A low-frequency material demand multi-scheme adaptive optimal prediction method and device

Abstract

The application relates to a low-frequency material demand multi-scheme adaptive optimal prediction method and device, which comprises the following steps: obtaining a historical demand sequence of a target material, and determining the current parameter state of each candidate prediction model in a candidate prediction model set; performing backtest calculation on the historical demand sequence by using each candidate prediction model in the current parameter state, to obtain a plurality of candidate prediction results; performing deviation calculation on the plurality of candidate prediction results and the actual demand value at the corresponding time point in the historical demand sequence, to obtain a basic score representing prediction accuracy and a sorting score representing recent performance stability, and performing weighted fusion on the basic score and the sorting score to determine the comprehensive score corresponding to each candidate prediction model; and determining a target prediction model from each candidate prediction model according to the comprehensive scores, and outputting the candidate prediction result corresponding to the target prediction model. The application can improve the dynamic sensitivity and business stability of the prediction result.

Description

Technical Field

[0001] This application relates to the field of logistics supply chain management and data forecasting and analysis technology, and in particular to a multi-scheme adaptive optimal prediction method and apparatus for low-frequency material demand. Background Technology

[0002] High-speed trains have a wide variety of key components, and demand exhibits typical intermittency and uncertainty due to the operating environment and maintenance tasks. Establishing an accurate spare parts forecasting mechanism is crucial for achieving a balance between high availability and low inventory costs in rail transit operations and maintenance.

[0003] Existing technologies typically employ fixed statistical models to backtest historical demand sequences and use a single numerical error (such as mean absolute error) as the evaluation criterion to select a prediction model and guide subsequent replenishment decisions.

[0004] However, because the evaluation logic relies too heavily on historical averages, the model often exhibits severe "response dulling" and distorted prediction signals when faced with complex fluctuations in maintenance conditions. This single-dimensional evaluation system cannot effectively capture the dynamic evolution of demand characteristics, making it difficult for the prediction results to provide reliable decision support in changing environments. Summary of the Invention

[0005] Therefore, it is necessary to provide a multi-scheme adaptive optimal prediction method and apparatus for low-frequency material demand that improves the dynamic sensitivity and operational stability of prediction results, in order to address the aforementioned technical problems.

[0006] Firstly, this application provides a multi-scheme adaptive optimal prediction method for low-frequency material demand, the method comprising: Obtain the historical demand sequence for the target material; The demand variation coefficient is calculated based on the historical demand sequence, and the relationship between the demand variation coefficient and the preset threshold is used to determine whether to trigger the parameter update procedure for the preset set of candidate prediction models. Based on the result of the trigger determination, determine the current parameter status of each candidate prediction model in the candidate prediction model set; Using each candidate prediction model in its current parameter state, backtesting calculations are performed on the historical demand sequence to obtain multiple candidate prediction results. By using the deviation calculation between multiple candidate prediction results and the actual demand values ​​at corresponding time points in the historical demand sequence, a basic score representing the prediction accuracy and a ranking score representing the stability of recent performance are obtained. The basic score and the ranking score are then weighted and fused to determine the comprehensive score corresponding to each candidate prediction model. Based on the comprehensive scores, the target prediction model is determined from the candidate prediction models, and the candidate prediction results corresponding to the target prediction model are output.

[0007] Secondly, this application provides a multi-scheme adaptive optimal prediction device for low-frequency material demand, the device comprising: The sequence acquisition module is used to acquire the historical demand sequence of the target material; The trigger determination module is used to calculate the demand variation coefficient based on the historical demand sequence, and determine whether to trigger the parameter update program for the preset set of candidate prediction models based on the relationship between the demand variation coefficient and the preset threshold. The parameter management module is used to determine the current parameter status of each candidate prediction model in the candidate prediction model set based on the result of the trigger judgment. The backtesting calculation module is used to perform backtesting calculations on the historical demand sequence using each candidate prediction model in its current parameter state, and obtain multiple candidate prediction results. The multidimensional scoring module is used to calculate the deviation between multiple candidate prediction results and the actual demand values ​​at corresponding time points in the historical demand sequence, to obtain a basic score that represents the prediction accuracy and a ranking score that represents the stability of recent performance. The basic score and the ranking score are then weighted and fused to determine the comprehensive score corresponding to each candidate prediction model. The model selection module is used to determine the target prediction model from the candidate prediction models based on the comprehensive scores, and output the candidate prediction results corresponding to the target prediction model.

[0008] The beneficial effects of this application are as follows: 1) This application achieves a keen capture of changes in the consumption pattern of high-speed train parts by dynamically triggering parameter updates through the coefficient of variation. This mechanism effectively overcomes the problem of feedback passivation in the prediction model in traditional technologies, ensuring that the model parameters can be autonomously optimized as the operating conditions evolve, and making the prediction logic highly consistent with the actual maintenance environment; (2) By constructing a dual evaluation system of basic accuracy and recent stability, this scheme effectively filters out inferior models that are prone to distorted signals due to recent drastic fluctuations. This multi-dimensional scoring weighted fusion mechanism significantly reduces the risk of deviation in replenishment decisions and provides more valuable decision support for the scientific storage and inventory balancing of intermittent materials. 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.

[0010] Figure 1 This is a flowchart illustrating a multi-scheme adaptive optimization prediction method for low-frequency material demand in an embodiment of this application. Figure 2 This is a schematic diagram of the structure of a multi-scheme adaptive optimal prediction device for low-frequency material demand in an embodiment of this application. Detailed Implementation

[0011] 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.

[0012] As highly complex rail transit equipment, the efficient operation and maintenance of high-speed trains heavily relies on a vast spare parts support system. Because the demand for train parts is often triggered by sudden malfunctions or random condition-based maintenance, material consumption exhibits a highly intermittent characteristic. In actual operation, different types of parts (such as mechanical fasteners, electronic components, and polymer seals) have varying consumption patterns and are affected by a combination of external factors, including train operating mileage, climate, and track conditions. This uncertainty in demand directly exacerbates the difficulty of inventory management. Ensuring a high availability rate while avoiding the capital tied up in slow spare parts turnover remains a long-standing challenge for railway companies.

[0013] In existing technologies, demand forecasting for intermittent materials typically employs standard statistical modeling methods. Technicians usually select one or more mathematical models to fit historical consumption records and use traditional error metrics (such as mean absolute deviation) to evaluate the accuracy of the forecast. This approach tends to select the model that numerically best matches historical trends as the basis for subsequent material procurement and inventory replenishment decisions.

