Product service data processing method and device for ERP service scene and medium

Abstract

The embodiment of the invention discloses a product business data processing method and device for an ERP business scene and a medium, and relates to the technical field of ERP, and the method comprises the steps: obtaining historical sales data, supply chain data and external environment data of a target product in an ERP system under the triggering of a business index prediction request, and obtaining a business index prediction result of the target product based on the supply chain data; determining a plurality of dynamic time scales corresponding to the target product to determine a multi-dimensional feature set corresponding to each dynamic time scale; determining a business index linear predicted value and a nonlinear residual sequence through an ARIMA model and each multi-dimensional feature set; according to the external environment data, performing decomposition prediction on the nonlinear residual error sequence to determine a reconstructed residual error sequence, and determining current service index prediction data corresponding to each dynamic time scale based on the service index linear prediction value and the reconstructed residual error sequence; and performing multi-scale fusion on the business index prediction data corresponding to each dynamic time scale, and determining target business index prediction data of the target product.

Description

Technical Field

[0001] This manual relates to the field of ERP technology, and in particular to a product business data processing method, device and medium for an ERP business scenario. Background Technology

[0002] In enterprise information management, Enterprise Resource Planning (ERP) systems have become a key platform for integrating core business operations such as procurement, sales, inventory, and production. Traditional ERP systems' sales forecasting modules often employ simple statistical models such as moving averages and exponential smoothing.

[0003] ERP systems contain various product categories, each with significantly different business cycles. It's important to note that the business cycle refers to the time-bound pattern of a product's flow within the ERP system, constrained by business rules; for example, the difference between fast-moving consumer goods (FMCG) and durable goods. Furthermore, the same product may be subject to multiple business scenarios such as procurement, replenishment, and promotions. Even within the same product category, different business scenarios can lead to variations in business cycles. For instance, a beverage might have a 7-day replenishment cycle under normal circumstances, but this might be reduced to 3 days during holiday promotions, requiring advance stocking to cope with peak sales. Traditional product forecasting models suffer from data processing limitations under these varying business cycle conditions. For example, a fixed time scale leads to unstable data sequence aggregation, failing to effectively capture non-linear patterns and causing model instability during data processing, making it unsuitable for adapting to the different business cycle characteristics of various products.

[0004] Therefore, in the process of processing product business data, traditional prediction models use a fixed time scale, which has data processing defects. The products cannot adapt to the dynamic business cycle characteristics of multiple products and multiple scenarios in the ERP system, and there is a risk of inaccurate analysis results. Summary of the Invention

[0005] This specification provides one or more embodiments of a product business data processing method, device, and medium for ERP business scenarios, which are used to solve the following technical problems: In the process of processing product business data, traditional prediction models use a fixed time scale, which has data processing defects and cannot adapt to the dynamic business cycle characteristics of multiple products and multiple scenarios in the ERP system, resulting in the risk of inaccurate analysis results.

[0006] One or more embodiments of this specification employ the following technical solutions: This specification provides one or more embodiments of a product business data processing method for an ERP business scenario. The method includes: upon triggering a business indicator prediction request, acquiring historical sales data, supply chain data, and external environment data of a target product in the ERP system; based on the supply chain data, determining multiple dynamic time scales corresponding to the target product to determine a multi-dimensional feature set corresponding to each dynamic time scale; determining a linear prediction value of the business indicator and a corresponding nonlinear residual sequence for each dynamic time scale using a pre-constructed ARIMA model and the multi-dimensional feature set corresponding to each dynamic time scale; decomposing and predicting the nonlinear residual sequence according to the external environment data to determine a reconstructed residual sequence; determining the current business indicator prediction data corresponding to each dynamic time scale based on the linear prediction value of the business indicator and the reconstructed residual sequence; and performing multi-scale fusion on the business indicator prediction data corresponding to each dynamic time scale to determine the target business indicator prediction data for the target product.

[0007] This specification provides one or more embodiments of a product business data processing device for an ERP business scenario, including: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the above-described method.

[0008] This specification provides one or more embodiments of a non-volatile computer storage medium storing computer-executable instructions configured to perform the above-described method.

[0009] The at least one technical solution adopted in the embodiments of this specification can achieve the following beneficial effects: Through the technical solutions in the embodiments of this specification, a dynamic time-scale prediction system that adapts to business scenarios is constructed, fundamentally solving the technical problem of the disconnect between fixed time scales and multiple business cycles in traditional ERP business indicator prediction. Traditional prediction models treat the time scale as a static hyperparameter; the embodiments of this specification, through supply chain data-driven scale generation, achieve accurate matching between business scenarios and time windows based on dynamic scale generation of supply chain data. Traditional solutions, due to the separation of linear models and residual processing, result in sudden signals (promotional impulses) being submerged by trend baselines, and supply chain disruption risks being diluted by environmental noise. The embodiments in this specification achieve a fundamental breakthrough through a dual-channel processing architecture. When the ARIMA model processes multi-dimensional feature sets, it uses stationarity transformation and adaptive difference order to strip away deterministic patterns such as procurement delivery rhythm and inventory turnover trends, transforming supply chain business parameters (such as supplier delivery stability) into model hyperparameters. This prevents the long-term correlation of industrial spare parts procurement cycles from being contaminated by short-term promotional fluctuations. Variational mode decomposition separates the intrinsic mode function components of the residual sequence according to spectral energy. The temporal convolutional network injects weighted external features through a gating mechanism to achieve targeted enhancement of sudden signals. The delivery stability index serves as a gating strength adjustment factor to block the transmission of high-risk supply chain signals. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments or prior art of this specification, the drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 A flowchart illustrating a product business data processing method for an ERP business scenario provided in an embodiment of this specification; Figure 2 This is a schematic diagram of the structure of a product business data processing device for an ERP business scenario provided in an embodiment of this specification. Detailed Implementation

[0011] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0012] This specification provides a product business data processing method for an ERP business scenario. It should be noted that the execution entity in this specification embodiment can be a server or any device with data processing capabilities. Figure 1 This is a flowchart illustrating a product business data processing method for an ERP business scenario provided in an embodiment of this specification, as shown below. Figure 1 As shown, the main steps include the following: Step S101: Triggered by the business indicator prediction request, historical sales data, supply chain data and external environment data of the target product are obtained from the ERP system. Based on the supply chain data, multiple dynamic time scales corresponding to the target product are determined to determine the multi-dimensional feature set corresponding to each dynamic time scale.

