Transportation price intelligent correction method and system based on hierarchical spreadback rule

Abstract

The invention discloses an intelligent transportation price correction method and system based on a hierarchical spreadback rule, and belongs to the technical field of transportation price optimization, and the method comprises the steps: calculating the order receiving probability of a driver according to the historical data of the driver through employing an improved Transform architecture model, calculating an initial quotation through the combination of an order and the data of the driver, and carrying out the calculation of the initial quotation; a probability weighted quotation is obtained through a nonlinear correction function according to order receiving probability adjustment; dividing price intervals on the basis of probability weighted quotation, setting a layered zero-rendering strategy, predicting the acceptance of the driver on the price after zero-rendering by using a gradient boosting tree model, judging whether secondary zero-rendering is carried out according to an acceptance threshold value, and determining a final quotation; recording the order receiving or rejecting behavior of the driver, and reversely updating the parameters of the order receiving probability prediction model of the driver and the layered zeroing strategy to form closed-loop optimization; through multi-model cooperation, dynamic parameter adjustment and feedback optimization, the transportation price is accurately corrected, the secondary payment probability is reduced, and the platform order transaction rate is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of transportation price optimization, and in particular to a method and system for intelligently correcting transportation prices based on a layered zeroing rule. Background Art

[0002] On existing transportation service platforms, after a user submits an order, the platform typically provides a transportation quote based on fixed pricing rules and partial data about the driver and vehicle. However, this approach presents numerous problems. For one thing, by only considering fixed pricing rules, drivers' historical data, such as the order completion rate and customer churn rate for specific order types, cannot be fully utilized. This makes it impossible to accurately estimate the driver's probability of accepting an order, resulting in quotes that may not be in line with actual conditions. This can lead to drivers rejecting orders or users abandoning orders due to excessive prices, resulting in customer churn. Furthermore, there is a lack of effective price processing mechanisms in the price display process. Directly displaying the exact price can make the price numerical value appear less concise and intuitive, increasing the user's decision-making costs and easily leading to double-billing. For example, users may later incur disputes over additional fees due to misunderstandings about price details. Therefore, a method that can comprehensively consider multiple factors and intelligently adjust transportation prices is urgently needed to address these issues.

[0003] For example, the Chinese patent application with publication number CN112862142A discloses a load and price forecasting and correction method, including: constructing a historical data set; constructing a second training set; constructing a forecasting model for forecasting, and correcting the forecast results. The construction of the second training set requires a first-level screening based on the day type of the day to be predicted and a second-level screening based on the average temperature of the day to be predicted. The steps of forecasting by the built forecasting model include two situations: point forecasting and interval forecasting. Both require first selecting a K-day data set that minimizes the error of the forecasting method result, and then predicting according to the forecasting algorithm, and correcting the forecast results. The parameter K selected from the historical data affected by time in the forecasting and correction method provided by this technical solution makes the forecast more accurate, and is applicable to the forecast of electricity prices and loads at the same time. It can perform both point forecasting and interval forecasting, and correct the forecast results, making it more accurate and flexible to use.

[0004] The above existing technologies have the following problems: the load, price forecasting and correction methods are mainly aimed at electricity price and load forecasting in the power sector, and the application scenarios are relatively single; when constructing the second training set, the load, price forecasting and correction methods are mainly screened based on day type and average temperature, and the use of historical data is relatively limited; although the prediction model used can perform point prediction and interval prediction, it is relatively fixed and only achieves different prediction methods by selecting different K-day data sets and prediction algorithms, so the model flexibility is poor; there is no mention of a mechanism for reverse updating the model or strategy based on actual results. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention proposes an intelligent transportation price correction method and system based on the hierarchical rounding rule. The method collects driver historical data, current order demand and idle driver data, and after preprocessing, uses the improved Transformer architecture model to calculate the driver's order acceptance probability, combines the order and driver data to calculate the initial quotation, and adjusts it according to the order acceptance probability through a nonlinear correction function to obtain a probability-weighted quotation; based on the probability-weighted quotation, the price range is divided, and a hierarchical rounding strategy is set. The gradient boosting tree model is used to predict the driver's acceptance of the rounded price, and whether to round the price twice is determined according to the acceptance threshold to determine the final quotation; the driver's order acceptance or rejection behavior is recorded, and the driver's order acceptance probability prediction model parameters and the hierarchical rounding strategy are reversely updated to form a closed-loop optimization; the present invention accurately corrects transportation prices through multi-model collaboration, dynamic parameter adjustment and feedback optimization, reduces the probability of secondary payment, and improves the platform's order transaction rate.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The intelligent transportation price correction method based on the layered zeroing rule includes:

[0008] Collect historical driver data to obtain data on current order demand and idle drivers on the platform; the driver historical data includes order completion rate, customer churn rate, and driver preference parameters;

[0009] Based on driver historical data and a driver acceptance probability prediction model built on an improved Transformer architecture, we calculate the probability of each driver accepting an order for the current order type.

[0010] The platform calculates an initial quote based on current order demand and available driver data, and uses a nonlinear correction function to adjust the initial quote based on the driver's probability of accepting the order, resulting in a probability-weighted quote.

[0011] Based on the price range of the probability-weighted quote, a layered zero-smoothing strategy is set. A gradient boosting tree model is used to predict drivers' acceptance of the probability-weighted quote after layered zero-smoothing. Combined with the acceptance threshold, a decision is made on whether to trigger a second zero-smoothing to obtain the final quote.

[0012] Record the driver's actual behavior of accepting or rejecting orders, and reversely update the order acceptance probability and tiered zeroing strategy based on the recorded results.