[0014] However, existing technologies suffer from a core technical flaw in practical applications: the forecasting mechanism lacks effective adaptive response capabilities to sudden "phased drifts" in demand patterns. Due to the inherent randomness of intermittent sequences, material consumption characteristics often fundamentally change with variations in train mileage, maintenance standards, or environmental conditions. Traditional technologies rely too heavily on long-term statistical averages, leading to severe feedback dulling and decision-making errors when faced with fundamental adjustments to the sequence structure. This lack of awareness of "mode switching" directly results in the inability to select the most suitable forecasting logic for the current operating conditions at critical time points, ultimately leaving material support plans in a state of "passive following" or "ineffective feedback."

[0015] Based on this, this embodiment provides a multi-scheme adaptive optimal prediction method for low-frequency material demand. By conducting parallel backtesting of multiple models and performing two-dimensional accuracy evaluation, it solves the technical challenge of accurately predicting the optimal demand for low-frequency materials when there are numerous zero-value disturbances. This method can be applied to material demand planning servers or intelligent warehousing decision-making platforms, such as… Figure 1 As shown, the method includes the following steps: S100: Obtain the historical demand sequence for the target material; "Target material" refers to industrial spare parts or long-tail SKUs with intermittent demand characteristics; "historical demand sequence" refers to the discrete consumption value vector generated by this material on a continuous time axis. The specific implementation process is as follows: Accessing the enterprise's underlying material management database via a data gateway, the system extracts the original outbound transaction records of the target material according to the SKU code. First, discrete outbound events are aggregated into time windows based on a preset time granularity (e.g., weekly, monthly). Then, missing value handling logic is executed: for time domains where no requisition record occurred within a specific statistical period, zero-padding is automatically performed to generate an evenly spaced sequence of non-negative values. .in, Indicates the first The actual demand value for each cycle, This is the total step size of the observation window. This sequence... This serves as the physical basis for subsequent statistical calculations and model inputs.

[0016] S200: Calculate the demand variation coefficient based on the historical demand sequence, and determine whether to trigger the parameter update procedure for the preset set of candidate prediction models based on the relationship between the demand variation coefficient and the preset threshold. The core of this step lies in dynamically deciding whether to reset the logical state of the prediction engine by quantifying the physical characteristics of demand fluctuations in real time. The specific implementation process is as follows: First, the statistical engine is called to process the sequence output in step S100. Perform fluctuation characteristic capture. Calculate the relative deviation of the discrete distribution of the sequence within the statistical window relative to its central position to obtain the demand variation coefficient. This coefficient reflects the heterogeneity level of intermittent demand sequences by characterizing the density and magnitude of demand pulses. Secondly, a preset fluctuation tolerance threshold is introduced. By comparing operator pairs and Perform a value determination and generate a Boolean trigger command based on the determination result. : Command activation ( ):when If the current physical characteristics of the material demand have been determined to have undergone "concept drift," meaning that the original parameters can no longer accurately fit the current random distribution, then a "parameter update" interrupt command is sent to the background task scheduler to force the subsequent retraining calculation program to start.

[0017] Instruction suppression ( ):when When the material demand pattern is determined to be within the existing stable fluctuation range, the model does not need to perform time-consuming calculations and updates, and can maintain the current logical operation state.

[0018] Based on this logic, an adaptive early warning mechanism driven by data features was established, which can accurately identify the key nodes in the transformation of material demand patterns from "low-frequency and stable" to "intense pulse".

[0019] S300: Based on the result of the trigger determination, determine the current parameter status of each candidate prediction model in the candidate prediction model set; "Parameter update procedure" refers to the process of updating parameters for each candidate prediction model. internal coefficient set Execute the re-optimized numerical computation logic. Based on the instructions generated in step S200. Different parameter generation strategies are employed: (1) In In the "triggered" state, execute the parameter update procedure: Each model in the candidate prediction model set (such as the improved TSB model, the Croston model, etc.) is placed in a parameter undetermined state. Then, the historical sequence obtained in step S100 is used... As a training sample set, an objective function (LossFunction) is constructed, which measures the cumulative deviation between the model's predicted values ​​and historical actual values ​​(e.g., using mean squared absolute scaling error).

[0020] By executing numerical optimization algorithms (such as gradient descent, heuristic search, or grid search), the optimal parameter combination that minimizes the objective function is iteratively searched within the parameter search space. Through this procedure, the coefficients within each model (such as coefficients reflecting demand frequency, coefficients reflecting demand scale, etc.) are reset to be relevant to the current sequence. The optimal estimate is used to form the current parameter state. .

[0021] (2) In In the "no trigger" state, the parameter inheritance strategy is executed: the optimal parameter combination used in the previous prediction cycle is directly retrieved from the preset parameter cache and used as the current execution parameter. .

[0022] (3) State determination and latching: To be determined (regardless of) still The parameters are mapped to the computational kernel of each candidate prediction model and used as immediate parameters for the subsequent S400 backtesting calculation.

[0023] The above procedure achieves a dynamic response of "updating on demand": when the demand pattern undergoes a qualitative change, new features are captured by reconstructing the objective function; when the demand pattern is stable, stability is ensured by inheriting historical parameters. This not only solves the problem of a step-like drop in accuracy caused by the "aging" of model parameters, but also avoids computational redundancy caused by frequent and unnecessary calculations.

[0024] S400: Using each candidate prediction model in its current parameter state, backtesting calculations are performed on the historical demand sequence to obtain multiple candidate prediction results. Using the model parameter snapshot locked in step S300, drive the set of candidate prediction models. Perform parallel backtesting calculations. The specific execution process is as follows: First, the length obtained in step S100 is Historical demand sequence According to the preset time cutoff point Divided into two mutually exclusive subsets: the training domain and the training domain. With the verification domain The length of the validation domain. The minimum statistical sample size requirement must be met to ensure the stability of subsequent bias calculations.

[0025] Secondly, each candidate prediction model based on After initializing the internal state, for each discrete time point within the verification domain... Perform simulations and predictions. For each time point... The model is only allowed to perceive Historical information is used to generate corresponding scalar prediction values. .

[0026] Finally, the predicted values ​​produced by each model in the validation domain are aligned along the time axis to generate the corresponding candidate prediction result sequence. .Depend on The sequences generated by the models together constitute the candidate prediction result matrix. .

[0027] Through the backtesting logic described above, the historical performance of the model in a real business environment was simulated, and the response differences of different algorithms when dealing with intermittent pulse demands were transformed into quantifiable data samples.

