An automobile accessory periodic sales prediction method, device, equipment and medium

Abstract

The application provides an automobile accessory periodic sales prediction method, device, equipment and medium, the method comprises the following steps: obtaining the historical periodic sales data of a target automobile accessory, and the basic prediction results generated by a plurality of basic prediction models according to the historical periodic sales data; according to the prediction step and the scene characteristics, the local prediction ability of each basic prediction model is evaluated; according to the basic prediction results of each basic prediction model, the multidimensional prediction divergence between the models is calculated; according to the local prediction ability of each basic prediction model itself, the multidimensional prediction divergence is conditionally punished and corrected to obtain the model credibility of each basic prediction model; the fusion weight is generated according to the model credibility, and the basic prediction results are weighted according to the fusion weight, and the final prediction result of the periodic sales is output. The application effectively improves the accuracy and stability of the periodic sales prediction of automobile accessories.

Description

Technical Field

[0001] This application relates to the field of automotive parts sales forecasting technology, and in particular to a method, device, equipment and medium for forecasting the cyclical sales of automotive parts. Background Technology

[0002] Automotive parts sales forecasting is a crucial aspect of automotive aftermarket supply chain management. Within the automotive parts supply chain, there is a wide variety of parts SKUs, and the monthly sales patterns of different SKUs vary significantly. Within the same market and channel, there are consistently high-selling SKUs, as well as SKUs with low-frequency sales, intermittent sales, and pronounced seasonality, and also new SKUs with a limited number of historically effective months.

[0003] Currently, monthly sales forecasts for auto parts typically employ single forecasting models such as time series models, exponential smoothing models, or machine learning models. However, due to significant differences in sales characteristics across different SKUs, a single model struggles to maintain stable forecast accuracy across all SKU types, often resulting in substantial forecast deviations for some SKUs. To overcome the limitations of single models, existing technologies have developed solutions that utilize multi-model fusion for sales forecasting.

[0004] Existing multi-model fusion schemes typically determine the fusion weights of each model in the following ways: First, they use the same fixed weights for all SKUs and all predicted months; second, they set the weights based on the global average error of each model across all historical samples; and third, when the prediction results of multiple models are inconsistent, they directly perform a simple average of the models.

[0005] The above method has the following problems in practical applications: First, the predictive capabilities of the same basic forecasting model vary across different forecasting steps. Monthly forecasts typically require outputting sales values ​​for multiple future calendar months simultaneously. Some models respond promptly to sales changes in recent months but tend to over-expose for distant months; others are more stable in grasping long-term seasonal structures but react more slowly to recent abrupt changes. If the same set of fixed weights is used for all future months, it cannot accommodate the differences in model capabilities across different forecasting steps.

[0006] Second, the weighting method based on the global average error is difficult to adapt to the local prediction scenarios of different SKUs. Setting weights based on the average error of all SKUs or all months may result in models with good overall performance still having a high weight for specific SKU types, while the predictive value of models that only perform well in scenarios with strong seasonality, low frequency, or recent trend changes is underestimated.

[0007] Third, when simply averaging the predictions of multiple models, the prediction bias of models that have failed in the current scenario can easily be introduced into the final result. When multiple basic prediction models produce significantly different sales volumes, month-on-month trends, or future multi-month trends for the same target month, simple averaging or fixed weights cannot identify whether a particular model has abnormally overestimated, underestimated, or failed trend extrapolation in the current scenario.

[0008] Fourth, simply reducing the weight of a model based on whether it deviates from the majority of results fails to distinguish whether the deviation is due to model failure or the model's greater sensitivity to real trend changes or seasonal inflection points. When sales experience real trend changes or seasonal inflection points, models with higher predictive power may give different predictions before other models. If a model is removed or reduced in weight simply because it deviates from the majority of results, a valid predictive signal will be lost.

[0009] Fifth, the historical effective months for new product SKUs or low-frequency SKUs are limited, making it impossible to provide sufficient rolling backtesting samples to evaluate the prediction errors of each model. Directly calculating model errors based on a small number of local samples can easily lead to abnormal concentration of model weights due to occasional fluctuations in a single month, affecting the stability of the prediction results.

[0010] In summary, existing monthly sales forecasts for auto parts suffer from low accuracy and poor stability. Summary of the Invention

[0011] To address the aforementioned problems in the prior art, this application provides a method, apparatus, device, and medium for predicting the cycle sales volume of automotive parts, which can improve the accuracy and stability of predicting the cycle sales volume of automotive parts.

[0012] To achieve the above objectives, the technical solution adopted in this application is as follows: Firstly, this application provides a method for forecasting the cyclical sales volume of automotive parts, including: Obtain historical sales data for the target auto parts, as well as basic prediction results generated by multiple basic prediction models based on the historical sales data; The local prediction capability of each basic prediction model is evaluated based on the prediction step size and scene characteristics. When there are insufficient local samples, a hierarchical grouping error with the same prediction step size is introduced to perform hierarchical shrinkage to obtain the local prediction capability. Based on the basic prediction results of each basic prediction model, calculate the multi-dimensional prediction discrepancies between the models. Based on the local prediction capabilities of each basic prediction model, conditional penalty correction is applied to the multi-dimensional prediction discrepancies to obtain the model credibility of each basic prediction model. Based on the model's credibility, a fusion weight is generated, and the basic prediction results are weighted according to the fusion weight to output the final prediction result of the periodic sales.

[0013] The beneficial effects of this application are as follows: By acquiring historical periodic sales data of target auto parts and basic prediction results generated by multiple basic prediction models, the local prediction capabilities of each basic prediction model are evaluated based on the prediction step size and scenario characteristics. This enables the model capability evaluation to be accurately aligned with the timeliness and scenario characteristics of the current prediction task. On this basis, by calculating the multi-dimensional prediction discrepancies between models and using the local prediction capabilities of each basic prediction model to perform conditional penalty correction on the multi-dimensional prediction discrepancies, the model with high local prediction capability receives a weaker penalty under the same discrepancy, while the model with low local prediction capability receives a stronger penalty under the same discrepancy. This retains the effective prediction signal of high-capability models while suppressing the abnormal deviation of low-capability models. Finally, fusion weights are generated based on model credibility and the final prediction result is output in a weighted manner. This achieves differentiated adaptive fusion for different SKUs and different prediction periods, effectively improving the accuracy and stability of periodic sales prediction for auto parts.

