Method and apparatus for seasonal decomposition forecasting of demand trends for medium-frequency materials

By generating medium-frequency material demand forecasting results through state-space model decomposition and physical constraints, the problem of multiple seasonal effects in medium-frequency material demand forecasting is solved, and high-precision and robust forecasting is achieved.

CN123070845APending Publication Date: 2026-11-10QINGDAO CRRC SIFANG RAIL TRANSIT EQUIP TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611245741.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-17
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies cannot accurately decouple multiple seasonal effects in medium-frequency material demand forecasting, resulting in forecast results that deviate from the actual business scenario and have low forecast accuracy.

Method used

The historical demand sequence is decomposed using a state-space model. By combining the damping factor decay trend component with error and seasonal components, the final prediction result is generated using physical property constraints, including branch construction, sequence decomposition, model synthesis, model selection, and prediction correction.

Benefits of technology

It improves the accuracy and robustness of medium-frequency material demand forecasting, ensures that the forecast results meet the physical boundary requirements, avoids model divergence and overfitting, and enhances the reliability of supply chain management.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123070845A_ABST
    Figure CN123070845A_ABST
Patent Text Reader

Abstract

This application relates to the fields of supply chain inventory management and big data forecasting technology, and in particular to a method and apparatus for trend and seasonal decomposition forecasting of medium-frequency material demand. The method includes: acquiring the historical demand sequence of the material to be measured, determining multiple execution cycle parameters and constructing corresponding candidate forecasting branches based on each execution cycle parameter; for each candidate forecasting branch, decomposing the historical demand sequence using a state-space model to obtain a component set consisting of error components, trend components, and seasonal components, and synthesizing a candidate forecasting model corresponding to each candidate forecasting branch; performing simulated forecasting using each candidate forecasting model within the verification interval of the historical demand sequence, and selecting a target forecasting model from the candidate forecasting models; generating initial forecast values ​​using the target forecasting model, and performing spatial mapping correction on the initial forecast values ​​based on the physical property constraints of the material to be measured, so as to output the final material demand forecasting result.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the fields of supply chain inventory management and big data forecasting technology, and in particular to a method and apparatus for trend seasonal decomposition forecasting of medium-frequency material demand. Background Technology

[0002] In the field of material supply chain management and demand forecasting, determining the demand trend of specific materials (such as raw materials, spare parts, etc.) in the future cycle usually requires statistical modeling and data analysis.

[0003] In related technologies, the processing method mainly uses conventional time series models to perform linear or nonlinear extrapolation on the historical consumption data of materials in order to calculate the expected consumption of materials in different time windows.

[0004] This approach often relies solely on historical averages or a single trend term for prediction, failing to accurately decouple the characteristic components of mid-frequency materials under multiple seasonal influences. This makes the prediction results prone to deviating from the actual business scenario, resulting in low prediction accuracy. Summary of the Invention

[0005] This application provides a method and apparatus for predicting the seasonality of demand trends for medium-frequency materials, which can accurately decouple the characteristic components of medium-frequency materials under multiple seasonal influences, and improve the matching degree and prediction accuracy between the prediction results and the real business scenario.

[0006] Firstly, this application provides a method for predicting the seasonality decomposition of demand trends for medium-frequency materials, the method comprising: Obtain the historical demand sequence of the material to be tested, extract the feature attributes of the historical demand sequence to determine multiple execution cycle parameters, and construct corresponding candidate prediction branches based on each execution cycle parameter; For each candidate prediction branch, the historical demand sequence is decomposed using a state-space model to obtain a set of components consisting of error components, trend components, and seasonal components. The trend component is decomposed into a smoothing term representing the current level and an incremental term representing the changing trend. The incremental term is attenuated using the damping factor preset in the candidate prediction branch to generate a restricted trend displacement. The smoothing term and the restricted trend displacement are then fused together, and the error component and seasonal component in the component set are combined to synthesize the candidate prediction model corresponding to each candidate prediction branch. Simulation predictions are performed using each candidate prediction model within the validation interval of the historical demand sequence, and the target prediction model is selected from the candidate prediction models based on the statistical index of the deviation between the simulation prediction results and the actual observed values. Initial predicted values ​​are generated using a target prediction model, and spatial mapping corrections are performed on the initial predicted values ​​based on the physical property constraints of the material to be measured, so as to output the final material demand prediction results.