[0013] In one embodiment of this specification, triggered by a business metric prediction request, historical sales data, supply chain data, and external environment data of the target product are obtained from the ERP system. It should be noted that the business metric here can be a product sales volume indicator. Significant cyclical differences exist between different products (such as fast-moving consumer goods and industrial spare parts) and multiple scenarios (such as procurement / promotion) for the same product within the ERP system. Historical sales data is used to reveal inherent product trends, supply chain data provides a basis for business rhythm, and external environment data such as economic indicators and promotional records are used to capture market disturbances. The absence of any dimension will cause the model to deviate from actual business logic; for example, predictions ignoring promotional activity records may misjudge short-term surges as noise.

[0014] Through the API interface of the ERP sales management module, timestamped transaction records are extracted based on the product's unique identifier, including fields such as transaction date, sales volume, and return volume, and associated with promotional activity IDs to mark promotional period data. Supply chain data includes procurement and inventory data. The order database of the procurement management module is accessed to extract supplier codes, order dates, planned delivery dates, and actual delivery dates, and the interval between consecutive orders is calculated. Integration with the warehouse management system (WMS) is used to obtain the average daily inventory and monthly outbound volume of target products, deriving inventory turnover rate and inventory age distribution. External environment data is obtained in real time by accessing public APIs from authoritative institutions, crawling time-series data of external economic indicators such as the Consumer Price Index (CPI) and Producer Price Index (PPI), and parsing the activity calendar of the enterprise's promotional management system to extract activity start and end times, discount rates, and coverage channel lists, and marking the weight of online / offline channels.

[0015] Using calendar days as the base unit, purchase order dates, inventory count dates, sales transaction dates, and promotional activity dates are mapped to a unified timeline. When data frequencies are inconsistent (e.g., economic indicators are updated monthly), a forward-filling method is used to expand monthly data into a daily dimension. Sales data and inventory data are linked through product SKU codes, purchase data and fulfillment records are linked through supplier IDs, and sales records and promotional rules are bound through promotional activity IDs.

[0016] Based on the supply chain data, multiple dynamic time scales corresponding to the product target are determined. Specifically, this includes: extracting multi-dimensional business cycles from the supply chain data to construct multi-dimensional cycle features, which include multiple process node cycle features; determining the dominant reference cycle in the multi-dimensional cycle features based on the multi-dimensional cycle features and the historical sales data of the product target; and expanding the time scale based on the dominant reference cycle to determine multiple dynamic time scales corresponding to the product target.

[0017] Based on the supply chain data, multi-dimensional business cycle extraction is performed to construct multi-dimensional cycle features. Specifically, this includes: extracting the order date sequence corresponding to historical purchase orders from the supply chain data; calculating the time interval between adjacent historical purchase orders corresponding to consecutive order dates based on the order date sequence; determining the purchase process node cycle features based on multiple such time intervals; obtaining inbound and outbound records from the supply chain data; calculating the ratio of annual sales volume to average inventory of a single product based on these records to determine the inventory turnover rate; determining the corresponding inventory turnover cycle based on the inventory turnover rate to determine the inventory process node cycle features; analyzing the promotional calendar data from the sales management system, marking the periodic distribution of promotional activity days, and determining the promotional process node cycle features; and extracting fulfillment records from the supplier management module, calculating the standard deviation of delivery time, and generating a supplier delivery stability index as a delivery process node cycle feature.

[0018] In existing technologies, fixed-time-scale aggregation ignores the interaction between product characteristics and scenarios, resulting in the forced compression of long-cycle procurement demand for industrial spare parts into the short-cycle framework of fast-moving consumer goods, or the smoothing and dilution of promotional period pulse signals by conventional replenishment models.

[0019] In one embodiment of this specification, historical order records are filtered by target product SKU through the order database interface of the ERP procurement management module, and a complete sequence containing the order creation date and planned delivery date is extracted. After sorting based on the order creation date, the time interval difference between adjacent orders is calculated, such as the time difference between the order creation date of order A and the order creation date of order B. After excluding invalid intervals caused by returns or cancellations, the arithmetic mean of consecutive valid intervals is taken as the procurement cycle benchmark value. In other words, the cycle characteristics of the procurement process nodes include the time interval differences corresponding to multiple historical orders, i.e., the procurement cycle duration of each procurement cycle, as well as the procurement cycle benchmark value.

[0020] Obtain the annual total sales volume and average inventory level of the target products from the inventory management system based on the product SKUs. Calculate the ratio of annual sales volume to average inventory level for each year to determine the inventory turnover rate. The unit of inventory turnover rate is times / year. Convert the inventory turnover rate into inventory turnover cycle by using the ratio of 365 to inventory turnover rate. Determine the cycle characteristics of inventory process nodes by using the annual inventory turnover cycle and the average of the inventory turnover cycles of previous years.

[0021] The activity calendar data of the promotion management system is analyzed to identify promotional activity records containing target products, and the start and end dates of the activities are marked. Based on the time difference between two adjacent activities, the interval between adjacent promotions is determined. The arithmetic mean of multiple adjacent promotion intervals is taken to obtain the promotion cycle benchmark value. Multiple promotion intervals and the promotion cycle benchmark value are used as the cycle characteristics of the promotion process nodes.

[0022] The system extracts order fulfillment records related to the target product from the supplier management module for the past 12 months, marking the number of days the actual delivery date deviates from the planned delivery date for each record. The standard deviation of these deviation days is calculated to obtain the delivery stability index for each supplier. The arithmetic mean of the delivery stability indices for multiple suppliers corresponding to the target product is then calculated to generate a quantitative delivery stability index for the target product, also known as the supplier delivery stability index. It should be noted that timeline alignment technology is used in the above process to map purchase order dates, inventory count timestamps, and promotional activity periods to a unified time coordinate system, eliminating periodic phase shifts caused by system time zone differences.