[0014] Optionally, the step of calculating the multi-dimensional prediction discrepancy between models based on the basic prediction results of each basic prediction model includes: For any two basic prediction models, calculate the magnitude divergence, variation divergence, and continuous path divergence respectively. The comprehensive divergence is obtained by weighting and combining the magnitude divergence, the change divergence, and the continuous path divergence.

[0015] As described above, by mapping the three divergences to a preset numerical range and then weighting and combining them to obtain a comprehensive divergence, the prediction differences between models can be fully characterized from multiple dimensions such as sales volume, direction of change, and trend, providing a more accurate basis for subsequent conditional penalty correction.

[0016] Optionally, the step of applying conditional penalty correction to the multi-dimensional prediction discrepancies based on the local prediction capabilities of each basic prediction model to obtain the model credibility of each basic prediction model includes: For each target automotive part SKU and each target prediction period, based on the local prediction capabilities of other basic prediction models, the overall divergence between the current basic prediction model and other basic prediction models is weighted to obtain the reliable model divergence strength: ; In the formula, A i R represents the reliability model divergence strength of the i-th basic prediction model. j For the local predictive capability of the j-th basic prediction model, D ijThe term represents the overall divergence, where ε is a smoothing constant. The conditional penalty Q is obtained by controlling the penalty intensity generated by equal divergence based on the local prediction capability of the current basic prediction model. i : Q i =(1-R i ) r A i ; In the formula, r is the reliability adjustment index; Based on the local prediction capability of the current basic prediction model and the conditional penalty, the model confidence level U is generated. i : U i =R i exp(-λQ i ); In the formula, λ is the divergence penalty intensity parameter.

[0017] As described above, the conditional penalty correction mechanism, which consists of the reliable model divergence strength formula, the conditional penalty formula, and the model credibility formula, achieves the technical effect of controlling the divergence penalty intensity by utilizing the model's own local predictive ability. This results in models with high local predictive ability receiving weaker penalties under the same divergence, while models with low local predictive ability receive stronger penalties under the same divergence. Thus, while retaining the effective prediction signals of a few high-ability models, it suppresses low-ability models that deviate significantly.

[0018] Optionally, evaluating the local prediction capability of each basic prediction model based on the prediction step size and the scene features includes: Construct a historical rolling sample based on the historical sales data of the aforementioned period; Based on the scene characteristics, similar samples are retrieved from the historical rolling samples. For each basic prediction model, an effective sample set is formed, and the local error and effective sample weight of each basic prediction model are calculated. When the effective sample weights are lower than a preset sample threshold, the hierarchical grouping error with the same prediction step size is obtained, and hierarchical shrinkage is performed based on the effective sample weights and the hierarchical grouping error to obtain the shrinkage error. : ; In the formula, n i For the first Effective sample weights for each basic prediction model Let κ be the local error, and κ be the shrinkage parameter. The hierarchical grouping error; The local prediction capability R is obtained based on the shrinkage error transformation. i ; ; In the formula, γ is the capacity conversion parameter.

[0019] As described above, by retrieving similar samples based on scene features, forming an effective sample set for each basic prediction model, and introducing hierarchical grouping error with the same prediction step size to shrink the hierarchy when the weight of the effective sample is lower than the preset sample threshold, the local prediction capability is obtained. This allows the evaluation of the local prediction capability to reflect the local prediction scene features of the current SKU, and to provide a stable estimation reference through hierarchical grouping error when there are insufficient samples.

[0020] Optionally, the step of retrieving similar samples from the historical rolling samples based on the scene features, forming an effective sample set for each basic prediction model, and calculating the local error and effective sample weights of each basic prediction model includes: Based on the distance between the scene features of the target auto parts and the scene features of the historical rolling samples, as well as the difference in the prediction step size, the sample weight is calculated, and similar samples are selected from the historical rolling samples. For each basic prediction model, similar samples with missing predictions are removed to form the effective sample set corresponding to each basic prediction model; Based on the effective sample set and the sample weights, calculate the local error and effective sample weights of each basic prediction model.

[0021] Optionally, obtaining the hierarchical grouping error with the same prediction step size includes: Following the order from fine-grained to coarse-grained levels, it is determined whether the current level meets the minimum sample size threshold. If it does, the grouping error corresponding to the current level is taken as the level grouping error. Otherwise, it is backed up to the next coarse-grained level for judgment until a level grouping error that meets the minimum sample size threshold is obtained.

[0022] Optionally, the prediction step size is calculated using the cross-period sequence difference, and the scenario features include the proportion of non-zero data, data volatility, recent change characteristics, and periodic characteristics.

[0023] Secondly, this application provides a device for predicting the cycle sales volume of automotive parts, comprising: The data acquisition module is used to acquire historical period sales data of the target auto parts, as well as basic prediction results generated by multiple basic prediction models based on the historical period sales data. The capability assessment module is used to evaluate the local prediction capability of each basic prediction model based on the prediction step size and scene characteristics. When there are insufficient local samples, a hierarchical grouping error with the same prediction step size is introduced to perform hierarchical shrinkage to obtain the local prediction capability. The divergence calculation module is used to calculate the multi-dimensional prediction divergence between models based on the basic prediction results of each basic prediction model. The divergence correction module is used to perform conditional penalty correction on the multi-dimensional prediction divergence based on the local prediction capabilities of each basic prediction model, so as to obtain the model credibility of each basic prediction model. The fusion output module is used to generate fusion weights based on the model's credibility, and to weight the basic prediction results based on the fusion weights to output the final prediction result of the periodic sales.

