Freight freight prediction method and system based on multi-dimensional feature fusion

Abstract

The invention provides a freight freight prediction method and system based on multi-dimensional feature fusion, and the method comprises the steps: constructing a multi-dimensional fusion feature based on an interaction relation between a multi-dimensional waybill feature in waybill data and a multi-dimensional external influence factor in external influence data; and obtaining a plurality of prediction sub-models based on multi-dimensional fusion feature training. And combining the plurality of prediction sub-models by adopting a weighted average method to obtain a prediction model. And under the condition that the query request is obtained, performing semantic analysis on the query request to obtain standardized query information, and inputting the query information into the prediction model to obtain freight freight prediction information corresponding to the query information. According to the scheme, the influence of external factors on freight prediction is introduced, the freight prediction accuracy is remarkably improved, the multiple prediction sub-models are combined according to the weight to obtain the prediction model, and the adaptability in different scenes is improved. The complete query information is obtained by performing semantic analysis on the query request, so that accurate execution of prediction can be guided, and the use experience of a user is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the application technology of artificial intelligence in the field of logistics transportation, in particular to the feature engineering and model integration technology in machine learning, and more particularly to a freight intelligent prediction method and system based on multi-dimensional feature fusion and multi-model integration. BACKGROUND

[0002] With the improvement of the informatization level of the logistics industry, the prediction of road freight is gradually developing from complete reliance on manual experience to the use of information systems for decision-making assistance. The current mainstream freight pricing and prediction technical solutions usually include data collection, feature processing, pricing model, manual intervention and result output. Among them, the data collection link mainly collects historical shipping data, including basic information such as starting point, ending point, cargo type, cargo weight, transportation distance, actual freight, etc. The feature processing link usually takes transportation distance and cargo weight as the core variables, and some systems will simply introduce limited features such as vehicle type, cargo type, seasonal coefficient, etc., and use linear combination or simple conditional judgment rules to perform feature processing. The pricing model link usually uses traditional algorithms such as linear regression, segmented pricing table, interval mapping, etc. to calculate and predict freight. The manual intervention link adjusts the model calculation results according to market conditions, holidays, oil prices and other subjective factors, usually accounting for 30%-50% of the weight of pricing decisions. Finally, in the result output link, the freight suggestion is output through the information system or manually for reference by business personnel.

[0003] In actual application, some large logistics platforms will introduce simple rule engines or interval recommendations based on historical average values, but the overall still focuses on distance and weight, lacking consideration of the impact of external environment or market factors on freight. Moreover, there may be a problem of weak adaptability of model types in model application, resulting in poor prediction accuracy. In addition, the existing prediction system usually has the defect of rigid interaction mode, requiring users to strictly input all parameters according to the preset format. Once the input is not standardized or incomplete, the prediction cannot be performed or the result is wrong, seriously affecting the usability and work efficiency of the system.

[0004] The existing technology has the following technical problems: First, the feature dimension is severely limited. The traditional method only considers 2-3 basic variables such as transportation distance and cargo weight, ignoring the key external factors affecting the freight: (1) geographical factors: such as whether the starting and ending points are logistics hub cities, road quality, etc.; (2) time factors: such as holiday transportation capacity shortage, seasonal demand fluctuations, etc.; (3) market factors: such as oil price fluctuations, freight index fluctuations, raw material price transmission effects, etc.; (4) supply and demand factors of transportation capacity: such as the number of regional trucks, load capacity distribution, return source probability, etc. This single-dimensional feature extraction leads to the inability of the prediction model to accurately reflect the real freight formation mechanism. Second, the model algorithm is not adaptable. The existing method mainly uses traditional algorithms such as linear regression and segmented pricing table, which cannot handle the non-linear relationship and complex interaction effect between multiple factors. For example, it cannot model the three-way interaction effect of "weight x distance x holiday coefficient", nor can it capture the dynamic correlation between "oil price rise → increased sensitivity of long-distance transportation costs". The single model has poor adaptability when facing different transportation scenarios, resulting in a sharp decline in prediction accuracy under certain conditions. Third, real-time and adaptive capabilities are lacking. The existing system lacks real-time data updating mechanism and cannot respond to market environment changes in time. For example, when a sudden weather event causes transportation capacity shortage, the system cannot automatically adjust the prediction model; when the oil price fluctuates greatly, the cost structure cannot be updated in real time. The model parameters are fixed, and there is no automatic learning and optimization capability based on new data, resulting in prediction results lagging behind the actual market situation. Fourth, the human-computer interaction experience is poor. The existing system requires users to strictly input all transportation parameters according to the preset format, and lacks intelligent completion capability when parameters are missing. When the user inputs incompletely or incorrectly, the system cannot make predictions or give incorrect results, seriously affecting the practicality and user experience. The above technical problems result in the prediction accuracy of the existing freight rate prediction method being generally lower than 70%, and a large amount of manual intervention is required for experience adjustment, which cannot meet the demand of modern logistics industry for high-precision, fast-response, and intelligent freight rate prediction. SUMMARY

[0005] The purpose of the embodiment of the present application is to provide a freight rate prediction method and system based on multi-dimensional feature fusion, to improve prediction accuracy and adaptability to different scenarios.

[0006] In a first aspect, the present application provides a freight rate prediction method based on multi-dimensional feature fusion, comprising: collecting waybill data and external influence data, the waybill data including multi-dimensional waybill features, and the external influence data including multi-dimensional external influence factors; based on the interaction between the multi-dimensional waybill features and the multi-dimensional external influence factors, a multi-dimensional fusion feature is constructed; training a plurality of basic machine learning models based on the multi-dimensional fusion features to obtain a plurality of prediction sub-models corresponding to the basic machine learning models, and combining the plurality of prediction sub-models by using a weighted average method to obtain a prediction model; In the case of obtaining a query request, performing semantic analysis on the query request to obtain standardized query information, inputting the query information into the prediction model to obtain freight rate prediction information corresponding to the query information.

[0007] In an optional implementation, the step of constructing the multi-dimensional fusion features based on the interaction relationship between the multi-dimensional waybill features and the multi-dimensional external influencing factors comprises: extracting multi-dimensional features from the multi-dimensional waybill features and the multi-dimensional external influencing factors in terms of a basic level, a geographical level, a time level, a market level, and a weather level; According to a business rule, the multi-dimensional fusion features are obtained by interacting and modeling the extracted multi-dimensional features based on the interaction relationship between the multi-dimensional features.

[0008] In an optional implementation, the method further comprises a step of preprocessing the multi-dimensional features, which comprises: performing feature standardization processing on the multi-dimensional features, and detecting and filtering out multi-dimensional features with abnormalities by using a quartile interval method; performing period transformation processing on periodic features in the multi-dimensional features; performing encoding processing on literal features in the multi-dimensional features.

[0009] In an optional implementation, the external influencing data comprises geographical data, market data, transport capacity data, weather data, time data, and policy data. The geographical data comprises geographical coordinates, administrative region division, and route planning. The market data comprises freight rate index, oil price, raw material price, and freight volume. The transport capacity data comprises real-time truck quantity in each region, load capacity distribution, and transport capacity supply and demand status. The weather data comprises weather status of the start point and the end point of each waybill and an impact score on transportation. The time data comprises holidays, workdays, and seasons. The policy data comprises environmental protection restriction information, road and bridge fee adjustment information, fuel tax information, and over-limit management standard information.

[0010] In an optional implementation, the multi-dimensional fusion features comprise a plurality of the following features: transport load related features, time period related features, market related features, supply and demand related features, and policy related features.

[0011] In an optional implementation, the transportation load-related features include a plurality of features constructed from interactions among load types, load weights, transportation distances, and transportation efficiencies. The time period-related features include a plurality of features constructed from linkages among holidays, seasons, periodic times, and market fluctuations.

[0012] In an optional implementation, the market-related features include a plurality of features constructed from correlations among regional economies, oil prices, raw material prices, and market fluctuations. The supply and demand-related features include a plurality of features constructed from linkages among transportation risks, transportation capacity matching, and regional markets. The policy-related features include a plurality of features constructed from linkages between environmental protection policies and cost fluctuations.

[0013] In an optional implementation, the method further includes: obtaining updated waybill data, and performing drift detection based on the updated waybill data in a case where the updated waybill data exceeds a preset threshold; in a case where the drift detection result exceeds a set threshold, triggering full training of the prediction model to obtain an updated prediction model; in a case where the drift detection result does not exceed the set threshold, triggering incremental training of the prediction model to obtain the updated prediction model; performing performance verification on the updated prediction model, and deploying the updated prediction model in a case where the performance verification is passed, or maintaining the original prediction model otherwise.

[0014] In an optional implementation, the step of performing semantic analysis on the query request to obtain standardized query information includes: performing semantic analysis on the query request to determine missing transportation parameters; generating inquiry information based on the missing transportation parameters, and obtaining feedback information obtained in response to the inquiry information; based on the feedback information, performing standardized processing according to business rules to supplement the missing transportation parameters to obtain standardized query information.

[0015] In a second aspect, the present application provides a freight transportation fare prediction system based on multi-dimensional feature fusion, which includes: a collection module configured to collect waybill data and external influence data, the waybill data including multi-dimensional waybill features, and the external influence data including multi-dimensional external influence factors; a construction module configured to construct multi-dimensional fusion features based on interactions between the multi-dimensional waybill features and the multi-dimensional external influence factors. a training module configured to train a plurality of basic machine learning models based on the multi-dimensional fusion features to obtain prediction sub-models corresponding to the basic machine learning models, and combine the prediction sub-models by using a weighted average method to obtain a prediction model; a query prediction module configured to, in a case where a query request is obtained, perform semantic analysis on the query request to obtain standardized query information, input the query information into the prediction model, and obtain freight rate prediction information corresponding to the query information.