[0028] S500: Calculate the deviation between multiple candidate prediction results and the actual demand values ​​at corresponding time points in the historical demand sequence to obtain a basic score that characterizes the prediction accuracy and a ranking score that characterizes the stability of recent performance. Then, perform a weighted fusion of the basic score and the ranking score to determine the comprehensive score corresponding to each candidate prediction model. By comparing the candidate prediction matrices Compared with the actual value of the validation domain Perform multi-criteria performance quantization: (1) Basic score Determination of: For each model Calculate its predicted sequence With the actual sequence The entire validation domain length global deviation Subsequently, a pre-defined monotonically decreasing mapping function is used. (For example ,in (As a scaling factor), converting the global bias into a state where Basic score of the interval This score reflects the model's average fitting accuracy on the full backtest sample.

[0029] (2) Ranking and scoring Determination of: To evaluate the model's sensitivity to capturing recent demand fluctuations, extract the values ​​immediately adjacent to the current prediction time from the validation domain. One historical evaluation cycle ( In each evaluation cycle Internally, a horizontal comparison of the single-point absolute deviations generated by each model is performed. and according to Rank each model in ascending order. The smaller the deviation, the higher the ranking. This is achieved by aggregating these... The ranking performance over each period is used to calculate a ranking score that characterizes the stability of recent performance. .

[0030] (3) Overall score Fusion: Setting weight vectors Through weighted linear fusion formula The comprehensive score corresponding to each model is calculated.

[0031] This step establishes a two-dimensional evaluation model physically by decoupling "long-term global bias" from "short-term relative ranking." Its technical significance lies in its ability to effectively identify and filter out models that, while having low historical average errors, exhibit drastic fluctuations or lags in response to recent changes in the demand environment. This overcomes the "accuracy trap" common in intermittent demand forecasting and improves the timeliness of model optimization conclusions.

[0032] S600: Based on the comprehensive scores, determine the target prediction model from the candidate prediction models and output the candidate prediction results corresponding to the target prediction model.

[0033] Iterate through the set of comprehensive scores corresponding to each model. The index number of the optimal model is determined using the maximum value retrieval algorithm. ,in .

[0034] Determine the target prediction model Then, the candidate prediction result matrix generated in step S400 is... Extract the corresponding sequence The final output is a globally optimal estimate based on the current material variation coefficient, historical data fitting and verification, and recent stability game theory, ensuring a high degree of fit between the prediction and the dynamic physical characteristics of material demand.

[0035] This solution automatically triggers model parameter reconstruction and optimization based on the dynamic evolution of material consumption patterns. This mechanism effectively overcomes the prediction lag caused by model feedback blunting, ensuring a high degree of synchronization between prediction parameters and the current actual maintenance environment, and significantly enhancing the adaptive response sensitivity under changing operating conditions. By constructing a multi-dimensional weighted evaluation system that includes basic scores and ranking scores, it achieves a synergistic evaluation of the prediction model in both "numerical accuracy" and "execution stability." This solution breaks through the limitations of existing technologies that rely solely on full-cycle error indicators, effectively filtering out inferior models with drastic recent performance fluctuations that are prone to inducing erroneous replenishment signals using ranking scores. This dual evaluation mechanism significantly reduces the business execution risk caused by prediction signal distortion, ensuring that the final selected target prediction model not only has higher fitting accuracy but also better recent operational stability, providing a more valuable scientific benchmark for the precise replenishment and inventory optimization of train parts.

[0036] In an exemplary embodiment, during the execution of step S200 above, this embodiment implements a business-aware dynamic parameter update triggering mechanism through the following refined sub-steps: The ratio of the standard deviation to the mean of the historical demand sequence within a preset period is calculated to obtain the demand variation coefficient. If the demand variation coefficient exceeds a preset threshold, it is determined that the demand pattern of the target material has drifted, and the full parameter retraining logic of each candidate prediction model in the candidate prediction model set is initiated to execute the parameter update program.

[0037] After determining the dynamic judgment criteria, the historical demand sequence generated in step S100 is... Perform the following statistical calculations: Mean calculation: Calculates the arithmetic mean of the sequence over a preset observation period. .

[0038] Standard deviation calculation: Calculate the sample standard deviation of the sequence. It is used to measure the absolute strength of the deviation of the demand impulse from the mean.

[0039] Calculate the ratio of the two to obtain the dimensionless coefficient of variation of demand. This coefficient eliminates the influence of dimensions and can purely reflect the fluctuation characteristics of the material demand structure.

[0040] Then, the output will be generated in real time. The result calculated in step S100 Compare them.

[0041] The preset threshold is a target dynamic threshold determined based on the business attributes of the target material. The determination process includes: obtaining pre-labeled business attribute tags for the target material; wherein the business attribute tags include at least ABC classification levels representing the material's value dimension and XYZ classification levels representing the material's turnover frequency dimension; matching the corresponding sensitivity correction factor in a preset sensitivity correction matrix based on the combination of ABC and XYZ classification levels; and using the sensitivity correction factor to perform a weighted correction calculation on the initial preset threshold to obtain the target dynamic threshold. The specific implementation process is as follows: First, differentiated judgment boundaries are set based on the operational attributes of the target material in order to balance "computation cost" and "prediction accuracy".

[0042] Extract pre-labeled business attribute tags from the material master data in the ERP system.

[0043] ABC classification levels: representing value dimensions. For example, Class A materials (high value / critical spare parts) account for 10% of the total SKUs but 70% of the value, while Class C materials (low value / general-purpose parts) have a very low value share.

[0044] XYZ classification level: Represents the turnover frequency dimension. For example, X-type demand is stable and frequent, while Z-type demand is extremely sparse and random.

[0045] Preset one The sensitivity correction matrix is ​​shown in Table 1 below as an example logic. The corresponding sensitivity correction factor is retrieved based on the combined coordinates [ABC, XYZ]. .

[0046] Table 1

[0047] Using the retrieved factors Global initial preset threshold The weighted calculation is performed using the following formula: This yields the target dynamic threshold for that specific material. .

[0048] For "AZ" type materials, due to their high value and extremely poor regularity, a small correction factor (0.75) is applied, thereby significantly reducing the trigger threshold. This makes it more sensitive to subtle shifts in demand patterns; conversely, for "CX" type materials, a larger correction factor (1.5) is applied to raise the threshold and filter out invalid calculations.

[0049] like If the demand pattern of the target material is found to have shifted, computing resources will be immediately allocated to initiate the full parameter retraining logic.

[0050] In an exemplary embodiment, during the execution of step S400, this embodiment constructs a set of prediction models (Improved TSB, Crosston, and SBA) containing three heterogeneous logics to process the historical demand sequence of materials in parallel. For the first type of model (Improved TSB), the implementation process is as follows: (i) For the first type of candidate prediction model in the candidate prediction model set: identify the demand occurrence points from the historical demand sequence, extract the demand values ​​corresponding to each demand occurrence point as demand scale subsequences, and extract the Boolean states representing whether demand has occurred in the historical demand sequence as demand probability subsequences; use the first type of candidate prediction model to perform smoothing prediction processing on the demand scale subsequences and demand probability subsequences respectively to obtain the corresponding scale prediction components and probability prediction components; the product of the scale prediction components and the probability prediction components is determined as the first candidate prediction result.