[0013] Specifically, the driver acceptance probability prediction model built based on the improved Transformer architecture is combined with the driver historical data to calculate each driver's acceptance probability for the current order type, including:

[0014] A1: Obtain historical driver data. Based on current order requirements, obtain order characteristics and use a boxplot to pre-process the driver data and order characteristics. Order characteristics include order type, transportation distance, and cargo weight.

[0015] A2: Load the improved Transformer architecture model and train it using the preprocessed driver history data and order features to obtain a trained improved Transformer architecture model; the improved Transformer architecture model is configured based on a multi-head attention mechanism;

[0016] A3: Real-time updated driver historical data and order features are input into the trained improved Transformer architecture model. After forward propagation calculations, the model uses a multi-head attention mechanism to extract contextual information about driver historical data and order features, capturing the dynamic correlations between different features. A feedforward neural network then performs nonlinear transformations to output the probability of each driver accepting an order for the current order type.

[0017] Specifically, the specific steps of A3 include:

[0018] A3.1: Obtain the latest driver status information and order characteristics in real time, and update driver historical data and order characteristics based on the latest status information and characteristics obtained in real time;

[0019] A3.2: Preprocess the updated driver history data and order features to obtain preprocessed real-time driver history data and order features. These preprocessed real-time driver history data and order features are organized into tensor form according to the model input format and then input into the trained improved Transformer architecture model.

[0020] A3.3: Map the pre-processed real-time driver history data and order feature data into a high-dimensional embedding space through linear transformation to generate query vectors, key vectors, and value vectors.

[0021] A3.4: Compute the similarity between the query vector and the key vector using the dot product operation and scale it by dividing by the dimension of the key vector to obtain the attention score.

[0022] A3.5: Apply the softmax function to the attention scores to convert them into probability distributions and obtain the attention weights.

[0023] A3.6: Take the weighted sum of the attention weights and the value vector to obtain the output of the multi-head attention mechanism;

[0024] A3.7: Concatenate the outputs of the multi-head attention mechanism and perform a linear transformation to obtain the final multi-head attention mechanism output.

[0025] A3.8: Input the final output of the multi-head attention mechanism into the first layer of the feedforward neural network and perform a linear transformation to obtain the first layer output.

[0026] A3.9: Input the output of the first layer into the second layer for a quadratic linear transformation to obtain the final output of the feedforward neural network.

[0027] A3.10: Map the final output of the feedforward neural network to the interval [0, 1] using the Sigmoid activation function to obtain the probability of each driver accepting the order for the current order type.

[0028] Specifically, the calculation of the initial quotation is based on the cumulative sum of mileage cost, time cost, service additional cost and environmental additional cost; the mileage cost is the product of the unit mileage cost and the estimated driving distance and road condition coefficient; the time cost is the product of the unit time cost and the estimated transportation time and time period coefficient.

[0029] Specifically, the calculation process of the probability weighted quotation includes:

[0030] Get the initial quote result and set the nonlinear correction function to the Sigmoid function;

[0031] The driver's order acceptance probability is used as the input of the nonlinear correction function to calculate the correction coefficient;

[0032] Calculate the product of the correction coefficient and the initial quotation to obtain the probability-weighted quotation.

[0033] Specifically, the specific process of setting the tiered zero-smoothing strategy based on the price range of the probability-weighted quotation includes:

[0034] If the probability-weighted quote is in the first price range, the unit digit is rounded off; the first price range is [0,100);

[0035] If the probability-weighted quote is in the second price range, it will be rounded down to zero; the second price range is [100,500);

[0036] If the probability-weighted quotation is in the third price range, it will be rounded to the nearest hundred; the third price range is [500,5000);

[0037] If the probability-weighted quotation is in the fourth price range, it will be rounded down to zero by rounding down the thousandth place; the fourth price range is [5000,+∞).

[0038] Specifically, the specific process of setting the tiered zeroing strategy based on the price range of the probability weighted quotation further includes: if the probability weighted quotation falls within the range of the price range threshold, then using fuzzy logic to make a judgment, and the using fuzzy logic to make a judgment includes:

[0039] If the driver's probability of accepting the order is greater than the preset probability threshold, a loose rounding method is used; the loose rounding method is to round up the corresponding price range;

[0040] If the probability of the driver accepting the order is less than or equal to the preset probability threshold of accepting the order, a strict rounding down method is adopted; the strict rounding down method is to round down the corresponding price range.

[0041] Specifically, the gradient boosting tree model is used to predict the driver's acceptance of the probability-weighted quotation after layered zeroing, and combined with the acceptance threshold, it is determined whether to trigger the second zeroing to obtain the final quotation, including:

[0042] B1: Obtain stratified, rounded-off probability-weighted quotes and driver historical data, and collect information on drivers' actual price acceptance.

[0043] B2: Load the pre-trained gradient boosting tree model;

[0044] B3: Input the probability-weighted quotes after stratification and the corresponding driver historical data into the trained gradient boosting tree model and output the predicted driver acceptance of the price.

[0045] B4: Set an acceptance threshold and compare the driver's predicted acceptance of the price with the acceptance threshold;

[0046] If the driver's predicted acceptance value for the price is greater than or equal to the acceptance threshold, the driver is deemed to have accepted the price and no secondary rounding is triggered. The current probability-weighted bid after rounding is the final bid.

[0047] If the driver's predicted acceptance value for the price is less than the acceptance threshold, it is considered that the driver does not accept the price, and a second round of price adjustment is triggered to obtain the new price after the second round of price adjustment.

[0048] B5: Input the new price after the second round of zeroing into the pre-trained gradient boosting tree model to predict the acceptance level. Repeat the judgment process of B3 and B4 until the acceptance threshold is met to obtain the final quotation.

[0049] Specifically, the triggering of the second round of zeroing and the second adjustment of the price are to adjust the probability-weighted quotes after the layered zeroing according to the rule of rounding down.