[0016] The application provides a freight rate prediction method and system based on multi-dimensional feature fusion. Multi-dimensional fusion features are constructed based on the interaction between multi-dimensional waybill features in waybill data and multi-dimensional external influencing factors in external influencing data. A plurality of basic machine learning models are trained based on the multi-dimensional fusion features to obtain prediction sub-models corresponding to the basic machine learning models. The prediction sub-models are combined by using a weighted average method to obtain a prediction model. In a case where a query request is obtained, semantic analysis is performed on the query request to obtain standardized query information. The query information is input into the prediction model to obtain freight rate prediction information corresponding to the query information.

[0017] The application combines the interaction between waybill data and external influencing data to construct multi-dimensional fusion features, introduces the influence of external factors on freight rate prediction, significantly improves the accuracy of freight rate prediction, and improves the adaptability in different scenarios by combining the prediction sub-models according to weights to obtain a prediction model. In addition, semantic analysis is performed on a query request to obtain complete query information, which can guide the accurate execution of prediction and improve the user experience. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiments of the application. It should be understood that the following drawings only show some embodiments of the application, and therefore should not be regarded as a limitation to the scope. Other related drawings can also be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0019] Figure 1 A flowchart of a freight rate prediction method based on multi-dimensional feature fusion provided by the embodiments of the application is shown in the figure. Figure 2 A hierarchical architecture schematic diagram of a freight rate prediction system based on multi-dimensional feature fusion provided by the embodiments of the application is shown in the figure. Figure 3 A flowchart of the sub-steps included in S12 in the figure. Figure 1 Figure 4 ​A flowchart of a preprocessing method in the freight rate prediction method provided by an embodiment of the present invention; Figure 5 This is a logic diagram of the raw material price transmission mechanism in an embodiment of the present invention; Figure 6 This is a logical diagram of the transport capacity supply and demand balance analysis characteristics, transport capacity asymmetry effect characteristics, and professional transport capacity matching characteristics in an embodiment of the present invention; Figure 7 A flow chart of a model updating method in the freight rate prediction method provided by an embodiment of the present invention; Figure 8 for Figure 1 Flowchart of the sub-steps included in S14; Figure 9 A functional module block diagram of a freight rate forecasting system provided by an embodiment of the present invention; Figure 10 This is a structural block diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present invention will be described below with reference to the accompanying drawings in the embodiments of the present invention.

[0021] See also Figure 1 The following is a flowchart of a freight rate prediction method based on multidimensional feature fusion according to an embodiment of the present invention. This method can be performed by a freight rate prediction system based on multidimensional feature fusion. This system can be implemented using software and / or hardware and can be configured in an electronic device such as a computer, server, or programmable logic controller. The detailed steps of this method are described below.

[0022] S11, collecting waybill data and external influence data, where the waybill data includes multi-dimensional waybill features, and the external influence data includes multi-dimensional external influence factors.

[0023] S12, based on the interactive relationship between multi-dimensional waybill features and multi-dimensional external influencing factors, a multi-dimensional fusion feature is constructed.

[0024] S13, training multiple basic machine learning models based on multi-dimensional fusion features to obtain prediction sub-models corresponding to each basic machine learning model, and combining multiple prediction sub-models using a weighted average method to obtain a prediction model.

[0025] S14, when a query request is obtained, performing semantic analysis on the query request to obtain standardized query information, inputting the query information into a prediction model, and obtaining freight rate prediction information corresponding to the query information.

[0026] Please refer to Figure 2 The hierarchical architecture of the freight cost prediction system provided by the embodiments of the present application is shown schematically, and the hierarchical architecture includes a data collection layer, a feature engineering layer, a model training layer, a prediction service layer, and a user interaction layer.

[0027] The data collection layer involves an electronic waybill system, a user reporting system, and external data sources. The feature engineering layer involves multi-dimensional feature extraction, automatic interaction discovery, and intelligent preprocessing. The model training layer involves multi-model fusion, real-time training scheduling, and parameter optimization. The prediction service layer involves high-performance services, model management, and cache optimization. The user interaction layer involves natural language processing, multi-end interfaces, and intelligent interaction.

[0028] The various levels are connected through bidirectional data flow and feedback mechanisms, forming a closed-loop automatic optimization system.

[0029] Data collection is performed at the data collection layer, and real-time data updating is achieved. In the prior art, when predicting freight costs, only basic variables such as distance and weight are considered, and the complex effects of geographic spatial differences, time periodicity, market dynamics, weather conditions, and cargo risks on freight rates are ignored, resulting in predicted results deviating from actual market conditions.

[0030] To address the above problems, unlike the prior art, which only collects waybill data, in the present embodiment, the collection of multi-source heterogeneous data is achieved. In the present embodiment, waybill data is collected from a historical waybill database and an electronic waybill system in real time. A real-time data interface is established to interface with the electronic waybill system, and when a new waybill is generated, data synchronization and incremental training are automatically triggered. Data timeliness needs to be ensured to ensure that newly generated waybill data is promptly incorporated into model training, so that the prediction model can reflect the latest market conditions.

[0031] In addition, in the present embodiment, a user-initiated reporting mechanism can also be established, such as establishing a user freight data reporting channel, to collect real-time market freight information. The data reported by the user needs to be subjected to multi-dimensional quality inspection, including inspection and processing such as data reasonableness inspection, outlier identification, and duplicate data removal, to ensure data quality. The above data also needs to be audited, and the auditing method can be, for example, an automatic preliminary review plus manual review method, to further ensure the quality of the reported data. The reported data that passes the audit can be automatically incorporated into the training data set to achieve dynamic updating of the model.

[0032] The waybill data includes multi-dimensional waybill features, such as origin, destination, cargo type, weight, and actual freight.

[0033] The external influence data includes geographic data, market data, transport capacity data, weather data, time data, and policy data. Each item of external influence data includes multi-dimensional external influence factors.

[0034] For example, the geographic data includes geographic coordinates, administrative division, route planning, etc. The market data includes freight rate index, oil price, raw material price (mainly including, such as aluminum price, steel price, coal price, etc.), freight volume, etc. Thus, it is beneficial to analyze the indirect influence of raw material price fluctuation on transportation demand through the price transmission mechanism subsequently.

[0035] The transport capacity data includes real-time truck quantity in each region, load capacity distribution, transport capacity supply and demand status, etc. For example, it includes multi-channel transport capacity information such as truck GPS position data, vehicle registration information, transport capacity platform data, logistics park vehicle statistics, etc.

[0036] The weather data includes weather conditions of the starting point and the ending point of each waybill and the influence score on transportation, etc. The time data includes holidays, weekdays, seasons, etc.

[0037] The policy data includes environmental protection restriction information, road and bridge fee adjustment information, fuel tax information, over-limit management standard information, etc., so as to help analyze the influence of policy changes on transportation cost and transport capacity supply.

[0038] In this embodiment, for the collected data, real-time data stream processing and batch data synchronization processing are supported to ensure the integrity and timeliness of the data. In addition, data quality evaluation indicators are established to monitor the data integrity, accuracy and consistency in real time. Further, data anomalies can be automatically identified and processed, including missing value filling, data cleaning, etc.

[0039] On the basis of realizing the above data collection and processing through the data collection layer, feature extraction and interactive fusion are realized through the feature engineering layer.

[0040] In this embodiment, in the feature engineering layer, in the process of constructing the multi-dimensional fusion features based on the interactive relationship between the multi-dimensional waybill features and the multi-dimensional external influencing factors, the following methods can be used, please refer to Figure 3 : S121, multi-dimensional feature extraction is performed on the multi-dimensional waybill features and the multi-dimensional external influencing factors from the basic level, the geographic level, the time level, the market level and the weather level.

[0041] S122, according to the business rules, multi-dimensional fusion features are obtained by interactive modeling based on the interactive relationship between the multi-dimensional features.

[0042] In this embodiment, the so-called basic level mainly refers to the basic features in the waybill data, including, such as starting point, ending point, cargo weight, distance, etc. The geographic level involves, such as coordinates, road conditions, hubs, etc. The time level involves, such as date, holiday, season, etc. The market level involves, such as oil price, freight rate index, raw material price, etc. The weather level involves specific weather conditions.

[0043] Based on this, specifically, the cargo weight, transportation distance features are extracted from the basic level, the geographic coordinates, hub city identification features are extracted from the geographic level, the holiday identification, season coding features are extracted from the time level, the oil price, freight index features are extracted from the market level, and the starting and ending weather score features are extracted from the weather level.

[0044] The specific implementation of the multi-dimensional feature extraction is as follows: Basic level feature extraction: Cargo weight feature: directly extract the cargo weight value (unit: ton) in the waybill.

[0045] Transportation distance feature: the actual driving distance (unit: kilometers) calculated based on the starting and ending point coordinates.

[0046] Cargo type feature: standardize the classification of cargo name and use one-hot encoding processing.

[0047] Transportation time efficiency requirement feature: divided into ordinary (1), urgent (2), and special urgent (3) according to customer requirements.