[0023] Based on this multidimensional cyclical feature and the historical sales data of the target product, the dominant reference cycle in the multidimensional cyclical feature is determined. Specifically, this includes: constructing regression models with historical sales data for multiple other cyclical features (excluding the delivery cycle feature) at multiple process node cyclical features, and calculating the adjusted R-squared value of the regression model for each other cyclical feature. 2 Value; Adjustment R based on each other periodic feature. 2 Value, adjusted to the maximum value of R2 The corresponding periodic characteristics determine the dominant reference period. In one embodiment of this specification, multiple regression models are constructed for the procurement cycle, inventory cycle, promotion cycle, and historical sales data, respectively. The procurement cycle model uses the purchase order interval as the independent variable and sales data as the dependent variable; the inventory cycle model uses the inventory turnover cycle as the independent variable and sales data as the dependent variable; the promotion cycle model uses the promotion decay cycle as the independent variable and sales data as the dependent variable. The adjusted R-values ​​for each model are calculated. 2 Value, adjust R 2 The value is solved by penalizing redundant independent variables. 2 The defect is calculated using the following formula: Where n is the sample size and p is the number of independent variables. This is the sum of squared residuals, which is the sum of squared errors between the model's predicted values ​​and the actual values. The sum of squares is the sum of the squares of the differences between the actual values ​​and the mean. The principle of the regression model lies in quantifying causal relationships through statistical methods. Each period's characteristics are used as independent variables, and sales data as the dependent variable. After constructing an independent model, the adjusted R-squared value is calculated. 2 The value essentially measures how much of the sales fluctuation can be explained by a particular cyclical characteristic. For example, the R-value in a purchasing cycle model adjustment... 2 =0.85, meaning 85% of sales changes can be explained by purchasing rhythm. Using adjusted R... 2 Value rather than R 2 To assess the explanatory power of cyclical characteristics on sales data, it's crucial to avoid misleading decisions. The number of independent variables in purchasing / inventory / promotion cycle models may differ; for example, a promotion model might include more advertising parameters. Using R... 2 The value will be biased towards complex models, while adjusting R... 2 It can eliminate the interference from the number of variables. Adjusting R... 2 To be more conservative, this approach prevents misjudging the dominance of the cycle due to accidental factors, such as the correlation between a coincidental promotion and sales fluctuations. Following the above method, the adjustment R corresponding to each of these other cyclical characteristics is determined. 2 According to the adjustment R corresponding to each of these other periodic characteristics 2 To maximize the adjustment of R 2 The periodic feature corresponding to the value is the dominant periodic feature, and the dominant reference period is determined based on the feature benchmark value in the dominant periodic feature.

[0024] The above technical solution addresses the limitations of traditional timescale determination methods that rely on fixed time windows or manual experience, failing to capture the reshaping effect of dynamic supply chain changes on business cycles. This solution utilizes a multiple regression model to construct and adjust R... 2The value-based decision-making mechanism enables dynamic scale generation driven by business needs. For example, when a car parts supplier experiences a sudden delivery delay, the system automatically detects the diminishing explanatory power of the procurement cycle model and adopts the inventory turnover cycle as the dominant time scale, avoiding the prediction distortion caused by traditional methods that forcibly compress delivery anomalies into a fixed time window. Scale decision-making based on statistical causality allows FMCG companies to accurately identify the pulse cycle characteristics of promotional activities, separating 48-hour limited-time flash sale signals from regular sales fluctuations and solving the inventory mismatch problem caused by cycle confusion in traditional technologies. Furthermore, compared to traditional R... 2 Blind bias towards complex models, adjusting R 2 Through a degree-of-freedom penalty mechanism, the value effectively isolates interference from redundant parameters. This dynamic adaptation capability enables industrial manufacturing enterprises to avoid the risk transmission of procurement cycles in real time during chip shortage crises and maintain the stability of production line forecasts.

[0025] In one embodiment of this specification, if there are at least two periodic characteristics, the adjustment R... 2 If the values ​​are the same, the dominant reference cycle is selected based on the delivery cycle characteristics corresponding to the same cycle characteristics (hereinafter referred to as candidate cycle characteristics). Adjust R 2 The same value means that these cyclical features are statistically indistinguishable in their explanatory power for sales data. In this case, additional business indicators (delivery stability) need to be introduced for selection. The role of the delivery stability indicator is to reflect the reliability of the supply chain (the smaller the standard deviation, the more stable). In cyclical conflict decisions, cyclical features less affected by supply chain risks should be prioritized because their predictions are more reliable. Calculate the correlation coefficient between each candidate cyclical feature and the delivery stability indicator, and select the cyclical feature with the smallest absolute correlation coefficient (or the weakest negative correlation) as the dominant reference period. For example, when supplier delivery is unstable (σ↑), the purchase order interval is passively lengthened, forming a false long-cycle feature. In this case, a high-adjusted R... 2 The values ​​represent the result of risk transmission, not actual business patterns. When calculating the correlation coefficient, it can be obtained by calculating the Pearson correlation coefficient between each candidate cycle feature and the delivery stability index (σ). First, time series data is constructed. Time series data for each cycle feature is extracted. From the procurement cycle feature, the time interval difference corresponding to each historical order is obtained to determine the procurement cycle sequence. From the inventory cycle feature, the annual inventory turnover cycle value is extracted to determine the inventory cycle sequence. From the promotion cycle feature, the adjacent promotion interval value is obtained to determine the promotion cycle sequence. The standard deviation sequence of delivery deviation days in the same time dimension as the above cycles is obtained to determine the delivery stability sequence.

