Steel product railway freight rate coupling prediction method considering line space-time heterogeneity

Abstract

The invention relates to the technical field of data analysis, space-time sequence prediction and multivariable coupling modeling, in particular to a steel product railway freight rate coupling prediction method considering line space-time heterogeneity, which comprises the following steps of: 1, judging key factors influencing freight rate addition and subtraction based on a causal test method, and establishing a data file; 2, extracting space-time heterogeneity, in a space dimension, defining origin and destination points of the transportation route as discrete classification variables, representing heterogeneity influence of the transportation route by adopting an embedded layer coding mode, in a time dimension, generating a core vector as a model parameter to be input by utilizing cyclic coding and a multi-layer perceptron technology, model parameters have time correlation and are used for identifying similar time periods and distinguishing different time periods; and step 3, carrying out freight rate prediction to obtain a final freight rate prediction result. The method has the advantage that the mechanism interpretability and prediction precision of freight rate prediction are further improved by adopting a coupling prediction framework and method.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis, spatiotemporal series prediction and multivariate coupling modeling, and in particular to a method for coupling prediction of railway freight rates for steel products taking into account spatiotemporal heterogeneity of lines. Background Art

[0002] The steel industry is one of the most resource-, energy- and emission-intensive material industries, accounting for nearly 22% of global industrial energy consumption. As the world's largest steel producer and consumer, China's efforts to promote the low-carbon transformation of its steel industry will determine the sustainable future of the global steel industry and are of great significance. In the context of bulk cargo logistics and transportation, it is of great practical significance to explore how to reduce costs and increase efficiency in steel logistics and transportation.

[0003] How to make causal judgments on factors that may affect steel freight rates, make periodic judgments based on existing historical steel demand data, and conduct seasonal and regional difference analysis, and based on this, establish a joint forecasting framework to predict future short-term steel prices and freight rates, thereby guiding actual production plans and making relative adjustments to transportation strategies is a major practical problem.

[0004] Existing methods DOI: 10.26989 / d.cnki.gdlhu.2023.001453 "Research on Railway Container Freight Rate Forecasting Based on Machine Learning Algorithm" Scholar Wei Huikang selected 1,176,255 original data from 2012 to 2020, and preliminarily determined 9 original features according to the "Railway Freight Rate Rules". Combined with the requirements of the machine learning algorithm for data features, the feature data was numerically converted to obtain 14 feature fields and data cleaning was completed for subsequent predictions; DOI: 10.27414 / d.cnki.gxnju.2022.002952 "Research on Railway Freight Rate Forecasting Based on Deep Learning" Scholar Huang Linjin used Pearson correlation analysis to quantify the degree of correlation between each factor and freight rate fluctuations and screened out the total amount of freight bills, warehousing price index, PMI, GDP, retail commodity price index, highway freight rate index, highway freight volume, iron ore production, and iron ore price index 9 The above method only calculates the correlation coefficient between each factor and the freight rate through correlation test, but fails to screen and identify the factors that have a key impact on the freight rate, and fails to explain the essential relationship between the influencing factors and the freight rate, which may lead to the neglect of unobserved confounding variables and easily lead to misleading conclusions.

[0005] Due to the influence of seasonal demand fluctuations and regional supply and demand imbalances, and the fact that bulk freight rates reflect the spatiotemporal value of bulk freight circulation, mining and revealing the spatiotemporal heterogeneity of bulk freight transportation and applying the spatiotemporal heterogeneity law to freight rate forecasting are important for improving the accuracy of bulk freight rate forecasting. However, the reported forecasting methods do not take into account the fluctuation law of freight rates with time and space changes when forecasting freight rates. For example, in the invention of application number 202311727505.3, a coal highway freight rate forecasting method based on machine learning algorithm is proposed. The historical coal highway freight rates are processed and the constructed feature data is divided into training data and test data, which are then input into the forecasting model for forecasting. The forecasting performance is checked by the mean absolute error (MAE value) of each round of training set and test set during the training process. This method performs well in highway freight rate forecasting, but railway freight rate forecasting is different from highway freight rate forecasting and is not suitable for railway freight rate forecasting. In DOI: 10.27414 / d.cnki.gxnju.2022.002952 "Research on Railway Freight Rate Forecasting Based on Deep Learning", scholar Huang Linjin attempts to convert the changes in the two base prices into the overall floating rate of the freight rate according to the railway freight rate structure and gives a freight rate fluctuation calculation formula to reflect the changes in the two base prices and facilitate the exploration of the freight rate fluctuation law. The LSTM multi-feature prediction model is introduced to extract the effective information of each influencing factor on the basis of retaining the time series characteristics of the data itself, and achieves high-precision prediction in the railway weekly freight rate fluctuation forecast.

[0006] Most existing studies use machine learning or deep learning methods to predict freight rates, but they are unable to accurately predict freight rates for specific ODs or specific time periods.