[0007] Secondly, this application provides a trend seasonality decomposition and forecasting device for medium-frequency material demand, the device comprising: The branch construction module is used to obtain the historical demand sequence of the material to be tested, extract the feature attributes of the historical demand sequence to determine multiple execution cycle parameters, and construct corresponding candidate prediction branches based on each execution cycle parameter. The sequence decomposition module is used to decompose the historical demand sequence for each candidate prediction branch using a state-space model to obtain a set of components consisting of error components, trend components, and seasonal components. The model synthesis module is used to decompose the trend component into a smoothing term representing the current level and an incremental term representing the changing trend, use the damping factor preset in the candidate prediction branch to attenuate the incremental term to generate a restricted trend displacement, and fuse the smoothing term and the restricted trend displacement, and combine the error component and seasonal component in the component set to synthesize the candidate prediction model corresponding to each candidate prediction branch. The model selection module is used to perform simulated predictions within the validation interval of the historical demand sequence using each candidate prediction model, and select the target prediction model from the candidate prediction models based on the statistical index of the deviation between the simulated prediction results and the actual observed values. The prediction correction module is used to generate initial prediction values ​​using the target prediction model, and to perform spatial mapping correction on the initial prediction values ​​based on the physical property constraints of the material to be measured, so as to output the final material demand prediction results.

[0008] This application has the following beneficial effects: By extracting characteristic attributes from historical demand sequences to construct candidate prediction branches, and using a state-space model to decompose each branch into components, a candidate model is synthesized by applying damping attenuation to the trend increment term and fusing error and seasonal components. Then, the target prediction model is selected within the validation interval, and the initial prediction values ​​are spatially mapped and corrected using physical attribute constraints to obtain the final result. This embodiment, by further superimposing a damping factor based on inventory turnover inertia correction and physical boundary constraints based on business logic on top of state-space decomposition, can directly optimize the values ​​of each sampling point in the prediction sequence while preserving the historical consumption trend of materials. This optimizes the prediction trajectory to meet physical boundary requirements at the quantitative level, improving the model's ability to capture the multi-seasonal characteristics of mid-frequency materials without significantly increasing statistical computational complexity, thus enhancing the robustness and application value of the prediction results in actual supply chain management. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the accompanying drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below; Figure 1 This is a flowchart illustrating a method for predicting the seasonal decomposition of medium-frequency material demand trends in one embodiment of this application. Figure 2 This is a schematic diagram of a device for predicting the seasonal decomposition of medium-frequency material demand trends, according to one embodiment of this application. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0011] In industrial manufacturing and supply chain management scenarios, demand forecasting for medium-frequency materials (such as electronic components or core parts with weekly or ten-day fluctuations) is central to production planning. Because these materials are driven by both macroeconomic market trends and complex time factors such as production cycles, seasonal promotions, and settlement cycles, traditional time series forecasting methods often present an inherent contradiction between "forecast accuracy" and "trend stability." On the one hand, simple linear regression cannot capture complex seasonal fluctuations; on the other hand, while "black box" models such as deep learning have high fitting rates, they often output negative values ​​or divergent trends due to a lack of physical constraints, severely misleading inventory decisions.

[0012] Therefore, how to introduce constraints that conform to physical common sense while ensuring the flexibility of the model, so as to achieve highly reliable decomposition and prediction of the demand for medium-frequency materials, has become a key problem that urgently needs to be solved in the field of smart supply chain.

[0013] Based on this, this embodiment provides a method for predicting the seasonal decomposition of medium-frequency material demand trends. This method can run on servers, cloud computing platforms, or supply chain scheduling processing units. Figure 1 As shown, the method specifically includes the following steps S101 to S105: S101. Obtain the historical demand sequence of the material to be tested, extract the feature attributes of the historical demand sequence to determine multiple execution cycle parameters, and construct corresponding candidate prediction branches based on each execution cycle parameter.

[0014] In this step, the time series data is first extracted from the database. To determine the execution cycle parameters, autocorrelation analysis was performed on the sequence. Specifically, different lag orders were calculated. The autocorrelation coefficient is used to extract peak shifts that significantly exceed the confidence interval. For example, if in and If a significant peak appears at a certain point, then the parameters of the two execution cycles are determined. .

[0015] Subsequently, different candidate branches are instantiated based on these parameters, such as the "weekly frequency model branch" and the "monthly frequency model branch". This logic of automatically locking the cycle based on the data autocorrelation attribute ensures that the model can adapt to the procurement rhythm of different materials.

[0016] S102. For each candidate prediction branch, the historical demand sequence is decomposed using a state-space model to obtain a set of components consisting of error components, trend components, and seasonal components.

[0017] In this step, the State Space Model (SSM) is recursively computed using the observation equations and state equations. Taking additive decomposition as an example, its mathematical expression is:

[0018] in, express Level of time. Indicates the trend. This indicates the seasonal term of the previous cycle. This represents the residual error. Using Kalman filtering or a smoothing exponential recursive algorithm, the state at each time step is updated, thereby separating out the pure trend component and seasonal component.