[0051] First, from the historical sequence Identify the point where the demand occurs (i.e. (Time position). Extract the values ​​at the occurrence points in sequence to form a demand scale subsequence. Convert the demand status at each time point into Boolean values ​​(1 for occurrence, 0 for non-occurrence), forming a demand probability subsequence. .

[0052] The first type of candidate forecasting model includes the improved TSB model. For the first type of candidate forecasting model, the corresponding scale prediction component and probability prediction component are obtained, including: calculating the ratio of the average prediction error to the average absolute deviation of the first type of candidate forecasting model in the current forecasting period to obtain an adaptive tracking signal; dynamically adjusting the scale smoothing parameter of the first type of candidate forecasting model when smoothing the demand scale subsequence according to the numerical trend of the adaptive tracking signal; performing exponential smoothing on the demand scale subsequence using the adjusted scale smoothing parameter to obtain the corresponding scale prediction component; obtaining a preset decay weight and using the decay weight to perform subtraction update on the demand probability smoothing parameter in the first type of candidate forecasting model; and performing exponential smoothing on the demand probability subsequence using the updated demand probability smoothing parameter to obtain the corresponding probability prediction component.

[0053] For the first type of candidate forecasting model (improved TSB model), this embodiment achieves accurate characterization of intermittent demand by applying differentiated dynamic parameter interventions to the demand size component and the demand probability component. The specific implementation steps are as follows: 1. Scale and Components of Demand Adaptive computation By monitoring the deviation of the model prediction in real time, the scaling smoothing parameters are dynamically adjusted. The specific process includes: (1) Iterative update of statistics: Two error statistics indicators are maintained and updated in real time: average prediction error (Used to measure the direction of deviation) and mean absolute deviation (Used to measure the intensity of fluctuations). The calculation formula is as follows:

[0054] in, The difference between the actual demand and the predicted demand for the current forecast period; : Preset error smoothing factor (e.g. ); , : respectively The mean of the rolling prediction error and the mean of the rolling absolute deviation at each time point.

[0055] (2) Adaptive tracking signal Generation: Using the ratio of the above statistics, a dimensionless adaptive tracking signal is obtained. :

[0056] The range of values ​​is .like Approaching This indicates that the prediction error has a strong directionality (continuous overestimation or underestimation), and it is determined that the demand pattern has undergone a trend drift.

[0057] (3) Scale smoothing parameter Dynamic adjustment: based on The numerical trend is used to generate adjusted scale smoothing parameters in real time. .like If an upward trend is observed (e.g., exceeding the threshold of 0.6), the corresponding adjustment should be made. The magnitude of the weights is used to enhance the model's absorption of new trends; if If the price is in a low and stable period (e.g., below 0.2), adjust accordingly downwards. This is to filter out random noise and enhance the smoothness of the prediction curve.

[0058] (4) Scale component generation, using the adjusted For demand size subsequence Perform exponential smoothing to obtain the scale prediction component. .

[0059] 2. Demand probability components Attenuation calculation For the demand probability component, this embodiment introduces the natural decay characteristics of the material lifecycle, the specific process of which includes: (1) Parameter subtraction update logic: obtain a preset decay weight. (For example And use this weight to smooth the original demand probability parameters of the model. Perform a subtraction update:

[0060] in, : The updated probability smoothing parameters used for the current period; The decay step size represents the rate at which demand is "passively forgotten" over time. : The preset lower limit of parameters to prevent parameters from being reduced to an invalid range.

[0061] (2) Generation of probability components: using the updated parameters Demand probability subsequence Perform exponential smoothing to obtain the probability prediction components. :

[0062] Through the Perform subtraction to reduce demand when demand does not occur for an extended period (i.e., ), probability prediction value It will move towards at a faster speed The approach effectively simulates the physical process of materials "transforming from dynamic to static" and even being eliminated.

[0063] Then, the scale smoothing parameters of the candidate prediction models of the first type are dynamically adjusted when performing smoothing calculations on the demand scale subsequence. This includes: extracting the actual demand value corresponding to the most recent sampling period from the historical demand sequence; calculating the proportional relationship between the actual demand value and the historical mean calculated based on the historical demand sequence to obtain the burst pulse intensity; if the burst pulse intensity exceeds a preset outlier threshold, the actual demand value is determined to be an occasional large fluctuation, and the corresponding compression coefficient is determined according to the magnitude of the burst pulse intensity; wherein, the compression coefficient is inversely proportional to the burst pulse intensity; and using the compression coefficient to perform upper limit suppression processing on the scale smoothing parameters in the candidate prediction models of the first type, so that the adjusted scale smoothing parameters are in a preset low sensitivity range.

[0064] In the above-mentioned adaptive tracking signal Complete the initial adjustment of the scaling smoothing parameters (to obtain) Following this, this embodiment further introduces a burst pulse upper limit suppression mechanism to perform a physical-level "sensitivity circuit breaker" for irregular large orders. The specific process includes: (1) Burst pulse intensity Real-time computation: from historical demand sequences Extract the actual demand value corresponding to the most recent sampling period. And obtain the historical mean calculated based on the historical demand sequence. (For example: using the moving average of the past 24 sampling periods). The burst pulse intensity is calculated using the following formula. :

[0065] in, : Indicates the intensity of the sudden pulse, reflecting the multiple by which the latest demand deviates from the normal level; : Represents the current actual observation value in the sequence; : Indicates the expected demand level in a historical context.

[0066] (2) Randomness determination and compressibility coefficient Determination of: the calculated Compared with the preset outlier threshold (For example Compare them. If they satisfy... Determine the current This is characterized by occasional large fluctuations. In this case, the intensity of the sudden pulse is considered. The value determines the corresponding compression coefficient. The formula is as follows:

[0067] Among them, due to the compression coefficient With burst pulse intensity Inversely proportional, and at this time Larger, therefore The value of will be much smaller than (For example, when the intensity is 10 times, This creates a "damper" at the physical level to proportionally reduce the smoothing weights.

[0068] (3) Upper limit suppression of scale smoothing parameters: using compression coefficient For the aforementioned steps by Scale smoothing parameters obtained from signal conditioning Perform upper limit suppression processing to obtain the final smoothed parameters used for calculation in the current period. :

[0069] Through this multiplicative suppression operation, the adjusted Forced back and placed in a preset low sensitivity range (e.g.) If no burst pulse is detected (i.e.) Then the compression factor will be... Set as Maintain the original constant.