[0026] Perform correlation analysis between each periodic feature sequence and the delivery stability sequence:

[0027] in, Let be the periodic characteristic value of the candidate period at time i. When the candidate period is the procurement period, it represents the procurement period duration characteristic of the procurement period at time i. Here, n represents the baseline value for the procurement cycle, and n represents the quantity at any given time point. Let Xi be the standard deviation of delivery deviation in days at time i. It should be noted that Xi represents a business cycle characteristic sequence that changes over time, such as the purchase order interval (purchase cycle characteristic) at time points t1=7 days, t2=10 days, t3=5 days, etc., and the delivery standard deviation at time points t1=±2 days, t2=±5 days, t3=±1 days, etc. These characteristic values ​​constitute a time series dataset (e.g., data from the past 36 months). The correlation coefficient between each cycle characteristic sequence and the delivery stability sequence is determined in the manner described above.

[0028] Through the above technical solution, when multiple periodic features are adjusted R... 2 When the values ​​cannot distinguish between good and bad, traditional technologies often resort to random selection or manual intervention. However, the embodiments in this specification construct an objective and quantitative secondary screening mechanism by dynamically coupling delivery stability indicators and correlation coefficient analysis. This effectively avoids the risk of prediction distortion caused by supplier delays, achieves dynamic decoupling of business cycles and supply chain risks, and reveals the inherent correlation between the characteristics of each cycle and delivery stability through correlation coefficient analysis. It can automatically isolate high-risk pseudo-cycle characteristics and maintain prediction accuracy.

[0029] Based on the dominant reference period, time scale expansion is performed to determine multiple dynamic time scales corresponding to the target product. Specifically, this includes: determining the autocorrelation coefficient sequence corresponding to the dominant reference period based on historical sales data to identify the lag order of the first zero crossing point in the autocorrelation coefficient sequence; generating a candidate scale expansion range based on the period length and lag order corresponding to the dominant reference period, and expanding by a specified multiple within this candidate scale expansion range to determine multiple candidate time scales; counting the number of consecutive segments and trend reversals of sales data with the same trend corresponding to each candidate time scale in the historical sales data to calculate the time series aggregation tightness coefficient corresponding to each candidate time scale; and filtering among multiple candidate time scales based on the time series aggregation tightness coefficient corresponding to each candidate time scale to determine multiple dynamic time scales corresponding to the target product.

[0030] In one embodiment of this specification, an autocorrelation coefficient sequence is calculated based on a subsequence of historical sales data corresponding to the dominant reference period. First, the sales data is standardized to eliminate dimensional differences. Then, the correlation strength between the current sales value and historical values ​​is calculated step-by-step according to the natural day lag order, generating an autocorrelation coefficient sequence with lag orders from 1 to the dominant period length L. The autocorrelation coefficient quantifies the persistence of the influence of historical data on the current value. When the sequence first turns from positive to negative (crossing zero), it marks the critical point where the historical influence disappears. Identifying this point requires sequentially scanning the sequence and recording the first lag order k with an autocorrelation coefficient ≤ 0. If the entire sequence is positive, the maximum lag order L is taken. This technical solution overcomes the shortcomings of traditional solutions that ignore business attenuation patterns. For example, the impact of a fast-moving consumer goods promotional pulse becomes zero after a lag of 3 days, while the impact of an industrial product procurement cycle can last up to a lag of 20 days.

[0031] When generating the candidate scale expansion range, the ratio of the zero-crossing lag order k to the dominant cycle length L (ρ=k / L) is used to quantify the cycle decay intensity, and the expansion factor α is dynamically adjusted accordingly. Specifically, the ρ value is first calculated, reflecting the persistence of the dominant cycle's influence in historical sales data. A smaller ρ value (e.g., ρ≤0.3) indicates faster cycle decay (e.g., promotional pulse signals), requiring a wider expansion to capture edge fluctuations; a larger ρ value (e.g., ρ>0.7) indicates stronger cycle inertia (e.g., long cycles of strategic procurement), requiring a narrower expansion to avoid noise interference. The expansion factor α is dynamically set based on the interval of the ρ value. When ρ≤0.3, α=0.5; when 0.3<ρ≤0.7, α=0.3; when ρ>0.7, α=0.1. The final generated candidate scale expansion range is [L×(1-α), L×(1+α)]. The above technical solution breaks through the limitations of traditional fixed-multiple expansion. For example, a beverage promotion pulse (ρ=0.2) automatically triggers wide-range expansion (α=0.5), extending its 48-hour dominant cycle to a range of [24,72] hours, ensuring coverage of possible pulse fluctuations.

[0032] When generating candidate timescales within the defined extended range [L×(1-α), L×(1+α)], a proportional step size design is adopted to ensure that the scale density is adapted to the business granularity. The step size calculation formula is S = L×β (β is 0.05), the principle of which is that the step size is proportional to the dominant cycle length L, thereby achieving a fine-grained step size (S=0.15 days ≈ 3.6 hours) for high-frequency goods (such as fresh produce L=3 days) and a coarse-grained step size (S=4.5 days) for low-frequency goods (such as equipment spare parts L=90 days). Within the extended range, an arithmetic sequence candidate set is generated with a step size S. For example, in the case of industrial spare parts, when L=28 days and ρ=0.82, α=0.1 generates a range of [25.2, 30.8] days, with a step size S=1.4 days, and the final candidate set is {25.2, 26.6, 28, 29.4, 30.8} days. Dynamic step size solves the resource waste problem caused by uniform step size in traditional solutions. For example, fresh produce requires hourly step size to capture the peak and trough of morning / evening market demand, while medical equipment only needs weekly step size to balance efficiency and accuracy.

[0033] For each candidate scale, historical sales data is scanned, and the number of consecutive segments with the same slope is counted (e.g., 5 consecutive days of monotonous increase counts as 1 segment). The number of abrupt changes in trend direction is also recorded (e.g., a change from increase to decrease counts as 1 instance). The density coefficient TC = 1 - (number of turning points / number of segments with the same trend), with a value range of [0,1]. The principle is to quantify data consistency within the scale; TC = 0.9 indicates a highly stable trend, while TC = 0.3 reveals noise interference. A density threshold T_min is set (usually ≥ 0.7), and candidate scales with TC ≥ T_min are retained. A set of qualified scales is output to determine multiple dynamic time scales corresponding to the target product.