[0024] Thirdly, this application provides a computer device, including: one or more processors, and a memory; the memory stores computer-readable instructions, which, when executed by the one or more processors, perform the steps of the method described in the first aspect.

[0025] Fourthly, this application provides a storage medium storing computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the method described in the first aspect.

[0026] The technical effects of the automotive parts cycle sales forecasting device provided in the second aspect, the computer equipment provided in the third aspect, and the storage medium provided in the fourth aspect are described in the relevant description of the automotive parts cycle sales forecasting method provided in the first aspect. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the main process of a method for predicting the cyclical sales volume of automotive parts according to an embodiment of this application; Figure 2 This is a schematic diagram of the structure of an automotive parts cycle sales forecasting device according to an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a computer device according to an embodiment of this application. Detailed Implementation

[0028] To better understand the above technical solutions, exemplary embodiments of this application will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application can be understood more clearly and thoroughly, and that the scope of this application can be fully conveyed to those skilled in the art.

[0029] The embodiments of this application are applied to the scenario of automotive parts sales forecasting, and are particularly suitable for automotive parts sales forecasting scenarios with multiple SKUs (Stock Keeping Units) and multiple forecasting periods. Existing multi-model fusion methods suffer from problems such as unreasonable weight allocation, high-capacity models being easily downgraded due to errors, and low-capacity models causing abnormal predictions that can skew the final result, resulting in insufficient accuracy and stability of sales forecasting.

[0030] To this end, in various embodiments of this application, historical periodic sales data of the target auto parts are obtained, along with basic prediction results generated by multiple basic prediction models based on the historical periodic sales data. The local prediction capability of each basic prediction model is evaluated based on the prediction step size and scenario characteristics. Specifically, when local samples are insufficient, hierarchical grouping errors with the same prediction step size are introduced for hierarchical contraction to obtain local prediction capability. Based on the basic prediction results of each basic prediction model, multi-dimensional prediction discrepancies between models are calculated. Conditional penalty correction is applied to the multi-dimensional prediction discrepancies based on the local prediction capability of each basic prediction model to obtain the model credibility of each basic prediction model. Fusion weights are generated based on the model credibility, and the basic prediction results are weighted according to the fusion weights to output the final prediction result of the periodic sales. Therefore, this application effectively improves the accuracy and stability of periodic sales prediction for auto parts.

[0031] The present application will now be described in further detail with reference to the accompanying drawings and embodiments.

[0032] This application provides a method for predicting the cycle sales volume of automotive parts, such as... Figure 1 As shown, the method includes: Step S1: Obtain historical sales data for the target auto parts, as well as basic prediction results generated by multiple basic prediction models based on the historical sales data.

[0033] In this embodiment, historical sales data is obtained from an enterprise resource planning system or data warehouse, with a monthly granularity. Specifically, for order-level, daily, or weekly sales records, the sales quantity of the same SKU is aggregated according to the natural month, and returns, cancellations, and write-offs are deducted to obtain the monthly net sales volume. Based on this, the forecast base month, future forecast window, target automotive parts SKU set, and market, channel, and category levels are pre-determined according to business needs. Historical months are continuously supplemented, and a valid observation identifier distinguishes between months with zero sales and months with missing data. Months before the forecast base month whose data has been completely aggregated constitute the set of historical valid months; months after the forecast base month and within the future forecast window constitute the set of future target months.

[0034] Next, establish no fewer than two basic forecasting models, each containing at least two different modeling mechanisms. These basic forecasting models may include exponential smoothing models, seasonal time series models, trend models, and nonlinear machine learning models; for intermittently sold SKUs, an intermittent demand model can also be added. Each basic forecasting model generates basic forecast values ​​for future target months based on the historical monthly sales series of the target SKU: ; In the formula, Y s Let i be the historical monthly sales sequence of the target SKU, s be the index of the base prediction model, t be the target SKU, and t be the target month. Model training and parameter determination use only data from the prediction base month and earlier.

[0035] To satisfy the non-negativity constraint on sales volume, if the output value of the basic forecasting model is less than zero, the output value will be set to zero; if the output value of the basic forecasting model is greater than or equal to zero, the original forecast value will be retained. If a basic forecasting model does not generate a valid result, it will not be included in the valid model set for the current SKU and the target month.

[0036] Step S2: Evaluate the local prediction capability of each basic prediction model based on the prediction step size and scene characteristics. When there are insufficient local samples, hierarchical grouping error with the same prediction step size is introduced to shrink the hierarchy and obtain the local prediction capability.

[0037] In this embodiment, the prediction step size is calculated using the cross-period sequence difference, and the scenario features include the proportion of non-zero data, data volatility, recent change characteristics, and periodic characteristics.

[0038] Specifically, the forecast step size refers to the difference in the number of months between the base month and the target month. For example, the forecast step size for the first month in the future is one. Since the forecast scenario may span multiple years, to avoid negative values ​​from December to January of the following year, the forecast step size is calculated using the difference in month numbers across years. The formula is as follows: ; In the formula, t is the target forecast month, and t0 is the forecast base month.

[0039] The scenario features are used to characterize the historical sales behavior patterns of the target SKU, including the proportion of non-zero months, sales fluctuations, recent changes, and seasonality. These features are calculated only based on the forecast base month and previous valid historical months. The formulas for calculating the proportion of non-zero months and the fluctuation coefficient are as follows: ; ; In the formula, z s H represents the proportion of non-zero months for the target SKU. sFor the set of historical valid months for the target SKU, y s,u Let |Hs| represent the monthly sales volume of the target SKU in historical month u, where |Hs| represents the total number of valid historical months, and the numerator is the number of valid historical months with sales greater than zero; v s std( ) represents the sales volatility coefficient of the target SKU. ) represents the standard deviation calculation, mean( ) represents the mean operation, and ε is a positive smoothing constant to prevent division by zero.