[0070] (4) Generation of scale components based on modified parameters: Finally, using the suppressed... For demand size subsequence Perform exponential smoothing to obtain the final scale prediction component. :

[0071] in, For the current period's scale prediction component (i.e. ), This represents the actual demand size for the current cycle. This is the historical forecast value from the previous period.

[0072] The probability prediction components obtained from the above calculations (See the aforementioned attenuation calculation section) and the scale prediction component after the above suppression processing. Perform multiplication and synthesis to determine the first candidate prediction result. :

[0073] This embodiment constructs a control system with both "dynamic following" and "outlier immunity" characteristics through nested logic of "TS adjustment" and "pulse suppression". When demand increases slowly, the TS signal is increased. To achieve rapid response; when demand surges suddenly (such as a large single order), the compression factor is adjusted. Rapid intervention achieves upper limit suppression. This dynamic game of physical parameters effectively solves the technical pain point of traditional TSB models when facing intermittent high-pulse materials, where the prediction curve is prone to violent oscillations and has difficulty returning to the baseline in the long term, significantly improving the robustness of inventory replenishment strategies.

[0074] "After completing the parallel computation of the first type of model (improved TSB), this embodiment of the invention simultaneously starts feature extraction and computation of the second type of model (Croston model). Unlike the first type of model, which continuously monitors the probability of demand occurrence, the second type of model focuses on the discrete time interval of demand occurrence to eliminate the interference of demand-free periods on prediction stability." Specifically, (ii) for the second type of candidate prediction model in the candidate prediction model set: extract the time span between two adjacent demand occurrence points in the historical demand sequence to obtain the demand occurrence interval sequence; use the second type of candidate prediction model to perform exponential smoothing prediction calculation on the demand values ​​in the historical demand sequence and the demand occurrence interval sequence to obtain the second candidate prediction result.

[0075] The second type of candidate forecasting model includes the Croston model. For the Croston model, this embodiment achieves balanced forecasting of intermittent demand by independently smoothing the demand quantity and demand intervals. The specific implementation steps are as follows: Historical demand sequence Perform feature deconstruction to generate two related subsequences: Demand numerical sequence Extract all actual required values ​​from the sequence. The sample points are denoted as .

[0076] Demand occurrence interval sequence Extract the time span between two adjacent demand occurrences. Specifically, if the first... This demand occurred At that moment, the This demand occurred At time, the corresponding interval sample This constitutes an interval sequence. .

[0077] The sequence reflects the "quantity" of materials each time they are requisitioned; The sequence then describes the average length of time the material is "resided" in the warehouse.

[0078] Then, using a preset smoothing parameter (For example Only during the demand occurrence cycle (i.e. Perform a synchronous smooth update on the two sequences mentioned above. The calculation formula is as follows: Demand numerical smoothing:

[0079] Demand interval smoothing:

[0080] in, : indicates up to the The smoothed demand size forecast at each demand occurrence point; : indicates up to the The smoothed average demand interval forecast at each demand occurrence point; : Represents the Croston smoothing coefficient, used to balance the weights of historical patterns and recent fluctuations.

[0081] The second candidate prediction result is obtained by performing a division operation using the smoothed predicted values ​​of the two dimensions mentioned above. :

[0082] This formula distributes the "single demand quantity" across the "average demand interval" at the physical level. For example, if the average demand is 10 units per instance, and it occurs on average every 5 cycles, then... Items / cycle. During the period of no demand between two points of demand occurrence. The value remains constant. This characteristic effectively avoids the drawback of traditional simple exponential smoothing (SES), which causes the predicted value to decay geometrically to zero during periods of no demand.

[0083] (iii) For the third type of candidate prediction model in the candidate prediction model set: obtain the second candidate prediction result, and use the preset deviation correction factor to perform down-adjustment calculation on the second candidate prediction result to obtain the third candidate prediction result.

[0084] The third type of candidate forecasting model includes the SBA model. For the SBA model, this embodiment aims to mathematically correct the original forecasts generated by the Croston model to prevent excessive inventory buildup when materials are extremely inactive. The specific implementation steps are as follows: First, retrieve the smoothing parameters used in the aforementioned second type of model (Croston) calculation process. Based on this parameter, a preset deviation correction factor is constructed. .

[0085] Calculation formula:

[0086] in, : refers to the common smoothing coefficients used for numerical and interval smoothing in the aforementioned Croston model (e.g. ); This is a mathematical correction operator for intermittent demand distribution characteristics. Mathematically, the Crosston model is suitable for extremely sparse demand (i.e., demand intervals are...). When the value is relatively large, the result of the division operation exhibits a statistically positive bias (i.e., the predicted value tends to be slightly higher than the actual mean). Introducing... This can effectively offset the bias, making the forecast results closer to the actual demand expectations.

[0087] Obtain the second candidate prediction result generated by the aforementioned second type of model. (Right now ), and utilize the deviation correction factor The downward adjustment calculation was performed to obtain the final third candidate prediction result. :

[0088] when At that time, the correction factor This means that the overall predictions from the Croston model have been lowered by 10%. (Smoothing coefficient) The larger the value (representing higher model sensitivity to recent fluctuations), the higher the correction factor. The smaller the value, the greater the downward adjustment. This logic ensures that the forecast remains prudent even in highly volatile scenarios.

[0089] "It is worth noting that, although" The calculation process depends on It represents an intermediate smoothing result, but physically it represents an 'asymmetric correction' to the intermittent demand distribution. In practical applications, when the demand interval sequence of materials... When it exhibits a long-tail distribution, This often leads to excessive recommendations for safety stock, while By adjusting the calculation downwards, it is possible to effectively suppress overestimation of the forecast. Therefore, It participates as an independent candidate component in the subsequent bias selection process, and... , Together, they form a predictive array covering all scenarios from trend to stable to sparse.

[0090] At this point, three candidate prediction results with different physical characteristics have been generated in parallel: (Improved TSB): It has the ability of "adaptive trend following" and "burst pulse suppression", and is suitable for materials with unstable demand intensity but obvious trend. (Croston): It has the ability of "interval-intensity decoupling", which is suitable for intermittent materials with stable demand frequency and no obvious decay trend; (SBA): In Based on this, a "prudent de-biasing" approach was implemented, making it suitable for materials with extremely sparse demand and high inventory holding costs.

[0091] Finally, the first candidate prediction result, the second candidate prediction result, and the third candidate prediction result are independently determined as the candidate prediction results corresponding to each candidate prediction model.