[0034] Determining the multi-dimensional feature set corresponding to each dynamic time scale specifically includes: performing seasonal frequency domain feature extraction and image feature transformation on the historical sales sequence corresponding to the historical sales data of each dynamic time scale to determine frequency domain features and two-dimensional spatiotemporal image features, wherein the frequency domain features include high-frequency fluctuation components and low-frequency trend components; extracting the inventory turnover rate trend component and supplier delivery stability index from the supply chain data to determine supply chain features; and using an attention mechanism to perform weighted fusion of the frequency domain features, the two-dimensional spatiotemporal image features, and the supply chain features to generate a multi-dimensional feature set.

[0035] Existing technologies, lacking integration of supply chain business parameters and spatiotemporal patterns, are prone to the following problems. For example, automotive parts forecasting ignores supplier delivery stability indicators and cannot respond to sudden supply disruption risks, such as a sharp increase in delivery standard deviation due to chip shortages. Furthermore, FMCG promotion forecasting fails to separate high-frequency fluctuation components, such as limited-time flash sales, from low-frequency trends, such as seasonal consumption upgrades, misjudging promotional bursts as noise.

[0036] In one embodiment of this specification, seasonal frequency domain feature extraction is performed on the historical sales series corresponding to historical sales data for each dynamic time scale. A Discrete Wavelet Transform is performed on the sales series, and the signal is decomposed using the Daubechies wavelet basis function to extract short-term fluctuations related to promotions / sudden events, determine high-frequency components, and extract low-frequency components, preserving seasonal consumption trends, such as the summer sales baseline increase for air conditioners. The intensity of fluctuations is quantified by the energy entropy value of the high-frequency components, and the long-term trend is indicated by the trend slope of the low-frequency components. Gramian Angular Field encoding is applied to convert the historical sales series for each dynamic time scale, i.e., the one-dimensional time series, into a two-dimensional polar coordinate matrix. Each sales value is mapped to an angle, the value range is normalized to 0~π, the time series point spacing is converted into polar radius, preserving time dependence, and a heat map is generated. Color temperature represents sales intensity (e.g., red = peak, blue = trough), and spatial distribution reflects the spatiotemporal relationship between promotional bursts and stockout cycles. Furthermore, based on supply chain data, inventory turnover rate trend components and supplier delivery stability indicators are extracted to determine supply chain characteristics.

[0037] An attention mechanism is employed to weightedly fuse the frequency domain features, the two-dimensional spatiotemporal image features, and the supply chain features, generating a multi-dimensional feature set. The frequency domain features (high / low frequency components), image features (heatmap blocks), and supply chain features (turnover rate trends, delivery metrics) are input into a two-layer fully connected network. Softmax is used to output normalized weights, with inventory turnover rate as an independent weight factor mandatory in the calculation. The three-modal features are then concatenated according to their weights, resulting in a tensor structure with dimensions compatible with the ARIMA model.

[0038] By enhancing weighting through delivery standard deviation, when an indicator exceeds a risk threshold, the weight of supply chain features is automatically increased, driving the predictive model to switch to the procurement cycle of alternative suppliers, effectively avoiding the risk of production stoppages. By locating outbreak blocks using Gram's field heatmap, promotional pulse features are given the highest weight in attention-weighted processing. Combined with inventory turnover trends to predict replenishment urgency, the stockout rate of goods during major promotional periods can be effectively reduced. Utilizing an attention mechanism to replace traditional feature stitching ensures the accuracy of predictions for long-cycle products such as auto parts while reducing the computational overhead for short-cycle predictions of FMCG products.

[0039] Step S102: Using the pre-built ARIMA model and the multi-dimensional feature set corresponding to each dynamic time scale, determine the linear prediction value of the business indicator and the corresponding nonlinear residual sequence for each dynamic time scale.

[0040] In one embodiment of this specification, after obtaining the multi-dimensional feature set corresponding to each dynamic time scale, the feature set needs to be processed by a pre-built ARIMA model to generate linear predicted values ​​of business indicators and nonlinear residual sequences. The input multi-dimensional feature set (including frequency domain features, two-dimensional spatiotemporal image features, and supply chain features) is normalized to eliminate dimensional differences. For the high-frequency fluctuation component and the low-frequency trend component in the frequency domain features, the difference order is adaptively adjusted. First-order difference is performed on the high-frequency component to smooth burst noise, and seasonal difference (the period length is taken from the seasonal marker in the feature set) is performed on the low-frequency component to separate long-term trends. The inventory turnover rate trend component in the supply chain features needs to have its logarithmic rate of change calculated, and the supplier delivery stability index is converted to a Z-score standardized form.

[0041] The multi-dimensional feature set is input into the ARIMA model according to the time scale. The model fits the deterministic patterns in the feature set through an autoregressive term (p-order historical value weighting) and a moving average term (q-order residual weighting), outputting linear predictions of business indicators at future time points. The deviation between actual historical sales and the model's fitted values ​​is calculated, generating a nonlinear residual sequence. This sequence contains information not explained by the linear model (such as sudden fluctuations in promotions and supply chain disruptions). The randomness of the residual sequence is verified using the Ljung-Box test to ensure no significant autocorrelation; the p-value must be greater than the significance threshold. If the test fails, the p / q parameters in the model are adjusted until the residuals meet the white noise characteristics.

[0042] For different time scales of the same product, the above process is executed independently to generate scale-specific linear predictions and residual sequences. The output results will be used as inputs for subsequent variational mode decomposition and multi-scale fusion.

[0043] Step S103: Based on external environment data, decompose and predict the nonlinear residual sequence to determine the reconstructed residual sequence. Based on the linear prediction value of the business indicator and the reconstructed residual sequence, determine the current business indicator prediction data corresponding to each dynamic time scale.

