Dynamic Material Demand Forecasting System and Methods in Logistics Support Management

Abstract

This invention relates to the field of logistics support management technology, specifically to a dynamic material demand forecasting system and method for logistics support management. The system includes a data acquisition module, a data analysis module, a material demand forecasting module, and a material demand forecasting display terminal. The data acquisition module is responsible for collecting historical usage data. The data analysis module calculates the consumption probability coefficient, comprehensive influence coefficient, mean, and standard deviation. Based on this data, the material demand forecasting module calculates future material demand using an adaptive multi-feature fusion forecasting framework. The display terminal is used to display the forecast results. This system, through multi-dimensional feature fusion and adaptive weight allocation, improves forecasting accuracy by 30%–45% compared to traditional methods, effectively reducing material shortages and excessive inventory, and improving the efficiency and accuracy of logistics support management.

Description

Technical Field

[0001] This invention relates to the field of logistics support management technology, specifically to a dynamic material demand forecasting system and method in logistics support management. Background Technology

[0002] Logistics support management, as a crucial link in modern organizational operations, has one of its core tasks being ensuring the timeliness and accuracy of material supply. Traditional methods for forecasting material demand mainly rely on simple statistical models, such as moving averages and exponential smoothing, or are entirely based on human experience. When faced with complex and ever-changing real-world environments, these methods often suffer from low forecasting accuracy, poor adaptability, and an inability to effectively consider multiple influencing factors.

[0003] With the increasing variety of goods and the growing complexity of supply chains, the limitations of traditional forecasting methods are becoming increasingly apparent. On the one hand, single-dimensional forecasting methods cannot fully capture the complex patterns of demand for goods; on the other hand, fixed-weight models cannot adapt to the dynamic changes in data characteristics. Furthermore, the cross-correlation between various influencing factors is ignored, leading to significant deviations between forecast results and actual demand.

[0004] Currently, although some improved forecasting methods exist, such as time series analysis and regression analysis, these methods still suffer from problems such as insufficient fusion of multi-dimensional features, lack of adaptive adjustment capabilities, and inability to effectively handle the forecasting needs of different types of materials. Therefore, developing a dynamic material demand forecasting system that can comprehensively consider multiple factors, possess adaptive capabilities, and is applicable to different types of materials is of great significance for improving the efficiency and accuracy of logistics support management. Summary of the Invention

[0005] The purpose of this invention is to provide a dynamic material demand forecasting system and method for logistics support management, aiming to solve the problems of low forecasting accuracy, poor adaptability, and inability to effectively consider multiple influencing factors in the existing technology.

[0006] This invention proposes a dynamic material demand forecasting system for logistics support management, comprising:

[0007] The data acquisition module is used to collect usage data of the materials to be predicted over the past several periods and send it to the data analysis module;

[0008] The data analysis module, connected to the data acquisition module, is used to calculate the consumption probability coefficient, comprehensive influence coefficient, mean usage of the material to be predicted, and standard deviation of the material usage, and sends the consumption probability coefficient, the comprehensive influence coefficient, the mean usage of the material to be predicted, and the standard deviation of the material usage to the material demand prediction module.

[0009] A material demand forecasting module, connected to the data analysis module, is used to receive the consumption probability coefficient, the comprehensive influence coefficient, the mean of the predicted material usage, and the standard deviation of the predicted material usage. Based on the received consumption probability coefficient, comprehensive influence coefficient, mean of the predicted material usage, and standard deviation of the predicted material usage, it calculates the predicted material demand value for the first future cycle of the predicted material using an adaptive multi-feature fusion forecasting framework.

[0010] The material demand forecasting display terminal is connected to the material demand forecasting module and is used to receive and display the material demand forecast value for the first future period of the material to be forecasted.

[0011] Preferably, the data analysis module includes the following processing when calculating the comprehensive impact coefficient of the material to be predicted:

[0012] Define influencing factors, including time influencing factors, spatial influencing factors, and event influencing factors, which are strongly correlated with the usage data of the material to be predicted;

[0013] The usage data of the material to be predicted over the past several periods and the usage data of all similar materials are collected. Cross-correlation analysis is then performed on the material to be predicted, the similar materials, and the influencing factors to determine the degree of correlation between the material to be predicted and the influencing factors.

[0014] Based on the correlation between the material to be predicted, the usage data of the same type of material, and all the influencing factors, the correlation between the usage data of the material to be predicted and the usage data of all the same type of material is calculated. The correlation includes strong correlation, medium correlation, and weak correlation.

[0015] The comprehensive influence coefficient of the material to be predicted is obtained by weighting the correlation between the usage data of the material to be predicted and the usage data of the same type of material that has a strong correlation with the usage data.

[0016] Preferably, the data analysis module includes the following processing when calculating the time influence factor:

[0017] Calculate the correlation coefficients between all the time-related factors and the data on the use of the materials to be predicted over all periods;

[0018] The average correlation coefficient over all periods is calculated.

[0019] Calculate the impact factor weight value of the time-related impact factor;

[0020] When the correlation coefficient between the time-influence factor and the data on the use of materials to be predicted is greater than a preset threshold, it is determined that there is a strong correlation between the time-influence factor and the data on the use of materials to be predicted.

[0021] Preferably, the data analysis module includes the following processing when calculating the consumption probability coefficient:

[0022] The consumption probability of the material to be predicted is defined as the ratio of the consumption amount per unit time to the usage amount in the previous period, i.e., the consumption probability value.

[0023] Calculate the average of the consumption probability values ​​of the historical usage cycles of the material to be predicted;

[0024] The consumption probability coefficient of the material to be predicted is calculated based on the average value.

[0025] Preferably, the adaptive multi-feature fusion prediction framework of the material demand prediction module includes:

[0026] The time-sensitive multi-level decay weighting unit is used to assign time weights to historical data based on the interval between adjacent time points, so that recent data has a higher weight and the weight of distant data decreases step by step.

[0027] The feature evaluation unit is used to evaluate the predictive value of each feature and generate a feature importance score.

[0028] A weight allocation unit is used to allocate initial weights based on the feature importance score;

[0029] The dynamic adjustment unit is used to adjust the feature weights in real time based on recent prediction error feedback;

[0030] The prediction generation unit is used to integrate the weighted feature prediction values ​​into the final prediction result.

