: Collect real-time business messages from the e-commerce platform buffer, extract the order timestamp and order type identifier from the message header node fields, obtain the timestamp representation value and the identifier character code, use the timestamp representation value as the primary key and the identifier character code as the value to perform association, and generate a temporal feature mapping table. S102: Call the temporal feature mapping table to obtain the window step parameter and span tolerance parameter, compare the timestamp representation value with the span tolerance parameter, filter the data items that fall within the range of the span tolerance parameter, and repeat the filtering according to the window step parameter shift range to obtain the time domain segmented state set. S103: Perform classification operation based on the time-domain segmented state set, extract the identifier character code within the interval, merge similar data items and sum the quantities to generate the total number of order categories in the interval segment, combine the total number of order categories in the interval segment into a one-dimensional array according to the time-transition sequence, and establish the order throughput sequence.

[0009] As a further aspect of the present invention, the specific steps of S2 are as follows: S201: Call the order throughput sequence, read the throughput corresponding to the discrete nodes in the sequence in sequence, extract the values ​​of two adjacent nodes in the order of time arrangement, perform differential operation to obtain the throughput difference, arrange the difference into a one-dimensional linear array structure in the unidirectional order of time progression, and establish a periodic throughput difference set. S202: Based on the periodic throughput difference set, obtain the preset sampling duration parameter, perform a division operation on multiple differences in the periodic throughput difference set with the sampling duration parameter, calculate the rate of change within the sampling interval, and sequentially concatenate the rates of change according to the order of the nodes where the throughput difference is located to obtain the time domain change mapping column. S203: For the time-domain change mapping column, set a gradient slope plane reference space, extract the rate of change in the time-domain change mapping column as the vertical coordinate scalar, use the relative displacement step size of the corresponding node as the horizontal coordinate scalar, construct two-dimensional coordinate mapping points and connect the points along the time axis to generate a throughput gradient slope sequence.

[0010] As a further aspect of the present invention, the specific steps of S3 are as follows: S301: Obtain the throughput gradient slope sequence, extract the original slope value of the node, read the sliding time window span variable and the node decay weight factor, multiply the original slope value and the node decay weight factor within the sliding time window span variable, sum them, and divide by the sum of the node decay weight factors to establish a dynamic weighted benchmark moving average series. S302: Call the dynamic weighted benchmark moving average column, subtract the reference benchmark mean in the dynamic weighted benchmark moving average column from the original slope value at the same sequential position, extract the time domain absolute deviation parameter, divide the time domain absolute deviation parameter by the corresponding reference benchmark mean, extract the positive mutation term, and generate the transient mutation ratio set. S303: For the set of transient change ratios, retrieve the upper and lower limit amplitude judgment parameters within the preset trigger state threshold interval set. If the judgment ratio falls into the target state threshold interval, read the target state threshold interval mapping classification characters, and concatenate the classification characters in order to obtain the peak pre-trigger level identifier.

[0011] As a further aspect of the present invention, the determination ratio falling into the target state threshold range refers to establishing multiple continuous numerical determination ranges based on the upper and lower limit amplitude determination parameters, comparing the ratio values ​​in the transient change ratio set with the lower limit boundary and the upper limit boundary in the numerical determination range respectively, extracting the numerical determination range where the ratio values ​​are within the lower limit boundary and the upper limit boundary, and setting it as the target state threshold range. The trigger state threshold interval set is obtained by extracting a historical mutation ratio sample set, calculating the expected mean and standard deviation of the historical mutation ratio sample set, multiplying the standard deviation by a fixed proportional constant to obtain the tolerance range parameter, and adding and subtracting the tolerance range parameter based on the expected mean of the data to obtain the upper and lower limits of fluctuation. The span between the lower and upper limits of fluctuation is divided equally, and the values ​​of the division nodes are extracted as upper and lower limit amplitude determination parameters. Adjacent upper and lower limit amplitude determination parameters are paired and combined to establish a numerical determination interval, and the numerical determination interval is integrated to determine the value.

[0012] As a further aspect of the present invention, the specific steps of S4 are as follows: S401: Obtain the peak pre-trigger level identifier, extract the aggregation period parameter, aggregate the peak pre-trigger level identifier within the time span of the aggregation period parameter, count and record the occurrence frequency parameter, associate each category identifier with the occurrence frequency parameter, and establish a level frequency mapping set. S402: Call the level frequency mapping set to traverse the occurrence frequency parameter, extract the maximum value corresponding to the identifier as the periodic dominant trigger level, search the upper and lower limit boundary identifiers in the level frequency mapping set, extract the preset danger quantization scale, combine the upper and lower limit boundary identifiers with the danger quantization scale to obtain the boundary extreme value state vector. S403: For the boundary extreme state vector, extract the upper distribution limit value and the lower distribution limit value, subtract the lower distribution limit value from the upper distribution limit value to extract the state span difference, read the dominant quantization value of the period dominant trigger level, divide the state span difference by the dominant quantization value, and generate the trigger fluctuation discrete parameter.

[0013] As a further aspect of the present invention, the specific steps of S5 are as follows: S501: Obtain the trigger fluctuation discrete parameters, extract the page jump sequence from the preset full-link log, count the jump frequency between each page node in the page jump sequence, divide the jump frequency by the total outflow frequency of the corresponding source page node, calculate the initial transition probability between nodes, and construct the initial state transition matrix. S502: Call the initial state transition matrix, extract the preset core transaction link node as the target correction object, perform nonlinear amplification mapping processing on the transition probability corresponding to the target correction object according to the trigger fluctuation discrete parameter, and perform corresponding deduction and row vector normalization processing on the transition probability of non-core transaction link nodes to generate a dynamic transition feature matrix. S503: Call the dynamic transition feature matrix to obtain the real-time online user distribution vector as the initial state vector, perform continuous iterative multiplication calculation, extract the converged steady-state probability vector as the traffic prediction confidence, match the quota correlation matrix to extract the corresponding elastic quota parameters, and construct e-commerce prediction and control instructions.

[0014] As a further aspect of the present invention, the step of performing nonlinear amplification mapping processing on the transition probability corresponding to the target correction object based on the trigger fluctuation discrete parameter, and performing corresponding deduction and row vector normalization processing on the transition probability of non-core transaction link nodes refers to performing division with a constant 1 as the dividend and the trigger fluctuation discrete parameter as the divisor, extracting the reciprocal as the mapping exponent parameter, using the transition probability pointing to the core transaction link node in the initial state transition matrix as the base, and performing a power function operation with the mapping exponent parameter as the exponent to obtain the amplified correction core probability, extracting the original transition probability of each non-core transaction link node in the same row, performing a summation with the correction core probability to obtain the amplified probability sum, using the correction core probability and the original transition probability of each non-core transaction link node as the dividend, and performing division normalization scaling with the amplified probability sum as the divisor to update the probability value of the current row.