[0050] The intelligent transportation price correction system based on the hierarchical zero-rounding rule includes: data collection module, order acceptance probability prediction module, quotation calculation module, zero-rounding decision module, and feedback update module;

[0051] The data collection module is used to collect driver historical data, current order requirements and data on idle drivers on the platform;

[0052] The order acceptance probability prediction module is used to predict the probability of each driver accepting an order for the current order type based on the collected data;

[0053] The quotation calculation module is used to calculate the initial quotation of the order and adjust it according to the probability of the driver accepting the order to obtain a probability-weighted quotation;

[0054] The zero-wiping decision module is used to set a hierarchical zero-wiping strategy based on the probability-weighted quotation and determine whether a second zero-wiping is required to determine the final quotation;

[0055] The feedback update module is used to record the driver's actual order acceptance or rejection behavior, and to reversely update the order acceptance probability and the layered zeroing strategy based on the recording results.

[0056] Specifically, the zero erasure decision module includes: a hierarchical zero erasure setting unit, an acceptance prediction unit, and a secondary zero erasure judgment unit;

[0057] The tiered zero-smoothing setting unit is used to set a tiered zero-smoothing strategy based on the price range of the probability-weighted quotation;

[0058] The acceptance prediction unit is used to predict the driver's acceptance of the probability-weighted quotation after layered zeroing using a gradient boosting tree model;

[0059] The secondary zero-wiping judgment unit is used to determine whether to trigger the secondary zero-wiping in combination with the acceptance threshold, so as to obtain the final quotation.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] 1. The present invention proposes an intelligent transportation price correction system based on the hierarchical zeroing rule, and optimizes and improves the architecture, operation steps and processes. The system has the advantages of simple process, low investment and operation costs, and low production work costs.

[0062] 2. The present invention proposes an intelligent transportation price correction method based on the layered zeroing rule. Through multi-dimensional data collection and in-depth analysis, combined with advanced machine learning models, it achieves accurate correction of transportation prices. It uses an improved Transformer architecture to calculate the probability of drivers accepting orders, fully tapping the value of drivers' historical data to make price adjustments more in line with the actual possibility of accepting orders. The nonlinear correction function optimizes the initial quotation based on the probability of accepting orders and generates a probability-weighted quotation to ensure that the price attracts drivers to accept orders while taking into account the interests of both the platform and users. The layered zeroing strategy is combined with the gradient boosting tree model to ensure that the price display is simple and intuitive, and can accurately predict driver acceptance, reduce the risk of secondary payment, and effectively improve user experience and platform operation efficiency.

[0063] 3. The present invention proposes an intelligent transportation price correction method based on the layered zeroing rule. After recording the driver's order acceptance or rejection behavior, the driver's order acceptance probability prediction model parameters and the layered zeroing strategy are reversely updated, so that the system can be continuously optimized according to actual operating conditions. With the accumulation of data and the iteration of the model, the order acceptance probability prediction becomes more accurate and the layered zeroing strategy becomes more reasonable, forming a virtuous cycle of continuous improvement, improving the order completion rate, and enhancing the platform's competitiveness in the market. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 Schematic diagram of the intelligent correction method for transportation prices based on the layered zeroing rule of the present invention;

[0065] Figure 2 This is a flow chart showing the principle of the intelligent transportation price correction method based on the layered zeroing rule of the present invention;

[0066] Figure 3 This is an architecture diagram of the intelligent transportation price correction system based on the layered zeroing rule of the present invention. DETAILED DESCRIPTION

[0067] Example 1

[0068] See also Figure 1 and Figure 2 The present invention provides an embodiment of a method for intelligently correcting transportation prices based on a hierarchical zeroing rule, the method comprising steps S1 to S5, including the following steps:

[0069] S1: Collect historical driver data to obtain data on current order demand and idle drivers on the platform; the driver historical data includes order completion rate, customer churn rate, and driver tendency parameters;

[0070] Among them, the methods of obtaining driver historical data include: obtaining it from the platform's database, which records the order information of each driver, including order status, customer feedback, and order acceptance preferences. Among them, the order completion rate is calculated by counting the number of orders successfully completed by the driver and the total number of orders received; the customer churn rate is calculated by counting the number of customers who canceled orders due to poor driver service, price issues, etc., and then comparing it with the total number of orders received; with the help of machine learning algorithms, the driver's historical order records are analyzed to find out their preferences in order type, time, location, etc., so as to determine the driver's tendency parameters.

[0071] The process of obtaining current order requirements includes: users submit orders on the platform, filling in detailed information such as departure place, destination, cargo type, transportation time, etc.; the platform verifies the order information submitted by users to ensure the completeness and accuracy of the information; and classifies orders according to order characteristics, such as cargo type, transportation distance, etc., so as to subsequently match them with suitable drivers.

[0072] The process of obtaining data on idle drivers on the platform includes: the platform monitors the status of drivers in real time, determines which drivers are idle, and then extracts basic information of idle drivers from the database, such as vehicle type, driving experience, and previously calculated order conversion rate, customer churn rate and driver tendency parameters.

[0073] S2: Based on the driver's historical data and the driver acceptance probability prediction model built on the improved Transformer architecture, the driver acceptance probability for the current order type is calculated for each driver;

[0074] S3: The platform calculates an initial quote based on current order demand and available driver data. It then uses a nonlinear correction function to adjust the initial quote based on the driver's probability of accepting the order, resulting in a probability-weighted quote.

[0075] The calculation of the initial quotation is based on the cumulative sum of mileage cost, time cost, service additional cost and environmental additional cost; the mileage cost is the product of the unit mileage cost and the estimated driving distance and road condition coefficient; the time cost is the product of the unit time cost and the estimated transportation time and time period coefficient.