[0007] Freight is an integral part of steel prices. As steel prices change, transportation prices will also change. When steel prices rise, freight rates will inevitably increase accordingly. This is a fact generally recognized in the industry. Reported literature generally believes that steel prices and freight rates have an obvious correlation, but the specific synergistic mechanism between the two is unclear. Therefore, if high-quality freight rate forecasts are required, coupled forecasts of steel prices and freight rates are required. Existing methods, such as the invention with application number 202210783351.9, disclose a railway freight rate fluctuation prediction method based on multivariate LSTM, which uniformly calculates the freight rate including base price 1 and base price 2, and proposes a calculation method for observing the fluctuation trend of the overall railway freight freight rate, taking into account the relationship between the fluctuation of the bulk commodity price index and transportation demand and freight rates, as well as the impact of transportation costs, macro-environment, market competition and multi-dimensional influencing factors of transported goods on the freight rate fluctuation rate. The LSTM model is used to predict the railway freight rate fluctuation rate, which reflects the actual situation of the current railway transportation price fluctuation, but fails to take into account the impact of the coupling effect of steel frame and freight rate. Summary of the invention

[0008] In view of the above problems, the purpose of the present invention is to provide a coupled prediction method for railway freight rates of steel products taking into account the temporal and spatial heterogeneity of lines, which is used to couple the prediction of steel prices and transportation prices, and provide more accurate prediction results for specific regions and times, so as to overcome the shortcomings of the above-mentioned prior art.

[0009] The present invention provides a method for coupling prediction of railway freight rates for steel products considering the temporal and spatial heterogeneity of lines, which specifically comprises the following steps: Step 1: Identify the key factors that affect the freight rate increase or decrease rate based on the causal test method. Establish a data file, pre-process the obtained data and conduct preliminary screening by correlation analysis, and then use the Granger causality test method to identify the factors that have a key impact on the freight rate increase or decrease rate for each influencing factor; Step 2: Extracting spatiotemporal heterogeneity, In the spatial dimension, the origin and destination points (OD pairs) of the transport routes are defined as discrete categorical variables, and the embedding layer coding method is used to represent the heterogeneous impact of the transport routes. In the temporal dimension, the core vector is generated by using the cyclic coding and multi-layer perceptron (MLP) technology as the input of the model parameters, so that the model parameters have time correlation, which is used to identify similar time periods and distinguish different time periods. Step 3: Freight rate forecast, Based on the number of cross-regional routes, the heterogeneous space-time routes are divided into multiple homogeneous subsets. An independent LSTM model is trained for each subset, and the steel freight rate addition and subtraction rate is predicted by weighted fusion. Twin components and interwoven components are constructed. The twin component is used to perform time series prediction on the freight rate addition and subtraction rate and steel price respectively, and the interwoven component is used to bridge the two time series so that they can exchange valuable information in the process of time-dependent modeling. Finally, the freight rate calculation formula is combined to output the prediction result of steel freight rate.

[0010] As the preferred embodiment of the present invention, step 1 also includes the following steps: A1: Correlation analysis, A11: Data preprocessing, Preprocess the obtained steel freight rates and possible influencing factors data, including quantification of influencing factor data, filling of missing values, deletion of duplicate values, and outlier processing, and finally obtain the time series of steel freight rate addition and subtraction rates and possible influencing factors; A12: Draw a scatter plot. The obtained data are plotted in a two-dimensional coordinate system in the form of scattered points, and the linear trend of the data is determined by the outline of the scatter plot; A13: Correlation calculation, The Pearson correlation coefficient is calculated to determine the correlation between each influencing factor and the freight rate. The correlation coefficient calculation formula is as follows: ; in represents the Pearson correlation coefficient, is the total number of samples, represents the sample mean, is the variable value. Here we need to explain: Indicates the freight rate increase or decrease rate. Indicates the value of the influencing factor. Each influencing factor needs to be independently correlated with the freight rate increase or decrease rate. The value of is between [-1,1]. The positive or negative value of indicates whether the influencing factor is positively correlated or negatively correlated with the freight rate. The absolute value of indicates the degree of correlation between the influencing factor and the freight rate. The larger the absolute value, the stronger the correlation. A14: test statistic, ; The test statistic constructed by T is subject to Student distribution with degrees of freedom, giving a significance level , determine whether the influencing factors are relevant; A2: Perform Granger causality test. A21: Assumptions: The null hypothesis of the Granger causality test is that sequence X is not the Granger cause of sequence Y, and the alternative hypothesis is that sequence X is the Granger cause of sequence Y. Here, sequence X is the optimal linear prediction function for constructing sequence Y, which is denoted as: ; At this point, the assumptions of the Granger causality test are equivalently expressed as; ; Here Indicates possible influencing factors. Indicates the rate of increase or decrease in steel freight rates. represents the current time step, represents the best linear fit hour, Time ago to current The maximum time series that has an impact, represents random error, Indicates except Other influencing factors, Indicates the corresponding maximum impact timing, is the corresponding weight coefficient; A22: test statistic, Using the test statistic F, two regression models are fitted, and the freight rate is fitted when the null hypothesis is established; ; ; When the alternative hypothesis is true, the fitted series Unconstrained prediction model for ; ; Calculate test statistics and make judgments; ; like ; Then the original hypothesis is rejected, and it is believed that sequence X has an impact on sequence Y. At the same time, the corresponding factors are considered to have an impact on steel freight rates. Otherwise, it is considered to have no impact, and the factors that affect the addition and subtraction rate of steel freight rates are obtained. time series.