[0015] A big data-based intelligent predictive management system for e-commerce includes: The order flow aggregation module acquires real-time order messages from the streaming data buffer of the e-commerce platform, parses and extracts order timestamps and order type identifiers, performs time span segmentation based on the sliding window algorithm, and performs classification and statistics based on order type identifiers to generate an order throughput sequence. The rate of change extraction module obtains the order throughput sequence, performs differential operation on the order throughput of adjacent sampling periods to extract the throughput difference, and performs time-domain rate of change mapping to construct a throughput gradient slope sequence. The traffic mutation sensing module acquires the throughput gradient slope sequence, inputs it into the weighted moving average model to perform baseline deviation analysis to extract the transient mutation ratio, performs matching and filtering based on a preset trigger threshold range, and generates a peak pre-trigger level identifier. The level quantification and evaluation module obtains the peak pre-trigger level identifier to construct a level convergence set, extracts the highest frequency identifier as the periodic dominant trigger level, extracts the boundary extreme value of the level convergence set and performs state span quantification and evaluation to generate trigger fluctuation discrete parameters. The quota elastic control module uses the trigger fluctuation discrete parameter to correct the initial state transition matrix constructed based on the preset full-link log, and performs steady-state probability distribution iterative calculation to obtain the traffic prediction confidence, extract the corresponding elastic quota parameters, and construct e-commerce prediction control instructions.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, real-time messages in the buffer are acquired and a throughput sequence is generated by segmentation and classification based on a sliding window. Differentiation and mapping are performed on the throughput of adjacent periods to construct a gradient slope sequence. The baseline deviation is analyzed using a weighted moving average model to extract the transient mutation ratio and generate a pre-trigger level identifier. The set extreme values ​​are extracted and quantitatively evaluated to generate discrete parameters and correct the initial transition matrix. The steady-state probability distribution is iteratively calculated to obtain confidence and construct control instructions. This accurately captures the business time-domain mutation trend and eliminates static deviation interference, achieving in-depth analysis and correction of dynamic transition features. This avoids the judgment distortion dilemma when facing sudden traffic and improves the link prediction accuracy and elastic control efficiency in complex scenarios. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the accompanying drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2 This is a detailed schematic diagram of S1 of the present invention; Figure 3 This is a detailed schematic diagram of S2 of the present invention; Figure 4 This is a detailed schematic diagram of S3 of the present invention; Figure 5 This is a detailed schematic diagram of S4 of the present invention; Figure 6 This is a detailed schematic diagram of S5 of the present invention; Figure 7 This is a system module diagram of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0020] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0021] Please see Figure 1This invention provides an intelligent predictive management method for e-commerce based on big data, comprising the following steps: S1: Obtain real-time order messages from the streaming data buffer of the e-commerce platform, parse and extract the order timestamp and order type identifier, perform time span segmentation based on the sliding window algorithm, classify and statistically analyze the order type identifier, and generate an order throughput sequence. S2: Obtain the order throughput sequence, perform differential operation on the order throughput of adjacent sampling periods to extract the throughput difference and perform time-domain change rate mapping to construct the throughput gradient slope sequence; S3: Obtain the throughput gradient slope sequence, input it into the weighted moving average model to perform baseline deviation analysis to extract the transient mutation ratio, perform matching and filtering based on the preset trigger threshold range, and generate peak pre-trigger level identifiers; S4: Obtain the peak pre-trigger level identifier to construct a level convergence set, extract the highest frequency identifier as the periodic dominant trigger level, extract the boundary extreme value of the level convergence set and perform state span quantitative evaluation to generate trigger fluctuation discrete parameters; S5: Use the triggered fluctuation discrete parameters to correct the initial state transition matrix built based on the preset full-link logs, and perform steady-state probability distribution iterative solution to obtain the traffic prediction confidence, extract the corresponding elastic quota parameters, and build e-commerce prediction and control instructions.

[0022] The order throughput sequence includes the order throughput of home appliances, apparel, and digital products. The throughput gradient slope sequence includes the first-order difference, second-order derivative, and rate of change. The peak pre-trigger level indicators include normal release, flow restriction queuing, and circuit breaker degradation indicators. The trigger fluctuation discrete parameters include the overall variance, sample standard deviation, and interquartile range. The e-commerce forecast control instructions include elastic scaling instructions, token bucket flow restriction instructions, and service degradation instructions.

[0023] Please see Figure 2 The specific steps of S1 are as follows: S101: Collect real-time business messages from the e-commerce platform buffer, extract the order timestamp and order type identifier from the message header node fields, obtain the timestamp representation value and the identifier character code, use the timestamp representation value as the primary key and the identifier character code as the value to perform association, and generate a temporal feature mapping table. The system reads the distributed message queue in the e-commerce platform server's memory via a data interface, directly extracting real-time business message data from the buffer. Lightweight data exchange format parsing is performed on the extracted business message data, and data cleaning preprocessing is completed by matching and removing null values ​​and abnormal garbled characters from the message content. Order timestamps and order type identifiers are precisely located from the cleaned message header fields. The extracted order timestamps are converted to a unified 10-digit decimal timestamp representation, and the order type identifier is mapped to a 2-digit identifier character code according to a pre-defined product category dictionary. Based on this, a hash mapping logic is used to set the 10-digit decimal timestamp representation as the primary key of the in-memory database, and the corresponding 2-digit identifier character code is set as the value under this primary key, performing strict key-value pair binding and dynamically generating a temporal feature mapping table in memory. For example, in actual large-scale promotional business monitoring, the converted representation value of the order occurrence time is extracted from the bottom-level nodes as 1774224000. Simultaneously, the order category is identified as home appliances, and its identifier character code is obtained as 12 from the classification dictionary. Then, 1774224000 is directly used as the primary key, and 12 as the corresponding value. A memory-level association operation is performed and the data is written into the temporal feature mapping table. The maximum resident capacity threshold of the mapping table is strictly set at 500,000 records. This specific value is calculated by referring to the maximum actual concurrent requests per second of a single node during the Double Eleven flash sales of e-commerce platforms over the past three years, multiplied by a redundancy factor of 1.5 for leniency. Its effective experimental value range during stress testing and optimization is between 300,000 and 800,000.