[0031] Preferably, the material demand forecasting module further includes a self-learning probability statistics and decision tree hybrid forecasting mechanism, wherein the hybrid forecasting mechanism includes:

[0032] The material characteristics analysis unit is used to analyze the regularity and fluctuation characteristics of historical material data;

[0033] A materials classification decision-maker is used to classify materials into predictable materials and unpredictable materials based on the aforementioned characteristics.

[0034] Consumption probability calculation unit, used to calculate the consumption probability coefficient of the material to be predicted;

[0035] Decision tree building unit, used to build predictive decision trees based on historical data;

[0036] The prediction method switching controller is used to automatically switch between probability and statistical prediction methods and decision tree prediction methods based on the set switching parameters.

[0037] Preferably, the material demand forecasting module further includes an adaptive fusion forecasting engine combining deep learning and ensemble learning, the forecasting engine comprising:

[0038] Deep learning model units are used to build time-series neural networks to capture complex patterns;

[0039] The autoregressive exogenous variable model unit is used to process the exogenous variable series of material requisition quantity;

[0040] The feature selection module is used to evaluate the correlation between features and the prediction target, and to select features with high stability and high predictive value.

[0041] The ensemble learning integration unit is used to train multiple prediction models in parallel and integrate the prediction results of multiple models through model combination to generate the final prediction value.

[0042] Preferably, the system further includes a material demand early warning module, connected to the material demand forecasting module and the material demand forecasting display terminal, for:

[0043] Receive the predicted material demand value for the first period of the future generated by the material demand forecasting module.

[0044] When the predicted demand for the material in the first future period is lower than the preset safety stock level, a red warning message is issued on the material demand prediction display terminal.

[0045] When the predicted demand for the material in the first future period is between the safety stock level and the preset target stock level, a yellow warning message is issued on the material demand prediction display terminal.

[0046] Preferably, the system further includes a feedback optimization module connected to the material demand forecasting module and the material demand forecasting display terminal, for:

[0047] Collect the deviation between the actual usage data of the material to be predicted and the predicted material demand value for the first future period;

[0048] Analyze the causes of the deviation and adjust the relevant parameters in the material demand forecasting module;

[0049] Update the model weights and feature importance in the material demand forecasting module, optimize the forecasting algorithm, and improve the accuracy of future forecasts.

[0050] The dynamic material demand forecasting method in logistics support management, applied to the aforementioned dynamic material demand forecasting system, includes the following steps:

[0051] The data acquisition module collects usage data of the materials to be predicted over several past periods.

[0052] The data analysis module calculates the consumption probability coefficient, comprehensive impact coefficient, mean usage of the predicted material, and standard deviation of the predicted material usage.

[0053] The material demand prediction module receives the consumption probability coefficient, the comprehensive influence coefficient, the mean of the usage of the material to be predicted, and the standard deviation of the usage of the material to be predicted. Based on the received consumption probability coefficient, comprehensive influence coefficient, mean of the usage of the material to be predicted, and standard deviation of the usage of the material to be predicted, the material demand prediction value for the first period of the future is calculated through an adaptive multi-feature fusion prediction framework.

[0054] The material demand forecast display terminal receives and displays the material demand forecast value for the first future period of the material to be predicted.

[0055] The step of calculating the predicted demand value of the material to be predicted for the first period using an adaptive multi-feature fusion prediction framework includes:

[0056] Historical data are assigned time weights based on the interval between adjacent time points by using a time-sensitive multi-level decay weighting unit.

[0057] The predictive value of each feature is evaluated through a feature evaluation unit, generating a feature importance score.

[0058] Initial weights are assigned based on the feature importance score by the weight allocation unit;

[0059] The feature weights are adjusted in real time based on recent prediction error feedback through a dynamic adjustment unit.

[0060] The prediction generation unit integrates the weighted predicted values ​​of each feature into a predicted value of the material demand for the first period of the future.

[0061] This invention achieves high accuracy, high adaptability, and high reliability in material demand forecasting through innovative technologies such as constructing an adaptive multi-feature fusion prediction framework, a time-sensitive multi-level decay weight mechanism, a multi-dimensional cross-correlation influence factor analysis framework, a self-learning probability statistics and decision tree hybrid prediction mechanism, and an adaptive fusion prediction engine combining deep learning and ensemble learning.

[0062] The present invention has the following beneficial effects:

[0063] 1. Improve prediction accuracy: By fusing multi-dimensional features and adaptive weight allocation, the prediction accuracy of the system of this invention is improved by 30% to 45% compared with traditional methods, which greatly reduces the situation of material shortage and excessive inventory.

[0064] 2. Enhanced system adaptability: This invention adopts a time-sensitive multi-level attenuation weight mechanism and dynamic adjustment unit, which can automatically adjust the prediction parameters according to the data characteristics, effectively adapting to the seasonal and sudden changes in the use of materials.

[0065] 3. Optimize resource utilization efficiency: Practical application results show that the system of the present invention can reduce inventory costs by 20% to 30%, reduce capital occupation by 15% to 25%, and increase material turnover rate by more than 20%.

[0066] 4. Enhanced decision support capabilities: The early warning mechanism and feedback optimization mechanism of the present invention can identify potential problems in advance and continuously optimize the prediction model, providing a reliable basis for management decisions.

[0067] 5. Wide range of applications: The system of this invention is applicable to a variety of scenarios such as military logistics, medical supplies, and industrial parts, and can meet the logistics support needs of organizations of different sizes. Attached Figure Description

[0068] Figure 1 This is a schematic diagram of the overall structure of the dynamic material demand forecasting system in logistics support management according to the present invention.

[0069] Figure 2 This is a schematic diagram of the data analysis module of the present invention;

[0070] Figure 3 This is a schematic diagram of the adaptive multi-feature fusion prediction framework structure of the material demand prediction module of the present invention.

[0071] Figure 4 This is a schematic diagram of the self-learning probability statistics and decision tree hybrid prediction mechanism of the present invention.

[0072] Figure 5 This is a schematic diagram of the adaptive fusion prediction engine for deep learning and ensemble learning of the present invention.

[0073] Figure 6 This is a flowchart of the dynamic material demand forecasting method in logistics support management according to the present invention. Detailed Implementation

[0074] Please refer to Figure 1 - Figure 6 The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the scope of the invention.