[0011] As the preferred embodiment of the present invention, step 2 also includes the following steps: B1: Use cyclic coding to indicate the time position of the current time slot. The steel freight rate addition and subtraction rates are similar in certain periods, showing continuity and cyclicality; continuity means that adjacent periods have similar dynamics due to gradual changes in market conditions, and cyclicality reflects recurring patterns, such as monthly and annual demand fluctuations; in order to facilitate modeling, the time slot indicators are converted to cyclic coding; ; ; in and Represent the ranking of time slots in a month and a year, respectively. and They represent their relative positions in a season and a year respectively. Considering that the cyclic coding needs to satisfy continuity and periodicity, and It is impossible to meet both requirements at the same time, so the loop encoding is defined as And it is represented by sine and cosine functions: ; B2: In addition to considering similarities, we also need to consider differences. The changing patterns of steel freight rates in different periods are not the same, and the fluctuations of steel freight rates in special periods are irregular. Therefore, it is necessary to build a time-aware model and make the parameter values ​​depend on the time and change with the input; B21: Using multi-layer perceptron (MLP) technology Transformed into a deeper level of time coding ; ; To facilitate subsequent calculations, is rearranged into a diagonal matrix, the core matrix , whose diagonal value is The value of B22: The core matrix is ​​fed into a parallel calculator to generate dynamic time-dependent model parameters. The calculation process is as follows: ; in, is an additional term calculated using the core matrix, is a learnable matrix with a shape suitable for matrix multiplication. Through matrix multiplication and nonlinear transformation, the core matrix is ​​incrementally transformed into the target matrix. is called the static component because it is the same for all inputs, is called the dynamic component because it varies with the input, represents the model parameters generated by the convolutional layer; since is transformed from the time slot indicator, so becomes time dependent; is a nonlinear activation function, is the Sigmoid activation function, and the calculation formula is as follows: ; B3: Number each different OD pair in the data and obtain the OD set: ; The discrete variables are encoded by using an embedding layer, which is used to map the discrete values ​​of each category into a low-dimensional continuous vector space. , the embedding layer will output a d-dimensional embedding vector; ; in Yes Category The corresponding embedding vector has a dimension of d, and C is the total number of OD categories.

[0012] As the preferred embodiment of the present invention, step 3 also includes the following steps: Freight rate calculation formula: The freight rate for each ton of full truckload, each 10kg of less-than-load cargo, and each container is called the freight rate. The current freight rate is composed of two parts: the base price 1 and base price 2: L; in Indicates the freight rate; Indicates the base price 1 of the goods. Base price 1 is the origin and destination operation fee that is not related to the transportation distance. This part of the fee is fixed; It represents the base price 2, which is the operating fee of the vehicle during operation; L is the transportation mileage; The freight rate calculation formula is as follows: ; in Represents freight rate, Indicates the tonnage of goods transported during full truck transportation. Indicates the corresponding freight rate The addition and subtraction rate, Indicates the capacity of the vehicle, symbol Indicates rounding up; C1: Determine base price 1 and base price 2, Base Price 1 and Base Price 2 are found by determining the railway transportation method and waybill number: C2: Forecast corresponding freight rate The addition and subtraction rate , First, the heterogeneous spatiotemporal lines are divided into multiple homogeneous subsets according to the number of cross-regions. An independent LSTM is trained for each subset. For each independent LSTM, the time series of key influencing factors obtained in steps 1 and 2 are used. , Dynamic time parameters And the embedding vector generated by the categorical variable OD , combined with the historical addition and subtraction rates of steel freight rates Make predictions and finally perform weighted fusion to get the prediction results ; The time series data of freight rate increase / decrease rate is , the input data of LSTM Depend on , Stitched together: ; The LSTM model processes time series data based on a gating mechanism; at each time step , the calculation of LSTM is divided into multiple steps: input gate, forget gate, output gate and unit state update; the state update formula of LSTM is as follows: Forget Gate : ; Input Gate : ; Candidate unit status : ; Unit Update Status : ; Output Gate : ; Hide status updates : ; in is the hidden state at the previous time step, is the Sigmoid activation function, tanh is the hyperbolic tangent activation function, is the weight matrix of each gate in LSTM, is the bias term; The final output is: ; in is the predicted value at time step t; C3: Build twin components, Each twin component, namely the freight rate addition / subtraction rate and steel price, is predicted using the above C2 step to obtain the result and , both are time series; C4: Construct interleaving components to interact with time information, The interleaved component is a bridge between two components, allowing them to exchange information and perform feature extraction at the same time. That is, each twin component collects relevant information and shares necessary information. The output layer of each twin component collects features extracted from multiple spatiotemporal blocks through skip connections and fuses them for specific prediction tasks. Although two independent LSTM models are used to capture the temporal dependencies from the two time series respectively, the changing temporal dependencies are ignored; to solve this problem, a time-interleaved module is proposed to bridge the two separated time series. and , which can exchange valuable information in the process of temporal dependency modeling; first, two additional convolutional layers TCL1 and TCL2 are set to extract temporal features as information, respectively. and Shared in time series; ; Secondly, in order to avoid the transmission of irrelevant information between different sequences, a self-gating mechanism is designed to filter the shared information. Considering that the Sigmoid function can assign a smooth weighted value to each feature, has good nonlinear modeling ability, can flexibly control the information flow, and can avoid the gradient disappearance problem during training, the Sigmoid function in the self-gating mechanism helps the model filter irrelevant or noisy information through this weighted operation, thereby improving the performance and stability of the model. To represent the Sigmoid function, , Indicates shared information and The weights of each are: ; ; ; ; The Sigmoid function calculation formula is as follows: ; This is the final result of the steel freight rate addition and subtraction rate prediction. Combined with the freight rate calculation formula, the final freight rate prediction result is obtained. .