[0019] S103, decompose the trend component into a smoothing term representing the current level and an incremental term representing the changing trend, use the damping factor preset in the candidate prediction branch to attenuate the incremental term to generate a restricted trend displacement, and fuse the smoothing term and the restricted trend displacement, and combine the error component and seasonal component in the component set to synthesize the candidate prediction model corresponding to each candidate prediction branch.

[0020] This is a key detail in controlling trend divergence in this embodiment. In the calculation... When calculating the predicted value after a step, the trend evolution logic is as follows: Let the current time be Its smoothing term is The incremental term is Introducing a damping factor Calculate the restricted trend displacement as The final synthesized prediction model formula is expressed as follows:

[0021] Through damping factor Exponential decay, incremental term The impact will vary with the prediction step size. The convergence occurs as the value increases, thus avoiding the infinite growth of predicted values ​​caused by traditional linear extrapolation.

[0022] S104. Perform simulation predictions within the verification interval of the historical demand sequence using each candidate prediction model, and select the target prediction model from the candidate prediction models based on the statistical index of the deviation between the simulation prediction results and the actual observations.

[0023] Historical data is divided into training and validation sets (e.g., the last 12 observations). On the validation set, the symmetric mean absolute percentage error (sMAPE) of each branch model is calculated:

[0024] Automatic comparison of each branch The model with the smallest error index is selected as the target model. This "horse race mechanism" ensures that the model has been verified for local optimality before actual output.

[0025] S105. Generate initial predicted values ​​using the target prediction model, and perform spatial mapping correction on the initial predicted values ​​based on the physical property constraints of the material to be measured, so as to output the final material demand prediction result.

[0026] Since material requirements cannot be negative physically, a zero-boundary physical threshold is set at the output layer. When the target prediction model outputs the initial predicted value At that time, the value mapping operator is invoked for correction.

[0027] In one basic implementation of this embodiment, the mapping operator employs a ReLU-style truncated mapping, i.e. .

[0028] In the basic prediction process described above, the "construction of candidate prediction branches" mentioned in S101 is the core of model adaptation. However, in actual business scenarios, due to the varying lengths of historical data for different materials, forcing long-term predictions onto short sequences (e.g., using only 3 months of data to train an annual cycle model) can lead to severe overfitting. To address this issue, this application introduces a sample reliability verification mechanism into the branch construction process in S101 in some preferred embodiments.

[0029] Regarding the process of "constructing the corresponding candidate prediction branch" in step S101 above, this embodiment provides a filtering logic aimed at improving the robustness of the model. The specific operation process is as follows: First, based on the determined multiple execution cycle parameters, a set of optional prediction branches corresponding to each one is constructed.

[0030] Based on the periodic parameters determined by the aforementioned ACF analysis or mapping table (such as...) day, moon, In [year], a preliminary set of optional prediction branches is generated. At this point, these branches are merely logical candidate states.

[0031] Secondly, based on the length of the historical demand sequence, the reliability of each branch in the set is verified.

[0032] Read the total length of the historical demand sequence for the current material The sequence length is then compared and correlated with the periodic attributes of each branch. The core criterion for validation is whether the sequence length is sufficient to support the model in learning parameters with the minimum number of iterations.

[0033] Finally, branches that do not meet the requirements are removed, and the remaining branches are retained as the final "candidate prediction branches".

[0034] If a certain branch (such as an annual cycle branch) If the required amount of historical data far exceeds the actual length of the current sequence, a "dynamic elimination" action will be performed. Ultimately, only branches that pass the verification can proceed to the subsequent SSM decomposition and horse racing stages (S102-S104). This mechanism ensures that subsequent simulation predictions are based on a solid statistical sample foundation, avoiding false trend judgments caused by insufficient samples.

[0035] To provide a quantifiable standard for the aforementioned "reliability verification," this embodiment further refines the specific verification algorithm. The specific calculation logic is as follows: The first step is to calculate the minimum sample size threshold for each optional prediction branch. .

[0036] For any branch in the set Using its corresponding execution cycle parameters Compared with the preset minimum modeling cycle multiple Perform the product operation.

[0037] The calculation formula is: In this embodiment, the multiple Typically, the value is set to 2 or 3 based on statistical significance requirements (i.e., it must cover at least 2 to 3 complete cycles). For example, if the monthly cycle parameter corresponding to a certain branch... Preset multiplier If so, the minimum sample size threshold for this branch is 60.