[0024] S102: Call the temporal feature mapping table to obtain the window step parameter and span tolerance parameter, compare the timestamp representation value with the span tolerance parameter, filter the data items that fall within the range of the span tolerance parameter, and repeat the filtering according to the window step parameter's shift range to obtain the time domain segmented state set; The system retrieves the established temporal feature mapping table from the memory data area, reads all primary keys and values ​​from internal storage, and obtains the window stepping parameter and span tolerance parameter for dividing time segments from the configuration center. Each timestamp representation value in the temporal feature mapping table is extracted and compared with the span tolerance parameter in turn, and a judgment logic is executed. Specifically, the difference between the current processing time reference point and the timestamp representation value is calculated. If the difference is less than or equal to the span tolerance parameter, the timestamp representation value is deemed to meet the condition, and its corresponding data item is filtered out and included in the current time window. If the difference is greater than the span tolerance parameter, it is discarded. After completing the filtering of the current time window, based on the value of the window stepping parameter, the time reference point is shifted forward, and the above difference calculation and filtering logic is repeated at the new time reference point. After multiple shifts and filterings, multiple sets of data items within discrete time periods are obtained, forming a time-domain segmented state set. For example, a time baseline is set to 1774224000, and the span tolerance parameter is obtained from the configuration center as 300 seconds, with a window step parameter of 60 seconds. When traversing the temporal feature mapping table, if a data item with a timestamp value of 1774224150 is encountered, its difference from the baseline is calculated, resulting in a difference of 150 seconds. Since 150 seconds is less than the span tolerance parameter of 300 seconds, this data item is filtered into the first time window. Subsequently, the baseline is shifted by 60 seconds according to the window step parameter, becoming 1774224060, and the filtering is repeated. The span tolerance parameter of 300 seconds is set based on the average operation dwell time of the user from browsing products to finally submitting the order, with an effective value range of 120 seconds to 600 seconds. The window step parameter of 60 seconds is optimized within the range of 30 seconds to 120 seconds using a grid search method to ensure the smoothness of monitoring.

[0025] S103: Perform classification operations based on the time-domain segmented state set, extract the identifier character code within the interval, merge similar data items and sum the quantities to generate the total number of order categories in the interval segment, combine the total number of order categories in the interval segment into a one-dimensional array according to the time-transition sequence, and establish the order throughput sequence. Fine-grained classification logic is executed on the extracted time-domain segmented state set, sequentially traversing the data items retained in each time span interval to accurately extract the identifier character code carried by each data item. A character matching comparison operation is performed on the extracted identifier character codes, grouping and merging data items with the same identifier character code. Then, an accumulation counting mechanism is initiated to sum the number of data items belonging to a certain identifier character code category, thus obtaining the total number of order categories within the current single interval. After completing the summation calculation for each independent interval, the total number of order categories calculated in each interval is sequentially concatenated according to the chronological order of the window reference point, forming a continuous one-dimensional array structure. This array ultimately serves as the order throughput sequence reflecting business traffic density. For example, in the first transition interval, after traversal and extraction, 150 data items have the identifier character code 12 representing home appliances, and another 80 data items have the identifier character code 15 representing clothing. Summation was performed on these two data categories separately, yielding total order categories of 150 and 80 for each interval. The total throughput for the current interval was accumulated to 230. Subsequently, in the adjacent second interval, the same category merging and summation operation was performed, resulting in a total of 250. The 230 from the first interval and the 250 from the second interval were stored in a one-dimensional array in chronological order, constructing an order throughput sequence such as array elements 230, 250, and 280. In the above summation calculation, the merging criteria for similar items required that the character encodings of both items be completely identical, with a fault-tolerant matching rate limited to 0.

[0026] Please see Figure 3 The specific steps of S2 are as follows: S201: Call the order throughput sequence, read the throughput corresponding to the discrete nodes in the sequence in sequence, extract the values ​​of two adjacent nodes in the order of time arrangement, perform differential operation to obtain the throughput difference, arrange the difference into a one-dimensional linear array structure in the unidirectional order of time progression, and establish a periodic throughput difference set. The system retrieves the previously constructed order throughput sequence from the data cache and sequentially reads the specific throughput values ​​recorded at each discrete node within this one-dimensional sequence. During the reading process, it strictly adheres to the temporal arrangement of sequence elements, synchronously extracting the values ​​corresponding to the two adjacent nodes in the time dimension at each step. A numerical difference operation is performed on the extracted values ​​of the two adjacent nodes. Specifically, the throughput value of the later time node is subtracted from the throughput value of the earlier time node to calculate the throughput difference between the two nodes. After obtaining all throughput differences, the calculated difference values ​​are rearranged and combined into a new one-dimensional linear array structure according to the unidirectional order of time window progression, thus establishing a periodic throughput difference set to reflect the fluctuation range of throughput. For example, the throughput values ​​corresponding to the first three discrete nodes in the order throughput sequence are 230, 250, and 280, respectively. First, extract the values ​​of the first adjacent node (230) and the second adjacent node (250), perform a difference operation, and subtract 230 from 250 to obtain a throughput difference of 20. Next, extract the values ​​of the second and third adjacent nodes (250 and 280) in chronological order, and perform the same difference operation, subtracting 250 from 280 to obtain the next throughput difference of 30. This process continues, sequentially filling the calculated differences of 20 and 30 into the contiguous memory space of a one-dimensional linear array according to the unidirectional temporal progression, generating a periodic throughput difference set containing elements 20 and 30. The difference operation here is strictly limited to adjacent nodes with a step size of 1, used to capture the finest-grained transient traffic jumps.

[0027] S202: Based on the periodic throughput difference set, obtain the preset sampling duration parameter, perform division operation on multiple differences in the periodic throughput difference set with the sampling duration parameter, calculate the rate of change within the sampling interval, and sequentially concatenate the rates of change according to the order of the nodes where the throughput difference is located to obtain the time domain change mapping column. Based on the established periodic throughput difference set in memory, a pre-set sampling duration parameter is obtained from the global configuration environment. For multiple consecutive difference values ​​recorded within the periodic throughput difference set, each is sequentially extracted and divided by the obtained sampling duration parameter. Specifically, the calculation mechanism involves using the individual throughput difference within the periodic throughput difference set as the dividend and the sampling duration parameter as the divisor to accurately calculate the flow change rate within each sampling interval. After all division operations are completed, the calculated change rate results are concatenated sequentially according to the original time sequence of each throughput difference within the periodic throughput difference set. For example, the first two difference values ​​retrieved from the preceding process are 20 and 30, and the preset sampling duration parameter is set to 60 seconds from the system environment variables. For the first difference, a division operation is performed, dividing the dividend 20 by the divisor 60, resulting in a change rate of 0.33 times per second within the sampling interval. Then, a division operation is performed for the second difference, dividing the dividend 30 by the divisor 60, resulting in a change rate of 0.50 times per second within the second sampling interval. Following the original order of the nodes containing the differences, the change rates of 0.33 and 0.50 are sequentially concatenated into a new sequence memory segment, obtaining the time-domain change mapping column. The sampling duration parameter of 60 seconds was set as the optimal balance point obtained through statistical regression testing of system stress logs from the most recent 12 major e-commerce promotional events, with its reasonable value strictly limited to between 30 and 120 seconds.