[0013] The beneficial effects of the present invention are as follows: 1. The present invention effectively distinguishes factors that have a real causal relationship with freight rates (such as iron ore prices, macroeconomic indicators, etc.) through the Granger causality test of time series, avoids the interference of pseudo-correlation factors in traditional correlation analysis, and reduces the risk of model overfitting; the modeling considering spatiotemporal heterogeneity enhances the adaptability of prediction, and adopts a prompter mechanism to inject time information into model parameters, so that the model has the ability to identify periodic continuity, and can dynamically adjust parameters to adapt to seasonal fluctuations (such as rising transportation costs in winter and differences in production peak and off-seasons), improve sensitivity to fluctuations in freight rates in specific periods, and by encoding OD pairs, construct spatial associations between regions, capture spatial heterogeneity of freight rates (such as differences in coastal and inland transportation costs), and achieve refined predictions for different regions; by establishing a mutual influence model between steel prices and freight rates (such as steel price increases triggering a surge in transportation demand and pushing up freight rates), the limitations of traditional single-variable predictions are broken through, and the coordinated change law of the two variables is used to improve prediction accuracy, while providing two-way decision support for production plans and transportation strategies. Through fine-grained predictions to support accurate decision-making, the forecast results of freight rates in specific regions and time windows can be output, which is better than existing macro-forecasting methods and helps enterprises formulate differentiated transportation route optimization and inventory allocation plans; it can dynamically adapt to the market environment, and by integrating spatiotemporal heterogeneity and coupling relationships, the model can effectively respond to the impact of nonlinear factors such as policy regulation (such as carbon emission restrictions) and emergencies (such as transportation interruptions caused by extreme weather) on the supply chain. It has achieved a double breakthrough in the accuracy and practicality of steel freight rate forecasts, and provided feasible technical support for logistics cost control and supply chain resilience improvement in the low-carbon transformation of the steel industry.

[0014] 2. The causal discrimination test in the present invention can clarify the internal law of freight rate changes and provide a mechanistic explanation for freight rate changes. Therefore, it is necessary to first discover the potential relationship between influencing factors and freight rates through correlation analysis, and then verify the mechanism through causal analysis to identify the key factors that affect freight rates. This method uses time series to first conduct correlation analysis on factors that may affect steel freight rates to conduct a preliminary screening of influencing factors, and then conducts a Granger causality test to identify the factors that really affect freight rates, which are used as input for subsequent freight rate addition and subtraction rate predictions, laying the foundation for ensuring high-quality predictions.

[0015] 3. The present invention considers the impact of changes in the departure station and the terminal station on the freight rate by taking into account the temporal and spatial heterogeneity, and converts it into a categorical variable, and then encodes it by an embedded layer method to explore the changing rules of the freight rate at different starting and ending points; in terms of time, a prompter mechanism for injecting time information into the model parameters is proposed, so that the model parameters have time correlation, and through the method of cyclic coding, sine function and cosine function are selected to represent the parameters, so that they have continuity in the three time gradients of week, month and year. At the same time, the time granularity can be adaptively adjusted according to the needs of each specific cycle, so as to identify similar and different time periods.