[0038] The second step is to obtain the total length of the sequence. and compare it with the threshold Perform a comparison of the measured values.

[0039] Automatically obtain the total number of observation points for the historical sequence of the material to be tested. For example, if there is only 45 days of historical data for the current material, then... .

[0040] The third step is to determine the reliability of the branch based on the comparison results and then remove it.

[0041] The following decision logic will be executed: like The branch is determined to be an unreliable branch. In the example above, Therefore, the monthly cycle branch is deemed unreliable and is removed from the candidate set.

[0042] like If the branch meets the minimum number of iterations required for modeling, it is retained.

[0043] By introducing this quantitative verification logic based on "period × multiple", this embodiment can automatically "tailor" the approach to the data availability of each type of material. It eliminates the risk of forcing long-period models onto short sequences at the source, ensuring that all candidate models participating in the "horse race" have sufficient statistical confidence, thereby significantly improving the robustness of the overall prediction framework.

[0044] Based on the basic prediction framework provided in the above embodiments, in order to further improve the degree of automation and prediction stability when handling complex material demands, this application provides a more targeted preferred implementation scheme on the basis of the above steps. The following will describe in detail the automated configuration of periodic identification and the refined processing of trend decay in conjunction with specific business logic.

[0045] In some preferred embodiments, for the process of "extracting feature attributes to determine periodic parameters" in step S101 above, this solution adopts an automated configuration mechanism based on sampling frequency matching, and the specific execution logic is as follows: First, analyze the timestamp intervals of historical demand sequences. Specifically, the time difference between two adjacent demand observations is calculated, and their mode or mean is statistically analyzed to determine the sampling frequency characteristics of the sequence. (For example, if the intervals are mostly 7 days, then) It is identified as "weekly frequency"; if the interval is 1 day, then... (For "daily frequency").

[0046] Secondly, based on the identified sampling frequency characteristics It automatically matches a preset periodicity mapping table. This mapping table stores the most likely seasonal patterns at different sampling frequencies. For example: when When using "daily frequency", the set of execution cycle parameters output by the mapping table is: , representing the weekly cycle, monthly cycle, and yearly cycle, respectively; when When using "weekly frequency", the set of execution cycle parameters output by the mapping table is: , representing monthly, quarterly, and annual cycles, respectively.

[0047] This mapping mechanism ensures that the model can cover complex combinations of non-seasonal dimensions (such as pure trends) and multiple seasonal cycles, avoiding the blindness of manually setting cycles.

[0048] Finally, based on the selected execution cycle parameters This involves instantiating and configuring the parameters of a pre-defined state-space model template. This includes initializing the seasonal dimension (e.g., ...) in the state vector. The model uses a dimensional vector and sets initial weights for each component. Through this process, the parameters of each execution cycle are transformed into a "candidate prediction branch" with specific physical meaning, achieving deep decoupling between the model architecture and material properties.

[0049] In step S103 above, to more accurately control the performance of the trend component in long-step prediction, this embodiment refines the mathematical process for generating the "constrained trend displacement". The specific operation flow is as follows: The first step is to utilize the preset damping factor in the candidate prediction branch. Construct a prediction step size Extended nonlinear decay weight In this embodiment, the weight follows an exponential decay law, that is... .because With the prediction step size The increase of weight It decreases rapidly and non-linearly.

[0050] The second step is to apply this non-linear decay weight to the incremental term in the trend component. The trend increment after decay is calculated. The calculation formula is as follows:

[0051] The physical significance of this step is that it is believed that the current growth (or decline) trend is losing momentum at every future time step, and the rate of loss is accelerating.

[0052] The third step is to combine the attenuated trend increments at each step size with the smoothing term in the trend component. The results are accumulated. The final constrained trend displacement represents the trend towards the future. The integral of the step increment. For example, in predicting the future step... When the step is taken, the restricted trend displacement is .

[0053] Through this nonlinear decay process, the prediction curve can maintain an incremental trend in the short term, while in the long term it will smoothly tend to a stable level, effectively solving the technical problem of "predicted values ​​diverging to unreasonable values ​​over time" that is common in medium-frequency material prediction.

[0054] In the aforementioned damping attenuation mechanism, the damping factor is typically set as a preset static constant. However, in real-world supply chain environments, material demand trends are often physically constrained by inventory buildup or turnover rates. To enable the forecasting model to perceive the "damping effect" of real-world operations, this application provides a dynamic adaptive correction scheme for the damping factor in some further preferred embodiments.