[0048] Geographic level feature extraction: Geographic coordinate feature: extract the latitude and longitude coordinates of the starting and ending points and normalize them to the [0, 1] interval.

[0049] Administrative region feature: extract the province, city, and county-level administrative division of the starting and ending points and use label encoding.

[0050] Hub city feature: based on the list of logistics hub cities, mark whether the starting and ending points are hubs (0 / 1 binary feature).

[0051] Route quality feature: calculate the route quality score based on the highway coverage rate, ranging from 0 to 1.

[0052] Time level feature extraction: Date feature: extract the year, month, day, and week of the query date and encode them respectively.

[0053] Holiday feature: based on the national statutory holiday calendar, mark whether it is a holiday and the type of holiday.

[0054] Season feature: according to the month, divide into spring (3-5), summer (6-8), autumn (9-11), and winter (12-2), and use one-hot encoding.

[0055] Periodic feature: perform sine and cosine transformation on the month and week, such as the month sine value = sin(2π×month / 12).

[0056] Market level feature extraction: Oil price feature: Extract the diesel price on the query date (unit: yuan / L) and perform standardization processing.

[0057] Freight rate index feature: Extract the highway freight price index corresponding to the route, with a base value of 100.

[0058] Raw material price feature: Extract the relevant raw material prices (such as aluminum price, steel price, etc.) according to the type of goods.

[0059] Freight volume feature: Extract the monthly freight volume statistics of the origin and destination areas.

[0060] Weather level feature extraction: Origin weather feature: Extract the weather conditions at the origin on the current day, divided into sunny (1), cloudy (2), rainy (3), and snowy (4).

[0061] Destination weather feature: Same as the origin weather feature extraction method.

[0062] Weather impact score: Based on the impact of weather on transportation, calculate the comprehensive impact score, ranging from 0 to 1.

[0063] Extreme weather indicator: Mark whether there is extreme weather such as heavy rain or heavy snow (0 / 1 binary feature).

[0064] From the above various levels, multi-dimensional features are extracted. The extracted multi-dimensional features can be preprocessed in the following way, please refer to Figure 4 : S21, standardize the multi-dimensional features and use the inter-quartile range method to detect and filter out abnormal multi-dimensional features.

[0065] S22, perform period transformation processing on the periodic features in the multi-dimensional features.

[0066] S23, encode the text features in the multi-dimensional features.

[0067] In order to ensure the unity and standardization of multi-dimensional features, multi-dimensional features can be standardized, for example, each feature can be normalized. In this way, subsequent features can be uniformly processed based on the feature values within the specified range.

[0068] In addition, the inter-quartile range method is used to filter multi-dimensional features, which can avoid the influence of extreme values and ensure the stability of the model.

[0069] Some multi-dimensional features have periodicity, such as week and month. Therefore, only the characteristics within a single period need to be extracted. Therefore, the periodicity of the multi-dimensional features is transformed, such as sine or cosine transformation.

[0070] In addition, some literal features such as provinces and cities are also included in the multi-dimensional features. Since machines are binary processing, such literal features cannot be directly processed, and therefore, the literal features can be encoded, for example, by using one-hot encoding.

[0071] In addition, for high-base-number features in the multi-dimensional features, mean encoding such as Target encoding can be used for encoding.

[0072] In addition to the above processing of the multi-dimensional features, importance analysis and correlation test can also be performed on each multi-dimensional feature to ensure that each multi-dimensional feature has a positive contribution to the prediction result.

[0073] In the embodiment, based on the above and according to business rules, multi-dimensional fusion features are obtained by interactive modeling of the extracted multi-dimensional features based on the interaction relationship between the multi-dimensional features.

[0074] The specific implementation of the interactive modeling includes the following four operations: (1) Feature cross operation: Weight-distance cross: generate three cross features of "weight x distance", "weight / distance", and "distance / weight".

[0075] Time-geography cross: generate cross features such as "holiday x hub city" and "season x geographical area".

[0076] Market-time cross: generate cross features such as "oil price x distance" and "freight index x season".

[0077] Specific calculation formula: Cross feature value = Feature A x Feature B (multiplication type).

[0078] Cross feature value = Feature A / Feature B (ratio type, handle the case where the denominator is 0).

[0079] (2) Feature combination operation: Transport efficiency feature: efficiency = weight / (distance x time coefficient).

[0080] Hub synergy feature: hub_effect = start hub weight x end hub weight x distance attenuation coefficient.

[0081] Cost pressure feature: cost_pressure = oil price change rate x distance sensitivity coefficient + holiday coefficient.

[0082] Specific calculation formula: Combined feature value = f (Feature A, Feature B,..., Feature N), where f is a predefined combination function.

[0083] (3) Feature transformation operations: Log transformation: Handle right-skewed distribution features like weight, distance, new feature = log (original feature + 1).

[0084] Trigonometric function transformation: Handle periodic features like month sin transformation = sin (2π × month / 12).

[0085] Binning transformation: Discretize continuous features, such as distance into short [0, 300), medium [300, 800), long [800, +∞).

[0086] (4) Business rule-driven feature construction: Based on the business logic of the freight industry, construct composite features: Transportation load efficiency feature: load_efficiency = (weight / 30) × (distance / 1000) × load utilization rate.

[0087] Where 30 tons is the standard load benchmark and 1000 kilometers is the standard distance benchmark.

[0088] Holiday space-time impact feature: holiday_impact = holiday coefficient × time attenuation coefficient × geographical impact coefficient.

[0089] Where the holiday coefficient is set according to the importance of the holiday (Spring Festival = 1.5, other statutory holidays = 1.2, ordinary holidays = 1.0).

[0090] Raw material price transmission feature: price_transmission = raw material price change rate × goods correlation coefficient × transmission time lag coefficient.

[0091] Where the correlation coefficient is set according to the degree of association between goods and raw materials.

[0092] Through the above four operations, the original basic features are transformed into fused features with business meaning and predictive ability, finally forming a multi-dimensional fused feature set containing 45 features.

[0093] Specifically, the multi-dimensional fused features include multiple items from the following features: transportation load related features, time period related features, market related features, supply and demand related features, and policy related features.

[0094] Among them, the transportation load related features include a plurality of features constructed by the interaction between load type, load weight, transportation distance and transportation efficiency.

[0095] Specifically, the transportation load related features can include transportation load efficiency features, long-distance scale effect features, cargo type segmentation effect features, cargo risk level pricing features, weight ladder pricing features, route quality weighting features, hub city synergy effect features, risk distance coupling effect features, transportation time efficiency level pricing features, loading and unloading condition complexity features and vehicle type specialization matching features.

[0096] The time period related features include a plurality of features constructed by the linkage relationship between holidays, seasons, periodic time and market fluctuations.

[0097] Specifically, the time period related features include holiday space-time influence features, seasonal market fluctuation features, periodic time pattern recognition features, data timeliness weighting features.

[0098] The market related features include a plurality of features constructed by the correlation relationship between regional economy, oil price, raw material price and market fluctuation.

[0099] Specifically, the market related features include regional economic difference features, oil price distance sensitivity features, raw material price transmission features and market index freight linkage features.

[0100] The supply and demand related features include a plurality of features constructed by the linkage relationship between transportation risk, transport capacity matching and regional market.

[0101] Specifically, the supply and demand related features include supply and demand pressure comprehensive evaluation features, transport capacity supply and demand balance features, transport capacity asymmetric effect features, transport capacity saturation interaction features, time transport capacity fluctuation features, hub transport capacity distribution effect features, distance transport capacity coupling features, cargo volume and truck matching degree features, regional transport capacity competition intensity features, transport capacity space-time distribution unevenness features, risk cargo transport capacity matching features and market competition intensity dynamic features.

[0102] The policy related features include a plurality of features constructed by the linkage relationship between environmental protection policy and cost fluctuation.

[0103] Specifically, the policy related features include environmental protection limit policy influence features, policy cost fluctuation features, multimodal transport synergy optimization features and intelligent freight rate expectation management features.

[0104] The following will detail the above-mentioned various features: Transportation load efficiency feature, comprehensive synergy effect of freight weight and transportation distance, establish transportation load efficiency evaluation system, adopt standardized formula: efficiency index = (weight / standard weight benchmark) x (distance / standard distance benchmark). The standard weight benchmark is preferably in the range of 30-40 tons, and the standard distance benchmark is preferably in the range of 800-1200 kilometers, and when the weight and distance exceed the corresponding benchmark at the same time, the scale effect advantage is presented.

[0105] Long-distance scale effect feature, for waybills exceeding the preset long-distance threshold, establish a non-linear scale discount coefficient, the calculation formula is: discount coefficient = (long-distance threshold / actual distance) ^ α, where the long-distance threshold is preferably 800-1200 kilometers, and the index α is preferably 0.2-0.5, determined according to specific business needs, reflecting the law of diminishing marginal cost of long-distance transportation, the longer the distance, the more obvious the scale economy effect.

[0106] Holiday space-time influence feature, establish a complete influence cycle system before and after holidays, including three stages: (1) pre-holiday warm-up period: the influence coefficient increases linearly from 0 to a peak value N1 days before the holiday (N1 is preferably in the range of 3-8 days, and major holidays such as the Spring Festival are preferably 15-25 days); (2) holiday peak period: the influence coefficient reaches a peak value on the holiday; (3) post-holiday decay period: the influence coefficient linearly decreases and returns to zero N2 days after the holiday (N2 is preferably in the range of 5-10 days, and major holidays are preferably 10-20 days), accurately capturing the complete time-varying law of holidays on freight rates.