[0016] 4. The present invention adopts a coupled prediction framework and method to further improve the mechanism interpretability and prediction accuracy of freight rate prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] By referring to the following description in conjunction with the accompanying drawings, and with a more comprehensive understanding of the present invention, other objects and results of the present invention will become more apparent and easy to understand. In the accompanying drawings: Figure 1 It is a technical roadmap of the steel price and freight price coupling prediction method in the present invention; Figure 2 This is a schematic diagram of the freight rate addition and reduction rate prediction considering the temporal and spatial heterogeneity in the present invention; Figure 3 It is a schematic diagram of the coupling prediction of the freight rate addition and subtraction rate and the steel price in the present invention. DETAILED DESCRIPTION

[0018] See also Figure 1-3 This embodiment provides a coupled prediction method for railway freight rates of steel products considering the temporal and spatial heterogeneity of lines based on the bottom-up idea. First, the factors that really affect the steel freight rate are screened out through the causal discrimination method; then, considering the temporal and spatial heterogeneity, time information of different granularities is injected into the prompt mechanism in time so that it can differentiate between similar and different time periods, and the impact of different departure stations and terminal stations on the freight rate is considered in space; finally, the freight rate and steel price are coupled predicted by the joint prediction method in order to obtain a higher accuracy freight rate prediction result.

[0019] The specific steps include: Step 1: Identify the key factors that affect the freight rate increase or decrease rate based on the causal test method; Establish a data file, pre-process the obtained data and conduct preliminary screening by correlation analysis, and then use the Granger causality test method to identify the factors that have a key impact on the freight rate increase or decrease rate for each influencing factor. The specific steps include A1 to A2; A1: Correlation analysis, A11: Data preprocessing, Preprocess the obtained steel freight rates and possible influencing factors data, including quantification of influencing factor data, filling of missing values, deletion of duplicate values, and outlier processing, and finally obtain the time series of steel freight rate addition and subtraction rates and possible influencing factors; A12: Draw a scatter plot. The obtained data are plotted in a two-dimensional coordinate system in the form of scattered points, and the linear trend of the data is determined by the outline of the scatter plot; A13: Correlation calculation, The Pearson correlation coefficient is calculated to determine the correlation between each influencing factor and the freight rate. The correlation coefficient calculation formula is as follows: ; in represents the Pearson correlation coefficient, is the total number of samples, represents the sample mean, is the variable value. Here we need to explain: Indicates the freight rate increase or decrease rate. Indicates the value of the influencing factor. Each influencing factor needs to be independently correlated with the freight rate increase or decrease rate. The value of is between [-1,1]. The positive or negative value of indicates whether the influencing factor is positively correlated or negatively correlated with the freight rate. The absolute value of indicates the degree of correlation between the influencing factor and the freight rate. The larger the absolute value, the stronger the correlation. A14: test statistic, ; The test statistic constructed by T is subject to Student distribution with degrees of freedom, giving a significance level , determine whether the influencing factors are relevant; A2: Perform Granger causality test. A21: Assumptions: The null hypothesis of the Granger causality test is that sequence X is not the Granger cause of sequence Y, and the alternative hypothesis is that sequence X is the Granger cause of sequence Y. Here, sequence X is the optimal linear prediction function for constructing sequence Y, which is denoted as: ; At this point, the assumptions of the Granger causality test are equivalently expressed as; ; Here Indicates possible influencing factors. Indicates the rate of increase or decrease in steel freight rates. represents the current time step, represents the best linear fit hour, Time ago to current The maximum time series that has an impact, represents random error, Indicates except Other influencing factors, Indicates the corresponding maximum impact timing, is the corresponding weight coefficient; A22: test statistic, Using the test statistic F, two regression models are fitted, and the freight rate is fitted when the null hypothesis is established; ; ; When the alternative hypothesis is true, the fitted series Unconstrained prediction model for ; ; Calculate test statistics and make judgments; ; like ; Then the original hypothesis is rejected, and it is believed that sequence X has an impact on sequence Y. At the same time, the corresponding factors are considered to have an impact on steel freight rates. Otherwise, it is considered to have no impact, and the factors that affect the addition and subtraction rate of steel freight rates are obtained. time series; Step 2: Extracting spatiotemporal heterogeneity, In the spatial dimension, the origin and destination points (OD pairs) of the transport routes are defined as discrete categorical variables, and the embedded layer coding method is used to represent the heterogeneous impact of the transport routes. In the temporal dimension, the core vector is generated by using the cyclic coding and multi-layer perceptron (MLP) technology as the input of the model parameters, so that the model parameters have time correlation, which is used to identify similar time periods and distinguish different time periods. The specific process includes B1 to B3: B1: Use cyclic coding to indicate the time position of the current time slot. The steel freight rate addition and subtraction rates are similar in certain periods, showing continuity and cyclicality; continuity means that adjacent periods have similar dynamics due to gradual changes in market conditions, and cyclicality reflects recurring patterns, such as monthly and annual demand fluctuations; in order to facilitate modeling, the time slot indicators are converted to cyclic coding; ; ; in and Represent the ranking of time slots in a month and a year, respectively. and They represent their relative positions in a season and a year respectively. Considering that the cyclic coding needs to satisfy continuity and periodicity, and It is impossible to meet both requirements at the same time, so the loop encoding is defined as And it is represented by sine and cosine functions: ; B2: In addition to considering similarities, we also need to consider differences. The changing patterns of steel freight rates in different periods are not the same, and the fluctuations of steel freight rates in special periods are irregular. Therefore, it is necessary to build a time-aware model and make the parameter values ​​depend on the time and change with the input; B21: Using multi-layer perceptron (MLP) technology Transformed into a deeper level of time coding ; ; To facilitate subsequent calculations, is rearranged into a diagonal matrix, the core matrix , whose diagonal value is The value of B22: The core matrix is ​​fed into a parallel calculator to generate dynamic time-dependent model parameters. The calculation process is as follows: ; in, is an additional term calculated using the core matrix, is a learnable matrix with a shape suitable for matrix multiplication. Through matrix multiplication and nonlinear transformation, the core matrix is ​​incrementally transformed into the target matrix. is called the static component because it is the same for all inputs, is called the dynamic component because it varies with the input, represents the model parameters generated by the convolutional layer; since is transformed from the time slot indicator, so becomes time dependent; is a nonlinear activation function, is the Sigmoid activation function, and the calculation formula is as follows: ; B3: Number each different OD pair in the data and obtain the OD set: ; The discrete variables are encoded by using an embedding layer, which is used to map the discrete values ​​of each category into a low-dimensional continuous vector space. , the embedding layer will output a d-dimensional embedding vector; ; in Yes Category The corresponding embedding vector has dimension d, and C is the total number of OD categories; Step 3: Freight rate forecast, Based on the number of cross-regional routes, the heterogeneous space-time routes are divided into multiple homogeneous subsets. An independent LSTM model is trained for each subset, and the steel freight rate addition and subtraction rate is predicted by weighted fusion. Twin components and interwoven components are constructed. The twin component is used to perform time series prediction on the freight rate addition and subtraction rate and steel price respectively, and the interwoven component is used to bridge the two time series so that they can exchange valuable information in the process of time-dependent modeling. Finally, the freight rate calculation formula is combined to output the prediction result of steel freight rate.