[0040] Among them, recent change features are calculated by the robust difference between the mean or median of the recent window and the previous window; seasonality features are calculated using 12th-order autocorrelation, seasonal decomposition strength, or functionally equivalent periodic indicators. Each feature is standardized according to the center and scale values ​​of the historical rolling samples before entering the subsequent similar sample retrieval step. It is important to note that the effective historical length is only used to determine the sufficiency of samples and is not included in the similarity vector; the prediction step size is used as an independent condition for sample selection and decay weights and is not involved in the similarity calculation of scene features.

[0041] In this embodiment, step S2, which evaluates the local prediction capability of each basic prediction model based on the prediction step size and scene features, includes: Step S21: Construct historical rolling samples based on historical periodic sales data.

[0042] Historical rolling samples refer to historical samples generated chronologically. For each historical forecast base month, only the data already obtained at that time is used to train or call a basic forecasting model to predict subsequent months. These samples, after actual sales are generated, contain predicted values, actual values, forecast step size, and monthly sales scenario characteristics. Specifically, multiple historical forecast base months are set in chronological order. In each historical base month, only the data already obtained at that time is used to call the same basic forecasting model as described above. After the sales of the historical target month are fully settled, a rolling sample is formed, containing the historical base month, target month, forecast step size, scenario characteristics at that time, model predicted values, and actual sales. Unsettled months cannot be used as training labels.

[0043] Step S22: Based on scene characteristics, retrieve similar samples from historical rolling samples, form an effective sample set for each basic prediction model, and calculate the local error and effective sample weight of each basic prediction model.

[0044] In this embodiment, step S22 includes: Step S221: Calculate sample weights based on the distance between the scene features of the target auto parts and the scene features of the historical rolling samples, as well as the difference in prediction step size, and select similar samples from the historical rolling samples.

[0045] In this embodiment, after obtaining historical rolling samples, samples similar to the current task are retrieved from them based on scene features. During retrieval, candidate samples are preferentially selected from historical tasks with the same prediction step size as the current task; if samples are insufficient, expansion to adjacent lead times is allowed. The similarity between the current scene and historical samples is measured by feature distance and sample weight. The formulas for calculating feature distance and sample weight are: ; ; In the formula, d n ω represents the feature distance between the current scene and historical sample n, p is the number of dimensions of the scene features, and ω is the feature distance between the current scene and historical sample n. k Let x be the weight of the k-th feature. k Let x be the value of the k-th feature in the current scene. n,k Let a be the value of the k-th feature in historical sample n. n Let τ be the sample weight of historical sample n, ρ be the distance decay parameter, ρ be the lead time decay parameter, and h be the sample weight of historical sample n. n h is the prediction step size for historical sample n. t This represents the prediction step size for the current task.

[0046] In this approach, the prediction step size is no longer included in the scene feature distance calculation. For each base prediction model, only the top few similar samples with valid historical predictions are retained to form the corresponding model-level sample set.

[0047] Therefore, the calculation of sample weights takes into account both the similarity of scene features and the closeness of prediction step sizes. Historical samples with more similar scene features and closer prediction step sizes have larger sample weights and contribute more to the evaluation of model capabilities in subsequent calculations.

[0048] Step S222: For each basic prediction model, remove the similar samples with missing predictions to form the effective sample set corresponding to each basic prediction model.

[0049] Different base prediction models may have varying historical success rates, and some models may not have produced valid predictions on certain historical samples. Therefore, after selecting similar samples, each base prediction model needs to be processed separately: similar samples with missing predictions for that model are removed, and only similar samples with valid predictions are retained, forming the valid sample set corresponding to that model. In this way, each base prediction model has its own independent valid sample set, avoiding capability estimation distortion caused by different models sharing the same error samples despite varying historical success rates.

[0050] Step S223: Calculate the local error and effective sample weight of each basic prediction model based on the effective sample set and sample weight.

[0051] After obtaining the effective sample set for each basic prediction model, the local error and effective sample weights of each model are calculated.

[0052] To balance high- and low-selling SKUs, and to avoid abnormally amplified relative errors when actual sales are close to zero, a smoothed relative error with a normal sales scale is adopted: ; In the formula, L i,n Let y be the smoothed relative error of the i-th basic prediction model on historical samples n. n For the actual sales volume of historical sample n, Let q be the predicted value of the i-th basic prediction model on historical sample n. s To ensure normal sales smoothing, the median of historical non-zero monthly sales for the target SKU is preferred, with ε being a positive smoothing constant to prevent division by zero.

[0053] Based on the smoothed relative error, the formulas for calculating the local error and effective sample weights of the model in the current scenario are as follows: ; ; In the formula, Let N be the local error of the i-th basic prediction model. i Let a be the effective sample set of the i-th basic prediction model. n L represents the sample weights of historical sample n. i,n Let n be the smoothed relative error of the i-th basic prediction model on historical samples n. i is the effective sample weight of the i-th basic prediction model, which is the sum of the weights of all samples in the effective sample set of the model.

[0054] Step S23: When the effective sample weights are lower than the preset sample threshold, obtain the hierarchical grouping error with the same prediction step size, and perform hierarchical shrinkage based on the effective sample weights and the hierarchical grouping error to obtain the shrinkage error. : ; In the formula, n i For the first Effective sample weights for each basic prediction model For local error, κ is the shrinkage parameter. This refers to the hierarchical grouping error.

[0055] Therefore, the larger the effective sample weight, the closer the shrinkage error is to the SKU local error; the smaller the effective sample weight, the greater the impact of the grouping error.

[0056] Specifically, when the effective sample weights are sufficient, the local error can reliably reflect the model's predictive ability in the current scenario. However, when the effective sample weights are lower than a preset sample threshold, relying solely on the local error may lead to unstable capability estimation. In this case, it is necessary to introduce hierarchical grouping errors with the same prediction step size for supplementation. In this embodiment, obtaining the hierarchical grouping error with the same prediction step size in step S23 includes: Following the order from fine-grained to coarse-grained levels, it is determined whether the current level meets the minimum sample size threshold. If it does, the grouping error corresponding to the current level is taken as the level grouping error. Otherwise, it is backed up to the next coarse-grained level for judgment until the level grouping error that meets the minimum sample size threshold is obtained.