[0075] like Figure 1 As shown, the dynamic material demand forecasting system for logistics support management provided by the present invention includes: a data acquisition module 1, a data analysis module 2, a material demand forecasting module 3, and a material demand forecasting display terminal 4. It may also include a material demand early warning module 5 and a feedback optimization module 6.

[0076] Data acquisition module 1 is used to collect usage data of the materials to be predicted over the past several periods and send it to data analysis module 2. In one embodiment of the present invention, data acquisition module 1 can collect data from an enterprise resource planning (ERP) system, a warehouse management system, or a material requisition record system. Preferably, the collected data includes structured information such as material ID, usage quantity, usage time, and inventory quantity. For example, for medical supplies of a hospital, data acquisition module 1 can collect the daily usage quantity, inventory quantity, and requisition records of each type of medical supply over the past 12 months.

[0077] Data analysis module 2 is connected to data acquisition module 1 and is used to calculate the consumption probability coefficient, comprehensive influence coefficient, mean usage of the material to be predicted, and standard deviation of the material usage. It then sends these parameters to material demand forecasting module 3. The detailed structure of data analysis module 2 will be described in detail later.

[0078] The material demand forecasting module 3 is connected to the data analysis module 2. It receives the consumption probability coefficient, the comprehensive influence coefficient, the mean and standard deviation of the predicted material usage, and calculates the predicted material demand for the first period using an adaptive multi-feature fusion forecasting framework. The detailed structure and working principle of the material demand forecasting module 3 will be explained in detail later.

[0079] The material demand forecasting display terminal 4 is connected to the material demand forecasting module 3 and is used to receive and display the material demand forecast value for the first period of the future. In a preferred embodiment of the present invention, the material demand forecasting display terminal 4 can be a computer monitor, tablet computer, or dedicated display device, which displays the forecast results in an intuitive chart and data format.

[0080] Furthermore, the system of the present invention may also include a material demand early warning module 5, which is connected to the material demand forecasting module 3 and the material demand forecasting display terminal 4. The material demand early warning module 5 is used to receive the material demand forecast value for the first period of the future generated by the material demand forecasting module 3; when the material demand forecast value for the first period of the future is lower than the preset safety stock level, a red early warning message is issued on the material demand forecasting display terminal 4; when the material demand forecast value for the first period of the future is between the safety stock level and the preset target stock level, a yellow early warning message is issued on the material demand forecasting display terminal 4.

[0081] The system of the present invention may further include a feedback optimization module 6, which is connected to the material demand forecasting module 3 and the material demand forecasting display terminal 4. The feedback optimization module 6 is used to collect the deviation between the actual usage data of the material to be forecasted and the predicted material demand value for the first future period; analyze the causes of the deviation, adjust the relevant parameters in the material demand forecasting module 3; update the model weights and feature importance in the material demand forecasting module 3, optimize the forecasting algorithm, and improve the accuracy of future forecasts.

[0082] like Figure 2 As shown, data analysis module 2 is mainly responsible for calculating the consumption probability coefficient, comprehensive influence coefficient, mean usage of the material to be predicted, and standard deviation of the material usage.

[0083] When calculating the comprehensive impact coefficient, Data Analysis Module 2 first defines the influencing factors, including time-related factors, spatial factors, and event-related factors. These factors are strongly correlated with the usage data of the materials to be predicted. Time-related factors can be seasonality, holidays, weekdays / restdays, etc.; spatial factors can be geographical location, warehousing conditions, transportation distance, etc.; and event-related factors can be promotional activities, policy changes, emergencies, etc.

[0084] Next, data analysis module 2 collects usage data of the material to be predicted over several past periods and usage data of all similar materials. It then performs cross-correlation analysis on the material to be predicted, similar materials, and influencing factors to determine the degree of correlation between the material to be predicted and the influencing factors. For example, regarding the mask usage of a hospital, data analysis module 2 will analyze the impact of the flu season (temporal influencing factor), different wards (spatial influencing factor), and infectious disease outbreaks (event influencing factor) on mask usage.

[0085] Based on the correlation between the usage data of the material to be predicted, similar materials, and all influencing factors, the data analysis module 2 calculates the correlation between the usage data of the material to be predicted and the usage data of all similar materials. This correlation includes strong correlation, medium correlation, and weak correlation. In a preferred embodiment of the present invention, the correlation can be quantified by the Pearson correlation coefficient. A correlation coefficient r greater than 0.7 indicates strong correlation, 0.4 ≤ r ≤ 0.7 indicates medium correlation, and r < 0.4 indicates weak correlation.

[0086] Finally, data analysis module 2 uses the correlation between the usage data of the material to be predicted and the usage data of similar materials with strong correlations to perform a weighted calculation to obtain the comprehensive impact coefficient of the material to be predicted. The weighted calculation can be performed using the following formula:

[0087] ,

[0088] in: The comprehensive impact coefficient of the material to be predicted. Let be the weighting coefficient for the i-th strongly correlated material of the same type. Let be the correlation coefficient between the material to be predicted and the i-th strongly correlated material of the same type. The total number of similar materials with strong correlation.

[0089] When calculating the time-related impact factor, data analysis module 2 employs the following processing steps:

[0090] First, calculate the correlation coefficients between all time-related factors and the data on the use of materials to be predicted over all periods. The correlation coefficients can be calculated using the following formula:

[0091] ,

[0092] in: The correlation coefficient between the time-related factors and the data on the use of materials to be predicted. The time influence factor value for the k-th period. This represents the average value of the time-related factors. For the predicted resource usage data in the k-th period, This represents the average value of the data on the use of materials to be predicted. This represents the total number of cycles.

[0093] Next, the average correlation coefficient over all periods is calculated:

[0094] ,

[0095] in: The average correlation coefficient over all periods. Let be the correlation coefficient for the k-th period. This represents the total number of cycles.

[0096] Then, calculate the impact factor weights of the time-related impact factors:

[0097] ,

[0098] in: The weighted values ​​of the time-related impact factor. The average correlation coefficient. This is the periodic influence coefficient, which is determined based on different periodic characteristics, and typically ranges from 0.8 to 1.2.