[0107] Regional economic difference feature, based on the freight volume and the number of employees at the origin and destination, construct a regional economic prosperity index, process the order of magnitude difference through logarithmic transformation, calculate the influence of the economic development level difference between the origin and destination on the freight rate.

[0108] Seasonal market fluctuation feature, establish a fine monthly freight rate cycle prediction system, determine the seasonal influence coefficient of each month by analyzing historical data, form an annual cycle fluctuation curve, where winter months (such as December and January) are usually peak periods (influence coefficient 0.8-1.0), spring months (such as April and May) are usually trough periods (influence coefficient 0-0.3), and summer and autumn months show an increasing trend (influence coefficient 0.3-0.8), accurately capturing the annual freight rate fluctuation law.

[0109] Route quality weighting feature, combine road condition score with actual driving distance to construct route quality weighted distance feature, actual transportation "perceived cost" is lower for routes with good road conditions.

[0110] The hub city synergy effect feature establishes a differentiated weight system for the starting point and the terminal hub city. By setting the starting point weight coefficient W1 and the terminal weight coefficient W2 (W1+W2=1, preferably W1 is 0.5-0.7, and W2 is 0.3-0.5), the double-hub route has better logistics infrastructure and cost advantage.

[0111] The cargo type segmentation effect feature establishes a differentiated pricing system for different cargos according to the degree of processing (precision processing, standard, bulk). Different pricing coefficients are set according to the complexity of processing and the level of added value. The processing degree of high-value goods is higher than that of low-value goods.

[0112] The cargo risk level pricing feature establishes a differentiated pricing system based on the safety risk of goods. According to the risk of goods, the complexity of operation, the insurance requirements and other factors, the risk level is divided. High-risk goods (such as aluminum rolls, chemicals, flammable materials, etc.) have high safety risks and low willingness of drivers to carry, resulting in a shortage of transport capacity and a significant premium on transport prices. Low-risk goods are fully competitive, and the pricing is relatively market-oriented. The risk premium coefficient is set according to the risk level of goods, and the premium coefficient of high-risk goods can reach 1.5-2.5 times.

[0113] The weight ladder pricing feature establishes a fine weight ladder premium system. According to the standard load capacity, multiple weight intervals are divided, and the corresponding premium coefficients are set in each interval. The light load interval (60%-80% below the standard load) sets the premium coefficient 1.2-1.4, the standard load interval sets the base coefficient 1.0, and the heavy load interval (more than 110%-130% of the standard load) sets the discount coefficient 0.85-0.95, identifying the optimal load efficiency interval.

[0114] The supply and demand pressure comprehensive evaluation feature combines transportation distance and hub city distribution to construct a market supply and demand pressure index. The supply and demand pressure coefficient of long-distance non-hub routes is higher.

[0115] The weather influence space aggregation feature integrates the weather conditions of the starting point and the terminal to construct the "maximum weather influence" and "average weather influence" features. Long-distance transportation is more susceptible to the cumulative impact of adverse weather.

[0116] The oil price distance sensitivity feature establishes an interactive influence system between oil price and transportation distance. The sensitivity coefficient K of long-distance transportation to oil price fluctuations is K times that of short-distance transportation (K is preferably 1.5-3.0), reflecting the non-linear influence of distance on fuel cost sensitivity.

[0117] The raw material price transmission feature, please refer to Figure 5, establish the indirect influence mechanism of raw material price fluctuations on transportation demand, and identify the complete transmission chain of "price-demand-capacity-freight". When the price of raw materials rises, the willingness of downstream manufacturers to purchase decreases, transportation orders decrease, capacity is oversupplied, and freight rates fall; when the price of raw materials falls, manufacturers anticipate cost reductions and rush to purchase, transportation demand surges, capacity is in short supply, and freight rates rise. Build demand price elasticity coefficients for different types of goods, such as bulk commodities, which have higher elasticity coefficients than finished products, reflecting the sensitivity of different goods to price fluctuations.

[0118] Market index freight weight linkage characteristics, combined with freight index and freight weight, establish a price elasticity difference system for heavy cargo in market fluctuations, with the market sensitivity coefficient of heavy cargo set as β times that of light cargo (β preferably 0.8-1.2).

[0119] Data timeliness weighting feature, based on the time decay characteristics of historical data, establish a data freshness weight function, W(t) is calculated according to the exponential decay law, where t is the time interval and T is the decay period (T is preferably 20-40 days), determined according to market fluctuation frequency and historical data update period, to ensure that the model can respond to market changes in a timely manner.

[0120] Periodic time pattern recognition feature, capture the periodicity of months, dates, and weeks through sine and cosine transformation, and identify the difference patterns of weekdays and weekends, the beginning and end of the month.

[0121] Please refer to Figure 6 , capacity supply and demand balance feature, based on the number of trucks at the starting point and ending point and freight demand, build a regional capacity supply and demand balance index, the calculation formula is: supply and demand balance = (truck quantity × average carrying capacity) / freight demand, when the supply and demand balance is less than 1, it is in short supply, and the freight rate rises; when the supply and demand balance is greater than 1, it is in oversupply, and the freight rate falls.

[0122] Capacity asymmetry effect feature, analyze the impact of starting point and ending point capacity difference on freight rates, establish a capacity asymmetry coefficient equal to the absolute value of the difference between the starting point and ending point truck density divided by the sum of the truck densities of the two places, the higher the capacity asymmetry, the greater the risk of empty return, and the more obvious the premium.

[0123] Capacity saturation interaction feature, combined with freight weight and regional carrying capacity distribution, build a heavy cargo capacity saturation index, heavy cargo is more sensitive to freight rates in areas with saturated capacity.

[0124] Time capacity fluctuation feature, establish capacity fluctuation patterns in different time periods, such as reduced truck availability during holidays (drivers returning home), tight capacity at the end of the year, and other seasonal capacity changes affecting freight rates.

[0125] HUB freight distribution effect characteristics, analyze the advantages of hub city freight distribution, hub city freight concentration is higher, and the freight allocation is more flexible, which has cost advantage compared with non-hub area.

[0126] Distance and freight coupling characteristics, long-distance transportation requires higher quality of freight, and the matching degree of distance and freight is established. The price elasticity of long-distance transportation is larger when the freight is tight.

[0127] Freight and truck matching degree characteristics, based on the ratio of historical same period freight volume and current available truck quantity, the matching degree index of freight and truck is constructed. When the matching degree is lower than the threshold, it indicates that the freight is tight, and the freight rate increases.

[0128] Regional freight competition intensity characteristics, calculate the truck density corresponding to unit freight volume, and construct the regional freight competition intensity index. The higher the competition intensity, the higher the freight rate, and the more market-oriented the pricing is.

[0129] Return freight source probability characteristics, based on the proportion of freight volume at the starting point and the terminal point and the truck flow direction statistics, the return freight source probability is predicted. The route with low return freight source probability needs to bear higher empty load cost.

[0130] Uneven distribution of freight space and time characteristics, analyze the distribution law of trucks in different time and different areas, and identify the dynamic influence of uneven distribution of freight space and time on freight rate, such as peak and off-peak, weekday and weekend freight distribution difference.

[0131] Risk freight and freight matching characteristics, establish the matching degree evaluation of high-risk freight and professional freight, consider the factors such as driver quality, vehicle condition and insurance coverage. High-risk freight needs higher quality professional freight, and the limited supply leads to more significant freight rate premium.

[0132] Risk distance coupling effect characteristics, analyze the interactive influence of freight risk level and transportation distance, the safety risk of long-distance transportation of high-risk freight increases exponentially, the willingness of drivers to carry is further reduced, and the freight rate premium coefficient increases nonlinearly with distance.

[0133] Transportation time efficiency level pricing characteristics, establish a differentiated pricing system based on transportation time efficiency requirements, and divide the time efficiency level according to the customer's urgency of transportation time. Urgent transportation needs priority scheduling, may need to drive at night or choose faster routes, and the freight resource is scarce, so the freight rate premium is significant; ordinary transportation time is flexible, and the freight selection is sufficient, so the pricing is relatively standard. The time efficiency premium coefficient is set according to the degree of urgency, and the premium coefficient of urgent transportation can reach 1.2-2.0 times.

[0134] Loading and unloading condition complexity features: based on factors such as loading and unloading difficulty, time cost, and equipment requirements, a loading and unloading complexity evaluation system is established. Complex loading and unloading conditions (such as high floor manual handling, special equipment requirements, time-limited loading and unloading, etc.) increase the driver's operation cost and time input, pushing up the freight rate; standard loading and unloading conditions have sufficient competition for transport capacity, and the pricing is market-oriented.

[0135] Market competition intensity dynamic features: based on factors such as transport capacity supply density, historical price fluctuation range, and number of competitors on similar routes, a market competition intensity index is constructed. In high-competition markets, freight rate fluctuations are relatively stable, and pricing tends to be more market-oriented; in low-competition markets, freight rate elasticity is greater, and there is more room for premium.

[0136] Environmental regulation policy impact features: an evaluation system for transport capacity supply constraints based on urban environmental regulation policies is established to analyze the impact of policy factors such as traffic restrictions, restricted time periods, and restricted areas on available transport capacity under different emission standards. Strict regulations reduce available transport capacity, pushing up freight rates; under lenient policies, transport capacity is sufficient, and freight rates are relatively stable.