[0044] Based on the external environment data, the nonlinear residual sequence is decomposed and predicted to determine the reconstructed residual sequence. Specifically, this includes: decomposing the nonlinear residual sequence into multiple intrinsic mode function (IMF) components through variational mode decomposition, and inputting each IMF component into a temporal convolutional network, wherein the input tensor dimension of the temporal convolutional network is consistent with the time step of the IMF component; performing feature mapping on the external environment data to generate a key-value pair matrix containing economic indicators and promotional activity records, wherein the key vector in the key-value pair matrix is ​​the economic indicator code data corresponding to the economic indicator, and the value vector is the promotional activity code data corresponding to the promotional activity record; determining the hidden layer output of the temporal convolutional network, using the hidden layer output as the query vector, calculating the attention distribution between the query vector and the key vector in the key-value pair matrix, and generating weighted external features based on the attention distribution and the value vector; injecting the weighted external features into the hidden layer of the temporal convolutional network through a gating mechanism, accumulating the predicted output of each IMF component, and generating the reconstructed residual sequence.

[0045] In one embodiment of this specification, a nonlinear residual sequence from an ARIMA model is first received as input. The residual sequence is then separated into intrinsic mode function components using variational mode decomposition (VMD) technology. This decomposition process automatically determines the number of components based on the spectral energy distribution characteristics of the residual signal, decomposing the original residual into several VMD components that characterize different frequency features. The high-frequency components correspond to short-term sudden fluctuations (such as promotional pulses), while the low-frequency components carry abnormal signals in the supply chain (such as delivery delays).

[0046] Each intrinsic mode function component is input into a temporal convolutional network for processing. The input tensor dimension of the temporal convolutional network strictly matches the time step of each component to ensure that the temporal characteristics of the components are fully preserved. For example, a three-dimensional input tensor with a time step of 48 is constructed for the components of a 48-hour promotional pulse. At the same time, the external environment data is structured and mapped. Economic indicators, including the month-on-month change rate of the consumer price index, the year-on-year slope of the producer price index, and the Z-Score of the consumer confidence index, are encoded into a set of key vectors. Promotional activity parameters, including discount percentage, logarithmic value of duration days, and channel coverage weight, are encoded into a set of value vectors, which are stacked vertically to generate a key-value pair matrix.

[0047] The state vectors of the hidden layers of the temporal convolutional network are extracted as query vectors. The dot product similarity between the query vector and the key matrix is ​​calculated to quantify the correlation strength between the network state and economic indicators. The similarity results are standardized and scaled to prevent gradient anomalies. An attention probability distribution is generated using the Softmax function, which reflects the importance weights of promotional features under specific economic conditions. The attention distribution is applied to the value matrix for weighted summation to generate weighted external features that incorporate the influence of the economic environment.

[0048] A gating mechanism is employed for targeted feature injection. The hidden states of the temporal convolutional network are concatenated with weighted external features, and a gating vector is generated using the hyperbolic tangent function, with a value range constrained to the [-1, 1] interval. A supplier delivery stability index is introduced to dynamically adjust the gating strength. When the delivery risk corresponding to the supplier delivery stability index is high, the gating value is compressed. Finally, the gating vector is applied to the hidden states of the temporal convolutional network through the Hadamard product to enhance or suppress key neurons. The processing results of each component are transformed by the output layer of the temporal convolutional network and then accumulated and integrated to generate a reconstructed residual sequence. Finally, this reconstructed residual sequence is superimposed with the linear sales forecast values ​​output by the ARIMA model, aligned along the time axis, to generate complete sales forecast data corresponding to the dynamic time scale.

[0049] It should be noted that in this embodiment, the variational mode decomposition uses adaptive spectral analysis to determine the number of components, the temporal convolutional network uses a causal dilated convolutional structure to capture long-term dependencies, the key-value mapping process needs to perform data standardization to eliminate dimensional differences, the scaling processing in the attention mechanism can avoid gradient anomalies in high-dimensional dot product operations, and the gated injection stage achieves supply chain risk linkage control through delivery stability indicators.

[0050] In automotive parts forecasting scenarios, this process can separate low-frequency signals of delivery delays from high-frequency promotional pulses; in FMCG applications, it can accurately correlate inflation indicators with discount strategies. Through precise spectral separation using Variational Mode Decomposition (VMD), environmental response using key-value attention mechanisms, and dynamic correction using gated targeted injection, the forecasting model can simultaneously capture sudden fluctuations (such as FMCG flash sales) and supply chain risks (such as chip delivery delays), addressing risks such as production stoppages in automotive parts due to accumulated linear deviations and inventory buildup in FMCG due to delayed environmental responses.

[0051] Step S104: Perform multi-scale fusion on the business indicator prediction data corresponding to each dynamic time scale to determine the target business indicator prediction data for the target product.

[0052] For each dynamic time scale, the predicted business indicators are fused across multiple scales to determine the predicted business indicators for the target product. Specifically, this involves: retrieving historical prediction results for each dynamic time scale from the historical prediction database and correlating them with actual sales data; calculating the mean absolute error for each dynamic time scale; determining the error contribution rate for each dynamic time scale based on the mean absolute error; analyzing the request source of the business indicator prediction request to determine the target business scenario corresponding to the request, and matching it with a preset scenario weight base factor. The target business scenario includes any one or more of the following: procurement, replenishment, and promotion scenarios; determining the scenario weight base for each dynamic time scale using the scenario weight base factor and the error contribution rate; determining the scale weight for each dynamic time scale based on the scenario weight base factors across multiple dynamic time scales; and weighting and fusing the predicted business indicators for each dynamic time scale according to its scale weight to determine the predicted business indicators for the target product.

[0053] The existing fixed linear weighted forecasting process ignores the reliability differences between scales and the dynamic nature of business scenarios. In the forecasting of auto parts, it fails to distinguish the difference in error distribution between the procurement cycle scale (30 days) and the inventory turnover scale (7 days), and imposes the historical forecasting error of the highly volatile procurement cycle onto the stable inventory cycle. In the scenario of fast-moving consumer goods promotions, it fails to dynamically adjust the scale weights according to the business type (such as e-commerce limited-time promotions vs. regular replenishment), causing short-term bursts of traffic to be smoothed by the long-cycle model.