[0020] Freight rate calculation formula: The freight rate for each ton of full truckload, each 10kg of less-than-load cargo, and each container is called the freight rate. The current freight rate is composed of two parts: the base price 1 and base price 2: ; in Indicates the freight rate; Indicates the base price 1 of the goods. Base price 1 is the origin and destination operation fee that is not related to the transportation distance. This part of the fee is fixed; It represents the base price 2, which is the operating fee of the vehicle during operation; L is the transportation mileage; The freight rate calculation formula is as follows: ; in Represents freight rate, Indicates the tonnage of goods transported during full truck transportation. Indicates the corresponding freight rate The addition and subtraction rate, Indicates the capacity of the vehicle, symbol Indicates rounding up; C1: Determine base price 1 and base price 2, Base Price 1 and Base Price 2 are found in the following table by determining the railway transportation method and waybill number:

[0021] C2: Forecast corresponding freight rate The addition and subtraction rate , First, the heterogeneous spatiotemporal lines are divided into multiple homogeneous subsets according to the number of cross-regions. An independent LSTM is trained for each subset. For each independent LSTM, the time series of key influencing factors obtained in steps 1 and 2 are used. , Dynamic time parameters And the embedding vector generated by the categorical variable OD , combined with the historical addition and subtraction rates of steel freight rates Make predictions and finally perform weighted fusion to get the prediction results ; The time series data of freight rate increase / decrease rate is , the input data of LSTM Depend on , Stitched together: ; The LSTM model processes time series data based on a gating mechanism; at each time step , the calculation of LSTM is divided into multiple steps: input gate, forget gate, output gate and unit state update; the state update formula of LSTM is as follows: Forget Gate : ; Input Gate : ; Candidate unit status : ; Unit Update Status : ; Output Gate : ; Hide status updates : ; in is the hidden state at the previous time step, is the Sigmoid activation function, tanh is the hyperbolic tangent activation function, is the weight matrix of each gate in LSTM, is the bias term; The final output is: ; in is the predicted value at time step t; C3: Build twin components, Each twin component, namely the freight rate addition / subtraction rate and steel price, is predicted using the above C2 step to obtain the result and , both are time series; C4: Construct interleaving components to interact with time information, The interleaved component is a bridge between two components, allowing them to exchange information and perform feature extraction at the same time. That is, each twin component collects relevant information and shares necessary information. The output layer of each twin component collects features extracted from multiple spatiotemporal blocks through skip connections and fuses them for specific prediction tasks. Although two independent LSTM models are used to capture the temporal dependencies from the two time series respectively, the changing temporal dependencies are ignored; to solve this problem, a time-interleaved module is proposed to bridge the two separated time series. and , which can exchange valuable information in the process of temporal dependency modeling; first, two additional convolutional layers TCL1 and TCL2 are set to extract temporal features as information, respectively. and Shared in time series; ; Secondly, in order to avoid the transmission of irrelevant information between different sequences, a self-gating mechanism is designed to filter the shared information. Considering that the Sigmoid function can assign a smooth weighted value to each feature, has good nonlinear modeling ability, can flexibly control the information flow, and can avoid the gradient disappearance problem during training, the Sigmoid function in the self-gating mechanism helps the model filter irrelevant or noisy information through this weighted operation, thereby improving the performance and stability of the model. To represent the Sigmoid function, , Indicates shared information and The weights of each are: ; ; ; ; The Sigmoid function calculation formula is as follows: ; This is the final result of the steel freight rate addition and subtraction rate prediction. Combined with the freight rate calculation formula, the final freight rate prediction result is obtained. .