[0028] S203: For the time-domain variation mapping column, set the gradient slope plane reference space, extract the rate of change in the time-domain variation mapping column as the vertical coordinate scalar, use the relative displacement step size of the corresponding node as the horizontal coordinate scalar, construct two-dimensional coordinate mapping points and connect the points along the time axis to generate the throughput gradient slope sequence. The generated time-domain change mapping column is retrieved, and a standard gradient slope plane reference space is initialized and set in the virtual memory space. In this two-dimensional coordinate system, the change rate values ​​stored within the time-domain change mapping column are extracted sequentially and used directly as the vertical coordinate scalars in the two-dimensional coordinate system. Simultaneously, the counting process reads the relative displacement step size experienced by the node corresponding to the current change rate and uses this displacement step size value directly as the horizontal coordinate scalar in the two-dimensional coordinate system. Based on the extracted horizontal and vertical coordinate scalar pairs, corresponding two-dimensional coordinate mapping points are constructed in the gradient slope plane reference space. After the mapping points of all nodes are constructed, adjacent two-dimensional coordinate mapping points are connected point-to-point using a straight-line equation, strictly following the horizontal time axis, to generate a throughput gradient slope sequence reflecting the overall flow acceleration trend. For example, when constructing the gradient slope plane reference space, if the change rate of the first node in the time-domain change mapping column is extracted to be 0.33, its first time-series displacement step size is recorded as 1. The vertical coordinate scalar 0.33 is then paired with the horizontal coordinate scalar 1 to construct the first two-dimensional coordinate mapping point, containing the horizontal coordinate 1 and the vertical coordinate 0.33. The rate of change of the second node is extracted to be 0.50, and the relative displacement step size is accumulated to 2, thus constructing the second two-dimensional coordinate mapping point, containing the horizontal coordinate 2 and the vertical coordinate 0.50. The difference between the vertical coordinates of these two points is then calculated and divided by the difference between the horizontal coordinates. Subtracting 0.33 from 0.50 yields 0.17, and subtracting 1 from the horizontal coordinate span of 2 yields 1. Finally, dividing these two values ​​yields the initial slope value of the line segment, which is 0.17. The relative displacement step size here is uniformly calculated using a dimensionless integer sequence, continuously accumulated starting from 1.

[0029] Please see Figure 4 The specific steps of S3 are as follows: S301: Obtain the throughput gradient slope sequence, extract the original slope value of the node, read the sliding time window span variable and the node decay weight factor, multiply the original slope value and the node decay weight factor within the sliding time window span variable, sum them, and divide by the sum of the node decay weight factors to establish a dynamic weighted benchmark moving average series. The generated throughput gradient slope sequence is obtained, and the original slope values ​​calculated from the connections between each node in the sequence are extracted one by one. The pre-defined sliding time window span variable and the node decay weight factor corresponding to each node are read from the system parameter library. Within the defined sliding time window span variable, all original slope values ​​included in the current window are multiplied by their corresponding node decay weight factors to obtain the weighted slope value for each node. Then, all weighted slope values ​​within the window are summed to obtain a weighted sum, and simultaneously, all node decay weight factors involved in the calculation are summed to obtain a weight factor sum. Finally, the weighted sum is divided by the weight factor sum to obtain the reference mean for the current time window, and a dynamic weighted reference mean line column is generated as the window moves. For example, the sliding time window span variable is set to include 3 consecutive nodes, and the original slope values ​​for the most recent 3 nodes are 0.17, 0.25, and 0.40, respectively. The corresponding node attenuation weight factors are set to 0.20, 0.30, and 0.50 according to the time distance, with the weight increasing closer to the current time. First, multiplication is performed to calculate 0.17 multiplied by 0.20 equals 0.034, 0.25 multiplied by 0.30 equals 0.075, and 0.40 multiplied by 0.50 equals 0.200. Then, these three products are summed to obtain a weighted sum of 0.309, and the attenuation weight factors 0.20, 0.30, and 0.50 are summed to obtain a total weight factor sum of 1.0. Finally, 0.309 is divided by 1.0 to obtain the reference baseline mean of 0.309, which is included in the dynamic weighted baseline moving average. The node attenuation weight factors are set using an exponential decay method, optimized based on the characteristics of 50 previous traffic surge events.

[0030] S302: Call the dynamic weighted benchmark moving average column, subtract the reference benchmark mean in the dynamic weighted benchmark moving average column from the original slope value at the same time sequence position, extract the time domain absolute deviation parameter, divide the time domain absolute deviation parameter by the corresponding reference benchmark mean, extract the positive mutation term, and generate the transient mutation ratio set. The system retrieves the pre-calculated dynamic weighted benchmark moving average column and performs node-level time series table column alignment and comparison. At identical and aligned time series nodes, the raw slope value is extracted and forcibly subtracted from the corresponding generated reference benchmark mean within the dynamic weighted benchmark moving average column. Through this subtraction operation, the time-domain absolute deviation parameter reflecting the magnitude of local abnormal mutations is accurately calculated and extracted. To eliminate measurement errors caused by different base values, the calculated time-domain absolute deviation parameter is then directly divided by the reference benchmark mean at its corresponding position. Through this division ratio operation, positive mutation terms with positive results are extracted separately. All filtered positive mutation terms are sequentially stored in a cache array according to the time pipeline's derivation order, thereby generating a set of transient mutation ratios used to characterize the degree of traffic surge anomalies. For example, at a certain identical time series node, the raw slope value extracted from real-time monitoring feedback is 0.450, while the corresponding reference benchmark mean of 0.300 is accurately retrieved from the dynamic weighted benchmark moving average column. First, a subtraction operation is performed on these two values: 0.450 minus 0.300, accurately calculating the time-domain absolute deviation parameter for this specific time point as 0.150. Next, a division ratio calculation is performed on this deviation result: the calculated time-domain absolute deviation parameter 0.150 is divided by the reference baseline mean of this node, 0.300, yielding a ratio of 0.500. Since this result 0.500 is greater than 0, it meets the criteria for a positive mutation term, and is therefore saved and added to the transient mutation ratio set. The time-domain absolute deviation parameter establishes the abnormal offset by calculating the absolute numerical deviation of the instantaneous slope from the mean.