[0054] In one embodiment of this specification, historical forecast result sequences for each dynamic time scale are first extracted from the ERP historical forecast database, correlated with actual sales data for the same period, and the mean absolute error (MAE) between the historical forecast values ​​and actual values ​​for each scale is calculated. An error contribution rate is then generated based on the proportional relationship between the sum of MAE across scales. For example, in this embodiment, the time scales are 3 days, 7 days, 14 days, and 30 days. The historical forecast data corresponding to 3 days, 7 days, 14 days, and 30 days under historical forecast conditions are obtained from the historical forecast database, and the mean absolute error (MAE) between the historical forecast values ​​and actual values ​​for each time scale (3 days, 7 days, 14 days, and 30 days) is calculated.

[0055] It should be noted that the numerator in the proportional relationship here is the mean absolute error at a certain dynamic time scale, and the denominator is the sum of the mean absolute errors at multiple dynamic time scales. The error contribution rate is used to quantify the impact of each scale on the overall prediction deviation, with a value range of 0 to 1. Simultaneously, the API call source of the business indicator prediction request is analyzed. If the request comes from the procurement management module (carrying the supplier ID and order type), it is marked as a procurement scenario; if it comes from the warehouse management system (associated with inventory warning thresholds), it is marked as a replenishment scenario; and if it comes from the promotion management platform (bound to an activity ID), it is marked as a promotion scenario.

[0056] The dominant scenario weighting factor is selected based on the current business scenario type (procurement / replenishment / promotion). This dominant scenario weighting factor includes procurement, replenishment, and promotion factors. The procurement scenario factor is determined based on the supplier ID carried when the procurement management module sends a request. It checks whether the supplier is a strategic supplier; if so, the procurement scenario factor is 1.2; otherwise, it is 0.8. The replenishment scenario factor is based on the inventory turnover rate coefficient extracted from the warehouse management system. When the turnover rate is greater than a preset threshold, the replenishment scenario factor is 1.1; otherwise, it is 0.9. The promotion scenario factor is extracted from the promotion management system. When the activity coverage rate is greater than 50%, the promotion scenario factor is 1.0; otherwise, it is 0.7.

[0057] The scenario weight base factor is generated by multiplying the scenario weight base factor by the error contribution rate, and then normalized to the scale weight using an exponential function to avoid negative weights. A weighted sum is then performed based on the predicted data and corresponding scale weights for each dynamic time scale to generate the predicted value of the target business indicator. By setting the scenario weight base factor, even for the same product and the same dataset, the determined fusion weights differ depending on the business scenario type, ensuring a strong correlation with the prediction process and the intended use of the predicted data. The output results are directly connected to the ERP decision-making module: in procurement scenarios, a supplier-tiered procurement list is generated; in replenishment scenarios, safety stock threshold alarms are triggered; and in promotional scenarios, channel-specific inventory recommendations are output.

[0058] In conventional multi-scale fusion techniques, the same static weighting formula, such as fixed linear weighting, is used for all scenarios, including procurement, replenishment, and promotions. This leads to a significant misalignment between the prediction results and the actual decision-making needs of the ERP system. By introducing a scenario weighting base factor, deep coupling between the prediction model and business decision-making objectives is achieved, solving the problem of the disconnect between prediction and application in traditional fusion methods.

[0059] The technical solutions implemented in this specification construct a dynamic time-scale prediction system that adapts to business scenarios, fundamentally solving the technical problem of the disconnect between fixed time scales and multiple business cycles in traditional ERP business indicator prediction. Traditional prediction models treat the time scale as a static hyperparameter; however, the embodiments in this specification achieve precise matching between business scenarios and time windows through supply chain data-driven scale generation and dynamic scale generation based on supply chain data. Traditional solutions, due to the separation of linear models and residual processing, result in sudden signals (promotional impulses) being submerged by trend baselines, and supply chain disruption risks being diluted by environmental noise. The embodiments in this specification achieve a fundamental breakthrough through a dual-channel processing architecture. When the ARIMA model processes multi-dimensional feature sets, it uses stationarity transformation and adaptive difference order to strip away deterministic patterns such as procurement delivery rhythm and inventory turnover trends, transforming supply chain business parameters (such as supplier delivery stability) into model hyperparameters. This prevents the long-term correlation of industrial spare parts procurement cycles from being contaminated by short-term promotional fluctuations. Variational mode decomposition separates the intrinsic mode function components of the residual sequence according to spectral energy. The temporal convolutional network injects weighted external features through a gating mechanism to achieve targeted enhancement of sudden signals. The delivery stability index serves as a gating strength adjustment factor to block the transmission of high-risk supply chain signals.

[0060] This specification also provides an embodiment of a product business data processing device for an ERP business scenario, such as... Figure 2 As shown, the device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described method.

[0061] This specification also provides a non-volatile computer storage medium storing computer-executable instructions configured to perform the above-described method.

[0062] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0063] The above description is merely one or more embodiments of this specification and is not intended to limit this specification. Various modifications and variations can be made to the one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of this specification.

Claims

1. A method for processing product business data in an ERP business scenario, characterized in that, The method includes: When triggered by a business indicator prediction request, historical sales data, supply chain data, and external environment data of the target product are obtained from the ERP system through a preset interface. Based on the supply chain data, multiple dynamic time scales corresponding to the target product are determined, so as to determine the multi-dimensional feature set corresponding to each dynamic time scale. By using a pre-built ARIMA model and a multi-dimensional feature set corresponding to each dynamic time scale, the linear prediction value of the business indicator and the corresponding nonlinear residual sequence for each dynamic time scale are determined. Based on the external environment data, the nonlinear residual sequence is decomposed and predicted to determine the reconstructed residual sequence. Based on the linear prediction value of the business indicator and the reconstructed residual sequence, the current business indicator prediction data corresponding to each dynamic time scale is determined. For each of the aforementioned dynamic time scales, the business indicator prediction data is fused across multiple scales to determine the target business indicator prediction data for the target product.

2. The product business data processing method for an ERP business scenario according to claim 1, characterized in that, Based on the supply chain data, multiple dynamic time scales corresponding to the target product are determined, specifically including: Based on the supply chain data, multi-dimensional business cycle extraction is performed to construct multi-dimensional cycle features, wherein the multi-dimensional cycle features include cycle features of multiple process nodes. Based on the multidimensional periodic characteristics and the historical sales data of the target product, the dominant reference period in the multidimensional periodic characteristics is determined; Based on the dominant reference period, a time scale extension is performed to determine multiple dynamic time scales corresponding to the target product.