[0099] Finally, when the correlation coefficient between the time-influence factor and the predicted material usage data is greater than a preset threshold, a strong correlation is determined between the time-influence factor and the predicted material usage data. In a preferred embodiment of the present invention, the preset threshold can be set to 0.7. This value is an empirical value derived from the analysis of a large amount of practical application data and can effectively distinguish between strong and weak correlations.

[0100] When calculating spatial impact factors, data analysis module 2 employs the following processing steps: First, it calculates the correlation coefficients between all spatial impact factors and the data on the use of the materials to be predicted across all regions. The correlation coefficients can be calculated using the following formula:

[0101] ,

[0102] in: This is the correlation coefficient between spatial impact factors and the data on the use of materials to be predicted. This represents the spatial influence factor value for the first region. This represents the average value of the spatial influence factor. This is the predicted resource usage data for the first region. This represents the average value of the data on the use of materials to be predicted. This represents the total number of regions.

[0103] Next, the average correlation coefficient across all regions is calculated:

[0104] ,

[0105] in: This represents the average correlation coefficient across all regions. Let l be the correlation coefficient of the l-th region. This represents the total number of regions.

[0106] Then, calculate the impact factor weights of the spatial impact factor:

[0107] ,

[0108] in: The impact factor weights of the spatial impact factor. The average correlation coefficient. This is the regional influence coefficient, which is determined according to the characteristics of different regions, and usually ranges from 0.7 to 1.3.

[0109] Finally, when the correlation coefficient between the spatial impact factor and the data on the use of materials to be predicted is greater than a preset threshold (usually 0.65), it is determined that there is a strong correlation between the spatial impact factor and the data on the use of materials to be predicted.

[0110] When calculating the event impact factor, data analysis module 2 uses the following processing steps:

[0111] First, identify significant events in historical data and calculate the rate of change in material usage before and after the events:

[0112]

[0113] in: The rate of change in material usage caused by event e. This represents the average amount of supplies used over a certain period of time following the incident. This represents the average resource usage over a certain period prior to the event. Next, the impact of each event on resource usage is calculated:

[0114] ,

[0115] in: To determine the degree of impact of event e, The rate of change in material usage caused by event e. This is a duration adjustment factor (a larger value is used when the event duration is long, typically ranging from 0.8 to 1.5). An event intensity adjustment factor is used (a larger value is taken for events with higher intensity, typically ranging from 0.9 to 1.8). Then, the weight values ​​of the event impact factors are calculated:

[0116] ,

[0117] in: Let e ​​be the weight value of event e. To determine the degree of impact of event e, The degree of influence of the i-th event, This represents the total number of events. Finally, the overall impact value of the event impact factors can be calculated using the following formula:

[0118] ,

[0119] in: This represents the combined impact value of the event's impact factors. Let e ​​be the weight value of event e. To determine the degree of impact of event e, The probability of event e occurring within the prediction period (determined based on historical data and expert evaluation, ranging from 0 to 1). The total number of events. The degree of influence of a particular event. When the value exceeds a preset threshold (usually 15%), it is determined that there is a strong correlation between the event's impact factor and the data on the use of materials to be predicted.

[0120] When calculating the consumption probability coefficient, data analysis module 2 first defines the consumption probability of the material to be predicted as the ratio of the consumption amount per unit time to the usage amount in the previous period, i.e., the consumption probability value. For example, if a hospital consumes 500 masks in one day, and the usage amount the previous day was 600, then the consumption probability value for that day is 500 / 600 = 0.833.

[0121] Then, calculate the average consumption probability value of the material to be predicted over its historical usage cycles:

[0122] ,

[0123] in: This represents the average probability of consumption. Let be the consumption probability value for the m-th historical usage cycle. This represents the total number of historical usage cycles.

[0124] Finally, the consumption probability coefficient of the material to be predicted is calculated based on the average value:

[0125] ,

[0126] in: The probability coefficient of consumption of the resource to be predicted. This represents the average probability of consumption. For reference purposes, the global average of historical data is usually taken.

[0127] In addition, data analysis module 2 calculates the mean and standard deviation of the predicted material usage. The mean can be calculated using the following formula:

[0128] ,

[0129] in: The average value of the predicted usage of materials. Let n be the usage data of the materials to be predicted in the i-th period, and n be the total number of periods.

[0130] The standard deviation can be calculated using the following formula:

[0131] ,

[0132] in: The standard deviation of the predicted material usage is... The data represents the usage of the materials to be predicted within the i-th period. Let n be the average of the amount of material usage to be predicted, and n be the total number of periods.

[0133] like Figure 3 As shown, the core of the material demand prediction module 3 is an adaptive multi-feature fusion prediction framework, which includes a time-sensitive multi-level decay weight unit 31, a feature evaluation unit 32, a weight allocation unit 33, a dynamic adjustment unit 34, and a prediction generation unit 35.

[0134] The time-sensitive multi-level decay weighting unit 31 is used to assign time weights to historical data based on the interval between adjacent time points, giving higher weights to recent data and progressively decreasing weights to older data. In one embodiment of the invention, the time weights can be calculated using an exponential decay function:

[0135] ,

[0136] in: As time weight, It is a natural constant. This is the attenuation rate parameter (usually between 0.1 and 0.5, adjusted according to the characteristics of the material). and These are the time points of the current cycle and the previous cycle, respectively.

[0137] Feature evaluation unit 32 is used to evaluate the predictive value of each feature and generate a feature importance score. The feature importance score can be calculated in the following way:

[0138] ,

[0139] in: Score the importance of feature f. Data is used for feature f and the i-th period The correlation coefficient, where n is the total number of periods and m is the total number of features. Let j be the j-th feature.

[0140] Weight allocation unit 33 is used to allocate initial weights based on feature importance scores. The initial weights can be obtained through normalization.

[0141] ,

[0142] in: The initial weights for feature f are... Score the importance of feature f. The importance score is given for the j-th feature, where m is the total number of features.

[0143] The dynamic adjustment unit 34 is used to adjust the feature weights in real time based on recent prediction error feedback. The adjustment formula can be expressed as:

[0144] ,

[0145] in: and These are the weights of feature f before and after adjustment, respectively. This is the learning rate (typically between 0.01 and 0.1). The prediction error E is set with respect to the weights. The partial derivatives of .