[0092] After completing the internal logic calculations of each candidate prediction model, this embodiment determines the calculated values ​​as the candidate prediction results for each model through the following steps, and constructs a heterogeneous prediction array set: The three values ​​generated in the above calculation process are mapped to their corresponding model IDs and stored independently in the candidate result buffer: Will pass through "adaptive tracking signal" "Adjustment" and "Sudden Pulse Upper Limit Suppression Coefficient" The value obtained after "correction" This is determined to be the final output of the improved TSB model. It carries the "probability of occurrence" characteristics of material demand and the sensitivity of "trend following".

[0093] The ratio obtained after "dual-track smoothing of demand value and occurrence interval" will be... This is determined to be the final output of the standard Croston model. It carries the characteristics of "average intensity" and "average cycle" of material demand, and has strong stability.

[0094] Based on "Bias correction factor" was applied The value obtained after "downward adjustment" This is determined to be the final output of the SBA model. It carries the characteristic of "prudent de-biasing" and is a conservative estimate for materials with long-tail distribution.

[0095] The above It is encapsulated as a three-dimensional prediction vector. Although The generation is reused in the cascaded architecture. Some intermediate variables (such as smoothing parameters) However, in the final determination stage, and These are considered as two competing solutions with completely independent physical meanings. This independent determination mechanism ensures that in the subsequent scoring process, the prediction bias of the three models can be calculated separately for the actual demand value at the same time point, thereby achieving "selecting the best from the best".

[0096] In one exemplary embodiment, after completing the candidate prediction result set... After encapsulation, this embodiment proceeds to step S500: the multi-model bias selection and scoring stage. The core logic of this stage is: instead of directly comparing the magnitude of the predicted values, it uses error fusion across three physical dimensions—probability, scale, and percentage—to perform a "health check" on the historical performance of each model, thereby selecting the most robust model under the current material characteristics. The following is a continuation of the previous in-depth embodiment: Specifically, this is achieved through the following sub-steps: Extracting the predicted demand probability for each prediction period from the candidate prediction results, and extracting the actual demand occurrence status for the corresponding prediction period from the historical demand sequence; where the actual demand occurrence status is a Boolean value; calculating the logical deviation between the predicted demand probability and the actual demand occurrence status to obtain the demand probability error; identifying demand sample points with actual demand values ​​greater than zero from the historical demand sequence, and calculating the magnitude deviation between the predicted value and the actual demand value at the demand sample points to obtain the demand scale error; calculating the percentage correction deviation of the predicted value relative to the actual demand value using a denominator correction algorithm; where the divisor of the denominator correction algorithm is the sum of the actual demand value and a preset smoothing constant to avoid interference from zero-value points in the historical demand sequence on the percentage calculation; performing a weighted summation calculation on the demand probability error, demand scale error, and percentage correction deviation, and performing inverse normalization on the summation result to obtain the basic score corresponding to each candidate prediction model.

[0097] First, evaluate the model's accuracy in predicting "when demand will occur": extract the corresponding predicted demand probabilities from each candidate prediction model (e.g., the output of the TSB model). Or the frequency of occurrence implicit in the Croston / SBA model From historical demand sequence The actual status of the demand for the corresponding period is extracted in the middle. If the actual demand value for this period ,but (Occurs); if ,but (Not occurred).

[0098] Logical deviation calculation:

[0099] This metric reflects the model's accuracy in capturing the rhythm of material "movement and stagnation." The smaller the error, the more sensitive the model is to judging the timing of replenishment.

[0100] Further evaluation of the model's accuracy in predicting "how much will occur" when demand arises: identifying only historical sequences. The demand sample points are used to exclude the interference of zero-value points and focus on the demand scale itself. Predicted values ​​are calculated for the identified sample points. Compared with actual demand value absolute deviation between :

[0101] To address the "division-to-zero anomaly" or percentage explosion problem caused by frequent zero values ​​in intermittent demand sequences, this embodiment introduces a denominator correction algorithm: For each demand sample point in the historical demand sequence, the absolute deviation between the corresponding predicted value and the actual demand value is calculated; each absolute deviation is used as the numerator, and the sum of the corresponding actual demand value and a preset smoothing constant is used as the denominator, and a division operation is performed to obtain the single-point correction deviation; wherein, the preset smoothing constant is a constant value greater than zero, used to maintain the numerical stability of the single-point correction deviation when the actual demand value is extremely small; the average value calculation is performed on all single-point correction deviations within a preset statistical interval to obtain the corrected percentage deviation.

[0102] Further utilizing the aforementioned "single-point absolute deviation" as the numerator, a denominator correction operation is performed: for each demand sample point in the historical demand sequence, the single-point absolute deviation is... As a molecule, the corresponding actual demand value With preset smoothing constant ( The sum of the constant values ​​(as the denominator) is used as the denominator, and division is performed to obtain the single-point correction deviation. :

[0103] Correction deviations for all single points within the preset statistical interval Perform mean normalization to obtain the final corrected percentage deviation. :

[0104] By uniformly using single-point absolute deviation as the numerator, it is ensured that and Consistency on the logical foundation, while avoiding percentage explosion caused by zero points by modifying the denominator.

[0105] Then, the errors from the three dimensions mentioned above are fused together to obtain the final base score. Weighted Accumulation: Based on the business's preference for "accurate timing" or "accurate quantity," a weight vector is set. Calculate the overall error Denormalization (calculating the base score): To ensure the scoring results conform to the intuitive logic of "higher score, better performance," denormalization is performed.

[0106] When the comprehensive error When approaching 0, Approaching 1 (full marks); the larger the error, The smaller the value, the better. This logarithmic inverse normalization can compress the influence of large biases and improve the robustness of the scoring.

[0107] In an exemplary embodiment, this embodiment captures the "recent state" of the candidate prediction model through step S600: stability ranking score calculation, specifically including the following detailed steps: First, from the multiple deviation results corresponding to each candidate prediction model obtained by the execution deviation calculation, extract the historical prediction deviation generated by each candidate prediction model at N consecutive corresponding time points. Here, each corresponding time point is defined as a historical evaluation period, and N is a positive integer; the candidate prediction models are retrieved from the memory cache in continuous... The deviation results generated within each historical evaluation period. To maintain logical consistency, this embodiment prioritizes extracting the aforementioned single-point correction deviations. (Right now This serves as the benchmark physical quantity for ranking. Each historical evaluation period corresponds to a time step of an actual demand point (e.g., the past 12 consecutive weeks).