[0055] Regarding the "constructing nonlinear decay weights" step mentioned in the above embodiments, in order to address the problem that static damping cannot respond to inventory fluctuations, this embodiment introduces a material flow inertia evaluation mechanism. Its specific logical implementation process is as follows: First, obtain the current inventory turnover rate of the material to be tested in real time. And calculate its comparison with the preset historical average turnover rate. Ratio offset .

[0056] By interfaceing with WMS (Warehouse Management System), the total amount of materials issued and the average inventory level within the current statistical period are extracted, and the current inventory turnover rate is calculated. The ratio offset is calculated as follows: This offset reflects the degree to which the current material flow rate deviates from the normal level.

[0057] Secondly, using the ratio offset Characterizing the material flow inertia coefficient of the test material at the current moment. .

[0058] This scheme establishes a negative correlation mapping function: .

[0059] The underlying physical logic is as follows: when When the current turnover rate is lower than the historical average, it means that inventory is piling up and the material turnover momentum is insufficient, resulting in a small inertia coefficient. In a physical sense, this means that inventory buildup has a "resistive effect" on demand trends, meaning that future demand growth should be more drastically compressed.

[0060] when Time (i.e., normal or accelerated flow), inertia coefficient Approaching 1, maintaining the original trend momentum.

[0061] Secondly, utilizing the material flow inertia coefficient As a weighting adjustment operator, it adjusts the preset damping factor. Perform product correction to obtain the dynamic adaptive damping parameters. .

[0062] The calculation formula is as follows: because When inventory is piling up, the corrected dynamic adaptive damping parameter is less than 1. It will be smaller than the basic damping factor. This means that in the trend prediction formula... In this context, a decrease in the base will lead to an increase in the increment term. With prediction step size It increases but decays even faster.

[0063] Finally, using dynamic adaptive damping parameters As a base, to predict the step size Perform a power function operation on the exponent to generate non-linear decaying weights. .

[0064] Right now: .

[0065] Through the above logic, this embodiment transforms the "inventory data" in ERP into "dynamic constraints" in the prediction algorithm. When a slowdown in material turnover is detected (i.e., a decrease in the ratio offset), the algorithm automatically adjusts the damping parameter by reducing the inertia coefficient, thereby forcing the prediction curve to converge to the horizontal direction more quickly. This prevents the model from giving aggressive growth predictions when inventory is high, significantly improving the reliability of the prediction results in the actual production environment.

[0066] After constructing the internal parameters of the model (such as the dynamic damping factor) as described above, several candidate prediction models with different periodic attributes and decay characteristics were generated. In order to select the optimal model that best matches the current material demand characteristics from these candidate models, this application has refined the "model selection" process in step S104 above in some preferred embodiments.

[0067] Regarding the process of "selecting the target prediction model from candidate prediction models" in step S104 above, this embodiment employs a "horse race mechanism" based on an independent validation set to ensure that the selected model possesses the strongest generalization ability. Its specific execution logic is as follows: First, validation interval data, independent of the modeling process, is extracted from the historical demand sequence.

[0068] The acquired historical demand sequence Divided chronologically into two non-overlapping subsets: the training interval and verification interval (For example, selecting the end of the sequence) (Each observation is used as a validation interval). Used for state initialization and component decomposition of each candidate prediction branch, and It serves as a truth reference for the "closed-book exam" and is not used in training the model parameters. This physical isolation ensures the objectivity of subsequent bias assessment.

[0069] Secondly, each candidate prediction model is used to simulate and predict the time corresponding to the data in the validation interval, generating a sequence of simulated prediction values.

[0070] For each candidate prediction model Based on its The horizontal, trend, and seasonal terms observed at the end are extrapolated into the future. By taking a step size, the corresponding simulated predicted value sequence is obtained. At this point, each candidate model has given its own prediction for the same "future".

[0071] Next, the dispersion index between the simulated predicted value sequence and the validation interval data is calculated, and the deviation statistics of each candidate prediction model are determined based on this index.

[0072] simulated predicted values With the true value of the verification interval A point-by-point comparison is performed. In this embodiment, the dispersion index is specifically characterized by root mean square error (RMSE) or mean absolute error (MAE).

[0073] Taking the root mean square error as an example, the calculation formula is as follows: This dispersion index It intuitively reflects the extent to which the model deviates from the actual physical trajectory during the simulation process. The smaller the value, the more accurately the model captures the recent fluctuations of the material.

[0074] Finally, the candidate prediction model with the smallest deviation statistical index was selected as the target prediction model.

[0075] Establish a decision evaluation matrix to rank all candidate models by their dispersion index. Automatically trigger the selection command: The model with the smallest deviation statistic is selected as the "winner" for the final output of the target predicted value.