[0146] The prediction generation unit 35 is used to integrate the weighted feature prediction values ​​into a final prediction result. The final prediction value can be calculated using the following formula:

[0147] ,

[0148] in: This represents the projected demand for the material in the first cycle of the future. The time weight for the i-th period. For the first The weights of each feature Features Based on the data used in the i-th cycle The prediction results The total number of cycles, The total number of features.

[0149] Feature prediction function The calculation method varies depending on the feature type:

[0150] ,

[0151] in: These are the autoregressive coefficients. Let the order be the autoregressive order. For random error term, and For regression coefficients, For the first The feature in the first The value of each period, This is a seasonal adjustment factor. This represents the length of the seasonal cycle. The parameters of the prediction functions for different feature types are obtained through training on historical data.

[0152] Time series characteristics refer to the patterns and trends of material usage over time, primarily capturing the autocorrelation and temporal continuity of material usage. Time series characteristics differ from the time-related factors mentioned earlier: time-related factors focus on the impact of external time factors such as seasonality, holidays, and workdays / restdays on material usage, while time series characteristics focus on the temporal variation patterns of material usage itself. The prediction function for time series characteristics uses an autoregressive model:

[0153] ,

[0154] in: For constant terms, The autoregressive coefficient represents the degree of influence of the usage in the ipth period on the usage in the ith period. This is the autoregression order, typically ranging from 3 to 7, depending on the strength of the temporal correlation of the data. This represents the material usage for the IP-th cycle. The random error term follows a normal distribution with a mean of 0. The autoregressive coefficients... This was obtained by fitting historical data using the least squares method, reflecting the weight of the impact of different time lags on the current prediction.

[0155] The influencing factor characteristics refer to the combined impact of the aforementioned time-related, spatial, and event-related influencing factors on the amount of materials used. The prediction function for the influencing factor characteristics uses a linear regression model:

[0156] ,

[0157] in: For the intercept term, Let be the regression coefficient, representing the th . The intensity of the influence of each influencing factor on the amount of materials used Let be the value of the j-th influencing factor in the i-th period, including time-related influencing factors (such as seasonal indices, holiday indices, etc.), spatial influencing factors (such as geographical location coefficients, warehousing condition scores, etc.), and event-related influencing factors (such as the intensity of promotional activities, the impact coefficient of policy changes, etc.). This represents the random error term. Regression coefficients. and It was obtained by fitting the historical data using the least squares method.

[0158] Seasonal characteristics refer to the periodic patterns of material consumption within a fixed period. They represent an important subset of time-related factors, but due to their unique periodicity, they require specialized prediction functions. The prediction function for seasonal characteristics employs a seasonally adjusted model:

[0159] ,

[0160] in: This is a seasonal adjustment factor, typically ranging from 0.8 to 1.2, and is dynamically adjusted according to seasonal intensity. For the length of the seasonal cycle, such as for annual seasonality, It can be 12 (months) or 52 (weeks). , and These represent the material usage at the previous, second, and third points in the same season, respectively. This function captures the impact of seasonal patterns on forecasts by using a weighted average of historical data from the same period.

[0161] These three feature types each have their own focus, but together they constitute a comprehensive description of resource usage patterns: time series features capture the temporal continuity and autocorrelation of resource usage; influencing factor features capture the impact of external factors (time, space, events) on resource usage; and seasonal features specifically handle periodic change patterns. In actual forecasting, the system dynamically adjusts the weights of these three feature types based on resource characteristics and data performance to achieve optimal forecasting results.

[0162] like Figure 4 As shown, the material demand forecasting module 3 also includes a self-learning probability statistics and decision tree hybrid forecasting mechanism, which includes a material characteristic analysis unit 41, a material classification decision maker 42, a consumption probability calculation unit 43, a decision tree construction unit 44, and a forecasting method switching controller 45.

[0163] The material characteristic analysis unit 41 is used to analyze the regularity, volatility, and other characteristics of historical material data. Regularity can be measured by the autocorrelation coefficient, while volatility can be measured by the coefficient of variation (standard deviation divided by mean).

[0164] The material classification decision-maker 42 is used to classify materials into predictable and unpredictable materials based on characteristics. In a preferred embodiment of the invention, when the coefficient of variation of a material is less than 0.3 and the autocorrelation coefficient is greater than 0.6, it can be classified as a predictable material; otherwise, it is classified as an unpredictable material. These thresholds are empirical values ​​derived from the analysis of a large amount of practical application data, which can effectively distinguish between different types of materials.

[0165] The consumption probability calculation unit 43 is used to calculate the consumption probability coefficient of the material to be predicted. The calculation method has been described in detail above.

[0166] Decision tree building unit 44 is used to build a predictive decision tree based on historical data. The decision tree building process includes the following steps:

[0167] 1. Obtain the training sample set T={(x1,y1),(x2,y2),...,(xn,yn)}, where x1, x2 and xn are the feature vectors of the 1st, 2nd and nth samples, respectively, and y1, y2 and yn are the predicted values ​​of the 1st, 2nd and nth samples, respectively.

[0168] 2. Select the optimal splitting feature and splitting point. The selection criteria can be information gain, Gini index, or reduction in mean square error.

[0169] 3. Divide the dataset into left and right subsets based on the selected splitting features and splitting points.

[0170] 4. Recursively execute steps 2 and 3 on the left and right subsets until the stopping condition is met (such as reaching the maximum depth, the number of node samples being less than the threshold, etc.).

[0171] The prediction method switching controller 45 is used to automatically switch between probabilistic statistical prediction methods and decision tree prediction methods according to a set switching parameter t. In one embodiment of the present invention, the switching rule can be defined as:

[0172] When t=1, the decision tree prediction method is used.

[0173] When t=2 or t=3, use the probability and statistical prediction method.

[0174] The t-value can be automatically adjusted based on factors such as the type of material and the required forecast accuracy.

[0175] like Figure 5 As shown, the material demand prediction module 3 may also include an adaptive fusion prediction engine of deep learning and ensemble learning, which includes a deep learning model unit 51, an autoregressive exogenous variable model unit 52, a feature selection module 53, and an ensemble learning integration unit 54.

[0176] Deep learning model unit 51 is used to construct a time-series neural network to capture complex patterns. In a preferred embodiment of the invention, a Long Short-Term Memory (LSTM) network or a Recurrent Neural Network (RNN) can be used to process time-series data. The network structure may include an input layer, an LSTM layer (64 neurons), a Dropout layer (dropout rate 0.2), a fully connected layer (32 neurons), and an output layer.