[0057] The hierarchical grouping error is obtained as follows: Following the hierarchical structure of market, channel, and category, the minimum sample size threshold is checked sequentially from fine-grained to coarse-grained levels. Specifically, the hierarchical grouping error is first calculated at the market, channel, and category levels. If the number of valid samples at that level is insufficient, it is regressed to the market and channel levels; if still insufficient, it is regressed to the global level. If the sample size for the same prediction step is still insufficient, it can be extended to adjacent prediction steps, using the aforementioned lead time attenuation weight.

[0058] Step S24: Obtain the local prediction capability R based on the shrinkage error transformation. i ; ; In the formula, γ is the capacity conversion parameter.

[0059] Among them, monthly local forecasting capability refers to the reliability of the basic forecasting model on historical rolling samples that are similar to the monthly sales characteristics of the current target SKU and have the same or adjacent forecasting steps.

[0060] In addition, this embodiment also records the actual fallback level and number of samples used to ensure that the source of the model's capabilities is traceable.

[0061] Step S3: Calculate the multi-dimensional prediction discrepancies between models based on the basic prediction results of each basic prediction model.

[0062] In this embodiment, step S3 includes: Step S31: For any two basic prediction models, calculate the magnitude divergence, change divergence, and continuous path divergence respectively.

[0063] The magnitude divergence is used to characterize the difference in the models' predictions for the target month's sales volume. The greater the difference between the two models' predictions for the same month, the greater the magnitude divergence. The calculation formula is as follows: ; In the formula, Let $\mathbf{i}$ be the magnitude difference between the $i$-th and $j$-th basic prediction models. Let be the predicted value of the target SKU in the target month t for the i-th basic prediction model. Let be the predicted value of the target SKU in the target month t by the j-th basic prediction model, and c be the positive smoothing constant.

[0064] The monthly variation divergence is used to characterize the difference in the magnitude of changes between adjacent months. Specifically, for the first target month, the baseline for variation is the most recently settled actual sales volume; for subsequent months, the baseline is the previous forecast month from the same basic forecast model. The formulas for calculating the standardized monthly variation and variation divergence are as follows:

[0065] ; In the formula, g i For the standardized monthly change of the i-th basic forecast model, Let b be the predicted value of the i-th base prediction model for the current target month. i Let c be the monthly variation baseline value of the i-th basic forecast model, and c be the positive smoothing constant. For the divergence between the i-th and j-th basic prediction models, g i and g j These are the standardized monthly changes for the i-th and j-th basic forecast models, respectively.

[0066] The consecutive monthly path divergence is used to characterize the differences in the trends of future consecutive months. During calculation, the consecutive monthly paths are first scaled and normalized according to the mean of each model's own path. ; In the formula, P i,k Let be the path-normalized value of the i-th basic prediction model in the k-th target month. Let be the mean of the predictions of the i-th basic prediction model within the path window, and c be the positive smoothing constant.

[0067] Based on this, the formula for calculating path divergence is: ; In the formula, Let K be the path divergence between the i-th and j-th basic prediction models, and P be the path window length. i,k and P j,k These are the path normalization values ​​of the i-th and j-th basic prediction models in the k-th target month, respectively.

[0068] Therefore, in this embodiment, the calculation results of the above three types of divergences have all been mapped to the zero-to-one interval, requiring no additional processing.

[0069] Step S32: Perform a weighted combination based on magnitude divergence, change divergence, and continuous path divergence to obtain a comprehensive divergence.

[0070] The final overall divergence is a weighted combination of the three types of divergences, calculated as follows: ; ; In the formula, D ij Let represent the comprehensive divergence between the i-th and j-th basic prediction models, where α, β, and χ are the weighting coefficients corresponding to the magnitude divergence, change divergence, and path divergence, respectively.

[0071] If the remaining months are less than the path window K, the calculation is based on the actual available months; if there is only one target month, the path divergence is closed, and the weights of the magnitude and monthly change divergence are renormalized.

[0072] Step S4: Based on the local prediction capabilities of each basic prediction model, conditional penalty correction is applied to the multi-dimensional prediction discrepancies to obtain the model credibility of each basic prediction model.

[0073] After obtaining the overall divergence, it is necessary to penalize and correct each model accordingly. However, the core idea here is not to directly use whether a model deviates from the majority result as the basis for weight reduction, but rather to first calculate the divergence strength of the model relative to other reliable models, and then control the divergence penalty by the model's own monthly local predictive ability—this is the basic meaning of reliability conditional divergence penalty. In this embodiment, conditional penalty correction is performed for each target automotive part SKU and each target prediction period. SKU-by-SKU and target-month-by-target fusion refers to generating a set of basic prediction model weights for each SKU and each future target month, enabling different model combinations to be used for different SKUs and different prediction step sizes.

[0074] In this embodiment, step S4 includes: Step S41: For each target automotive part SKU and each target prediction period, based on the local prediction capabilities of other basic prediction models, weight the overall divergence between the current basic prediction model and other basic prediction models to obtain the reliable model divergence strength: ; In the formula, A i R represents the strength of reliable model divergence for the i-th basic prediction model. j For the local predictive capability of the j-th basic prediction model, D ijTo account for the divergence, ε is a smoothing constant.

[0075] In the specific calculation, the local predictive capabilities of other basic prediction models are first utilized to weight the overall divergence between the current basic prediction model and other basic prediction models, thus obtaining the strength of the divergence between reliable models. The purpose of this is to give models with higher capabilities greater say in "judging whether the current model deviates from the consensus".

[0076] Step S42: Based on the local prediction capability of the current basic prediction model, control the penalty intensity generated by the equal divergence to obtain the conditional penalty Q. i : Q i =(1-R i ) r A i ; In the formula, r is the reliability adjustment index; In this process, after obtaining the reliable model divergence strength, the penalty strength generated by equal divergence is controlled by the local predictive ability of the current base prediction model, resulting in a conditional penalty. This formula shows that when the model's local predictive ability is high, the conditional penalty decreases; conversely, when R... i When the conditional penalty Q is low, i Increase.