[0076] In the present invention, the road condition coefficient is a parameter used to reflect the impact of different road conditions on transportation costs. Road conditions, such as congestion level, road quality, and weather conditions, can significantly affect transportation efficiency and costs. Therefore, it is necessary to adjust the coefficient. The setting method includes:

[0077] (1) Data collection: Based on historical order data, analyze the differences between actual driving speed and standard speed for different routes and time periods;

[0078] (2) Road condition classification: road conditions are divided into multiple levels, for example:

[0079] If the road condition coefficient is less than 1, it means that the road is clear and the driving speed is higher than the standard speed;

[0080] If the road condition coefficient is equal to 1, it means standard road conditions and is calculated according to normal costs;

[0081] If the road condition coefficient is greater than 1, it means that the driving speed is lower than the standard speed and the cost increases;

[0082] (3) Coefficient calculation: The road condition coefficient value corresponding to each road condition level is determined by calculating the ratio of actual driving time to standard time.

[0083] The time factor reflects the impact of different time periods, such as daytime, nighttime, and peak hours, on transportation costs. Driver labor costs, market supply and demand, traffic conditions, and other factors vary in different time periods, so prices need to be adjusted. The setting methods include:

[0084] (1) Time division: divided into morning and evening commuting time, normal daytime hours and late night hours;

[0085] During the morning and evening working hours, set the time period coefficient to be greater than 1;

[0086] During normal daytime hours, set the time period coefficient equal to 1;

[0087] During the late night period, set the period coefficient to less than 1;

[0088] (2) Coefficient determination: Based on historical order data, the order volume, driver response rate and transaction price in different time periods are analyzed, and the optimal coefficient for each time period is determined through a machine learning model. The formula is: time period coefficient = MODE (time period demand, driver, historical order transaction rate), where MODE represents the machine learning model. The machine learning model is the existing technical content in this field and is not an inventive solution of this application, so it will not be elaborated here.

[0089] S4: Based on the price range of the probability-weighted quote, a tiered zero-smoothing strategy is set. A gradient boosting tree model is used to predict the driver's acceptance of the probability-weighted quote after tiered zero-smoothing. Combined with the acceptance threshold, a decision is made as to whether to trigger a second zero-smoothing process to obtain the final quote.

[0090] S5: Record the driver's actual behavior of accepting or rejecting orders, and reversely update the order acceptance probability and stratified zeroing strategy based on the recorded results.

[0091] Furthermore, the specific steps of S5 include:

[0092] (1) After the driver receives the quotation, the driver's actual behavior of accepting or rejecting the order is recorded to obtain the driver's behavior data. This process can be completed automatically through the platform's interactive interface. When the driver clicks the accept or reject button, the system immediately captures the operation information;

[0093] (2) Storing the recorded driver behavior data in the database and associating it with order information, such as order type, transportation distance, cargo weight, probability-weighted quotation, final quotation, driver history data, etc., for subsequent analysis;

[0094] (3) Filter the historical records of the current driver and order type from the database, including the previous probability of accepting orders and the actual order acceptance or rejection behavior;

[0095] (4) Calculate the update factor of the order acceptance probability based on the driver's behavior data and historical data. If the driver accepts the order, it means that the previous order acceptance probability was too low and needs to be improved; if the driver refuses the order, it means that the previous order acceptance probability was too high and needs to be reduced.

[0096] (5) Use the update factor of the order acceptance probability to update the current driver’s order acceptance probability for the order type;

[0097] (6) Based on the drivers’ actual behavior of accepting or rejecting orders, analyze the drivers’ acceptance of different rounded-off prices. For the cases of accepting orders, it indicates that the rounded-off price is within the driver’s acceptable range; for the cases of rejecting orders, it indicates that the rounded-off price may be too high or unreasonable.

[0098] (7) Based on the results of the acceptance analysis, adjust the zero-off interval and the zero-off ratio in the tiered zero-off strategy. For example, if it is found that the driver rejection rate within a price range is high, increase the zero-off ratio in that range to improve the driver's acceptance.

[0099] (8) Update the adjusted zero-rounding rule to the system for use in the quotation calculation of subsequent orders.

[0100] Based on the driver's historical data and the driver acceptance probability prediction model built based on the improved Transformer architecture, the driver acceptance probability of each driver for the current order type is calculated, including:

[0101] A1: Obtain historical driver data. Based on current order requirements, obtain order characteristics and use a boxplot to pre-process the driver data and order characteristics. Order characteristics include order type, transportation distance, and cargo weight.

[0102] Furthermore, the specific steps for preprocessing driver historical data and order characteristics using boxplots in A1 include:

[0103] (1) Sort each numerical variable in the driver history data and order feature data in ascending order;

[0104] (2) Determine the 25th, 50th, and 75th percentiles of the sorted data, which are recorded as the lower quartile, median, and upper quartile, respectively;

[0105] (3) The interquartile range is obtained by calculating the difference between the upper quartile and the lower quartile;

[0106] (4) The lower limit of the outlier range is obtained by calculating the difference between the lower quartile and 1.5 times the interquartile range;

[0107] (5) The upper limit of the outlier range is obtained by calculating the sum of the upper quartile and 1.5 times the interquartile range;

[0108] (6) Compare each numerical variable in the driver history data and order feature data with the lower limit of the outlier range and the upper limit of the outlier range. If it is not within the outlier range, it is an outlier and the median is used to replace the original numerical variable.

