A marine shipping freight rate estimation method based on fusion time sequence convolution network

Abstract

The present application relates to the technical field of machine learning, in particular to a kind of marine shipping freight rate estimation method fusing time series convolution network, in the present application, deviated path is screened out according to boundary threshold, mapping weight is selected according to jump amplitude, form periodic mapping weight tensor, time series convolution network algorithm is superimposed residual connection, and with flexible positive function, low, medium, high three grades of subinterval increment are accumulated, form ordered subinterval freight rate interval, make freight rate estimation result expand from single point value to interval result with boundary constraint and order constraint, calculate combination affinity with ordered subinterval freight rate interval and reference freight rate vector, obtain deviated candidate value using multilayer perception machine algorithm, select residual upper limit according to the number of quotes and truncate deviation, and quotation verification extends from error evaluation to interval landing, residual boundary, risk level joint determination, improve estimation boundary explainability and quotation risk identification stability.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method for estimating ocean freight rates by incorporating temporal convolutional networks. Background Technology

[0002] The field of machine learning technology aims to drive model parameter updates through sample data, enabling the model to generate predicted values ​​of target variables under given input variables. It generates estimated ocean freight rates for a specified billing period based on historical freight rates, route attributes, container type attributes, cost items, and supply and demand status, and uses mean absolute error, root mean square error, and mean absolute percentage error as performance evaluation indicators for the model.

[0003] A new method for estimating ocean freight rates using fusion temporal convolutional networks aims to establish a nonlinear mapping relationship between historical freight rate sequences, route attributes, container type attributes, cost items, supply and demand status, and ocean freight rates for the target billing period through causal convolution, dilated convolution, residual connections, and gating fusion mechanisms. This ensures that the estimation results meet preset thresholds for mean absolute error, root mean square error, or mean absolute percentage error, and provides numerical basis for quotation verification, freight rate trend calculation, and space revenue management.

[0004] Existing technologies mainly establish a nonlinear mapping relationship between input variables and ocean freight rates for the target billing period based on historical freight rates, route attributes, container type attributes, cost items, and supply and demand status. The output is usually a single freight rate estimate, and performance is measured by mean absolute error, root mean square error, and mean absolute percentage error. In practice, single-point estimates are insufficient to express the relationship between low-level, mid-level, and high-level boundaries. Quotation verification personnel can only judge risk based on the difference between the predicted value and the quoted value, and cannot know whether the quoted value is within the reasonable range, in the lower or higher range. When there are fluctuations in the operating area, offsets in the effective date of the fee item, or cross-month billing cycles, fixed time step inputs can easily include data from different periods in the same convolution window, causing the estimated value to respond prematurely or lagging behind. For new routes, new container type combinations, and low-frequency quotation combinations, the number of continuous historical samples is insufficient, and model parameter updates are easily affected by a small number of abnormal quotations. The offset magnitude lacks constraints on the number of quotations, which may result in excessively large deviation values ​​in scenarios with small changes near the benchmark freight rate. Furthermore, the existing estimation process may only output a single value based on a limited number of samples, which cannot limit the offset boundary of low-frequency combinations relative to the benchmark freight rate, thereby affecting the identification of abnormal quotations in quotation verification, trend calculation, and cabin revenue management. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for estimating ocean freight rates by incorporating temporal convolutional networks.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for estimating ocean freight rates by fusing temporal convolutional networks, comprising the following steps: S1: Based on historical freight rate sequences, route numbers, ports of origin, ports of destination, carrying areas, trade directions, container type codes, container capacity levels, applicable fee types, and fee item amounts, the route embedding, container type embedding, and fee item embedding are cross-weighted according to carrying areas, container type families, and fee structures to generate a benchmark freight rate reference vector. S2: Based on the benchmark fare reference vector, the billing cycle coordinates are matched with soft dynamic time warping, offset paths are filtered out according to boundary thresholds, and mapping weights are selected according to jump amplitude to generate a cycle mapping weight tensor. S3: Based on the periodic mapping weight tensor and the benchmark freight rate reference vector, the mapping sequence is input into the causal dilation convolution, and the temporal convolutional network algorithm is used to superimpose residual connections and accumulate three quantile increments according to the flexible positive value function to generate an ordered quantile freight rate range. S4: Based on the ordered quantile freight rate range and the benchmark freight rate reference vector, the bilinear affinity score is used, and a multilayer perceptron algorithm is employed to select the upper limit of the residual based on the number of quotations and truncate the offset to generate the affinity residual freight rate. S5: Based on the affinity residual freight value and the ordered quantile freight rate range, the risk ratio is obtained by dividing the range width by the median quantile freight rate. The level is selected according to the landing area status and the threshold range, and a bid risk level label is generated.