[0177] Autoregressive exogenous variable model unit 52 is used to process exogenous variable sequences such as material requisition quantity. In one embodiment of the present invention, an autoregressive moving average model (ARMA) or an autoregressive integral moving average model (ARIMA) can be used to process the exogenous variable sequence.

[0178] The feature selection module 53 is used to evaluate the correlation between features and the prediction target, and to screen features with high stability and high predictive value. Feature screening can be based on methods such as correlation analysis, analysis of variance, or principal component analysis. In a preferred embodiment of the present invention, features with an absolute correlation coefficient greater than 0.3 are retained, while the variance stability of the features (coefficient of variation less than 0.5) is also required.

[0179] The ensemble learning unit 54 is used to train multiple prediction models in parallel and integrates the prediction results of multiple models through model combination to generate a final prediction value. In one embodiment of the invention, two ensemble learning methods can be used: Bagging (e.g., random forest) and Boosting (e.g., gradient boosting tree). Model weights can be determined by their performance on the validation set.

[0180] ,

[0181] in: The weights of model m, Let be the mean squared error of model m on the validation set. This represents the total number of models.

[0182] The final predicted value can be calculated using a weighted average:

[0183] ,

[0184] in: This is the final predicted value. The weights of model m, The predicted value of model m. This represents the total number of models.

[0185] Model predictions The calculation method depends on the type of model used:

[0186] For deep learning models (such as LSTM):

[0187] ,

[0188] in: For activation function, This is the output layer weight matrix. Let be the hidden state at time t. This is the output layer bias vector.

[0189] For the random forest model:

[0190] ,

[0191] in: For the number of decision trees, Let be the prediction function of the t-th decision tree. The input feature vector. For the gradient boosting tree model:

[0192] ,

[0193] in: For the number of weak learners, Let k be the prediction function of the k-th weak learner. This is the input feature vector.

[0194] The material demand early warning module 5 is connected to the material demand forecast module 3 and the material demand forecast display terminal 4, and is used to receive the material demand forecast value for the first period of the future of the material to be predicted generated by the material demand forecast module 3.

[0195] When the projected demand for a material in the first future cycle is lower than the preset safety stock level, the material demand early warning module 5 issues a red warning message on the material demand forecast display terminal 4. The safety stock level can be determined based on factors such as material importance, procurement cycle, and historical volatility. In a preferred embodiment of the invention, the safety stock level can be set as follows:

[0196] ,

[0197] in: To maintain a safety stock level, This represents the average daily demand. This is the service level coefficient (typically ranging from 1.65 to 2.33, corresponding to a service level of 95% to 99%). The coefficient of variation of demand. Procurement lead time (days).

[0198] When the projected demand for the material in the first future cycle falls between the safety stock level and the preset target stock level, the material demand early warning module 5 issues a yellow warning message on the material demand forecast display terminal 4. The target stock level is typically set as follows:

[0199] ,

[0200] in: For the target inventory level, This represents the average daily demand. Procurement lead time (days) This is for safety stock levels.

[0201] The feedback optimization module 6 is connected to the material demand forecasting module 3 and the material demand forecasting display terminal 4, and is used to collect the deviation between the actual usage data of the material to be forecasted and the material demand forecast value of the material to be forecasted in the first future cycle.

[0202] The deviation can be calculated using the following formula:

[0203] ,

[0204] in: The percentage of deviation. For actual use data, These are predicted values.

[0205] Feedback optimization module 6 analyzes the causes of deviations and adjusts relevant parameters in material demand forecasting module 3. For example, if the forecast values ​​for several consecutive periods are lower than the actual values, it may be necessary to increase the decay rate parameter in time decay weighting unit 31. To increase the impact of recent data; if the predicted value fluctuates too much, it may be necessary to adjust the feature weights and increase the weight ratio of stable features.

[0206] Feedback optimization module 6 is also responsible for updating the model weights and feature importance in the material demand forecasting module 3, optimizing the forecasting algorithm, and improving the accuracy of future forecasts. Weight updates can be achieved using gradient descent.

[0207] ,

[0208] in: and These are the weights before and after the update, respectively. The learning rate (typically between 0.001 and 0.01). For loss function Weights The gradient.

[0209] loss function Calculated using mean squared error (MSE):

[0210] ,

[0211] in: For the sample size, Let i be the actual amount of materials used for the i-th sample. Let be the predicted material usage for the i-th sample.

[0212] gradient of loss function The calculation is as follows:

[0213] ,

[0214] in: This is the partial derivative of the predicted value with respect to the weights, and its calculation method varies depending on the prediction model.

[0215] like Figure 6 As shown, the present invention also provides a dynamic material demand forecasting method in logistics support management. This method is applied to the aforementioned dynamic material demand forecasting system and includes the following steps:

[0216] Step S1: Collect usage data of the material to be predicted over the past several periods through data acquisition module 1.

[0217] In one embodiment of the invention, the periods may be the past 12 months, and the usage data includes daily or weekly material usage. For example, for components of a manufacturing company, weekly component usage data over the past 12 months can be collected.

[0218] Step S2: Calculate the consumption probability coefficient, comprehensive influence coefficient, mean usage and standard deviation of the material to be predicted through data analysis module 2.

[0219] The calculation method has been described in detail above and will not be repeated here.

[0220] Step S3: Receive the consumption probability coefficient, comprehensive influence coefficient, mean value of the predicted material usage, and standard deviation of the predicted material usage through the material demand prediction module 3. Based on the received consumption probability coefficient, comprehensive influence coefficient, mean value of the predicted material usage, and standard deviation of the predicted material usage, calculate the predicted material demand value for the first period of the future through an adaptive multi-feature fusion prediction framework.

[0221] The steps for calculating the predicted demand value of the materials to be predicted in the first cycle using the adaptive multi-feature fusion prediction framework include:

[0222] Step S31: Assign time weights to historical data based on the interval between adjacent time points using the time-sensitive multi-level decay weighting unit 31.

[0223] The method for calculating time weights has been described in detail above and will not be repeated here.

[0224] Step S32: Evaluate the predictive value of each feature through the feature evaluation unit 32 and generate a feature importance score.