[0137] Vehicle type specialization matching features: an evaluation of the specialization matching degree between cargo characteristics and vehicle types is established, taking into account factors such as carriage type (van, flatbed, tank car, etc.), load capacity, and special equipment. Highly specialized trucks are relatively scarce, and freight rates are significantly premium; general-purpose trucks have sufficient competition, and pricing is market-oriented.

[0138] Policy-related cost fluctuation features: a mechanism for the transmission of policy-related cost changes such as road and bridge fee concessions, fuel tax adjustments, and over-limit management standards to freight rates is established. When policy-related costs rise, transportation costs increase, and freight rates rise; when policy-related concessions are implemented, costs decrease, and freight rates fall. The impact of different policies and the transmission time of the impact vary.

[0139] Multi-modal transport coordination and optimization features: for combined scenarios of highway transportation and other transportation modes such as rail and water transportation, a multi-modal transport cost comparison and path optimization model is established. When other transportation modes have a significant cost advantage, highway transportation demand decreases, and freight rates are under pressure; when highway transportation has a time efficiency or flexibility advantage, freight rates remain premium.

[0140] Intelligent freight rate expectation management features: based on historical freight rate fluctuation patterns, market supply and demand trends, seasonal factors, and other factors, a freight rate expectation analysis system is established. By identifying deviations between market expectations and actual freight rates, a correction factor is provided for freight rate forecasting. When excessive optimism or pessimism is detected in the market, the model can make appropriate adjustments to improve forecasting accuracy. The above are multi-dimensional fusion features obtained through multi-dimensional feature interaction modeling based on business rules in the application scenarios of freight transportation.

[0141] The above multi-dimensional fusion features will affect the freight prediction, and the final freight is calculated under the comprehensive influence. The influence of each feature on the freight is as follows: Transportation load efficiency feature: solves the inaccuracy of traditional linear pricing by weight, and the prediction error of scale effect of heavy load long-distance transportation is significantly reduced.

[0142] Long-distance scale effect feature: accurately captures the non-linear cost reduction law of long-distance transportation, which can significantly improve the prediction accuracy of long-distance routes compared with linear distance model.

[0143] Holiday space-time influence feature: accurately captures the complete cycle of freight fluctuation before and after holidays (5-20 days preheat before holidays + 7-13 days decay after holidays), which can significantly improve the prediction accuracy of time sensitivity compared with fixed holiday coefficient.

[0144] Regional economic difference feature: identifies the price difference between "economically developed areas to underdeveloped areas" and "underdeveloped areas to developed areas", which significantly improves the prediction accuracy of cross-regional transportation.

[0145] Seasonal market fluctuation feature: effectively predicts the annual freight cycle (highest in December, lowest in April), effectively reducing seasonal prediction bias.

[0146] Route quality weighting feature: distinguishes the cost difference between expressway and ordinary road, which significantly improves the prediction accuracy of complex routes.

[0147] Hub city synergy effect feature: identifies the cost advantage of "double hub" routes, which can significantly improve the prediction accuracy compared with single distance model.

[0148] Cargo type subdivision effect feature: accurately distinguishes aluminum product subdivision market pricing, effectively reducing the prediction bias caused by "one-size-fits-all" cargo classification.

[0149] Cargo risk level pricing feature: accurately identifies the safety premium of high-risk goods (such as aluminum coils), and predicts the freight price rise caused by the low willingness of drivers to carry, which can significantly improve the prediction accuracy of high-risk goods freight.

[0150] Risk cargo capacity matching feature: accurately predicts the impact of professional capacity shortage on high-risk goods freight, effectively improving the accuracy of special goods pricing.

[0151] Risk distance coupling effect feature: effectively predicts the combined premium effect of long-distance high-risk transportation, which significantly improves the prediction accuracy of complex risk scenarios.

[0152] Weight step pricing feature: discovers the premium law of "golden load interval" of 28-32 tons, effectively improving the weight sensitivity prediction accuracy.

[0153] Supply-demand pressure comprehensive evaluation feature: Predict the impact of supply-demand imbalance in remote areas for long-distance transportation, effectively improving the prediction accuracy of special routes.

[0154] Weather impact spatial aggregation feature: Effectively predict the cumulative impact of severe weather on long-distance transportation, reducing weather-related prediction errors.

[0155] Oil price distance sensitivity feature: Accurately capture the differentiated impact of oil price fluctuations on transportation of different distances, significantly improving the prediction accuracy of fuel cost sensitivity.

[0156] Raw material price transmission feature: Accurately predict the indirect impact of raw material price fluctuations on transportation costs through demand transmission, identify transportation cost declines during price increases and transportation cost increases during price declines, and significantly improve the prediction accuracy of macroeconomic sensitivity.

[0157] Market index freight weight linkage feature: Identify the price elasticity of heavy cargo in market fluctuations, effectively improving the accuracy of market change response.

[0158] Data timeliness weighting feature: Ensure that the model responds to the latest market changes in a timely manner, reducing prediction bias caused by historical data lag.

[0159] Periodic time pattern recognition feature: Accurately capture the price rules of weekdays / weekends, beginning of the month / end of the month, effectively improving the prediction accuracy of time sensitivity.

[0160] Supply and demand balance of transport capacity feature: Accurately identify the state of regional transport capacity supply and demand imbalance, predict the price increase during the period of transport capacity shortage, and significantly improve the prediction of supply and demand sensitivity.

[0161] Transport capacity asymmetry effect feature: Effectively predict the return empty cost of one-way transportation, identify the price premium of asymmetric route transport capacity, and significantly improve the prediction accuracy of asymmetric routes.

[0162] Transport capacity saturation interaction feature: Accurately predict the price fluctuations of heavy cargo during the transport capacity saturation period, significantly improving the prediction accuracy of heavy cargo transport price sensitivity.

[0163] Time and transport capacity fluctuation feature: Accurately capture the impact of transport capacity fluctuations during holidays, year-end, and other special periods, significantly reducing seasonal transport capacity prediction bias.

[0164] Hub transport capacity distribution effect feature: Identify the transport capacity allocation advantage of hub cities, improve the differentiated prediction accuracy of hub vs. non-hub routes.

[0165] Distance and transport capacity coupling feature: Effectively predict the price elasticity changes of long-distance transportation during transport capacity shortage, significantly improving the prediction accuracy of long-distance transport capacity sensitivity.

[0166] Freight volume truck matching feature: Accurately predict the critical state of supply and demand of transport capacity, identify the signal of freight rate increase in advance, and significantly improve the prediction accuracy of the critical point of supply and demand.

[0167] Regional transport capacity competition intensity feature: Accurately identify the pricing difference between high-competition and low-competition regions, and effectively improve the prediction of regional competition sensitivity.

[0168] Return freight source probability feature: Effectively predict the impact of empty load cost on freight rates, and significantly improve the prediction accuracy of one-way transportation cost.

[0169] Uneven distribution of transport capacity in space and time feature: Accurately capture the impact of transport capacity distribution changes at different times on freight rates, and effectively improve the prediction accuracy of time-based freight rates.

[0170] Transportation time efficiency level pricing feature: Accurately identify the impact of time efficiency requirements on freight rates, predict the premium level of urgent transportation, and significantly improve the prediction accuracy of time efficiency sensitivity.

[0171] Loading and unloading condition complexity feature: Effectively predict the impact of loading and unloading conditions on freight rates, identify the cost premium of complex operation scenarios, and effectively improve the prediction accuracy of operation complexity.

[0172] Market competition intensity dynamic feature: Accurately identify the pricing strategy differences under different competition environments, and effectively improve the prediction accuracy of competition-sensitive routes.

[0173] Environmental protection limit policy impact feature: Accurately predict the constraint effect of limit policy on transport capacity supply, and significantly improve the prediction accuracy of policy-sensitive regional freight rates.

[0174] Vehicle type specialization matching feature: Effectively identify the scarcity premium of specialized trucks, and effectively improve the prediction accuracy of special cargo transportation.

[0175] Policy-based cost fluctuation feature: Accurately capture the transmission effect of policy adjustments on freight rates, and significantly improve the prediction accuracy of policy sensitivity.

[0176] Multi-modal transport coordination optimization feature: Effectively predict changes in competition between different transportation modes, and improve the prediction accuracy of transportation mode selection.

[0177] Intelligent freight rate expectation management feature: Accurately capture the feedback impact of market expectations on actual freight rates, and significantly improve the prediction accuracy of market sentiment-driven freight rate fluctuations.

[0178] It should be noted that the numerical parameters involved in the above embodiments can be adjusted according to the needs of the actual application scene, and the embodiments do not limit this, for example, the division threshold of light load, standard load and heavy load in the weight threshold parameter can be adjusted according to different industry standards and vehicle capacity, such as any demarcation point in the range of 20-50 tons. In the distance classification parameter, the distance thresholds of short, middle and long distances can be flexibly set in the range of 200-2000 kilometers. In the time period parameter, the number of days before and after the holiday influence and the monthly distribution of the seasonal coefficient can be adjusted according to the characteristics of different regions and industries. In the weight coefficient parameter, the weight coefficients of various interaction features can be optimized and adjusted within a reasonable range. In the attenuation coefficient parameter, the data timeliness and long-distance scale effect index parameters can be adjusted in the range of 0.1-1.0.

[0179] The data collection and feature processing method provided by the embodiments is helpful to improve the fitting, accuracy and response real-time of the subsequent prediction results and actual scenarios compared with the method used in the prior art.