[0076] Through the division, simulation, comparison, and optimization processes described in the above embodiments, this solution overcomes the instability of a single model when facing different materials. By utilizing the "real-world simulation" of the validation interval, it can automatically identify which periodic parameter configuration and which damping attenuation weight best reproduces the true demand trend of the material, thereby fundamentally ensuring the reliability of the final output prediction results.

[0077] After selecting the target prediction model and generating initial prediction values ​​using the aforementioned "horse race mechanism," the algorithm enters the most crucial "physical implementation" stage. Since mathematical models (such as state-space models) may produce physically meaningless negative values ​​at the purely computational level, while real material demand is irreversible (i.e., the demand quantity cannot be negative), spatial correction is necessary.

[0078] Regarding the application of "physical attribute constraints" in step S105 above, this embodiment first provides a correction scheme based on hard constraints of business logic. The specific execution logic is as follows: First, based on the irreversible nature of the demand for the material to be tested, a zero-boundary physical threshold is set at the business logic level. .

[0079] In supply chain management, the requisition or purchase quantity of materials is unidirectional in the physical space; negative values ​​usually represent returns or abnormal write-offs, which are not within the scope of forecasting. Therefore, Set the physical boundary constraints as the initial predicted values.

[0080] Secondly, the initial predicted value is compared with the zero-boundary physical threshold for out-of-bounds detection, and the quantity mapping operator is called to perform spatial correction.

[0081] The initial prediction sequence output by the target model Perform a point-by-point scan. If the initial predicted value at a certain moment is detected... If the value exceeds the limit, it is determined that an "out-of-bounds" event has occurred. At this point, the value mapping operator is triggered. In the basic form of this embodiment, this operator is manifested as a truncation function:

[0082] This operator forces all negative outliers that exceed the physical boundary to be mapped to the zero-boundary physical threshold.

[0083] Finally, the spatially corrected predicted values ​​are determined as the final required results and output.

[0084] This process ensures that the data output to the production planning system is secure and executable from a business logic perspective.

[0085] In practical applications, the simple truncation described above may result in a "hard inflection point" on the predicted curve at 0, disrupting the continuity of the signal. To achieve a smoother correction effect that better conforms to the laws of physical evolution, this application introduces a correction mechanism based on a potential barrier function in some preferred embodiments. The specific implementation process is as follows: First, the zero-boundary physical threshold is defined as the potential boundary, and an inverse barrier function is established that strengthens as the predicted value approaches the potential boundary. .

[0086] Introducing the concept of a potential barrier from physics, we define a function of the reverse force. For example, we can use the following exponential potential barrier function: in, This is the sensing coefficient. The characteristic of this function is that when the predicted value... When far from 0, the potential barrier value is extremely small; when When the barrier value rapidly approaches or even attempts to penetrate the zero point, it increases exponentially.

[0087] Secondly, during the recursive process of the state-space model, the mapping correction vector is calculated based on the displacement of the initial predicted value relative to the potential boundary. .

[0088] Instead of truncating the result directly, the model is monitored in real time at each step of the recursion. When the predicted trajectory shows a downward trend of penetrating the potential boundary, a mapping correction vector that pushes back in the positive direction is calculated based on the current displacement gradient using the inverse barrier function. The magnitude of this vector is positively correlated with the depth to which the predicted value penetrates the boundary.

[0089] Finally, the initial predicted values ​​are reconstructed nonlinearly using the mapping correction vector.

[0090] The correction vector is superimposed into the evolutionary logic of the state equation, causing the predicted trajectory to be subjected to a "reverse thrust" before touching the zero point, resulting in an upward arc. This reconstruction process smoothly maps the predicted trajectory in the negative value region to the positive convergence interval of the zero-boundary physical threshold.

[0091] By introducing this nonlinear reconstruction based on the inverse barrier function, this embodiment not only solves the physical constraint that demand cannot be negative, but more importantly, it maintains the mathematical continuity and smoothness of the prediction curve near the critical point. This avoids unnecessary and frequent adjustments triggered by downstream automated scheduling due to precipitous data fluctuations, achieving a deep harmony between algorithm output and engineering implementation.

[0092] Based on the same inventive concept, this application also provides a device for predicting the seasonal decomposition of medium-frequency material demand trends to implement the aforementioned method for predicting the seasonal decomposition of medium-frequency material demand trends. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the device for predicting the seasonal decomposition of medium-frequency material demand trends provided below can be found in the limitations of the method for predicting the seasonal decomposition of medium-frequency material demand trends described above, and will not be repeated here.