[0225] The calculation method for feature importance scores has been described in detail above and will not be repeated here.

[0226] Step S33: Assign initial weights based on feature importance scores through weight allocation unit 33.

[0227] The method for allocating initial weights has been described in detail above and will not be repeated here.

[0228] Step S34: The feature weights are adjusted in real time by the dynamic adjustment unit 34 based on the recent prediction error feedback.

[0229] The method for adjusting feature weights has been described in detail above and will not be repeated here.

[0230] Step S35: The weighted predicted values ​​of each feature are integrated into the predicted value of the material demand for the first period of the future by the prediction generation unit 35.

[0231] The method for calculating the final predicted value has been described in detail above and will not be repeated here.

[0232] Step S4: Receive and display the predicted material demand value for the first period of the future for the material to be predicted through the material demand forecast display terminal 4.

[0233] Preferably, the material demand forecasting display terminal 4 can intuitively display the forecast results in the form of charts, data tables or dashboards, so that managers can quickly understand the material demand situation.

[0234] In addition, the method of the present invention may also include the following steps:

[0235] Step S5: Receive the predicted material demand value for the first period of the future generated by the material demand forecasting module 3 through the material demand early warning module 5; when the predicted material demand value for the first period of the future is lower than the preset safety stock level, issue a red warning message on the material demand forecasting display terminal 4; when the predicted material demand value for the first period of the future is between the safety stock level and the preset target stock level, issue a yellow warning message on the material demand forecasting display terminal 4.

[0236] Step S6: Collect the deviation between the actual usage data of the materials to be predicted and the predicted demand value of the materials in the first future cycle through the feedback optimization module 6; analyze the reasons for the deviation, adjust the relevant parameters in the material demand prediction module 3; update the model weights and feature importance in the material demand prediction module 3, optimize the prediction algorithm, and improve the accuracy of future predictions.

[0237] To better illustrate the practical application effects of the present invention, a specific application example is given below.

[0238] The medical supplies management department of a hospital used the dynamic supply demand forecasting system of this invention to predict the demand for disposable medical masks. First, the data acquisition module 1 collected daily mask usage data from the hospital's supply management system over the past 12 months. Then, the data analysis module 2 calculated the mask consumption probability coefficient (1.2), comprehensive influence coefficient (0.85), average usage (2000 masks / day), and standard deviation of usage (300 masks / day).

[0239] The material demand forecasting module 3 receives these parameters and calculates the predicted mask demand for the next week using an adaptive multi-feature fusion forecasting framework. During the calculation, the time-sensitive multi-level decay weighting unit 31 assigns higher weights (0.6-0.9) to the data from the most recent month. The feature evaluation unit 32 identifies seasonality and departmental traffic as important features affecting mask usage. The weight allocation unit 33 assigns initial weights accordingly. The dynamic adjustment unit 34 fine-tunes the weights based on the prediction error (around 5%) from the past two weeks. Finally, the prediction generation unit 35 calculates the predicted mask demand for the next week to be 15,000 masks.

[0240] Because the predicted value was higher than the safety stock (10,000 units) but lower than the target stock (20,000 units), the material demand early warning module 5 issued a yellow warning on the material demand forecast display terminal 4, reminding managers to pay attention to the mask inventory status. One week later, the feedback optimization module 6 collected actual usage data of 14,200 units, which deviated from the predicted value by 5.33%. Based on this, the model parameters were adjusted, and the prediction algorithm was further optimized.

[0241] After three months of operation, the system reduced the hospital's mask inventory costs by 25% and the shortage rate of supplies from 4% to 0.5%, greatly improving the efficiency and accuracy of medical supply management.

[0242] The dynamic material demand forecasting system and method for logistics support management provided by this invention achieves high accuracy, high adaptability and high reliability in material demand forecasting through innovative technologies such as constructing an adaptive multi-feature fusion forecasting framework, a time-sensitive multi-level decay weight mechanism, a multi-dimensional cross-correlation influence factor analysis framework, a self-learning probability statistics and decision tree hybrid forecasting mechanism, and an adaptive fusion forecasting engine of deep learning and ensemble learning.

[0243] Compared to existing technologies, the system of this invention has the following significant advantages: it breaks through the limitations of single prediction to multi-dimensional fusion prediction, achieves a leap from static weights to dynamic adaptive weights, innovatively introduces a cross-correlation analysis framework, develops a hybrid prediction mechanism for different material types, and integrates the latest deep learning and ensemble learning technologies. These innovations enable the system of this invention to significantly improve prediction accuracy, optimize resource utilization efficiency, enhance decision support capabilities, and is applicable to various logistics support management scenarios.

[0244] Practical application results show that the system of the present invention can improve prediction accuracy by 30% to 45%, reduce inventory costs by 20% to 30%, reduce capital occupation by 15% to 25%, and increase material turnover by more than 20%, bringing significant economic and management value to logistics support management.

[0245] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of protection of the present invention.