[0108] Secondly, for each of the N consecutive historical evaluation periods, the magnitude of the historical prediction deviation is sorted among the candidate prediction models, and the period ranking score is assigned to each candidate prediction model in the historical evaluation period according to the ranking result. Among them, the smaller the historical prediction deviation, the better the corresponding cycle ranking score; for Each evaluation cycle in the evaluation cycle ( A "horizontal comparison" is performed among all candidate predictive models during this period. Within, compare the single-point correction biases generated by each model. . The smaller the value, the more accurate the model's prediction at that specific point in time. Let the total number of models be [value missing]. Based on the ranking results, period rankings are assigned to each model. The model with the smallest bias (ranked 1st) receives the highest ranking score (e.g., (score); largest deviation (ranked) The model obtains the lowest possible ranking score (e.g., 1 point). If the ranking in this period is... , This ranking mechanism shields the impact of differences in demand bases across different cycles, focusing only on the relative merits of the models, thus exhibiting strong robustness.

[0109] Then, based on the time span of each historical evaluation period and the current period to be predicted, time decay weights are matched to the corresponding period rankings. The shorter the time span, the greater the corresponding time decay weight; to reflect that "recent performance is more important than long-term performance," a time decay weight is set for each period. Calculate the evaluation cycle Time span between the current forecast period (i.e., the target forecast date) and the current forecast period. . The smaller (i.e., the closer to the present) The larger the value, the more likely it is to decay. This embodiment employs the following exponential decay model or linear decay model:

[0110] in This is the preset attenuation factor. (Through...) This forces the model to perform well in the most recent period in order to obtain a high score, thereby effectively identifying the "recent decline" in model performance.

[0111] Finally, a weighted summation calculation is performed on the ranking scores for each of the N historical evaluation periods and the corresponding time decay weights to obtain the ranking score. The ranking scores within each evaluation period are weighted and summed with their corresponding weights to obtain the final ranking score.

[0112] This score will serve as an important component of the "multidimensional bias score," and together with the aforementioned baseline score, will determine the final model selection result.

[0113] Based on the same inventive concept, this application also provides a multi-scheme adaptive optimal prediction device for low-frequency material demand, used to implement the multi-scheme adaptive optimal prediction method for low-frequency material demand described above. 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 multi-scheme adaptive optimal prediction device for low-frequency material demand provided below can be found in the limitations of the multi-scheme adaptive optimal prediction method for low-frequency material demand described above, and will not be repeated here.

[0114] like Figure 2 As shown, this embodiment provides a multi-scheme adaptive optimization prediction device for low-frequency material demand. The device includes: Sequence acquisition module 11 is used to acquire the historical demand sequence of the target material; The trigger determination module 12 is used to calculate the demand variation coefficient based on the historical demand sequence, and determine whether to trigger the execution of a parameter update program for the preset set of candidate prediction models based on the relationship between the demand variation coefficient and the preset threshold. The parameter management module 13 is used to determine the current parameter status of each of the candidate prediction models in the candidate prediction model set based on the result of the trigger determination. Backtesting calculation module 14 is used to perform backtesting calculations on the historical demand sequence using each of the candidate prediction models in the current parameter state, and obtain multiple candidate prediction results. The multidimensional scoring module 15 is used to perform deviation calculation between the multiple candidate prediction results and the actual demand values ​​at corresponding time points in the historical demand sequence to obtain a basic score representing the prediction accuracy and a ranking score representing the stability of recent performance. The basic score and the ranking score are then weighted and fused to determine the comprehensive score corresponding to each candidate prediction model. The model selection module 16 is used to determine the target prediction model from the candidate prediction models based on the comprehensive scores, and output the candidate prediction results corresponding to the target prediction model.

[0115] The modules in the aforementioned multi-scheme adaptive optimization prediction device for low-frequency material demand 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 memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0116] 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 multi-scheme adaptive optimal prediction method for low-frequency material demand.

[0117] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the above-described multi-scheme adaptive optimal prediction method for low-frequency material demand.

[0118] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the aforementioned multi-scheme adaptive optimal prediction method for low-frequency material demand.

[0119] 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.

[0120] 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.

[0121] 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 multi-scheme adaptive optimal prediction method for low-frequency material demand, characterized in that, The method includes: Obtain the historical demand sequence for the target material; The demand variation coefficient is calculated based on the historical demand sequence, and the relationship between the demand variation coefficient and a preset threshold is used to determine whether to trigger a parameter update procedure for the preset set of candidate prediction models. Based on the result of the trigger determination, the current parameter state of each of the candidate prediction models in the candidate prediction model set is determined; Using each of the candidate prediction models in the current parameter state, backtesting calculations are performed on the historical demand sequence to obtain multiple corresponding candidate prediction results; By performing deviation calculations between the multiple candidate prediction results and the actual demand values ​​at corresponding time points in the historical demand sequence, a basic score representing prediction accuracy and a ranking score representing recent performance stability are obtained. The basic score and the ranking score are then weighted and fused to determine the comprehensive score corresponding to each candidate prediction model. Based on the comprehensive scores, a target prediction model is determined from the candidate prediction models, and the candidate prediction results corresponding to the target prediction model are output.

2. The method according to claim 1, characterized in that, The step of determining whether to trigger a parameter update procedure for a preset set of candidate prediction models includes: The ratio of the standard deviation to the mean of the historical demand sequence within a preset period is calculated to obtain the demand variation coefficient; If the demand variation coefficient exceeds a preset threshold, it is determined that the demand pattern of the target material has drifted, and the full parameter retraining logic of each candidate prediction model in the candidate prediction model set is initiated to execute the parameter update program.

3. The method according to claim 2, characterized in that, The preset threshold is a target dynamic threshold determined based on the business attributes of the target material, and its determination process includes: Obtain pre-labeled business attribute tags for the target material; wherein, the business attribute tags include at least ABC classification levels representing the material value dimension and XYZ classification levels representing the material turnover frequency dimension; Based on the combination of the ABC classification level and the XYZ classification level, the corresponding sensitivity correction factor is matched in the preset sensitivity correction matrix; The initial preset threshold is weighted and corrected using the sensitivity correction factor to obtain the target dynamic threshold.

4. The method according to claim 1, characterized in that, The prediction calculations are performed on the historical demand sequences respectively to obtain multiple candidate prediction results, including: For the first type of candidate prediction model in the set of candidate prediction models: identify demand occurrence points from the historical demand sequence, extract the demand values ​​corresponding to each demand occurrence point as a demand scale subsequence, and extract the Boolean states representing whether demand has occurred in the historical demand sequence as a demand probability subsequence; use the first type of candidate prediction model to perform smoothing prediction processing on the demand scale subsequence and the demand probability subsequence respectively to obtain the corresponding scale prediction component and probability prediction component; the product of the scale prediction component and the probability prediction component is determined as the first candidate prediction result; wherein, the first type of candidate prediction model includes the improved TSB model; For the second type of candidate prediction model in the candidate prediction model set: extract the time span between two adjacent demand occurrence points in the historical demand sequence to obtain the demand occurrence interval sequence; use the second type of candidate prediction model to perform exponential smoothing prediction calculation on the demand values ​​in the historical demand sequence and the demand occurrence interval sequence to obtain the second candidate prediction result; wherein, the second type of candidate prediction model includes the Croston model. For the third type of candidate prediction model in the set of candidate prediction models: obtain the second candidate prediction result, and perform a down-adjustment calculation on the second candidate prediction result using a preset bias correction factor to obtain the third candidate prediction result; wherein, the third type of candidate prediction model includes the SBA model; The first candidate prediction result, the second candidate prediction result, and the third candidate prediction result are each independently determined as the candidate prediction result corresponding to each of the candidate prediction models.