[0007] As a further embodiment of the present invention, the benchmark freight reference vector includes route reference value, container type reference value, cost item reference value, and billing cycle weight value; the cycle mapping weight tensor includes target cycle weight array, historical cycle weight array, starting point boundary weight, ending point boundary weight, and path jump identifier; the ordered quantile freight range includes low quantile freight rate, median quantile freight rate, high quantile freight rate, quantile interval width, and quantile skew direction; the affinity residual freight value includes combined affinity value, residual upper limit value, offset residual value, benchmark superimposed freight value, and quotation frequency tier identifier; and the quotation risk level marker includes freight rate landing status, interval width ratio, risk level number, quantile deviation direction, and quotation verification marker.

[0008] As a further aspect of the present invention, the specific steps for generating the benchmark freight reference vector are as follows: Based on historical freight rate sequences, route numbers, ports of origin, ports of destination, carrying areas, trade directions, container type codes, container capacity levels, applicable fee types, and fee item amounts, the quotations are divided into groups according to carrying areas, container families, and fee structures. The freight amount within the same quotation group is multiplied by the weight of the billing cycle and then summed to obtain a weighted table of combined freight rates. Based on the combined fare weighting table, the route embedding, container type embedding, and cost item embedding are multiplied by the combined fare weight value, and the regional component, container family component, and cost structure component are summed to generate a benchmark fare reference vector.

[0009] As a further aspect of the present invention, the specific steps for generating the periodic mapping weight tensor are as follows: Based on the benchmark fare reference vector, the distance value is calculated between the target billing cycle location and the historical billing cycle location, and locations exceeding the boundary threshold are filtered out to generate a cycle location matching table.

[0010] Based on the periodic position matching table, the weights of continuous positions are retained according to the jump magnitude, the weights of historical positions are assigned to the target periodic coordinates, and a periodic mapping weight tensor is generated.

[0011] As a further aspect of the present invention, the specific steps for generating the ordered quantile freight rate range are as follows: Based on the cycle mapping weight tensor and the benchmark fare reference vector, the historical location weight is read according to the target billing cycle coordinates. The historical location weight, the benchmark fare component, and the cost item amount component are multiplied one by one. The summation is performed on multiple product values ​​under the same target coordinates to obtain the cycle weighted response sequence. Based on the aforementioned periodically weighted response sequence, a temporal convolutional network algorithm is adopted. With the target coordinate as the output position, only the sequence values ​​of the target coordinate and the sequence values ​​before the target coordinate are selected. Three consecutive values ​​and five consecutive values ​​are extracted at intervals of one, two, four, eight, and sixteen, respectively. The extracted values ​​are multiplied and added to the corresponding convolution parameters one by one. The results of the multiplication and addition at multiple intervals are added to the retained values ​​of the same dimension to obtain the freight rate temporal response vector. Based on the aforementioned freight rate time-series response vector, the low-order output value, median increment value, and high-order increment value are read. The median increment value and the high-order increment value are restricted to non-negative values. The median increment value is added to the low-order output value, and the high-order increment value is added to the median output value. The difference between the high-order output value and the low-order output value is calculated to generate an ordered quantile freight rate interval.

[0012] As a further aspect of the present invention, the temporal convolutional network algorithm first reads the periodically weighted response sequence, sets the target coordinates as the output coordinates, establishes a sequence buffer window according to the coordinate order, and retains only the target coordinates and the values ​​before the target coordinates. The interval values ​​are set sequentially as 1, 2, 4, 8, and 16. Under each interval value, a position group of length 3 and a position group of length 5 are selected respectively. The position group of length 3 consists of the target coordinates, the position one interval ahead, and the position two intervals ahead. The position group of length 5 consists of the target coordinates, the position one interval ahead, the position two intervals ahead, the position three intervals ahead, and the position four intervals ahead. The values ​​of the position group of length 3 are multiplied and added item by item with the convolution parameters of length 3, and the values ​​of the position group of length 5 are multiplied and added item by item with the convolution parameters of length 5. The two multiplication and addition results under the same interval value are summed to obtain the interval response value. Multiple interval response values ​​are accumulated in layer order, and the residual retention value is called and the accumulated result is summed to output the freight rate temporal response vector.

[0013] As a further aspect of the present invention, the historical location weights are read according to the target billing cycle coordinates, the target billing cycle coordinate sequence number and the historical billing cycle location sequence number are determined, the target billing cycle coordinate sequence number is used as a tensor column index, the historical billing cycle location sequence number is used as a tensor row index, the row and column intersection cell weight value is read to obtain the historical location weights, when the same target billing cycle coordinate corresponds to multiple historical billing cycle locations, multiple weight values ​​are read sequentially from small to large according to the historical billing cycle location sequence number to form a historical location weight column vector, when the row and column intersection cell is empty, the weight value is recorded as 0, when the sum of multiple historical billing cycle location weight values ​​is not equal to 1, the multiple weight values ​​are normalized to obtain the historical location weights under the target billing cycle coordinates.