Claims

1. A dynamic material demand forecasting system in logistics support management, characterized in that, The method comprises the following steps: a data collection module is configured to collect usage data of the to-be-predicted material in a plurality of past periods and send the data to a data analysis module; the data analysis module is connected to the data collection module and is configured to calculate a consumption probability coefficient, a comprehensive influence coefficient, a mean value of the usage amount of the to-be-predicted material, and a standard deviation of the usage amount of the to-be-predicted material, and send the consumption probability coefficient, the comprehensive influence coefficient, the mean value of the usage amount of the to-be-predicted material, and the standard deviation of the usage amount of the to-be-predicted material to a material demand prediction module; the material demand prediction module is connected to the data analysis module and is configured to receive the consumption probability coefficient, the comprehensive influence coefficient, the mean value of the usage amount of the to-be-predicted material, and the standard deviation of the usage amount of the to-be-predicted material, and calculate a material demand prediction value of the to-be-predicted material in a first future period based on the received consumption probability coefficient, comprehensive influence coefficient, mean value of the usage amount of the to-be-predicted material, and standard deviation of the usage amount of the to-be-predicted material, by using an adaptive multi-feature fusion prediction framework; a material demand prediction display terminal is connected to the material demand prediction module and is configured to receive and display the material demand prediction value of the to-be-predicted material in the first future period; when calculating the comprehensive influence coefficient of the to-be-predicted material, the data analysis module comprises the following processing: defining influence factors, the influence factors including time influence factors, space influence factors, and event influence factors, the influence factors having a strong correlation with the usage data of the to-be-predicted material; collecting the usage data of the to-be-predicted material in a plurality of past periods and the usage data of all the same type of materials, and performing cross-correlation analysis on the to-be-predicted material, the same type of materials, and the influence factors to determine the correlation degree between the to-be-predicted material and the influence factors; based on the correlation degree between the to-be-predicted material, the usage data of the same type of materials, and all the influence factors, the correlation degree between the usage data of the to-be-predicted material and the usage data of all the same type of materials is calculated, the correlation degree including strong correlation, medium correlation, and weak correlation; using the correlation degree between the usage data of the to-be-predicted material and the usage data of the same type of materials having strong correlation, a weighted operation is performed to obtain the comprehensive influence coefficient of the to-be-predicted material; the adaptive multi-feature fusion prediction framework of the material demand prediction module comprises: a time-sensitive multi-level decay weight unit configured to assign time weights to historical data based on the interval of adjacent time points, so that recent data obtains a higher weight and the weight of long-term data decreases gradually; a feature evaluation unit configured to evaluate the prediction value of each feature and generate a feature importance score; a weight allocation unit configured to allocate an initial weight based on the feature importance score; a dynamic adjustment unit configured to adjust the feature weight in real time according to the recent prediction error feedback; a prediction generation unit configured to integrate the weighted feature prediction values into a final prediction result; the material demand prediction module further comprises a self-learning probability statistics and decision tree hybrid prediction mechanism, the hybrid prediction mechanism comprising: ​ The material characteristic analysis unit is configured to analyze regularity and volatility characteristics of the material historical data; The material classification decision maker is configured to classify the material into predictable material and unpredictable material according to the characteristics; The consumption probability calculation unit is configured to calculate a consumption probability coefficient of the material to be predicted; The decision tree construction unit is configured to construct a prediction decision tree based on the historical data; The prediction method switching controller is configured to automatically switch between the probability statistical prediction method and the decision tree prediction method according to a set switching parameter.

2. The dynamic material demand forecasting system of claim 1, wherein, When calculating the time influence factor, the data analysis module includes the following processing: Calculate the correlation coefficient between all the time influence factors and the use data of the material to be predicted in all periods; Calculate the average value of the correlation coefficients in all periods; Calculate the influence factor weight value of the time influence factor; When the correlation coefficient between the time influence factor and the use data of the material to be predicted is greater than a preset threshold, it is determined that there is strong correlation between the time influence factor and the use data of the material to be predicted.

3. The dynamic material demand forecasting system of claim 1, wherein, When calculating the consumption probability coefficient, the data analysis module includes the following processing: Define the consumption probability of the material to be predicted as the ratio of the consumption amount in a unit time to the use amount in the last period, that is, the consumption probability value; Calculate the average value of the consumption probability values of the historical use periods of the material to be predicted; Calculate the consumption probability coefficient of the material to be predicted based on the average value.

4. The dynamic material demand forecasting system of claim 1, wherein, The material demand prediction module further includes a self-adaptive fusion prediction engine of deep learning and ensemble learning, and the prediction engine includes: A deep learning model unit is configured to construct a time series neural network to capture complex patterns; An autoregressive exogenous variable model unit is configured to process exogenous variable sequences of material use amount; A feature selection module is configured to evaluate the correlation between features and prediction targets, and filter features with high stability and large prediction value; An ensemble learning integration unit is configured to train multiple prediction models in parallel, integrate the prediction results of multiple models through model combination, and generate a final prediction value.

5. The dynamic material demand forecasting system of claim 1, wherein, The system further includes a material demand early warning module connected with the material demand prediction module and the material demand prediction display terminal, configured to: Receive the material demand prediction value of the material to be predicted in the future first period generated by the material demand prediction module; When the material demand prediction value of the material to be predicted in the future first period is lower than a preset safety stock amount, issue a red early warning information on the material demand prediction display terminal; When the material demand prediction value of the material to be predicted in the future first period is between the safety stock amount and a preset target stock amount, issue a yellow early warning information on the material demand prediction display terminal.

6. The dynamic material demand forecasting system of claim 1, wherein, The system further includes a feedback optimization module connected with the material demand prediction module and the material demand prediction display terminal, configured to: Collect the deviation between the actual use data of the material to be predicted and the material demand prediction value of the material to be predicted in the future first period; Analyze the deviation reason and adjust the related parameters in the material demand prediction module; Updating the model weights and feature importance in the material demand prediction module, optimizing the prediction algorithm, and improving the accuracy of future predictions.

7. A method for dynamic material demand forecasting in logistics support management, characterized in that, The dynamic material demand prediction system of any one of claims 1 to 6, comprising the following steps: Collecting usage data of the material to be predicted over a number of past periods through the data collection module; Calculating the consumption probability coefficient, the comprehensive influence coefficient, the mean of the usage amount of the material to be predicted, and the standard deviation of the usage amount of the material to be predicted through the data analysis module; Receiving the consumption probability coefficient, the comprehensive influence coefficient, the mean of the usage amount of the material to be predicted, and the standard deviation of the usage amount of the material to be predicted through the material demand prediction module, and calculating the material demand prediction value of the first future period of the material to be predicted based on the received consumption probability coefficient, comprehensive influence coefficient, mean of the usage amount of the material to be predicted, and standard deviation of the usage amount of the material to be predicted through an adaptive multi-feature fusion prediction framework; Receiving and displaying the material demand prediction value of the first future period of the material to be predicted through the material demand prediction display terminal; The step of calculating the material demand prediction value of the first future period of the material to be predicted through the adaptive multi-feature fusion prediction framework comprises: Assigning time weights to historical data based on the interval of adjacent time points through a time-sensitive multi-level decay weight unit; Evaluating the prediction value of each feature through a feature evaluation unit to generate a feature importance score; Assigning initial weights based on the feature importance score through a weight allocation unit; Adjusting the feature weights in real time according to recent prediction error feedback through a dynamic adjustment unit; Integrating the weighted feature prediction values into the material demand prediction value of the first future period of the material to be predicted through a prediction generation unit.

Images