[0031] S303: For the set of transient change ratios, retrieve the upper and lower limit amplitude judgment parameters within the preset trigger state threshold interval set. If the judgment ratio falls into the target state threshold interval, read the target state threshold interval mapping classification character, and concatenate the classification characters in order to obtain the peak pre-trigger level identifier. For the transient mutation ratio set, the upper and lower limit amplitude judgment parameters within the trigger state threshold interval set are actively retrieved. The mechanism for establishing this interval set involves extracting a historical mutation ratio sample set and calculating the expected mean and standard deviation parameters of this sample set. The standard deviation parameter is multiplied by a fixed proportional constant configured in the configuration to obtain the tolerance range parameter. Based on the expected mean of the data, the tolerance range parameter is accumulated and subtracted to accurately calculate the upper and lower limits of the fluctuation. The span between the two boundaries is divided into equal intervals, and the values ​​of the division nodes are extracted as the upper and lower limit amplitude judgment parameters. Adjacent judgment parameters are paired and combined to establish numerical judgment intervals. The real-time ratio values ​​within the transient mutation ratio set are compared with the lower and upper limits of each judgment interval. When the ratio value is greater than the lower limit and less than the upper limit, the corresponding numerical judgment interval is extracted and set as the target state threshold interval. Then, the pre-mapped classification characters of this target state threshold interval are read, and the classification characters are concatenated in chronological order to obtain the peak pre-trigger level identifier. For example, the expected mean of the historical samples is calculated to be 0.20, and the standard deviation parameter is 0.05. A fixed ratio constant is set to 2.0. Multiplying 0.05 by 2.0 yields a tolerance range parameter of 0.10. Adding 0.10 to 0.20 gives an upper limit boundary of 0.30, and subtracting 0.10 gives a lower limit boundary of 0.10. The span from 0.10 to 0.30 is divided into interval judgment parameters of 0.10, 0.20, and 0.30, establishing two intervals: 0.10 to 0.20 and 0.20 to 0.30. The real-time ratio value is 0.25. Comparison shows it falls within the 0.20 to 0.30 range, and the corresponding mapped character is set to a high-risk warning level.

[0032] Please see Figure 5 The specific steps of S4 are as follows: S401: Obtain the peak pre-trigger level identifier, extract the aggregation period parameter, collect the peak pre-trigger level identifier within the time span of the aggregation period parameter, count and record the occurrence frequency parameter, associate each category identifier with the occurrence frequency parameter, and establish a level frequency mapping set. The process involves acquiring the concatenated peak pre-trigger level identifier sequence string and extracting the aggregation period parameter from the control unit to define the data aggregation span. Within the time span corresponding to this set aggregation period parameter, all peak pre-trigger level identifiers falling within that time period are comprehensively aggregated. For the aggregated level identifier strings, a classification counter accumulator is activated to perform a sequential accumulation count operation on the frequency of occurrence of various types of level identifiers, accurately outputting the occurrence frequency parameter corresponding to each category. Subsequently, a two-dimensional mapping data structure is constructed, performing corresponding association operations between each category identifier string and the statistically obtained occurrence frequency parameter. Based on this, a level frequency mapping set reflecting the distribution density of various warning states over a period of time is established. For example, the aggregation period parameter extracted from the control unit is set to a duration of 3600 seconds. By searching all peak pre-trigger level identifiers generated within this 3600-second period, it is found that they contain multiple different types of warning characters. The data collected by the counter accumulator, after strict matching and accumulation calculation, recorded the frequency of the ordinary alarm level indicator as 45 times, the moderate danger level indicator as 12 times, and the high danger warning level indicator as 3 times. Subsequently, the ordinary alarm level was directly mapped to the value 45, the moderate danger level to the value 12, and the high danger warning level to the value 3, generating a level frequency mapping set containing key-value relationships in memory. The specific value of the aggregation period parameter 3600 seconds was set based on the average duration of a typical small-scale live-streaming traffic-generating activity in e-commerce, and its laboratory stress testing and optimization range was controlled between 1800 seconds and 7200 seconds to ensure the completeness of feature capture throughout the entire time period.

[0033] S402: Call the level frequency mapping set to traverse the occurrence parameter, extract the maximum value corresponding to the identifier as the periodic dominant trigger level, search the upper and lower limit boundary identifiers in the level frequency mapping set, extract the preset danger quantization scale, combine the upper and lower limit boundary identifiers with the danger quantization scale to obtain the boundary extreme value state vector. The completed level frequency mapping set is retrieved from memory, and a full sequence numerical comparison operation is performed on all occurrence parameters stored internally. A bubble sort logic is used to extract the occurrence parameter with the largest value in the entire mapping set, and the identifier type directly corresponding to this maximum value is explicitly determined as the dominant trigger level for the current period. After locking the dominant level, the level frequency mapping set is searched again to accurately extract the highest and lowest warning level identifiers that have occurred as upper and lower limit boundary identifiers. Subsequently, a preset danger quantification scale is extracted from an external security policy library. A numerical mapping conversion is performed on the obtained upper and lower limit boundary identifiers in combination with this preset danger quantification scale, quantizing the originally character-based boundary identifiers into precise numerical representations, constructing a boundary extreme value state vector representing the current period's extreme state. For example, traversing the level frequency mapping set, 45 ordinary alarms, 12 moderate danger alarms, and 3 high danger alarms are read. After horizontal comparison of numerical values, 45 is confirmed as the maximum value, and the ordinary alarm corresponding to 45 is established as the dominant trigger level for the period. Next, for the overcounted identifiers generated within the search set, high-risk warnings are extracted as upper limit boundary identifiers, and ordinary warnings as lower limit boundary identifiers. A preset hazard quantization scale is extracted, which specifies that ordinary warnings are mapped to a quantization value of 1.0, moderate hazard to 2.0, and high hazard to 3.0. The lower limit boundary identifier, ordinary warning, is converted to the value 1.0, and the upper limit boundary identifier, high hazard, is converted to the value 3.0. Finally, these are concatenated to generate a boundary extreme value state vector containing elements 1.0 and 3.0. The span difference of the preset hazard quantization scale is set to 1.0. The advantage of this operational logic is that by extracting the most frequent state to establish the dominant state and integrating the quantization scale to transform the boundary identifiers, category labels that originally could not participate in mathematical operations are endowed with numerical boundary measurement capabilities.