[0022] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for coupling prediction of railway freight rates for steel products considering the temporal and spatial heterogeneity of lines, characterized in that: The following steps are involved: Step 1: Identify the key factors that affect the freight rate increase or decrease rate based on the causal test method. Establish a data file, pre-process the obtained data and conduct preliminary screening by correlation analysis, and then use the Granger causality test method to identify the factors that have a key impact on the freight rate increase or decrease rate for each influencing factor; Step 2: Extracting spatiotemporal heterogeneity, In the spatial dimension, the starting and ending points of the transport routes are defined as discrete categorical variables, and the embedded layer coding method is used to represent the heterogeneous impact of the transport routes. In the temporal dimension, the core vector is generated by using the cyclic coding and multi-layer perceptron technology as the input of the model parameters, so that the model parameters have time correlation, which is used to identify similar time periods and distinguish different time periods. Step 3: Freight rate forecast, Based on the number of cross-regional routes, the heterogeneous space-time routes are divided into multiple homogeneous subsets. An independent LSTM model is trained for each subset, and the steel freight rate addition and subtraction rate is predicted by weighted fusion. Twin components and interwoven components are constructed. The twin component is used to perform time series prediction on the freight rate addition and subtraction rate and steel price respectively, and the interwoven component is used to bridge the two time series so that they can exchange valuable information in the process of time-dependent modeling. Finally, the freight rate calculation formula is combined to output the prediction result of steel freight rate.

2. The method for coupling prediction of railway freight rates for steel products considering the temporal and spatial heterogeneity of lines according to claim 1 is characterized in that: Step 1 also includes the following steps: A1: Correlation analysis, A11: Data preprocessing, Preprocess the obtained steel freight rates and possible influencing factors data, including quantification of influencing factor data, filling of missing values, deletion of duplicate values, and outlier processing, and finally obtain the time series of steel freight rate addition and subtraction rates and possible influencing factors; A12: Draw a scatter plot. The obtained data are plotted in a two-dimensional coordinate system in the form of scattered points, and the linear trend of the data is determined by the outline of the scatter plot; A13: Correlation calculation, The Pearson correlation coefficient is calculated to determine the correlation between each influencing factor and the freight rate. The correlation coefficient calculation formula is as follows: ； in represents the Pearson correlation coefficient, is the total number of samples, represents the sample mean, is the variable value. Here we need to explain: Indicates the freight rate increase or decrease rate. Indicates the value of the influencing factor. Each influencing factor needs to be independently correlated with the freight rate increase or decrease rate. The value of is between [-1,1]. The positive or negative value of indicates whether the influencing factor is positively correlated or negatively correlated with the freight rate. The absolute value of indicates the degree of correlation between the influencing factor and the freight rate. The larger the absolute value, the stronger the correlation. A14: test statistic, ； The test statistic constructed by T is subject to Student distribution with degrees of freedom, giving a significance level , determine whether the influencing factors are relevant; A2: Perform Granger causality test. A21: Assumptions: The null hypothesis of the Granger causality test is that sequence X is not the Granger cause of sequence Y, and the alternative hypothesis is that sequence X is the Granger cause of sequence Y. Here, sequence X is the optimal linear prediction function for constructing sequence Y, which is denoted as: ； At this point, the assumptions of the Granger causality test are equivalently expressed as; ； Here Indicates possible influencing factors. Indicates the rate of increase or decrease in steel freight rates. represents the current time step, represents the best linear fit hour, Time ago to current The maximum timing of the impact, represents random error, Indicates except Other influencing factors, Indicates the corresponding maximum impact timing, is the corresponding weight coefficient; A22: test statistic, Using the test statistic F, two regression models are fitted, and the freight rate is fitted when the null hypothesis is established; ； ； When the alternative hypothesis is true, the fitted series Unconstrained prediction model for ； ； Calculate test statistics and make judgments; ； like ; Then the original hypothesis is rejected, and it is believed that sequence X has an impact on sequence Y. At the same time, the corresponding factors are considered to have an impact on steel freight rates. Otherwise, it is considered to have no impact, and the factors that affect the addition and subtraction rate of steel freight rates are obtained. time series.