[0109] A2: Load the improved Transformer architecture model and train it using the preprocessed driver history data and order features to obtain a trained improved Transformer architecture model; the improved Transformer architecture model is configured based on a multi-head attention mechanism;

[0110] Furthermore, the specific steps of A2 include:

[0111] (1) Divide the preprocessed driver history data and order features into a training set and a validation set in a ratio of 7:3;

[0112] (2) Constructing a model structure based on a pre-designed improved Transformer architecture; the improved Transformer architecture adds a multi-head attention mechanism to the traditional Transformer architecture and performs adaptive weight optimization. The traditional Transformer architecture is the prior art content in this field and is not the inventive solution of this application, and will not be described in detail here;

[0113] (3) defining a loss function and an optimizer, wherein the loss function selects a binary cross entropy loss function; the optimizer selects stochastic gradient descent, and the binary cross entropy loss function and stochastic gradient descent are prior art contents in this field and are not the inventive solution of this application, and are not described in detail here;

[0114] (4) Input a batch of data from the training set into the model, perform calculations through each layer of the model, and finally obtain the model output, which is the predicted probability of each driver accepting the order for the current order type;

[0115] (5) Based on the model output and the true label, the loss value is calculated using the defined loss function, and based on the loss value, the gradient of the loss function with respect to the model parameters is calculated using the chain rule. The chain rule is a prior art in this field and is not an inventive solution of the present application, and is not described in detail here.

[0116] (6) Use the optimizer to update the model parameters according to the calculated gradient. At the same time, after each training cycle, the model is evaluated using the validation set to obtain the trained improved Transformer architecture model.

[0117] A3: Real-time updated driver historical data and order features are input into the trained improved Transformer architecture model. After forward propagation calculations, the model uses a multi-head attention mechanism to extract contextual information about driver historical data and order features, capturing the dynamic correlations between different features. A feedforward neural network then performs nonlinear transformations to output the probability of each driver accepting an order for the current order type.

[0118] The specific steps of A3 include:

[0119] A3.1: Obtain the latest driver status information and order characteristics in real time, and update driver historical data and order characteristics based on the latest status information and characteristics obtained in real time;

[0120] A3.2: Preprocess the updated driver history data and order features to obtain preprocessed real-time driver history data and order features. These preprocessed real-time driver history data and order features are organized into tensor form according to the model input format and then input into the trained improved Transformer architecture model.

[0121] The specific process of organizing the pre-processed real-time driver history data and order features into tensor form according to the model input format includes:

[0122] (1) Converting the pre-processed real-time driver history data and order features into numerical features, and standardizing the numerical features to make them have the same scale, so as to avoid certain features having too large an impact on the model due to their large numerical range. The standardization process adopts Z-score standardization, which is the existing technical content in this field and is not the inventive solution of this application, and will not be described in detail here;

[0123] (2) Further transform and combine the standardized numerical features according to business needs and model requirements. For example, calculate new features such as the average order acceptance interval in the driver's historical data and the unit weight transportation cost in the order features.

[0124] (3) Integrate the processed driver historical data and order features to form a complete data set in tensor form.

[0125] A3.3: Map the pre-processed real-time driver history data and order feature data into a high-dimensional embedding space through linear transformation to generate query vectors, key vectors, and value vectors.

[0126] For example, assume that the pre-processed real-time driver history data and order feature data are input as X, and the weight matrix of the embedding layer is 、 、 , then the query vector Q, key vector K and value vector V are: .

[0127] A3.4: Calculate the similarity between the query vector and the key vector using the dot product operation and scale it by dividing by the dimension of the key vector to get the attention score , where T represents transpose, represents the dimension of the key vector;

[0128] A3.5: Applying a softmax function to the attention scores , convert the attention score into a probability distribution and get the attention weight ;

[0129] A3.6: Take the weighted sum of the attention weights and the value vector to get the output of the multi-head attention mechanism , where i represents the number of heads computed in parallel in the multi-head attention mechanism;

[0130] A3.7: Concatenate the outputs of the multi-head attention mechanism and perform a linear transformation to obtain the final multi-head attention mechanism output ,in, represents the learnable weight matrix;

[0131] A3.8: The output of the final multi-head attention mechanism Input into the first layer of the feedforward neural network, perform linear transformation, and obtain the output result of the first layer ,in, represents the first layer weight matrix, represents the first layer bias vector, represents the maximum value function;

[0132] A3.9: Output the first layer Input to the second layer for quadratic linear transformation to obtain the final output of the feedforward neural network ,in, represents the second layer weight matrix, represents the second layer bias vector;

[0133] A3.10: Pass the final output of the feedforward neural network through the Sigmoid function Mapped to the interval [0,1], we can get the probability of each driver accepting the order for the current order type. , where the Sigmoid function is: , and e is an exponential constant.

[0134] The calculation process of the probability weighted quotation includes:

[0135] Get the initial quote result and set the nonlinear correction function to the Sigmoid function;

[0136] The driver's order acceptance probability p is used as the input of the nonlinear correction function to calculate the correction coefficient , where c represents the scaling variable parameter and b represents the offset variable parameter;

[0137] In this invention, in order to enhance the flexibility and adaptability of the function and enable it to better fit the data distribution and requirements in actual business scenarios, a scaling variable parameter c and an offset variable parameter b are introduced. The scaling variable parameter c is mainly used to control the steepness of the Sigmoid function curve. The larger the c value, the more sensitive the function is to changes in the input value. In the scenario of adjusting the quotation, if , can make a small change in the probability of a driver accepting an order have a greater impact on the quote, thereby emphasizing the adjustment effect of the probability of accepting an order on the quote. For example, in areas or time periods where drivers choose less, a small change in the probability of accepting an order may require a larger quote adjustment to attract drivers to accept orders. On the contrary, if The function curve is smooth, and the change of input value has relatively little impact on the output value. This can smooth the adjustment of quotation due to order acceptance probability to a certain extent and avoid excessive fluctuation of quotation. For example, in areas where supply and demand are relatively balanced and driver resources are relatively sufficient, there is no need to make overly sensitive adjustments to the quotation to keep the price relatively stable.