[0034] S403: For the boundary extreme state vector, extract the upper and lower distribution values, subtract the lower distribution value from the upper distribution value to extract the state span difference, read the dominant quantization value of the period dominant trigger level, divide the state span difference by the dominant quantization value, and generate the trigger fluctuation discrete parameters. For the boundary extreme value state vector output from the pre-processing stage, the upper distribution limit value representing the highest risk level and the lower distribution limit value representing the lowest risk level within the current cycle are directly extracted from this vector through addressing. Then, the extracted upper distribution limit value is subtracted from the lower distribution limit value, and a difference calculation operation is performed to accurately extract the state span difference between the two. Next, the internal mapping dictionary is queried to read the dominant quantization value corresponding to the dominant trigger level of the current cycle. The previously calculated state span difference value is used as the dividend, and the dominant quantization value is used as the divisor; a division operation is performed between the two, directly generating the trigger fluctuation discrete parameter used to measure the severity of fluctuations. For example, the boundary extreme value state vector is first accurately located and read to contain elements 1.0 and 3.0. The upper distribution limit value is then extracted and set to 3.0, and the lower distribution limit value is extracted and set to 1.0. A subtraction operation is performed on the two values, subtracting 1.0 from 3.0, accurately calculating the state span difference within the current aggregation cycle to be 2.0. After obtaining the difference, the ordinary alarm that was established as the dominant trigger level in the previous operation is retrieved, and its corresponding dominant quantization value is retrieved from the quantization scaling dictionary, which is confirmed to be 1.0. Then, the final division calculation instruction is executed, dividing the state span difference dividend 2.0 by the dominant quantization value divisor 1.0. This division operation accurately yields the current trigger fluctuation discrete parameter result as 2.0. During this calculation, all upper limit values, lower limit values, and dominant quantization values ​​involved in the calculation must be positive real numbers of the same dimension.

[0035] Please see Figure 6 The specific steps of S5 are as follows: S501: Obtain the trigger fluctuation discrete parameters, extract the page jump sequence from the preset full-link log, count the jump frequency between each page node in the page jump sequence, divide the jump frequency by the total outflow frequency of the corresponding source page node, calculate the initial transition probability between nodes, and construct the initial state transition matrix. The system retrieves the trigger fluctuation discrete parameter, whose specific value has been assigned a value of 2.0, from the preceding calculation output. It then directly reads the preset end-to-end logs from the underlying distributed storage nodes. Character matching and cleaning operations are performed on the read log data to remove incomplete data packet records. Subsequently, the system parses and extracts a coherent page jump sequence recording the user's page access trajectory in chronological order. For the extracted page jump sequence, a window counting mechanism is activated to scan and count the actual jump frequency between each specified page node in the sequence, line by line. Simultaneously, the number of jumps from a single source page node to all external associated nodes is summed to obtain the total outflow frequency. The specified jump frequency is used as the dividend, and the total outflow frequency of the corresponding source page node is used as the divisor. Division is performed to accurately calculate the initial transition probability between nodes. After calculating the probabilities of all associated nodes, the initial transition probabilities between each node are filled into a 2D array in memory, with the source page node as the horizontal row coordinate and the target page node as the vertical column coordinate, constructing the initial state transition matrix. The matrix dimension is set to match the total number of core page nodes across the entire site, and the sum of the elements in the internal row vectors is strictly constrained to 1.0. For example, if the trigger fluctuation discrete parameter is set to 2.0, and the log scan reveals that the frequency of user jumps from the homepage to the core transaction link node is 20,000, and the total outflow frequency from the homepage to all pages across the site is 100,000, a division calculation is performed. Dividing the dividend 20,000 by the divisor 100,000 yields an initial transition probability of 0.20 between nodes pointing to the core node. Similarly, the jump frequency to the other four non-core transaction link nodes is calculated to be 20,000, and dividing each by 100,000 yields an initial transition probability of 0.20 for each. These values ​​are then sequentially filled into a 2D array to establish the initial state transition matrix. The advantage of this operational logic is that it uses the frequency ratio under the law of large numbers to reduce the dimensionality of unordered access logs and transform them into a probability matrix with coherent properties.

[0036] S502: Call the initial state transition matrix, extract the preset core transaction link nodes as target correction objects, perform nonlinear amplification mapping processing on the transition probability corresponding to the target correction object according to the trigger fluctuation discrete parameters, and perform corresponding deduction and row vector normalization processing on the transition probability of non-core transaction link nodes to generate a dynamic transition feature matrix. The process involves invoking the pre-stored initial state transition matrix in memory and extracting the preset core transaction link nodes from the business configuration file, directly setting them as the target correction objects. The transition probability values ​​pointing to the core transaction link nodes within the initial state transition matrix are located, and the trigger fluctuation discrete parameter with a value of 2.0 is read. Based on this, a nonlinear amplification mapping process is performed on the transition probabilities corresponding to the target correction objects. A constant 1 is used as the dividend, and the trigger fluctuation discrete parameter is used as the divisor to perform division, extracting the reciprocal as the mapping exponent parameter. The transition probability pointing to the core transaction link node is used as the base, and the mapping exponent parameter is used as the exponent to perform a power function operation, obtaining the amplified corrected core probability. The original transition probabilities of each non-core transaction link node in the same row are extracted and summed with the corrected core probability to obtain the amplified probability sum. The corrected core probability and the original transition probabilities of each non-core transaction link node are used as dividends, and the sum of amplified probabilities is used as the divisor to perform division, normalization, and scaling, updating the in-row probability values ​​to generate a dynamic transition feature matrix. For example, the initial transition probability from the homepage to the core transaction link node is 0.20, and the initial transition probability of the four non-core nodes is also 0.20, with a trigger fluctuation discrete parameter of 2.0. Dividing 1 by 2.0 yields a mapping exponent parameter of 0.50. A power function operation is performed, raising the base 0.20 to the power of 0.50 to obtain a corrected core probability of 0.447. Summing the four 0.20s with 0.447 yields a total amplified probability of 1.247. Normalization is then performed, dividing 0.447 by 1.247 yields a final core node probability of 0.358, and dividing 0.20 by 1.247 yields probabilities of 0.160 for each non-core node, thus overwriting the current row to generate a dynamic transition feature matrix. The advantage of this operational logic is that it uses a power function mapping to fit the traffic saturation characteristics when users converge towards the core node.