[0077] Step S43: Based on the local prediction capability and conditional penalty of the current basic prediction model, generate the model confidence level U. i : U i =R i exp(-λQ i ); In the formula, λ is the divergence penalty intensity parameter.

[0078] Therefore, the greater the divergence between the model and other reliable models, the stronger the penalty; the higher the model's local capability, the weaker the penalty for the same level of divergence. This mechanism is used to retain a small number of effective predictive signals that have historically high local capabilities, while suppressing models with low capability and significant deviations.

[0079] Step S5: Generate fusion weights based on model credibility, and weight the basic prediction results according to the fusion weights to output the final prediction result of the periodic sales.

[0080] In this embodiment, after obtaining the model confidence scores of each basic prediction model, they can be converted into fusion weights. Specifically, non-negative normalization is performed on the model confidence scores to form the model fusion weights for the current SKU and the current target month: ; In the formula, w iU represents the fusion weights of the i-th basic prediction model. i and U j Let represent the model confidence of the i-th and j-th basic prediction models, M represent the total number of effective basic prediction models, and ε represent the smoothing constant.

[0081] After obtaining the fusion weights, the basic prediction values ​​of each basic prediction model are weighted and summed to obtain the final monthly sales forecast value: ; In the formula, w represents the final predicted value of the target SKU in the target month t. i,s,t The fusion weights for the i-th basic prediction model based on the target SKU and the target month t are: This represents the base forecast value for the i-th base forecast model for the target SKU and the target month t.

[0082] By repeating the above steps for all months in the target month set, a continuous monthly sales forecast sequence can be formed.

[0083] In the embodiments of this application, processing of various boundary scenarios is also supported.

[0084] For handling zero sales, low sales, and negative forecasts, a positive smoothing constant is added to the calculation of magnitude divergence, smoothing relative error, and monthly change divergence. The median of historical non-zero monthly sales for the target SKU is preferred as the normal sales metric to avoid abnormal amplification of relative error when actual sales are close to zero. Negative forecast values ​​output by the basic forecast model are converted to zero before being included in the divergence and fusion calculations, while the original values ​​are retained as diagnostic fields.

[0085] For handling unsettled months and future information isolation, historical target months must meet the requirement of complete settlement of actual sales. Months that are not yet settled or whose data has not been fully collected cannot be used as true labels. Historical scenario features can only be calculated using data from the corresponding historical forecast baseline month and earlier.

[0086] For handling new products, low-frequency products, and insufficient local samples, when the model-level local effective sample weight is lower than the local sample threshold, the influence of the grouping error with the same prediction step size is enhanced by hierarchical shrinkage; when there are no local samples at all, the finest hierarchical error that meets the minimum grouping sample size is directly used.

[0087] For handling insufficient grouped samples, if the samples at the market, channel, and category levels are insufficient, the process will fall back to the market, channel, and global same-prediction step size levels in sequence; if the global same-prediction step size is still insufficient, it is allowed to expand to adjacent prediction step sizes and use lead time decay weights.

[0088] For handling missing or failed models, if a base prediction model does not produce a valid prediction, it will not be included in the divergence matrix and weight normalization for the current SKU and target month. Its historical local error will only be calculated using the sample set predicted by the model that is valid. If there are fewer than two valid models, the valid model with the highest local capability will be used as an auxiliary output and an anomaly marker will be recorded.

[0089] For cases where the path length is insufficient, if the remaining months are less than the path window, the actual available months will be used for calculation; if there is only one target month, the path divergence will be closed, and the magnitude and weight of the divergence will be renormalized.

[0090] When the overall model capability is low, if the local prediction capability of all effective models is lower than the preset threshold, select the top-ranked models in terms of local capability to form a reliability-weighted auxiliary result and record the low confidence level.

[0091] In the embodiments of this application, the implementation parameters in each step can be configured using the following exemplary value ranges. The number of base prediction models can be selected within the range of [2, 10], and must include at least two different modeling mechanisms. The historical monthly window can be selected within the range of [12, 72], using all valid months and triggering hierarchical rollback when historical data is insufficient. The future prediction window can be selected within the range of [1, 24], with weights calculated separately for each target month. The number of similar samples can be selected within the range of [15, 120], filtering valid historical prediction samples according to the model. The path window can be selected within the range of [2, 8], calculating based on the actual available months when insufficient. The contraction parameter κ can be selected within the range of [6, 35], the distance decay parameter τ can be selected within the range of [0.3, 2.4], and the lead time decay parameter ρ can be selected within the range of [0.4, 0.99]. The divergence weights α, β, and χ can be 0.35, 0.30, and 0.35, respectively. The capability conversion parameter γ can be selected, for example, within the range of [0.3, 3.5]; the reliability adjustment index r can be selected, for example, within the range of [1, 3]; and the divergence penalty strength parameter λ can be selected, for example, within the range of [0.3, 2.5]. These parameters can be determined through historical rolling verification, grid search, Bayesian optimization, or other parameter selection methods that do not use future information.

[0092] In the embodiments of this application, traceable prediction interpretation is also supported. Using SKU, target month, and base prediction model as the smallest granularity, the following fields are output: Sample Basis field, recording the number of similar samples at the model level, effective sample weights, prediction step size range, and the actual hierarchical backtracking path used; Capability Assessment field, recording local errors, grouping errors, hierarchical shrinkage errors, and the model's local prediction capability derived from these; Divergence Correction field, recording magnitude divergence, monthly change divergence, consecutive monthly path divergence, overall divergence strength, and conditional divergence penalty; Fusion Output field, recording the overall credibility of each model, fusion weights, and the final monthly sales forecast value; Data Boundary field, recording the prediction base month, data truncation month, actual sales settlement identifier, and calculation time, to prove that information after the target month was not used. These fields are used for both prediction result verification and subsequent model iteration and anomaly investigation.