[0180] Specifically, the prior art only considers basic variables such as distance and weight, ignores the complex influence of geographical spatial differences (such as logistics hub effect), time periodicity (such as holiday, seasonal fluctuation), market dynamics (such as oil price change, supply and demand of transport capacity), weather conditions, risk level of goods, etc. on the freight rate, especially the lack of recognition of the indirect influence of raw material price fluctuation on freight through demand transmission mechanism, and the mechanism modeling of the willingness of drivers to carry high-risk goods (such as aluminum rolls, chemical products) due to safety hazards, transport capacity shortage and freight price premium, resulting in deviation of the prediction results from the actual market situation.

[0181] The traditional linear regression or segmented model used in the prior art is difficult to capture the non-linear relationship and interaction effect between multiple factors, and cannot handle the complex dependence relationship between features, such as the synergistic effect of distance and weight, the interactive influence of holiday and route, etc.

[0182] The prior art lacks dynamic adaptability and cannot automatically adjust the prediction results according to real-time data such as market conditions, oil price fluctuations and weather changes. The model parameters are fixed, the response to market changes is lagging, and it is difficult to adapt to the rapid changes of the logistics market.

[0183] The prior art has high artificial dependence and needs a lot of manual intervention and experience correction. The pricing process is highly subjective and inefficient, and it is difficult to realize standardization and scaling, and the manual experience has deviation and inconsistency. The system has poor scalability and cannot support intelligent prediction of nationwide, different types of goods and complex routes. It lacks multi-terminal access capability and real-time interaction function, and cannot meet the diversified needs of modern logistics business.

[0184] The prior art has insufficient data utilization, lacks effective integration of multi-source heterogeneous data, cannot utilize rich data resources such as external market data, weather data, geographic information, and in particular lacks real-time transport capacity data (such as the number of trucks in each region and the supply and demand status of transport capacity), cannot accurately reflect the dynamic changes of the transport market, and the value of data has not been fully tapped.

[0185] The above shortcomings result in that the accuracy rate of the freight prediction of the prior art is generally lower than 70% in actual application, the response time is long (usually requiring manual intervention and taking several minutes), and it is difficult to adapt to the complex and changeable logistics market environment.

[0186] The multi-dimensional feature fusion and intelligent modeling method proposed in the embodiment can systematically solve the above problems and realize intelligent freight prediction with high precision and fast response.

[0187] In the model training layer in the embodiment, model training is performed based on the above multi-dimensional fusion features. Compared with the single linear model used in the prior art, a multi-model fusion strategy is used in the embodiment. Specifically, a plurality of base machine learning models are trained using multi-dimensional fusion features to obtain a plurality of prediction sub-models. Then, the plurality of prediction sub-models are fused according to weights to obtain a prediction model. In this way, the prediction model obtained can integrate the advantages of different algorithms, for example, gradient boosting algorithm has advantages in processing nonlinear relationship and feature interaction, and lightweight algorithm provides fast training and inference ability.

[0188] When combining the plurality of prediction sub-models, the weight values of the weights in the weight combination can be dynamically adjusted based on the prediction performance obtained under different weight combinations. Moreover, the weights of the prediction sub-models can be adaptively adjusted according to the prediction confidence to improve the prediction stability.

[0189] The plurality of prediction sub-models can be combined by using the weighted average method in the following manner: Calculate the root mean square error RMSE_i of each prediction sub-model on the validation data set; calculate the weight coefficient of each prediction sub-model according to the formula w_i = (1 / RMSE_i) / ∑(1 / RMSE_j); calculate the final prediction result according to the formula P =∑(w_i×P_i), where P_i is the prediction value of the i-th prediction sub-model.

[0190] In the embodiment, each prediction sub-model can use different types of machine learning models, such as neural network models, random forest models, gradient boosting tree models, etc.

[0191] When fusing the plurality of prediction sub-models, different model fusion strategies can be used, such as weighted average method, voting mechanism, stacked ensemble, etc.

[0192] Please refer to Figure 7In this embodiment, a real-time automatic training scheduler is also designed in the model training layer, and the real-time updating of the model is realized through the following steps: S31, obtaining updated waybill data, and performing drift detection based on the updated waybill data in a case where the updated waybill data exceeds a preset threshold.

[0193] S32, in a case where the drift detection result exceeds a set threshold, triggering full training of the prediction model to obtain an updated prediction model.

[0194] S33, in a case where the drift detection result does not exceed the set threshold, triggering incremental training of the prediction model to obtain an updated prediction model.

[0195] S34, verifying the performance of the updated prediction model.

[0196] S35, in a case where the performance verification is passed, deploying the updated prediction model.

[0197] S36, in a case where the performance verification is not passed, keeping the original prediction model.

[0198] In this embodiment, a real-time data triggering mechanism is set, and the updated waybill data can be obtained by monitoring the addition of the electronic waybill system and the user reported data, and the updating training of the model is automatically triggered when the updated waybill data exceeds the preset threshold.

[0199] In order to ensure the effectiveness of the data, the updated waybill data needs to be audited, and only the waybill data that passes the quality verification is included in the process.

[0200] The updating training of the model includes full training and incremental training. The full training is to retrain the model based on all samples in the current training set, and the incremental training is to train the model based on the new samples generated.

[0201] The drift detection can be performed based on the updated waybill data, and the full training or the incremental training is determined according to the detection result. The drift detection is used to detect the distribution difference between the new waybill data and the original waybill data, for example, the PSI (Population Stability Index) drift detection, the KL (Kullback-Leibler) divergence detection and the like. If the difference is too large, that is, the drift detection result exceeds the set threshold, the model needs to be adjusted, that is, the prediction model is updated by using the full training. If the drift detection result is less than the set threshold, the incremental training can be triggered.

[0202] Wherein, in the case of triggering incremental training, in order to avoid too frequent update, it is also detected whether the time length from the last time incremental training is performed exceeds the preset time length, if not, it is waited, and when the preset time length is reached, incremental training is performed again.

[0203] After obtaining the updated prediction model through any of the above methods, performance evaluation and detection of the updated prediction model are still needed, if the performance of the updated prediction model is improved, the updated prediction model can be deployed, and hot update online. If the performance of the updated prediction model is not improved, the original prediction model can be maintained, and the reason why the performance of the updated prediction model is poor can be analyzed.

[0204] In this embodiment, in the process of performing model parameter optimization, model parameters can be automatically tuned through methods such as grid search, Bayesian optimization, genetic algorithm, etc. without manual intervention.

[0205] In this process, a model robustness guarantee mechanism is also provided, including abnormality detection and early warning, specifically, a multi-level abnormality detection system can be established to identify input data abnormalities, prediction result abnormalities, model performance abnormalities, etc. and provide intelligent early warning and processing suggestions.

[0206] In addition, for data scarce scenarios such as new routes and new goods types, a prediction mechanism based on similarity matching is established.

[0207] In addition, a confidence evaluation mechanism is provided, for example, each prediction result of the prediction model has a confidence score, when the data quality is insufficient or the market environment is abnormal, the prediction confidence is reduced, thereby providing a risk prompt.

[0208] The above is the implementation method of the data collection, data feature processing, feature fusion, model training, etc. The trained prediction model can be used for automatic freight prediction in the actual application stage.

[0209] In this embodiment, the query request can be obtained based on the user interaction layer, and the freight prediction based on the prediction model under the query request can be performed in the prediction service layer.

[0210] In actual application scenarios, the query request information input by the user is often incomplete, and it is difficult to directly make accurate prediction based on the query request, therefore, in this embodiment, the query request is subjected to semantic analysis to obtain standardized query information, specifically, please refer to Figure 8 This step can be implemented in the following way: S141, the query request is subjected to semantic analysis to determine the missing transportation parameters.

[0211] S142, inquiry information is generated based on the missing transportation parameters, and feedback information obtained in response to the inquiry information is obtained.

[0212] S143, based on the feedback information, performing normalization processing according to the business rules, supplementing the missing transportation parameters, to obtain standardized query information.

[0213] In this embodiment, the query request can be analyzed by a large language model to determine the transportation parameters contained in the query request. Based on the complete transportation parameters specified and the transportation parameters contained in the query request, the missing transportation parameters can be determined, for example, the missing transportation parameters are the type of goods, the total amount of goods, etc.

[0214] After determining the missing transportation parameters, the corresponding inquiry information is generated, for example, if the type of goods is missing, the inquiry information "missing type of goods, please supplement the type of goods" is returned to the user. In the case of obtaining inquiry information by the user, the transportation parameters are supplemented based on the guidance in the inquiry information, and fed back to the large language model as feedback information. In combination with the feedback information, the missing transportation parameters are obtained by normalization processing in combination with the requirements in the business rules, so that the standardized query information can be obtained.

[0215] The obtained standardized query information is input into the prediction model, and the freight rate prediction information is output.

[0216] Specifically, the semantic analysis of the query request is specifically implemented as follows: (1) Input preprocessing: Text cleaning: remove special characters, extra spaces, and unify encoding format.

[0217] Word segmentation processing: use Chinese word segmentation tool to segment the query text.

[0218] Entity recognition: identify key entities such as location name, goods name, weight value, and time expression.

[0219] (2) Parameter extraction: Based on predefined parameter templates and regular expression patterns: Origin extraction: identify location information in patterns such as "from XX", "XX shipment".