[0037] S503: Call the dynamic transition feature matrix to obtain the real-time online user distribution vector as the initial state vector, perform continuous iterative multiplication calculation, extract the converged steady-state probability vector as the traffic prediction confidence, match the quota correlation matrix to extract the corresponding elastic quota parameters, and construct e-commerce prediction and control instructions. The system invokes the constructed dynamic transition feature matrix and extracts a real-time online user distribution vector representing the proportion of currently online users on each page from the front-end gateway node using a data probe. This vector is then directly set as the initial state vector. Continuous iterative multiplication is performed between the initial state vector and the dynamic transition feature matrix. The 1D initial state vector and the 2D dynamic transition feature matrix are multiplied to obtain the first-generation result vector. This first-generation result vector is then multiplied again by the dynamic transition feature matrix to obtain the second-generation result vector. This multiplication operation is continuously looped, extracting the difference between corresponding positions of adjacent result vectors. The iteration terminates when all absolute differences are less than the convergence tolerance variable, and the converged steady-state probability vector is extracted. The convergence values ​​corresponding to the core transaction link nodes in the steady-state probability vector are read as the traffic prediction confidence level. Using the traffic prediction confidence level as the retrieval basis, a pre-set quota association matrix is ​​matched. By comparing the magnitudes, the specific span range is determined, and the elastic quota parameters within the corresponding row are extracted. The elastic quota parameters are then combined with the routing address identifier of the target physical resource pool through segment concatenation to construct and generate e-commerce prediction and control instructions. For example, the initial state vector and the dynamic transition feature matrix are multiplied together 120 times consecutively. The difference between the result vectors of the 119th and 120th generations is less than the convergence tolerance variable of 0.001, triggering a stop command. The value of 0.88 for the core order page is extracted from the steady-state probability vector, establishing a traffic prediction confidence level of 0.88. The quota correlation matrix is ​​scanned, and 0.88 is compared with the interval boundary, determining it falls within the extremely high response range of 0.80 to 1.00, extracting the elastic quota parameter as 3000. The quota value of 3000 is nested and concatenated with the hexadecimal routing gateway identifier, outputting the e-commerce prediction and control command.

[0038] Table 1: Correlation Matrix of Predicted Quotas

[0039] As shown in Table 1, by setting discrete confidence thresholds, the continuous steady-state probabilities obtained from dynamic calculations can be quickly converted into tiered physical resource allocation parameters.

[0040] Please see Figure 7 The big data-based e-commerce intelligent prediction management system includes: The order flow aggregation module acquires real-time order messages from the streaming data buffer of the e-commerce platform, parses and extracts order timestamps and order type identifiers, performs time span segmentation based on the sliding window algorithm, and performs classification and statistics based on order type identifiers to generate an order throughput sequence. The rate of change extraction module obtains the order throughput sequence, performs differential operation on the order throughput of adjacent sampling periods to extract the throughput difference, and performs time-domain rate of change mapping to construct the throughput gradient slope sequence. The traffic mutation sensing module acquires the throughput gradient slope sequence, inputs it into the weighted moving average model to perform baseline deviation analysis to extract the transient mutation ratio, performs matching and filtering based on the preset trigger threshold range, and generates a peak pre-trigger level identifier. The level quantification and evaluation module obtains the peak pre-trigger level identifier to construct a level convergence set, extracts the highest frequency identifier as the periodic dominant trigger level, extracts the boundary extreme value of the level convergence set and performs state span quantification and evaluation to generate trigger fluctuation discrete parameters. The quota elastic control module uses the trigger fluctuation discrete parameter to correct the initial state transition matrix based on the preset full-link log, and performs steady-state probability distribution iterative solution to obtain the traffic prediction confidence, extract the corresponding elastic quota parameters, and construct e-commerce prediction control instructions.

[0041] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection of the described technical solutions.

Claims

1. A big data-based intelligent predictive management method for e-commerce, characterized in that: Includes the following steps: S1: Obtain real-time order messages from the streaming data buffer of the e-commerce platform, parse and extract the order timestamp and order type identifier, perform time span segmentation based on the sliding window algorithm, classify and statistically analyze the order type identifier, and generate an order throughput sequence. S2: Obtain the order throughput sequence, perform differential operation on the order throughput of adjacent sampling periods to extract the throughput difference and perform time-domain rate of change mapping to construct a throughput gradient slope sequence; S3: Obtain the throughput gradient slope sequence, input it into the weighted moving average model to perform baseline deviation analysis to extract the transient mutation ratio, perform matching and filtering based on the preset trigger threshold interval, and generate peak pre-trigger level identifier; S4: Obtain the peak pre-trigger level identifier to construct a level convergence set, extract the highest frequency identifier as the periodic dominant trigger level, extract the boundary extreme value of the level convergence set and perform state span quantitative evaluation to generate trigger fluctuation discrete parameters; S5: Use the triggering fluctuation discrete parameters to correct the initial state transition matrix constructed based on the preset full-link logs, and perform steady-state probability distribution iterative calculation to obtain the traffic prediction confidence, extract the corresponding elastic quota parameters, and construct e-commerce prediction and control instructions.

2. The e-commerce intelligent predictive management method based on big data according to claim 1, characterized in that, The order throughput sequence includes the order throughput of home appliances, apparel, and digital products. The throughput gradient slope sequence includes the first-order difference, the second-order derivative, and the rate of change. The peak pre-trigger level identifier includes the normal release identifier, the flow-limiting queuing identifier, and the circuit breaker degradation identifier. The trigger fluctuation discrete parameters include the population variance, the sample standard deviation, and the interquartile range. The e-commerce predictive control instructions include the elastic scaling instruction, the token bucket flow-limiting instruction, and the service degradation instruction.

3. The e-commerce intelligent predictive management method based on big data according to claim 1, characterized in that, The specific steps of S1 are as follows: S101: Collect real-time business messages from the e-commerce platform buffer, extract the order timestamp and order type identifier from the message header node fields, obtain the timestamp representation value and the identifier character code, use the timestamp representation value as the primary key and the identifier character code as the value to perform association, and generate a temporal feature mapping table. S102: Call the temporal feature mapping table to obtain the window step parameter and span tolerance parameter, compare the timestamp representation value with the span tolerance parameter, filter the data items that fall within the range of the span tolerance parameter, and repeat the filtering according to the window step parameter shift range to obtain the time domain segmented state set. S103: Perform classification operation based on the time-domain segmented state set, extract the identifier character code within the interval, merge similar data items and sum the quantities to generate the total number of order categories in the interval segment, combine the total number of order categories in the interval segment into a one-dimensional array according to the time-transition sequence, and establish the order throughput sequence.

4. The e-commerce intelligent predictive management method based on big data according to claim 3, characterized in that, The specific steps of S2 are as follows: S201: Call the order throughput sequence, read the throughput corresponding to the discrete nodes in the sequence in sequence, extract the values ​​of two adjacent nodes in the order of time arrangement, perform differential operation to obtain the throughput difference, arrange the difference into a one-dimensional linear array structure in the unidirectional order of time progression, and establish a periodic throughput difference set. S202: Based on the periodic throughput difference set, obtain the preset sampling duration parameter, perform a division operation on multiple differences in the periodic throughput difference set with the sampling duration parameter, calculate the rate of change within the sampling interval, and sequentially concatenate the rates of change according to the order of the nodes where the throughput difference is located to obtain the time domain change mapping column. S203: For the time-domain change mapping column, set a gradient slope plane reference space, extract the rate of change in the time-domain change mapping column as the vertical coordinate scalar, use the relative displacement step size of the corresponding node as the horizontal coordinate scalar, construct two-dimensional coordinate mapping points and connect the points along the time axis to generate a throughput gradient slope sequence.