[0138] The offset variable parameter b is used to control the center position of the Sigmoid function curve. In the scenario of adjusting the quotation, the offset variable parameter b can adjust the central probability value that has a greater impact on the quotation due to the probability of accepting an order based on actual business needs. For example, if b is set to 0.6, it means that when the probability of the driver accepting an order is close to 0.6, the adjustment range of the quotation will be relatively large, and when the probability of accepting an order is far from 0.6, the adjustment range will be relatively small.

[0139] Calculate the product of the correction coefficient and the initial quotation to obtain the probability-weighted quotation.

[0140] The specific process of setting the tiered zero-smoothing strategy based on the price range of the probability-weighted quotation includes:

[0141] If the probability-weighted quote is in the first price range, the unit digit is rounded off; the first price range is [0,100);

[0142] For example, the price is 43 yuan, which is rounded to 40 yuan; the price is 47 yuan, which is rounded to 50 yuan.

[0143] If the probability-weighted quote is in the second price range, it will be rounded down to zero; the second price range is [100,500);

[0144] For example, the price is 123 yuan, which is rounded off to 120 yuan.

[0145] If the probability-weighted quotation is in the third price range, it will be rounded to the nearest hundred; the third price range is [500,5000);

[0146] For example, the price is 1,234 yuan, which is rounded to 1,200 yuan; the price is 1,268 yuan, which is rounded to 1,300 yuan.

[0147] If the probability-weighted quotation is in the fourth price range, it will be rounded down to zero by rounding down the thousandth place; the fourth price range is [5000,+∞).

[0148] For example, the price is 12,345 yuan, which is rounded off to 12,000 yuan.

[0149] The specific process of setting the tiered zeroing strategy based on the price range of the probability weighted quotation further includes: if the probability weighted quotation falls within the range of the price range critical value, then using fuzzy logic to make a judgment. For example, if the probability weighted quotation is 492 yuan, which is a value close to the price range of 500 yuan, then using fuzzy logic to make a judgment. The using fuzzy logic to make a judgment includes:

[0150] If the probability of the driver accepting the order is greater than the preset probability threshold, a loose rounding method is used; the loose rounding method is to round up the corresponding price range. For example, if the probability-weighted quote is 492 yuan, and the probability of the driver accepting the order is greater than the preset probability threshold, the rounded price is 500 yuan;

[0151] If the probability of the driver accepting the order is less than or equal to the preset probability threshold of accepting the order, the strict rounding-down method is adopted; the strict rounding-down method is to round down the corresponding price range. For example, if the probability-weighted quotation is 483 yuan, and the probability of the driver accepting the order is less than or equal to the preset probability threshold of accepting the order, the rounded-down price is 400 yuan.

[0152] The method uses the gradient boosting tree model to predict the driver's acceptance of the probability-weighted quotation after layered zeroing, combines the acceptance threshold, and determines whether to trigger the second zeroing to obtain the final quotation, including:

[0153] B1: Obtain stratified, rounded-off probability-weighted quotes and driver historical data, and collect information on drivers' actual price acceptance.

[0154] B2: Loading a pre-trained gradient boosting tree model. The gradient boosting tree model is a prior art in this field and is not an inventive solution of the present application, so it will not be described in detail here.

[0155] B3: Input the probability-weighted quotes after stratification and the corresponding driver historical data into the trained gradient boosting tree model and output the predicted driver acceptance of the price.

[0156] B4: Set an acceptance threshold and compare the driver's predicted acceptance of the price with the acceptance threshold;

[0157] If the driver's predicted acceptance value for the price is greater than or equal to the acceptance threshold, the driver is deemed to have accepted the price and no secondary rounding is triggered. The current probability-weighted bid after rounding is the final bid.

[0158] If the driver's predicted acceptance value for the price is less than the acceptance threshold, it is considered that the driver does not accept the price, and a second round of price adjustment is triggered to obtain the new price after the second round of price adjustment.

[0159] B5: Input the new price after the second round of zeroing into the pre-trained gradient boosting tree model to predict the acceptance level. Repeat the judgment process of B3 and B4 until the acceptance threshold is met to obtain the final quotation.

[0160] The triggering of the second round of zeroing and the second adjustment of the price are to adjust the probability weighted quotes after the layered zeroing according to the rule of rounding down.

[0161] Example 2

[0162] See also Figure 3 Another embodiment provided by the present invention is a transportation price intelligent correction system based on a hierarchical zeroing rule, comprising:

[0163] Data collection module, order probability prediction module, quotation calculation module, zero-clearing decision module, feedback update module;

[0164] The data collection module is used to collect historical data of drivers, current order demands, and data of idle drivers on the platform;

[0165] The order acceptance probability prediction module is used to predict the probability of each driver accepting an order for the current order type based on the collected data;

[0166] The quotation calculation module is used to calculate the initial quotation of the order and adjust it according to the probability of the driver accepting the order to obtain a probability-weighted quotation;

[0167] The zero-smoothing decision module is used to set a hierarchical zero-smoothing strategy based on probability-weighted quotes and determine whether a second zero-smoothing is required to determine the final quote;

[0168] The feedback update module is used to record the driver's actual order acceptance or rejection behavior, and reversely update the order acceptance probability and layered zeroing strategy based on the recorded results, enabling the system to continuously learn and optimize.

[0169] The order acceptance probability prediction module includes: a model building unit and a probability calculation unit;

[0170] A model building unit, which is used to build a driver acceptance probability prediction model based on an improved Transformer architecture. This model can handle complex sequence data and capture long-term dependencies in the data to improve prediction accuracy.

[0171] The probability calculation unit is used to use the constructed model, combined with the driver's historical data and current order information, to calculate the probability of each driver accepting an order for the current order type.