[0220] Destination extraction: identify location information in patterns such as "to XX", "to XX".

[0221] Weight extraction: identify numerical information in patterns such as "XX tons", "weight XX".

[0222] Goods type extraction: match goods name based on goods dictionary.

[0223] (3) Missing parameter identification: Check the completeness of necessary parameters: Essential parameter list = [origin, destination, cargo type, cargo weight].

[0224] Missing parameters = essential parameter list - extracted parameters.

[0225] (4) Intelligent inquiry generation: Generate natural language inquiries based on missing parameters: Missing origin: Generate "Please provide the shipment origin location."

[0226] Missing destination: Generate "Please provide the shipment destination."

[0227] Missing weight: Generate "Please provide the shipment weight (unit: tons)."

[0228] Missing cargo type: Generate "Please provide the shipment type (e.g., aluminum products, steel, chemicals, etc.)."

[0229] (5) Feedback information processing: Parse and verify user feedback information: Location verification: Verify the existence of the location through a geocoding API, and obtain the standard name and coordinates.

[0230] Weight verification: Check the reasonableness of the value (0 < weight < 50 tons).

[0231] Cargo type verification: Match the predefined cargo classification dictionary and perform standardized mapping.

[0232] (6) Standardized query information construction: Organize the extracted and completed parameters into a standard format: { "origin": "standard origin name", "destination": "standard destination name", "cargo_type": "standard cargo type code", "weight": weight value, "query_date": "query date", "urgency": "time requirement (default normal)" } In this embodiment, the user interaction layer supports intelligent interaction processing, which can perform semantic understanding on query requests and automatically identify and standardize geographic location information. In addition, it can also perform parameter reasoning based on the dialogue history, and automatically fill in missing transportation parameters according to business rules, etc.

[0233] The user interaction layer provides a standardized calling interface, supports unified calling of multiple applications, and can intelligently check the format of input parameters and the rationality of business logic. Business-friendly error diagnosis and correction suggestions can be provided.

[0234] In this embodiment, the user interaction layer can be implemented based on different types of client applications developed, such as web applications, mobile applications, desktop applications, etc. Different communication protocols and data formats can be used, and different interaction platforms and interfaces can be integrated.

[0235] The prediction service layer can be implemented based on different service architectures, such as microservices, monolithic applications, distributed systems, etc. Different interface protocols can be used, such as REST interfaces, GraphQL interfaces, remote procedure calls, etc. Different caching and storage solutions can be used. The prediction service layer has high performance, dynamic model management, etc., and can achieve millisecond-level response time, which is a significant improvement over the second-level response speed of existing technologies. It also supports high-concurrency request processing and adaptive caching mechanisms to avoid repeated calculations.

[0236] In dynamic management, automatic model version management based on timestamps can be implemented, online switching of models can be achieved through hot updates without interrupting service. Model state and prediction performance can be monitored in real time to ensure stable model operation.

[0237] The freight prediction scheme provided by the present application has been put into practical prediction application and has achieved good prediction results in practical prediction application.

[0238] For example, a certain logistics limited company used the method of the present application to predict the transportation cost of aluminum rods from Longxi County, Dingxi City, Gansu Province to Sihong County, Suqian City, Jiangsu Province on July 13, 2025. The system collected multi-dimensional data including cargo weight (32.5 tons), transportation distance (1472 kilometers), current oil price, weather conditions, holiday information, raw material price, etc. Through multi-dimensional feature fusion, the prediction system output the freight as 229.9 yuan / ton, and the actual transaction price was 230 yuan, with a prediction error of only 0.1 yuan.

[0239] A certain logistics limited company used the method of the present application to predict the transportation cost of aluminum rods from Datong Hui and Tu Autonomous County, Xining City, Qinghai Province to Wen'an County, Langfang City, Hebei Province on July 13, 2025. The system collected multi-dimensional data including cargo weight (32.5 tons), transportation distance (1669.75 kilometers), current oil price, weather conditions, holiday information, raw material price, etc. Through multi-dimensional feature fusion, the prediction system output the freight as 221.8 yuan / ton, and the actual transaction price was 220 yuan, with a prediction error of only 1.8 yuan.

[0240] A logistics company uses the method of the application to predict the aluminum rod transportation cost from Datong Hui Autonomous County of Xining City in Qinghai Province to Yinan District of Tongling City in Anhui Province on June 30, 2025. The system collects multi-dimensional data including the weight of the goods (32.5 tons), the transportation distance (1961.64 kilometers), the current oil price, the weather condition, the holiday information, the raw material price and the like. Through multi-dimensional feature fusion, the prediction system outputs the transportation cost of 259.6 yuan / ton, and the actual transaction price is 260 yuan, with a prediction error of only 0.4 yuan.

[0241] A logistics company uses the method of the application to predict the aluminum rod transportation cost from Yuncheng City in Shanxi Province to Linyi City in Shandong Province on June 28, 2025. The system collects multi-dimensional data including the weight of the goods (32.5 tons), the transportation distance (839.53 kilometers), the current oil price, the weather condition, the holiday information, the raw material price and the like. Through multi-dimensional feature fusion, the prediction system outputs the transportation cost of 146.4 yuan / ton, and the actual transaction price is 150 yuan, with a prediction error of only 3.5 yuan.

[0242] The above is only an illustrative description of the application in actual prediction application, and it can be seen that the application can obtain accurate prediction results that meet the actual scene in actual application.

[0243] The freight transportation cost prediction scheme provided by the application first proposes a multi-dimensional feature fusion architecture, which organically combines the internal features of the shipping order and the external environmental factors. Moreover, the raw material price transmission mechanism and the cargo risk level are innovatively introduced. The technical breakthrough from a single model to multi-model fusion is realized.

[0244] The scheme solves the limitations of traditional methods that only consider distance and weight, and the problem that the model cannot adapt to market changes. Moreover, the problem of user-unfriendly interaction and tedious parameter input in traditional methods is solved.

[0245] The technical scheme provided by the embodiment has at least the following advantages: The prediction accuracy is significantly improved: through multi-dimensional feature fusion and multi-model integration method, the mean absolute percentage error (MAPE) can be controlled within 3%, the determination coefficient R2 can be above 0.98, and the correlation coefficient between the predicted value and the actual value is above 0.99, which greatly improves the prediction accuracy compared with the traditional method.

[0246] The response speed is greatly improved: based on the high-performance prediction service architecture, the response is realized within milliseconds, and the response efficiency is significantly improved compared with the traditional method which needs human participation.

[0247] The automation level is greatly improved: the whole process automation is realized from data collection, feature extraction, model training to prediction output, which greatly reduces the demand for human intervention.​

[0248] Comprehensive application scenario coverage: Supports freight rate prediction for any starting and ending points, multiple cargo types, and complex route conditions nationwide, significantly improving over traditional methods limited to specific regions or cargo types.

[0249] Optimized user experience: Supports natural language queries and intelligent parameter completion, with an automatic completion accuracy rate of over 95% and a user input efficiency improvement of over 5 times.

[0250] Strong model adaptation capability: Through real-time data updates and automatic model iteration mechanisms, the system can quickly adapt to market changes and maintain prediction accuracy.

[0251] In summary, the multi-dimensional feature fusion-based freight rate prediction method provided by the embodiment adopts a multi-dimensional feature fusion technology, including systematic extraction and fusion architecture of four dimensions (geography, time, market, and weather) features, and especially includes modeling of the indirect impact of raw material prices on freight rates through demand transmission mechanisms.

[0252] The solution supports an automatic feature interaction discovery mechanism and a multi-element interaction feature generation algorithm based on business rules, including multi-dimensional interaction modeling of cargo risk level, transportation time requirement, loading and unloading complexity, environmental policy constraints, vehicle specialization matching degree, transportation distance, and transportation supply and demand.

[0253] Adopting a multi-model fusion prediction method, the system uses a dynamic weight fusion strategy of multiple machine learning algorithms.

[0254] Supporting real-time data access and processing mechanisms, the system supports real-time data stream processing and automatic incremental training of electronic waybill systems.

[0255] Setting up a crowdsourcing data quality control system, the system supports intelligent auditing, hierarchical verification, and automatic training mechanisms for user-reported data.

[0256] Supporting multi-dimensional automatic training scheduling, the system uses an intelligent training strategy based on multiple trigger conditions such as data volume, quality, time, and drift.

[0257] Supporting a multi-end intelligent interaction architecture, the system supports a multi-modal interaction system with natural language understanding.

[0258] The solution is applicable to various transportation scenarios and business models, and parameters can be adjusted according to different industry standards, regional characteristics, and cargo types, making it widely applicable and expandable, and possessing continuous learning and self-optimization capabilities.

[0259] Based on the same inventive concept, please refer to Figure 9The embodiment of the present application also provides a functional module schematic diagram of the freight transport freight prediction system based on multi-dimensional feature fusion. The embodiment can divide the functional modules of the freight transport freight prediction system based on multi-dimensional feature fusion according to the method embodiment. For example, each functional module can be divided according to each function, or two or more functions can be integrated in one processing module. The integrated module can be realized in the form of hardware or in the form of a software functional module. It should be noted that the division of the module in the embodiment of the present application is illustrative, and is only a logical function division. In actual implementation, another division mode can be used.