5. The e-commerce intelligent predictive management method based on big data according to claim 4, characterized in that, The specific steps for S3 are as follows: S301: Obtain the throughput gradient slope sequence, extract the original slope value of the node, read the sliding time window span variable and the node decay weight factor, multiply the original slope value and the node decay weight factor within the sliding time window span variable, sum them, and divide by the sum of the node decay weight factors to establish a dynamic weighted benchmark moving average series. S302: Call the dynamic weighted benchmark moving average column, subtract the reference benchmark mean in the dynamic weighted benchmark moving average column from the original slope value at the same sequential position, extract the time domain absolute deviation parameter, divide the time domain absolute deviation parameter by the corresponding reference benchmark mean, extract the positive mutation term, and generate the transient mutation ratio set. S303: For the set of transient change ratios, retrieve the upper and lower limit amplitude judgment parameters within the preset trigger state threshold interval set. If the judgment ratio falls into the target state threshold interval, read the target state threshold interval mapping classification characters, and concatenate the classification characters in order to obtain the peak pre-trigger level identifier.

6. The e-commerce intelligent predictive management method based on big data according to claim 5, characterized in that, The determination ratio falling into the target state threshold range refers to establishing multiple continuous numerical determination ranges based on the upper and lower limit amplitude determination parameters, comparing the ratio values ​​in the transient change ratio set with the lower limit boundary and upper limit boundary in the numerical determination range respectively, and extracting the numerical determination range where the ratio values ​​are inside the lower limit boundary and upper limit boundary, which is set as the target state threshold range. The trigger state threshold interval set is obtained by extracting a historical mutation ratio sample set, calculating the expected mean and standard deviation of the historical mutation ratio sample set, multiplying the standard deviation by a fixed proportional constant to obtain the tolerance range parameter, and adding and subtracting the tolerance range parameter based on the expected mean of the data to obtain the upper and lower limits of fluctuation. The span between the lower and upper limits of fluctuation is divided equally, and the values ​​of the division nodes are extracted as upper and lower limit amplitude determination parameters. Adjacent upper and lower limit amplitude determination parameters are paired and combined to establish a numerical determination interval, and the numerical determination interval is integrated to determine the value.

7. The e-commerce intelligent predictive management method based on big data according to claim 5, characterized in that, The specific steps of S4 are as follows: S401: Obtain the peak pre-trigger level identifier, extract the aggregation period parameter, aggregate the peak pre-trigger level identifier within the time span of the aggregation period parameter, count and record the occurrence frequency parameter, associate each category identifier with the occurrence frequency parameter, and establish a level frequency mapping set. S402: Call the level frequency mapping set to traverse the occurrence frequency parameter, extract the maximum value corresponding to the identifier as the periodic dominant trigger level, search the upper and lower limit boundary identifiers in the level frequency mapping set, extract the preset danger quantization scale, combine the upper and lower limit boundary identifiers with the danger quantization scale to obtain the boundary extreme value state vector. S403: For the boundary extreme state vector, extract the upper distribution limit value and the lower distribution limit value, subtract the lower distribution limit value from the upper distribution limit value to extract the state span difference, read the dominant quantization value of the period dominant trigger level, divide the state span difference by the dominant quantization value, and generate the trigger fluctuation discrete parameter.

8. The e-commerce intelligent predictive management method based on big data according to claim 7, characterized in that, The specific steps of S5 are as follows: S501: Obtain the trigger fluctuation discrete parameters, extract the page jump sequence from the preset full-link log, count the jump frequency between each page node in the page jump sequence, divide the jump frequency by the total outflow frequency of the corresponding source page node, calculate the initial transition probability between nodes, and construct the initial state transition matrix. S502: Call the initial state transition matrix, extract the preset core transaction link node as the target correction object, perform nonlinear amplification mapping processing on the transition probability corresponding to the target correction object according to the trigger fluctuation discrete parameter, and perform corresponding deduction and row vector normalization processing on the transition probability of non-core transaction link nodes to generate a dynamic transition feature matrix. S503: Call the dynamic transition feature matrix to obtain the real-time online user distribution vector as the initial state vector, perform continuous iterative multiplication calculation, extract the converged steady-state probability vector as the traffic prediction confidence, match the quota correlation matrix to extract the corresponding elastic quota parameters, and construct e-commerce prediction and control instructions.

9. The e-commerce intelligent predictive management method based on big data according to claim 8, characterized in that, The process involves performing nonlinear amplification mapping on the transition probability corresponding to the target correction object based on the trigger fluctuation discrete parameter, and performing corresponding deduction and row vector normalization on the transition probability of non-core transaction link nodes. This process involves using a constant 1 as the dividend and the trigger fluctuation discrete parameter as the divisor to perform division, extracting the reciprocal as the mapping exponent parameter, using the transition probability pointing to the core transaction link node in the initial state transition matrix as the base, and using the mapping exponent parameter as the exponent to perform a power function operation to obtain the amplified correction core probability. The process also involves extracting the original transition probabilities of each non-core transaction link node in the same row, summing them with the correction core probability to obtain the amplified probability sum, using the correction core probability and the original transition probabilities of each non-core transaction link node as the dividends, and using the sum of amplified probabilities as the divisor to perform division normalization scaling, thereby updating the probability value of the current row.

10. A big data-based intelligent predictive management system for e-commerce, characterized in that: The system is used to implement the big data-based intelligent predictive management method for e-commerce as described in any one of claims 1-9, and the system comprises: The order flow aggregation module acquires real-time order messages from the streaming data buffer of the e-commerce platform, parses and extracts order timestamps and order type identifiers, performs time span segmentation based on the sliding window algorithm, and performs classification and statistics based on order type identifiers to generate an order throughput sequence. The rate of change extraction module obtains the order throughput sequence, performs differential operation on the order throughput of adjacent sampling periods to extract the throughput difference, and performs time-domain rate of change mapping to construct a throughput gradient slope sequence. The traffic mutation sensing module acquires the throughput gradient slope sequence, inputs it into the weighted moving average model to perform baseline deviation analysis to extract the transient mutation ratio, performs matching and filtering based on a preset trigger threshold range, and generates a peak pre-trigger level identifier. The level quantification and evaluation module obtains the peak pre-trigger level identifier to construct a level convergence set, extracts the highest frequency identifier as the periodic dominant trigger level, extracts the boundary extreme value of the level convergence set and performs state span quantification and evaluation to generate trigger fluctuation discrete parameters. The quota elastic control module uses the trigger fluctuation discrete parameter to correct the initial state transition matrix constructed based on the preset full-link log, and performs steady-state probability distribution iterative calculation to obtain the traffic prediction confidence, extract the corresponding elastic quota parameters, and construct e-commerce prediction control instructions.

Images