[0093] In the embodiments of this application, the experiment can use a rolling time window for verification, rather than random shuffling. Each historical month is used sequentially as the prediction baseline month, and the prediction results for the first, third, sixth, and twelfth months are evaluated using only previously obtained data to train or call the basic prediction model. Overall metrics may include Weighted Mean Absolute Percentage Error (WMAPE), Mean Absolute Error (MAE), and Bias. Suggested comparison dimensions include overall accuracy, different prediction step sizes, different SKU types, high-discrepancy samples, and a few reliable models. The actual effect of the technical solution of this application is verified by comparing it with the best single model, simple averaging, global error weighting, same-lead-period error weighting, and ordinary stacking, and by conducting ablation experiments.

[0094] In summary, the various embodiments of this application have the following technical effects: By evaluating the local prediction capabilities of each basic prediction model based on the prediction step size and scene characteristics, the model capability evaluation can be accurately aligned with the current prediction task, solving the problem that the global average error cannot be adapted to different SKUs and different prediction step sizes.

[0095] By constructing a three-dimensional model prediction divergence—scale divergence, change divergence, and continuous path divergence—and weighting and combining them to obtain a comprehensive divergence, we can comprehensively characterize the prediction differences between models from multiple dimensions, providing a more accurate basis for conditional penalty correction.

[0096] By applying conditional penalty correction to each SKU and each target prediction period, the model's local prediction capability is used to control the intensity of the divergence penalty. This results in a weaker penalty for high-capacity models under the same divergence and a stronger penalty for low-capacity models, thereby suppressing abnormal deviations of low-capacity models while preserving the effective prediction signals of high-capacity models.

[0097] By retrieving similar samples based on scene features and forming a model-level effective sample set, hierarchical grouping error is introduced to shrink the sample when the effective sample weights are insufficient, thereby obtaining local prediction capabilities. This not only reflects the local prediction scene features of the current SKU, but also provides a stable estimation reference through hierarchical grouping error when samples are insufficient.

[0098] By normalizing the model credibility into fusion weights and weighting the output of the final prediction results, differentiated adaptive fusion for different SKUs and different prediction periods is achieved, which effectively improves the accuracy and stability of the cycle sales forecast for auto parts.

[0099] Therefore, this application evaluates the local prediction capabilities of each basic prediction model under the current prediction task and uses these local prediction capabilities to control the penalty intensity of prediction discrepancies between models, generating fusion weights for each SKU and each target prediction period. This allows for the retention of effective prediction signals from high-capability models while suppressing abnormal deviations from low-capability models, achieving differentiated adaptive fusion for different SKUs and prediction periods, and effectively improving the accuracy and stability of automotive parts cycle sales forecasting.

[0100] In one embodiment, such as Figure 2 As shown, this application also provides a device for predicting the cycle sales volume of automotive parts, comprising: The data acquisition module 201 is used to acquire historical periodic sales data of the target auto parts, as well as basic prediction results generated by multiple basic prediction models based on the historical periodic sales data.

[0101] The capability assessment module 202 is used to assess the local prediction capability of each basic prediction model based on the prediction step size and scene characteristics. When there are insufficient local samples, hierarchical grouping error with the same prediction step size is introduced to perform hierarchical shrinkage to obtain the local prediction capability.

[0102] The divergence calculation module 203 is used to calculate the multi-dimensional prediction divergence between models based on the basic prediction results of each basic prediction model.

[0103] The divergence correction module 204 is used to perform conditional penalty correction on the multi-dimensional prediction divergence based on the local prediction capabilities of each basic prediction model, so as to obtain the model credibility of each basic prediction model.

[0104] The fusion output module 205 is used to generate fusion weights based on the model's credibility, and to weight the basic prediction results based on the fusion weights to output the final prediction result of the periodic sales.

[0105] In one embodiment, such as Figure 3 As shown, this application also provides a computer device 300, comprising: The communication interface 301 allows for information exchange with other devices or network nodes.

[0106] One or more processors 302 are connected to a communication interface 301 to enable information interaction with other devices or network nodes, and to execute the methods provided by one or more technical solutions in the above embodiments when running computer programs.

[0107] Memory 303 is used to store computer-readable instructions that can be executed on processor 302. When executed by one or more processors 302, the computer-readable instructions perform the steps of a method for predicting the periodic sales volume of automotive parts as described in the above embodiment.

[0108] The computer device 300 in this embodiment only shows a block diagram of a portion of the structure related to the present application solution, and does not constitute a limitation on the computer device to which the present application solution is applied. A specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements, such as power supply, input / output interfaces, etc. Furthermore, the computer device 300 of this embodiment can operate an operating system, such as Windows Server, stored in memory 303. TM Mac OS X TM Unix TM Linux TM FreeBSD TM Or similar.

[0109] In one embodiment, this application also provides a storage medium storing computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the automotive parts cycle sales forecasting method described in the above embodiment.

[0110] The computer-readable instructions in the aforementioned computer device 300 and storage medium may be application programs.

[0111] In addition, the specific descriptions of the technical effects and steps corresponding to the automotive parts cycle sales forecasting device, the computer device 300, and the storage medium in the above embodiments are all based on the relevant descriptions of the embodiments of the automotive parts cycle sales forecasting method.

[0112] Since the systems / devices described in the above embodiments of this application are systems / devices used to implement the methods of the above embodiments of this application, those skilled in the art can understand the specific structure and modifications of the system / devices based on the methods described in the above embodiments of this application, and therefore will not be repeated here. All systems / devices used in the methods of the above embodiments of this application fall within the scope of protection of this application.

[0113] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0114] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions.

[0115] It should be noted that any reference numerals within parentheses in this application should not be construed as limiting the technical solution. The word "comprising" does not exclude the presence of components or steps not listed in the technical solution. The word "a" or "an" preceding a component does not exclude the presence of multiple such components. This application can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In technical solutions that list several devices, several of these devices may be embodied by the same hardware. The use of the terms first, second, third, etc., is merely for convenience and does not indicate any order. These terms can be understood as part of the component names.