[0260] For example, in the case of dividing each functional module according to each function, Figure 9 The freight transport freight prediction system based on multi-dimensional feature fusion shown is only a functional module schematic diagram. The freight transport freight prediction system based on multi-dimensional feature fusion can include a collection module, a construction module, a training module and a query prediction module. The functions of each functional module of the freight transport freight prediction system based on multi-dimensional feature fusion are described in detail below.

[0261] The collection module is used for collecting waybill data and external influence data. The waybill data includes multi-dimensional waybill features, and the external influence data includes multi-dimensional external influence factors. The construction module is used for constructing multi-dimensional fusion features based on the interaction relationship between the multi-dimensional waybill features and the multi-dimensional external influence factors. The training module is used for training a plurality of basic machine learning models based on the multi-dimensional fusion features, to obtain a prediction sub-model corresponding to each basic machine learning model, and combining the plurality of prediction sub-models by using a weighted average method to obtain a prediction model. The query prediction module is used for, in the case of obtaining a query request, performing semantic analysis on the query request to obtain standardized query information, inputting the query information into the prediction model, and obtaining freight transport freight prediction information corresponding to the query information.

[0262] The freight transport freight prediction system based on multi-dimensional feature fusion provided in the embodiment can be used to execute the freight transport freight prediction method based on multi-dimensional feature fusion in any of the implementation manners in the above embodiments. For details not described in the embodiment, refer to the corresponding description in the above embodiments, which will not be described herein again.

[0263] Please refer to Figure 10, is a block diagram of the structure of an electronic device provided in an embodiment of the present invention. This electronic device can be a computer device, server, or other device in a backend prediction platform. The electronic device includes a memory, a processor, and a communication module. The memory, processor, and communication module are electrically connected to each other, directly or indirectly, to enable data transmission or exchange. For example, these components can be electrically connected via one or more communication buses or signal lines.

[0264] Memory is used to store computer programs or data. Memory can include, but is not limited to, random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), and electrically erasable programmable read-only memory (EEPROM).

[0265] The processor is used to read / write data or programs stored in the memory and execute the freight rate prediction method based on multi-dimensional feature fusion provided by any embodiment of the present invention.

[0266] The communication module is used to establish a communication connection between the electronic device and other communication terminals through the network, and is used to send and receive data through the network.

[0267] It should be understood that Figure 10 The structure shown is only a schematic diagram of the structure of the electronic device. The electronic device may also include Figure 10 More or fewer components than shown, or with Figure 10 Different configurations shown.

[0268] Furthermore, an embodiment of the present invention also provides a computer-readable storage medium, which stores machine-executable instructions. When the machine-executable instructions are executed, the freight rate prediction method based on multi-dimensional feature fusion provided in the above embodiment is implemented.

[0269] Specifically, the computer-readable storage medium can be a general-purpose storage medium, such as a removable disk or hard disk. When the computer program on the computer-readable storage medium is executed, the above-described freight rate prediction method based on multi-dimensional feature fusion can be executed. Regarding the processes involved in executing the executable instructions on the computer-readable storage medium, please refer to the relevant description in the above-mentioned method embodiment and will not be further described here.

[0270] In summary, the application provides a freight transport fare prediction method and system based on multi-dimensional feature fusion. Transport order data containing the starting point, ending point, cargo type and weight, and six types of external influence data containing geography, market, transport capacity, weather, time and policy are collected. 45 kinds of multi-dimensional fusion features are constructed through feature cross, combination and transformation operations, including transport load efficiency features, holiday space-time influence features, raw material price transmission features and transport supply and demand balance features. Gradient boosting tree and random forest are used to train sub-models, and the performance reflected by the RMSE of the validation set is used to distribute the weighted coefficient. The final prediction model is generated after fusion. The system supports natural language query and intelligent parameter completion, identifies missing parameters through semantic analysis and guides users to complete them. Compared with traditional methods, the prediction accuracy is improved from 70% to more than 97%, the MAPE is reduced from 15% to less than 3%, and the response time is improved from minutes to milliseconds, significantly improving the accuracy, speed and user experience of freight transport fare prediction.

[0271] In the embodiments of the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of units is only a logical function division, and actual implementation can have another division manner. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some communication interface, device or unit, which can be electrical, mechanical or other forms.

[0272] In addition, the units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or distributed on a plurality of network units. According to actual needs, some or all of the units can be selected to achieve the purpose of the embodiment.

[0273] Furthermore, the functional modules in each embodiment of the present application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0274] It should be noted that if the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the parts of the present application that essentially contribute to the prior art or the parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0275] The above is only an embodiment of the present application and is not used to limit the protection scope of the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A freight rate prediction method based on multi-dimensional feature fusion, characterized in that: The method comprises: Collecting waybill data and external influence data, wherein the waybill data includes multi-dimensional waybill features and the external influence data includes multi-dimensional external influence factors; Based on the interactive relationship between the multidimensional waybill features and the multidimensional external influencing factors, a multidimensional fusion feature is constructed; Training multiple basic machine learning models based on the multidimensional fusion features to obtain prediction sub-models corresponding to each of the basic machine learning models, and combining the multiple prediction sub-models using a weighted average method to obtain a prediction model; When a query request is obtained, semantic analysis is performed on the query request to obtain standardized query information, and the query information is input into the prediction model to obtain freight rate prediction information corresponding to the query information.

2. The freight rate prediction method based on multi-dimensional feature fusion according to claim 1 is characterized in that: The step of constructing a multidimensional fusion feature based on the interactive relationship between the multidimensional waybill feature and the multidimensional external influencing factors includes: Extracting multidimensional features from the multidimensional waybill features and the multidimensional external influencing factors at the basic level, geographical level, time level, market level, and weather level; According to business rules and based on the interactive relationship between multidimensional features, the extracted multidimensional features are interactively modeled to obtain multidimensional fusion features.

3. The freight rate prediction method based on multi-dimensional feature fusion according to claim 2 is characterized in that: The method further comprises the step of preprocessing the multidimensional features, the step comprising: Performing feature standardization on the multidimensional features, and using the interquartile range method to detect and filter out abnormal multidimensional features; Performing periodic transformation processing on the periodic features in the multidimensional features; The text features in the multi-dimensional features are encoded.

4. The freight rate prediction method based on multi-dimensional feature fusion according to claim 1 is characterized in that: The external influencing data includes geographical data, market data, capacity data, weather data, time data and policy data; The geographic data includes geographic coordinates, administrative divisions, and route planning; The market data includes freight rate index, oil prices, raw material prices, and freight volume; The transport capacity data includes the real-time number of trucks in each region, load capacity distribution, and transport capacity supply and demand status; The weather data includes the weather conditions at the starting point and the ending point of each waybill and the impact score on transportation; The time data includes holidays, working days, and seasons; The policy data includes environmental protection restriction information, road and bridge fee adjustment information, fuel tax information, and overload management standard information.

5. The freight rate prediction method based on multi-dimensional feature fusion according to claim 1 is characterized in that: The multi-dimensional fusion features include multiple features of the following features: transportation load related features, time period related features, market related features, supply and demand related features, and policy related features.

6. The freight rate prediction method based on multi-dimensional feature fusion according to claim 5 is characterized in that: The transport load related features include a plurality of features constructed by the interactive relationship between load type, load weight, transport distance and transport efficiency; The time period related features include multiple features constructed by the linkage relationship between holidays, seasons, periodic time and market fluctuations.

7. The freight rate prediction method based on multi-dimensional feature fusion according to claim 5 is characterized in that: The market-related characteristics include a plurality of characteristics constructed by correlations between regional economy, oil prices, raw material prices and market fluctuations; The supply and demand related characteristics include multiple characteristics constructed by the linkage relationship between transportation risk, capacity matching and regional markets; The policy-related features include multiple features constructed by the linkage relationship between environmental protection policies and cost fluctuations.

8. The freight rate prediction method based on multi-dimensional feature fusion according to claim 1 is characterized in that: The method further comprises: Obtain updated waybill data, and if the updated waybill data exceeds a preset threshold, perform drift detection based on the updated waybill data; When the drift detection result exceeds a set threshold, full training of the prediction model is triggered to obtain an updated prediction model; When the drift detection result does not exceed the set threshold, triggering incremental training of the prediction model to train an updated prediction model; The updated prediction model is subjected to performance verification. If the performance verification passes, the updated prediction model is deployed; otherwise, the original prediction model is retained.

9. The freight rate prediction method based on multi-dimensional feature fusion according to claim 1 is characterized in that: The step of performing semantic analysis on the query request to obtain standardized query information includes: Performing semantic analysis on the query request to determine missing transportation parameters; generating query information based on the missing transport parameters and obtaining feedback information in response to the query information; Based on the feedback information, normalization processing is performed according to business rules, and missing transportation parameters are supplemented to obtain standardized query information.

10. A freight rate prediction system based on multi-dimensional feature fusion, characterized in that: The system comprises: A collection module, configured to collect waybill data and external influence data, wherein the waybill data includes multi-dimensional waybill features and the external influence data includes multi-dimensional external influence factors; A construction module, configured to construct a multidimensional fusion feature based on the interactive relationship between the multidimensional waybill feature and the multidimensional external influencing factors; A training module is used to train multiple basic machine learning models based on the multidimensional fusion features to obtain prediction sub-models corresponding to each of the basic machine learning models, and combine the multiple prediction sub-models using a weighted average method to obtain a prediction model; The query prediction module is used to perform semantic analysis on the query request to obtain standardized query information when a query request is obtained, input the query information into the prediction model, and obtain freight rate prediction information corresponding to the query information.

Images