3. The method for coupled prediction of railway freight rates for steel products considering the temporal and spatial heterogeneity of lines according to claim 1 is characterized in that: Step 2 also includes the following steps: B1: Use cyclic coding to indicate the time position of the current time slot. The steel freight rate addition and subtraction rates are similar in certain periods, showing continuity and cyclicality; continuity means that adjacent periods have similar dynamics due to gradual changes in market conditions, and cyclicality reflects recurring patterns, such as monthly and annual demand fluctuations; in order to facilitate modeling, the time slot indicators are converted to cyclic coding; ； ； in and Represent the ranking of time slots in a month and a year, respectively. and They represent their relative positions in a season and a year respectively. Considering that the cyclic coding needs to satisfy continuity and periodicity, and It is impossible to meet both requirements at the same time, so the loop encoding is defined as And it is represented by sine and cosine functions: ； B2: In addition to considering similarities, we also need to consider differences. The changing patterns of steel freight rates in different periods are not the same, and the fluctuations of steel freight rates in special periods are irregular. Therefore, it is necessary to build a time-aware model and make the parameter values ​​depend on the time and change with the input; B21: Using multi-layer perceptron technology Transformed into a deeper level of time coding ; ； To facilitate subsequent calculations, is rearranged into a diagonal matrix, the core matrix , whose diagonal value is The value of B22: The core matrix is ​​fed into a parallel calculator to generate dynamic time-dependent model parameters. The calculation process is as follows: ； in, is an additional term calculated using the core matrix, is a learnable matrix with a shape suitable for matrix multiplication. Through matrix multiplication and nonlinear transformation, the core matrix is ​​incrementally transformed into the target matrix. is called the static component because it is the same for all inputs, is called the dynamic component because it varies with the input, represents the model parameters generated by the convolutional layer; since is transformed from the time slot indicator, so becomes time dependent; is a non-linear activation function, is the Sigmoid activation function, and the calculation formula is as follows: ； B3: Number each different OD pair in the data and obtain the OD set: ； The discrete variables are encoded by using an embedding layer, which is used to map the discrete values ​​of each category into a low-dimensional continuous vector space. , the embedding layer will output a d-dimensional embedding vector; ； in Yes Category The corresponding embedding vector has a dimension of d, and C is the total number of OD categories.

4. The method for coupling prediction of railway freight rates for steel products considering the temporal and spatial heterogeneity of lines according to claim 1 is characterized in that: Step 3 also includes the following steps: Freight rate calculation formula: The freight rate for each ton of full truckload, each 10kg of less-than-load cargo, and each container is called the freight rate. The current freight rate is composed of two parts: the base price 1 and base price 2: ； in Indicates the freight rate; Indicates the base price 1 of the goods. Base price 1 is the origin and destination operation fee that is not related to the transportation distance. This part of the fee is fixed; It represents the base price 2, which is the operating fee of the vehicle during operation; L is the transportation mileage; The freight rate calculation formula is as follows: ； in Represents freight rate, Indicates the tonnage of goods transported during full truck transportation. Indicates the corresponding freight rate The addition and subtraction rate, Indicates the capacity of the vehicle, symbol Indicates rounding up; C1: Determine base price 1 and base price 2, Base Price 1 and Base Price 2 are found by determining the railway transportation method and waybill number: C2: Forecast corresponding freight rate The addition and subtraction rate , First, the heterogeneous spatiotemporal lines are divided into multiple homogeneous subsets according to the number of cross-regions. An independent LSTM is trained for each subset. For each independent LSTM, the time series of key influencing factors obtained in steps 1 and 2 are used. , Dynamic time parameters And the embedding vector generated by the categorical variable OD , combined with the historical addition and subtraction rates of steel freight rates Make predictions and finally perform weighted fusion to get the prediction results ; The time series data of freight rate increase / decrease rate is , the input data of LSTM Depend on , Stitched together: ； The LSTM model processes time series data based on a gating mechanism; at each time step , the calculation of LSTM is divided into multiple steps: input gate, forget gate, output gate and unit state update; the state update formula of LSTM is as follows: Forget Gate : ; Input Gate : ; Candidate unit status : ; Unit Update Status : ; Output Gate : ; Hide status updates : ； in is the hidden state at the previous time step, is the Sigmoid activation function, tanh is the hyperbolic tangent activation function, is the weight matrix of each gate in LSTM, is the bias term; The final output is: ； in is the predicted value at time step t; C3: Build twin components, Each twin component, namely the freight rate addition / subtraction rate and steel price, is predicted using the above C2 step to obtain the result and , both are time series; C4: Construct interleaving components to interact with time information, Propose a time-interleaving module to bridge two separated time series and , which can exchange valuable information in the process of temporal dependency modeling; first, two additional convolutional layers TCL1 and TCL2 are set to extract temporal features as information, respectively. and Shared in time series; ； Secondly, in order to avoid the transmission of irrelevant information between different sequences, a self-gating mechanism is designed to filter the shared information. Considering that the Sigmoid function can assign a smooth weighted value to each feature, has good nonlinear modeling ability, can flexibly control the information flow, and can avoid the gradient disappearance problem during training, the Sigmoid function in the self-gating mechanism helps the model filter irrelevant or noisy information through this weighted operation, thereby improving the performance and stability of the model. To represent the Sigmoid function, , Indicates shared information and The weights they respectively hold; ； ； ； ； The Sigmoid function calculation formula is as follows: ; This is the final result of the steel freight rate addition and subtraction rate prediction. Combined with the freight rate calculation formula, the final freight rate prediction result is obtained. .

Images