[0116] Furthermore, it should be noted that in the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0117] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the scope of protection of this application should be interpreted to include the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0118] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if these modifications and variations fall within the scope of protection of this application and its equivalents, this application should also include these modifications and variations.

Claims

1. A method for predicting the cyclical sales volume of automotive parts, characterized in that, include: Obtain historical sales data for the target auto parts, as well as basic prediction results generated by multiple basic prediction models based on the historical sales data; The local prediction capability of each basic prediction model is evaluated based on the prediction step size and scene characteristics. When there are insufficient local samples, a hierarchical grouping error with the same prediction step size is introduced to perform hierarchical shrinkage to obtain the local prediction capability. Based on the basic prediction results of each basic prediction model, calculate the multi-dimensional prediction discrepancies between the models. Based on the local prediction capabilities of each basic prediction model, conditional penalty correction is applied to the multi-dimensional prediction discrepancies to obtain the model credibility of each basic prediction model. Based on the model's credibility, a fusion weight is generated, and the basic prediction results are weighted according to the fusion weight to output the final prediction result of the periodic sales.

2. The method for predicting the cyclical sales volume of automotive parts according to claim 1, characterized in that, The calculation of multi-dimensional prediction discrepancies between models based on the basic prediction results of each basic prediction model includes: For any two basic prediction models, calculate the magnitude divergence, variation divergence, and continuous path divergence respectively. The comprehensive divergence is obtained by weighting and combining the magnitude divergence, the change divergence, and the continuous path divergence.

3. The method for predicting the cyclical sales volume of automotive parts according to claim 2, characterized in that, The step of applying conditional penalty correction to the multi-dimensional prediction discrepancies based on the local prediction capabilities of each basic prediction model to obtain the model credibility of each basic prediction model includes: For each target automotive part SKU and each target prediction period, based on the local prediction capabilities of other basic prediction models, the overall divergence between the current basic prediction model and other basic prediction models is weighted to obtain the reliable model divergence strength: ； In the formula, A i R represents the reliability model divergence strength of the i-th basic prediction model. j For the local predictive capability of the j-th basic prediction model, D ij The comprehensive divergence is given by ε, where ε is a smoothing constant. The conditional penalty Q is obtained by controlling the penalty intensity generated by equal divergence based on the local prediction capability of the current basic prediction model. i : Q i =(1-R i ) r A i ; In the formula, r is the reliability adjustment index; Based on the local prediction capability of the current basic prediction model and the conditional penalty, the model confidence level U is generated. i : U i =R i exp(-λQ i ); In the formula, λ is the divergence penalty intensity parameter.

4. The method for predicting the cyclical sales volume of automotive parts according to claim 1, characterized in that, The evaluation of the local prediction capability of each basic prediction model based on the prediction step size and the scene features includes: Construct a historical rolling sample based on the historical sales data of the aforementioned period; Based on the scene characteristics, similar samples are retrieved from the historical rolling samples. For each basic prediction model, an effective sample set is formed, and the local error and effective sample weight of each basic prediction model are calculated. When the effective sample weights are lower than a preset sample threshold, the hierarchical grouping error with the same prediction step size is obtained, and hierarchical shrinkage is performed based on the effective sample weights and the hierarchical grouping error to obtain the shrinkage error. : ； In the formula, n i For the first Effective sample weights for each basic prediction model Let κ be the local error, and κ be the shrinkage parameter. The hierarchical grouping error; The local prediction capability R is obtained based on the shrinkage error transformation. i ; ； In the formula, γ is the capacity conversion parameter.

5. The method for predicting the cyclical sales volume of automotive parts according to claim 4, characterized in that, The step of retrieving similar samples from the historical rolling samples based on the scene features, forming an effective sample set for each basic prediction model, and calculating the local error and effective sample weights for each basic prediction model includes: Based on the distance between the scene features of the target auto parts and the scene features of the historical rolling samples, as well as the difference in the prediction step size, the sample weight is calculated, and similar samples are selected from the historical rolling samples. For each basic prediction model, similar samples with missing predictions are removed to form the effective sample set corresponding to each basic prediction model; Based on the effective sample set and the sample weights, calculate the local error and effective sample weights of each basic prediction model.

6. The method for predicting the cyclical sales volume of automotive parts according to claim 4, characterized in that, The step of obtaining the hierarchical grouping error with the same prediction step size includes: Following the order from fine-grained to coarse-grained levels, it is determined whether the current level meets the minimum sample size threshold. If it does, the grouping error corresponding to the current level is taken as the level grouping error. Otherwise, it is backed up to the next coarse-grained level for judgment until a level grouping error that meets the minimum sample size threshold is obtained.

7. The method for predicting the cyclical sales volume of automotive parts according to any one of claims 1 to 6, characterized in that, The prediction step size is calculated using the difference in serial numbers across cycles, and the scenario features include the proportion of non-zero data, data volatility, recent change characteristics, and periodic characteristics.

8. A device for predicting the cyclical sales volume of automotive parts, characterized in that, The device includes: The data acquisition module is used to acquire historical period sales data of the target auto parts, as well as basic prediction results generated by multiple basic prediction models based on the historical period sales data. The capability assessment module is used to evaluate the local prediction capability of each basic prediction model based on the prediction step size and scene characteristics. When there are insufficient local samples, a hierarchical grouping error with the same prediction step size is introduced to perform hierarchical shrinkage to obtain the local prediction capability. The divergence calculation module is used to calculate the multi-dimensional prediction divergence between models based on the basic prediction results of each basic prediction model. The divergence correction module is used to perform conditional penalty correction on the multi-dimensional prediction divergence based on the local prediction capabilities of each basic prediction model, so as to obtain the model credibility of each basic prediction model. The fusion output module is used to generate fusion weights based on the model's credibility, and to weight the basic prediction results based on the fusion weights to output the final prediction result of the periodic sales.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images