5. The method according to claim 4, characterized in that, For the candidate prediction model of the first type, obtaining the corresponding scale prediction component and probability prediction component includes: Calculate the ratio of the average prediction error to the average absolute deviation of the candidate prediction model of the first type in the current prediction period to obtain the adaptive tracking signal; Based on the numerical change trend of the adaptive tracking signal, the scale smoothing parameter is dynamically adjusted when the first type of candidate prediction model performs smoothing calculation on the demand scale subsequence. The demand size subsequence is subjected to exponential smoothing using the adjusted size smoothing parameter to obtain the corresponding size prediction component; Obtain a preset decay weight, and use the decay weight to perform a subtraction update on the demand probability smoothing parameter in the candidate prediction model of the first type; The demand probability subsequence is subjected to exponential smoothing using the updated demand probability smoothing parameters to obtain the corresponding probability prediction components.

6. The method according to claim 5, characterized in that, The dynamic adjustment of the scale smoothing parameters when the first type of candidate prediction model performs smoothing calculations on the demand scale subsequence includes: Extract the actual demand value corresponding to the most recent sampling period from the historical demand sequence; The ratio between the actual demand value and the historical average value calculated based on the historical demand sequence is used to obtain the burst pulse intensity. If the burst pulse intensity exceeds a preset outlier threshold, the actual demand value is determined to be an occasional large fluctuation, and the corresponding compression coefficient is determined based on the magnitude of the burst pulse intensity; wherein, the compression coefficient is inversely proportional to the burst pulse intensity. The compression coefficient is used to perform upper limit suppression on the scale smoothing parameter in the candidate prediction model of the first type, so that the adjusted scale smoothing parameter is in a preset low sensitivity range. Specifically, the scale smoothing parameter, which is in the low sensitivity range, is used to perform smoothing calculations on the demand scale subsequence extracted from the historical demand sequence, so as to suppress the excessive shift of the scale prediction component towards the direction of the burst pulse intensity.

7. The method according to claim 1, characterized in that, The step of calculating the deviation between the multiple candidate prediction results and the actual demand values ​​at corresponding time points in the historical demand sequence to obtain a basic score characterizing the prediction accuracy includes: Extract the predicted demand probability corresponding to each prediction period from the candidate prediction results, and extract the actual demand occurrence state corresponding to the prediction period from the historical demand sequence; wherein, the actual demand occurrence state is a Boolean value; The logical deviation between the predicted demand probability and the actual state of demand occurrence is calculated to obtain the demand probability error. From the historical demand sequence, identify demand sample points where the actual demand value is greater than zero, and calculate the magnitude deviation between the predicted value of the candidate prediction result and the actual demand value at the demand sample points to obtain the demand scale error. The percentage deviation of the predicted value relative to the actual demand value is calculated using a denominator correction algorithm; wherein the divisor of the denominator correction algorithm is the sum of the actual demand value and a preset smoothing constant, so as to avoid the interference of zero points in the historical demand sequence on the percentage calculation; The demand probability error, the demand size error, and the correction percentage deviation are weighted and accumulated, and the accumulated result is inversely normalized to obtain the basic score corresponding to each candidate prediction model.

8. The method according to claim 7, characterized in that, The step of calculating the percentage deviation of the predicted value relative to the actual demand value using the denominator correction algorithm includes: For each demand sample point in the historical demand sequence, calculate the absolute deviation between the corresponding predicted value and the actual demand value; The absolute deviations are used as numerators, and the sum of the corresponding actual demand value and the preset smoothing constant is used as the denominator. A division operation is performed to obtain the single-point correction deviation. The preset smoothing constant is a constant value greater than zero, which is used to maintain the numerical stability of the single-point correction deviation when the actual demand value is extremely small. The mean value is calculated for all the single-point correction deviations within the preset statistical interval to obtain the correction percentage deviation.

9. The method according to claim 1, characterized in that, By calculating the deviation between the multiple candidate prediction results and the actual demand values ​​at corresponding time points in the historical demand sequence, a ranking score characterizing the stability of recent performance is obtained, including: From the multiple deviation results corresponding to each candidate prediction model obtained by performing the deviation calculation, extract the historical prediction deviation generated by each candidate prediction model at N consecutive corresponding time points; wherein, each corresponding time point is defined as a historical evaluation period, and N is a positive integer; For each of the N consecutive historical evaluation periods, the candidate prediction models are sorted according to the magnitude of the historical prediction deviation, and each candidate prediction model is assigned a period ranking score in that historical evaluation period based on the ranking result; wherein, the smaller the historical prediction deviation, the better the corresponding period ranking score. Based on the time span between each historical evaluation period and the current period to be predicted, a time decay weight is assigned to the corresponding period ranking; wherein, the smaller the time span, the greater the corresponding time decay weight. The ranking score is obtained by performing a weighted summation calculation on the ranking score of each of the N historical evaluation periods and the corresponding time decay weight.

10. A multi-scheme adaptive optimal prediction device for low-frequency material demand, characterized in that, The device includes: The sequence acquisition module is used to acquire the historical demand sequence of the target material; The trigger determination module is used to calculate the demand variation coefficient based on the historical demand sequence, and determine whether to trigger the execution of a parameter update program for a preset set of candidate prediction models based on the relationship between the demand variation coefficient and a preset threshold. The parameter management module is used to determine the current parameter status of each of the candidate prediction models in the candidate prediction model set based on the result of the trigger determination. The backtesting calculation module is used to perform backtesting calculations on the historical demand sequence using each of the candidate prediction models in the current parameter state, to obtain multiple corresponding candidate prediction results; The multidimensional scoring module is used to perform deviation calculations between the multiple candidate prediction results and the actual demand values ​​at corresponding time points in the historical demand sequence to obtain a basic score that characterizes the prediction accuracy and a ranking score that characterizes the stability of recent performance. The basic score and the ranking score are then weighted and fused to determine the comprehensive score corresponding to each candidate prediction model. The model selection module is used to determine the target prediction model from the candidate prediction models based on the comprehensive scores, and output the candidate prediction results corresponding to the target prediction model.

Images