3. The product business data processing method for an ERP business scenario according to claim 2, characterized in that, Based on the multidimensional periodic characteristics and the historical sales data of the target product, the dominant reference period in the multidimensional periodic characteristics is determined, specifically including: For each of the multiple process node cycle characteristics, excluding the delivery cycle characteristic, a regression model is constructed with the historical sales data, and the adjusted R-squared value of the regression model for each other cycle characteristic is calculated. 2 value; According to the adjustment R corresponding to each of the other periodic features 2 Value, adjusted to the maximum value of R 2 The periodic characteristics corresponding to the value determine the dominant reference period.

4. The product business data processing method for an ERP business scenario according to claim 2, characterized in that, Based on the dominant reference period, a timescale extension is performed to determine multiple dynamic timescales corresponding to the target product, specifically including: Based on the historical sales data, determine the autocorrelation coefficient sequence corresponding to the dominant reference period, so as to identify the lag order of the first zero crossing point in the autocorrelation coefficient sequence; Based on the period length and hysteresis order corresponding to the dominant reference period, a candidate scale expansion range is generated to expand by a specified multiple within the candidate scale expansion range, thereby determining multiple candidate time scales. In the historical sales data, the number of consecutive segments of sales data with the same trend and the number of trend reversals corresponding to each candidate time scale are counted to calculate the time series aggregation density coefficient corresponding to each candidate time scale. Based on the temporal aggregation density coefficient corresponding to each candidate time scale, multiple dynamic time scales corresponding to the target product are determined by screening among multiple candidate time scales.

5. The product business data processing method for an ERP business scenario according to claim 1, characterized in that, Determining the multi-dimensional feature set corresponding to each of the aforementioned dynamic time scales specifically includes: Seasonal frequency domain feature extraction and image feature transformation are performed on the historical sales sequence corresponding to the historical sales data of each dynamic time scale to determine frequency domain features and two-dimensional spatiotemporal image features, wherein the frequency domain features include high-frequency fluctuation components and low-frequency trend components; From the supply chain data, the inventory turnover rate trend component and the supplier delivery stability index are extracted to determine the supply chain characteristics. The inventory turnover rate trend component is determined based on the inbound and outbound records, and the supplier delivery stability index is determined based on the standard deviation of the supplier's fulfillment time. An attention mechanism is used to weight and fuse the frequency domain features, the two-dimensional spatiotemporal image features, and the supply chain features to generate a multi-dimensional feature set.

6. The product business data processing method for an ERP business scenario according to claim 1, characterized in that, Based on the external environment data, the nonlinear residual sequence is decomposed and predicted to determine the reconstructed residual sequence, specifically including: The nonlinear residual sequence is decomposed into multiple intrinsic mode function components by variational mode decomposition, and each intrinsic mode function component is input into a temporal convolutional network, wherein the input tensor dimension of the temporal convolutional network is consistent with the time step of the intrinsic mode function component. The external environment data is feature-mapped to generate a key-value pair matrix containing economic indicators and promotional activity records. The key vectors in the key-value pair matrix are the economic indicator codes corresponding to the economic indicators, and the value vectors in the key-value pair matrix are the promotional activity codes corresponding to the promotional activity records. The hidden layer output of the temporal convolutional network is determined, and the hidden layer output is used as a query vector. The attention distribution between the query vector and the key vector in the key-value pair matrix is ​​calculated, so as to generate weighted external features based on the attention distribution and the value vector. The weighted external features are injected into the hidden layer of the temporal convolutional network through a gating mechanism, and the predicted output of each intrinsic mode function component is accumulated to generate a reconstructed residual sequence.

7. The product business data processing method for an ERP business scenario according to claim 1, characterized in that, For each of the aforementioned dynamic time scales, the predicted business metrics data are fused using multi-scale methods to determine the predicted target business metrics data for the target product. Specifically, this includes: Historical prediction results for each dynamic time scale are obtained from the historical prediction database and associated with actual sales data. The mean absolute error for each dynamic time scale is calculated, and the error contribution rate for each dynamic time scale is determined based on the mean absolute error. The request source of the business indicator prediction request is analyzed to determine the target business scenario corresponding to the business indicator prediction request, so as to match the preset scenario weight base factor. The target business scenario includes any one or more of the following: procurement scenario, replenishment scenario and promotion scenario. By using the scene weight base factor and the error contribution rate corresponding to each dynamic time scale, the scene weight base corresponding to each dynamic time scale is determined, so as to determine the scale weight corresponding to each dynamic time scale based on the scene weight base corresponding to multiple dynamic time scales. Based on the scale weight corresponding to each dynamic time scale, the business indicator prediction data corresponding to the dynamic time scale are weighted and fused to determine the target business indicator prediction data for the target product.

8. The product business data processing method for an ERP business scenario according to claim 2, characterized in that, Based on the aforementioned supply chain data, multi-dimensional business cycle extraction is performed to construct multi-dimensional cycle features, specifically including: Extract the order date sequence corresponding to historical purchase orders from the supply chain data, and calculate the time interval between adjacent historical purchase orders corresponding to consecutive order dates based on the order date sequence. Determine the cycle characteristics of the purchase process nodes based on multiple time intervals. The system obtains inbound and outbound records from the supply chain data, calculates the ratio of annual sales volume of a single product to average inventory level based on the inbound and outbound records, determines the inventory turnover rate, and determines the corresponding inventory turnover cycle based on the inventory turnover rate, so as to determine the cycle characteristics of the inventory process nodes. Analyze the promotional calendar data of the sales management system, mark the periodic distribution of promotional activity days, and determine the periodic characteristics of promotional process nodes; The system extracts performance records from the supplier management module, calculates the standard deviation of delivery time, and generates a supplier delivery stability index as a periodic feature of the delivery process nodes.

9. A product business data processing device for an ERP business scenario, characterized in that, The device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method as described in any one of claims 1-8.

10. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are configured to perform the method as described in any one of claims 1-8.

Images