[0093] In one exemplary embodiment, such as Figure 2 As shown, a trend seasonality decomposition and forecasting device for medium-frequency material demand includes: Branch construction module 11 is used to obtain the historical demand sequence of the material to be tested, extract the feature attributes of the historical demand sequence to determine multiple execution cycle parameters, and construct corresponding candidate prediction branches based on each execution cycle parameter. The sequence decomposition module 12 is used to decompose the historical demand sequence using a state-space model for each candidate prediction branch, so as to obtain a set of components consisting of error components, trend components and seasonal components. The model synthesis module 13 is used to decompose the trend component into a smoothing term representing the current level and an incremental term representing the changing trend, use the damping factor preset in the candidate prediction branch to attenuate the incremental term to generate a restricted trend displacement, and fuse the smoothing term and the restricted trend displacement, and combine the error component and seasonal component in the component set to synthesize the candidate prediction model corresponding to each candidate prediction branch. The model selection module 14 is used to perform simulated predictions within the verification interval of the historical demand sequence using each candidate prediction model, and select the target prediction model from the candidate prediction models based on the statistical index of the deviation between the simulated prediction results and the actual observed values. The prediction correction module 15 is used to generate initial predicted values ​​using the target prediction model, and to perform spatial mapping correction on the initial predicted values ​​based on the physical property constraints of the material to be measured, so as to output the final material demand prediction result. Each module in the aforementioned seasonal decomposition and forecasting device for medium-frequency material demand trends can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0094] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method for predicting the trend seasonality decomposition of medium-frequency material demand.

[0095] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which determines the implementation of the above-described method for predicting the seasonal decomposition of medium-frequency material demand trends when the program is executed by a processor.

[0096] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the aforementioned method for predicting the trend of seasonal decomposition of medium-frequency material demand.

[0097] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile deterministic machine-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0098] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0099] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for predicting the seasonal decomposition of demand trends for medium-frequency materials, characterized in that, The method includes: Obtain the historical demand sequence of the material to be tested, extract the feature attributes of the historical demand sequence to determine multiple execution cycle parameters, and construct corresponding candidate prediction branches based on each of the execution cycle parameters; For each candidate prediction branch, the historical demand sequence is decomposed using a state-space model to obtain a set of components consisting of error components, trend components, and seasonal components. The trend component is decomposed into a smoothing term representing the current level and an incremental term representing the changing trend. The incremental term is attenuated using the damping factor preset in the candidate prediction branch to generate a restricted trend displacement. The smoothing term and the restricted trend displacement are fused together and combined with the error component and seasonal component in the component set to synthesize the candidate prediction model corresponding to each candidate prediction branch. Simulation prediction is performed using each of the candidate prediction models within the validation interval of the historical demand sequence, and a target prediction model is selected from the candidate prediction models based on the deviation statistics between the simulation prediction results and the actual observed values. The target prediction model is used to generate initial prediction values, and the initial prediction values ​​are spatially mapped and corrected based on the physical property constraints of the material to be measured, so as to output the final material demand prediction result.

2. The method according to claim 1, characterized in that, The step of extracting feature attributes from the historical demand sequence to determine multiple execution cycle parameters and constructing candidate prediction branches corresponding one-to-one with each of the execution cycle parameters specifically includes: Analyze the timestamp intervals of the historical demand sequence to obtain the sampling frequency characteristics of the historical demand sequence; Based on the sampling frequency characteristics, a preset period mapping table is matched to determine multiple execution period parameters covering non-seasonal dimensions and multiple seasonal periods; Based on the execution cycle parameters, the model template corresponding to the preset state space model is instantiated and configured to generate each candidate prediction branch.

3. The method according to claim 2, characterized in that, The step of using a preset damping factor in the candidate prediction branch to attenuate the incremental term and generate a constrained trend displacement specifically includes: The candidate prediction branch is preset with a damping factor, which is used to construct a nonlinear decay weight that extends with the prediction step size. The nonlinear decay weight is applied to the incremental term in the trend component to calculate the decayed trend increment. The attenuated trend increment is added to the smoothing term in the trend component to obtain the constrained trend displacement.

4. The method according to claim 3, characterized in that, The construction of the nonlinear decay weights that extend with the prediction step size specifically includes: Obtain the current inventory turnover rate of the material to be tested, and calculate the offset of its ratio to the preset historical average turnover rate; The ratio offset is used to characterize the material flow inertia coefficient of the material under test at the current moment; wherein the material flow inertia coefficient is negatively correlated with the ratio offset, which is used to reflect the hindering effect of inventory accumulation on demand trends. Using the material flow inertia coefficient as a weight adjustment operator, the preset damping factor is multiplied and corrected to obtain the dynamic adaptive damping parameter; Using the dynamic adaptive damping parameter as the base, a power function operation is performed with the prediction step size as the exponent to generate a nonlinear decay weight.