[0172] The quotation calculation module includes: an initial quotation calculation unit and a nonlinear correction unit;

[0173] The initial quotation calculation unit is used to calculate an initial quotation based on the current order demand and idle driver data, which is the basis for the quotation;

[0174] The nonlinear correction unit is used to use a nonlinear correction function to adjust the initial quotation according to the probability of the driver accepting the order, thereby obtaining a probability-weighted quotation, so that the quotation can better reflect the actual situation and the possibility of the driver accepting the order.

[0175] The zero erasure decision module includes: a hierarchical zero erasure setting unit, an acceptance prediction unit, and a secondary zero erasure judgment unit;

[0176] The tiered zero-smoothing setting unit is used to set the tiered zero-smoothing strategy based on the price range of the probability-weighted quotation and determine the zero-smoothing rules for different price ranges;

[0177] The acceptance prediction unit is used to predict the driver's acceptance of the probability-weighted quote after layered zeroing using the gradient boosting tree model. The gradient boosting tree model can handle complex nonlinear relationships and improve the accuracy of prediction;

[0178] The secondary zero-wiping judgment unit is used to determine whether to trigger the secondary zero-wiping based on the acceptance threshold, so as to obtain the final quotation.

[0179] The feedback update module includes: behavior recording unit, order acceptance probability update unit, and zero-clearing strategy update unit;

[0180] The behavior recording unit is responsible for recording the driver's actual behavior of accepting or rejecting orders, providing a basis for feedback updates of the system;

[0181] The order acceptance probability update unit is used to update the previously calculated order acceptance probability based on the driver's actual behavior, so that the order acceptance probability is more consistent with the actual situation;

[0182] The zero-rounding strategy update unit is used to adjust and optimize the layered zero-rounding strategy based on the drivers' acceptance of prices to improve the performance and adaptability of the system.

[0183] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific embodiments. The above-mentioned specific embodiments are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also change, modify, replace and modify the above-mentioned embodiments without departing from the purpose and scope of protection of the present invention. These are all protected by the present invention.

[0184] If the technical solution disclosed herein involves personal information, the product using the technical solution disclosed herein has clearly informed the individual of the personal information processing rules and obtained the individual's voluntary consent before processing the personal information. If the technical solution disclosed herein involves sensitive personal information, the product using the technical solution disclosed herein has obtained the individual's separate consent before processing the sensitive personal information and at the same time meets the "explicit consent" requirement. For example, on personal information collection devices such as cameras, a clear and prominent sign is set to inform that the individual has entered the personal information collection scope and that personal information will be collected. If the individual voluntarily enters the collection scope, it is deemed that they agree to the collection of their personal information; or on the personal information processing device, when the personal information processing rules are notified by obvious signs / information, the individual's authorization is obtained through pop-up information or by asking the individual to upload their personal information. The personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the type of personal information processed.

Claims

1. The intelligent correction method of transportation price based on the layered zeroing rule is characterized by: include: Collect historical data of drivers, obtain data on current order demand and idle drivers on the platform; The driver's historical data includes order completion rate, customer churn rate and driver tendency parameters; Based on driver historical data and a driver acceptance probability prediction model built on an improved Transformer architecture, we calculate the probability of each driver accepting an order for the current order type. The platform calculates an initial quote based on current order demand and available driver data, and uses a nonlinear correction function to adjust the initial quote based on the driver's probability of accepting the order, resulting in a probability-weighted quote. Based on the price range of the probability-weighted quote, a layered zero-smoothing strategy is set. A gradient boosting tree model is used to predict drivers' acceptance of the probability-weighted quote after layered zero-smoothing. Combined with the acceptance threshold, a decision is made on whether to trigger a second zero-smoothing to obtain the final quote. Record the driver's actual behavior of accepting or rejecting orders, and reversely update the order acceptance probability and tiered zeroing strategy based on the recorded results.

2. The intelligent transportation price correction method based on the layered zeroing rule according to claim 1 is characterized in that: Based on the driver's historical data and the driver acceptance probability prediction model built based on the improved Transformer architecture, the driver acceptance probability of each driver for the current order type is calculated, including: A1: Obtain historical driver data. Based on current order requirements, obtain order characteristics and use a boxplot to pre-process the driver data and order characteristics. Order characteristics include order type, transportation distance, and cargo weight. A2: Load the improved Transformer architecture model and train it using the preprocessed driver history data and order features to obtain a trained improved Transformer architecture model; the improved Transformer architecture model is configured based on a multi-head attention mechanism; A3: Real-time updated driver historical data and order features are input into the trained improved Transformer architecture model. After forward propagation calculations, the model uses a multi-head attention mechanism to extract contextual information about driver historical data and order features, capturing the dynamic correlations between different features. A feedforward neural network then performs nonlinear transformations to output the probability of each driver accepting an order for the current order type.

3. The intelligent correction method for transportation prices based on the layered zeroing rule according to claim 2 is characterized in that: The specific steps of A3 include: A3.1: Obtain the latest driver status information and order characteristics in real time, and update driver historical data and order characteristics based on the latest status information and characteristics obtained in real time; A3.2: Preprocess the updated driver history data and order features to obtain preprocessed real-time driver history data and order features. These preprocessed real-time driver history data and order features are organized into tensor form according to the model input format and then input into the trained improved Transformer architecture model. A3.3: Map the pre-processed real-time driver history data and order feature data into a high-dimensional embedding space through linear transformation to generate query vectors, key vectors, and value vectors. A3.4: Compute the similarity between the query vector and the key vector using the dot product operation and scale it by dividing by the dimension of the key vector to obtain the attention score. A3.5: Apply the softmax function to the attention scores to convert them into probability distributions and obtain the attention weights. A3.6: Take the weighted sum of the attention weights and the value vector to obtain the output of the multi-head attention mechanism; A3.7: Concatenate the outputs of the multi-head attention mechanism and perform a linear transformation to obtain the final multi-head attention mechanism output. A3.8: Input the final output of the multi-head attention mechanism into the first layer of the feedforward neural network and perform a linear transformation to obtain the first layer output. A3.9: Input the output of the first layer into the second layer for a quadratic linear transformation to obtain the final output of the feedforward neural network. A3.10: Map the final output of the feedforward neural network to the interval [0, 1] using the Sigmoid activation function to obtain the probability of each driver accepting the order for the current order type.

4. The intelligent transportation price correction method based on the layered zeroing rule according to claim 3 is characterized in that: The calculation of the initial quotation is based on the cumulative sum of mileage cost, time cost, service additional cost and environmental additional cost; the mileage cost is the product of the unit mileage cost and the estimated driving distance and road condition coefficient; the time cost is the product of the unit time cost and the estimated transportation time and time period coefficient.

5. The intelligent correction method for transportation prices based on the layered zeroing rule according to claim 4 is characterized in that: The calculation process of the probability weighted quotation includes: Get the initial quote result and set the nonlinear correction function to the Sigmoid function; The driver's order acceptance probability is used as the input of the nonlinear correction function to calculate the correction coefficient; Calculate the product of the correction coefficient and the initial quotation to obtain the probability-weighted quotation.

6. The intelligent transportation price correction method based on the layered zeroing rule according to claim 5 is characterized in that: The specific process of setting the tiered zero-smoothing strategy based on the price range of the probability-weighted quotation includes: If the probability-weighted quote is in the first price range, the unit digit is rounded off; the first price range is [0,100); If the probability-weighted quote is in the second price range, it will be rounded down to zero; the second price range is [100,500); If the probability-weighted quotation is in the third price range, it will be rounded to the nearest hundred; the third price range is [500,5000); If the probability-weighted quotation is in the fourth price range, it will be rounded down to zero by rounding down the thousandth place; the fourth price range is [5000,+∞).

7. The intelligent transportation price correction method based on the layered zeroing rule according to claim 6 is characterized in that: The specific process of setting the tiered zeroing strategy based on the price range of the probability weighted quotation further includes: if the probability weighted quotation falls within the range of the price range threshold, then using fuzzy logic to make a judgment, and the using fuzzy logic to make a judgment includes: If the driver's probability of accepting the order is greater than the preset probability threshold, a loose rounding method is used; the loose rounding method is to round up the corresponding price range; If the probability of the driver accepting the order is less than or equal to the preset probability threshold of accepting the order, a strict rounding down method is adopted; the strict rounding down method is to round down the corresponding price range.

8. The intelligent transportation price correction method based on the layered zeroing rule according to claim 7 is characterized in that: The method uses the gradient boosting tree model to predict the driver's acceptance of the probability-weighted quotation after layered zeroing, combines the acceptance threshold, and determines whether to trigger the second zeroing to obtain the final quotation, including: B1: Obtain stratified, rounded-off probability-weighted quotes and driver historical data, and collect information on drivers' actual price acceptance. B2: Load the pre-trained gradient boosting tree model; B3: Input the probability-weighted quotes after stratification and the corresponding driver historical data into the trained gradient boosting tree model and output the predicted driver acceptance of the price. B4: Set an acceptance threshold and compare the driver's predicted acceptance of the price with the acceptance threshold; If the driver's predicted acceptance value for the price is greater than or equal to the acceptance threshold, the driver is deemed to have accepted the price and no secondary rounding is triggered. The current probability-weighted bid after rounding is the final bid. If the driver's predicted acceptance value for the price is less than the acceptance threshold, it is considered that the driver does not accept the price, and a second round of price adjustment is triggered to obtain the new price after the second round of price adjustment. B5: Input the new price after the second round of zeroing into the pre-trained gradient boosting tree model to predict the acceptance level. Repeat the judgment process of B3 and B4 until the acceptance threshold is met to obtain the final quotation.

9. The intelligent transportation price correction method based on the layered zeroing rule according to claim 8 is characterized in that: The triggering of the second round of zeroing and the second adjustment of the price are to adjust the probability weighted quotes after the layered zeroing according to the rule of rounding down.

10. A transportation price intelligent correction system based on a hierarchical zero-rounding rule, which is used to implement the transportation price intelligent correction method based on a hierarchical zero-rounding rule according to any one of claims 1 to 9, characterized in that: include: Data collection module, order probability prediction module, quotation calculation module, zero-clearing decision module, feedback update module; The data collection module is used to collect driver historical data, current order requirements and data on idle drivers on the platform; The order acceptance probability prediction module is used to predict the probability of each driver accepting an order for the current order type based on the collected data; The quotation calculation module is used to calculate the initial quotation of the order and adjust it according to the probability of the driver accepting the order to obtain a probability-weighted quotation; The zero-wiping decision module is used to set a hierarchical zero-wiping strategy based on the probability-weighted quotation and determine whether a second zero-wiping is required to determine the final quotation; The feedback update module is used to record the driver's actual order acceptance or rejection behavior, and to reversely update the order acceptance probability and the layered zeroing strategy based on the recording results.

11. The intelligent transportation price correction system based on the hierarchical zeroing rule according to claim 10 is characterized in that: The zero erasure decision module includes: a hierarchical zero erasure setting unit, an acceptance prediction unit, and a secondary zero erasure judgment unit; The tiered zero-smoothing setting unit is used to set a tiered zero-smoothing strategy based on the price range of the probability-weighted quotation; The acceptance prediction unit is used to predict the driver's acceptance of the probability-weighted quotation after layered zeroing using a gradient boosting tree model; The secondary zero-wiping judgment unit is used to determine whether to trigger the secondary zero-wiping in combination with the acceptance threshold, so as to obtain the final quotation.

Images