5. The method according to claim 1, characterized in that, The step of performing simulated predictions using each of the candidate prediction models within the validation interval of the historical demand sequence, and selecting a target prediction model from the candidate prediction models based on the statistical index of the deviation between the simulated prediction results and the actual observed values, specifically includes: Extract validation interval data from the historical demand sequence, independent of the modeling process; The candidate prediction models are used to simulate and predict the time corresponding to the verification interval data, thereby generating a sequence of simulated prediction values ​​for each candidate prediction model. Calculate the dispersion index between the simulated predicted value sequence and the verification interval data, and determine the deviation statistics index of each candidate prediction model based on the dispersion index; The candidate prediction model with the smallest deviation statistical index is selected as the target prediction model.

6. The method according to claim 1, characterized in that, The step of spatially mapping and correcting the initial predicted value based on the physical property constraints of the material to be measured, in order to output the final material demand prediction result, specifically includes: Based on the irreversible demand attribute of the material to be tested, a zero-boundary physical threshold is set at the business logic level as a physical boundary constraint for the initial predicted value. The initial predicted value and the zero-boundary physical threshold are subjected to boundary detection, and the value mapping operator is called to map the negative outliers that exceed the physical boundary constraints to the zero-boundary physical threshold, so as to spatially correct the initial predicted value. The predicted value after spatial correction is determined as the final material demand forecast result and output.

7. The method according to claim 6, characterized in that, The invoked value mapping operator maps negative outliers exceeding the physical boundary to the zero-boundary physical threshold, specifically including: The zero-boundary physical threshold is defined as the potential boundary, and an inverse barrier function is established that strengthens as the predicted value approaches the potential boundary. During the recursive process of the state-space model, the mapping correction vector is calculated using the reverse barrier function based on the displacement of the initial predicted value relative to the potential boundary. The initial predicted value is nonlinearly reconstructed using the mapping correction vector, and the predicted trajectory in the negative region is mapped to the positive convergence interval of the zero-boundary physical threshold.

8. The method according to claim 1, characterized in that, The construction of corresponding candidate prediction branches based on each of the execution cycle parameters specifically includes: Based on the determined multiple execution cycle parameters, construct a set of optional prediction branches that correspond one-to-one with each parameter; Based on the length of the historical demand sequence, sample reliability verification is performed on each branch in the set of optional prediction branches; Eliminate optional prediction branches that do not meet the minimum number of iterations required for modeling, and determine the remaining optional prediction branches as candidate prediction branches; wherein each candidate prediction branch corresponds to its respective execution cycle parameter.

9. The method according to claim 8, characterized in that, Based on the length of the historical demand sequence, sample reliability verification is performed on each branch in the set of optional prediction branches, specifically including: For each of the optional prediction branches, the corresponding execution cycle parameter is multiplied by the preset minimum modeling cycle multiple to obtain the minimum sample capacity threshold of the corresponding branch. Obtain the total sequence length of the historical demand sequence, and compare the total sequence length with the minimum sample capacity threshold corresponding to each of the optional prediction branches; Based on the comparison results, the optional prediction branches whose total sequence length is less than the minimum sample size threshold are determined to be unreliable branches that do not meet the minimum number of iterations required for modeling.

10. A device for predicting the seasonality of demand trends for medium-frequency materials, characterized in that, The device includes: The branch construction module is used to obtain the historical demand sequence of the material to be tested, extract the feature attributes of the historical demand sequence to determine multiple execution cycle parameters, and construct corresponding candidate prediction branches based on each of the execution cycle parameters. The sequence decomposition module is used to decompose the historical demand sequence using a state-space model for each candidate prediction branch to obtain a component set consisting of error components, trend components, and seasonal components. The model synthesis module is used to decompose the trend component into a smoothing term representing the current level and an incremental term representing the changing trend, use the damping factor preset in the candidate prediction branch to attenuate the incremental term to generate a restricted trend displacement, and fuse the smoothing term and the restricted trend displacement, and combine the error component and seasonal component in the component set to synthesize the candidate prediction model corresponding to each candidate prediction branch. The model selection module is used to perform simulated predictions using each of the candidate prediction models within the verification interval of the historical demand sequence, and select a target prediction model from the candidate prediction models based on the deviation statistics between the simulated prediction results and the actual observed values. The prediction correction module is used to generate an initial prediction value using the target prediction model, and to perform spatial mapping correction on the initial prediction value based on the physical property constraints of the material to be measured, so as to output the final material demand prediction result.