[0014] As a further aspect of the present invention, the specific steps for generating the affinity residual transport value are as follows: Based on the ordered quantile fare range and the benchmark fare reference vector, the median quantile fare, range width, route component, container type component, and cost structure component are read. The route component, the row and column values ​​of the weight matrix, and the container type component are multiplied and added. The multiplied and added values ​​are added to the cost structure component to generate a combined affinity scoring table. Based on the combined affinity scoring table, a multilayer perceptron algorithm is used to read the combined affinity value, the number of bids identifier, and the benchmark freight rate component. When the number of bids is less than five, 8% of the benchmark freight rate component is selected. When the number of bids is not less than five and less than twenty, 15% of the benchmark freight rate component is selected. When the number of bids is not less than twenty, the interval width ratio limit is selected to generate a residual limit table. Based on the residual limit table, the offset candidate value, residual upper limit value, and combined affinity value are read. The absolute value of the offset candidate value is compared with the residual upper limit value. If it exceeds the residual upper limit value, the residual upper limit value with the same sign is selected. If it does not exceed the residual upper limit value, the offset candidate value is retained. The selected offset value is added to the base freight rate component to generate the affinity residual freight value.

[0015] As a further aspect of the present invention, the multilayer perceptron algorithm is employed. First, the combined affinity value, the number of quotes identifier, and the benchmark freight rate component are read. These components are then arranged in the order of combined affinity value first, number of quotes identifier in the middle, and benchmark freight rate component last, yielding a residual input vector. The first-layer weight matrix and the first-layer bias vector are then used to perform a weighted summation on the residual input vector, resulting in the first-layer node values. Values ​​less than 0 in the first-layer node values ​​are replaced with 0, yielding the first hidden layer vector. Finally, the second-layer weight matrix and the second-layer bias vector are used to perform a weighted summation on the first hidden layer... The vector is weighted and summed to obtain the second-level node value. The values ​​less than 0 in the second-level node value are replaced with 0 to obtain the second hidden layer vector. The output layer weight vector and output layer bias value are called, and the second hidden layer vector is weighted and summed to obtain the offset candidate value. The quotation count indicator is compared with 5 and 20. When the quotation count indicator is less than 5, 8% of the benchmark freight rate component is recorded as the upper limit of the residual. When the quotation count indicator is not less than 5 and less than 20, 15% of the benchmark freight rate component is recorded as the upper limit of the residual. When the quotation count indicator is not less than 20, the interval width ratio limit is recorded as the upper limit of the residual, thus obtaining the residual limit table.

[0016] As a further aspect of the present invention, the specific steps for generating the quoted risk level marker are as follows: Based on the aforementioned affinity residual freight value and ordered quantile freight rate range, the range width is divided by the median quantile freight rate, and the residual freight value is compared with the quantile boundary to generate a quotation drop ratio table. Based on the aforementioned quotation landing area ratio table, a quotation risk level label is generated by selecting a level number according to the risk ratio threshold, low boundary, and high boundary.

[0017] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, cross-weighting is performed on historical freight rate sequences, route numbers, ports of origin, ports of destination, carrying areas, trade directions, container type codes, container capacity levels, applicable fee types, and fee item amounts to form a benchmark freight rate reference vector. This makes ocean freight rate estimation no longer dependent on fitting a single historical sequence, but incorporates differences in routes, container types, and fee structures into the same estimation chain. In this invention, offset paths are filtered out according to boundary thresholds, and mapping weights are selected according to jump amplitudes to form a periodic mapping weight tensor. This allows different start and end dates, different period lengths, and different cost effective positions to correspond to a unified periodic coordinate. Based on the periodic mapping weight tensor and the benchmark freight reference vector, a temporal convolutional network algorithm is used to superimpose residual connections, and a flexible positive value function is used to accumulate the increments of the low, middle, and high quantiles to form an ordered quantile freight interval. This expands the freight estimation result from a single-point value to an interval result with boundary constraints and order constraints. In this invention, the affinity of a combination is calculated using ordered quantile fare intervals and a benchmark fare reference vector. A multilayer perceptron algorithm is used to obtain offset candidate values. The upper limit of the residual is selected according to the number of quotations and the offset is truncated to form the affinity residual fare value. This ensures that low-frequency routes, low-frequency container types, and new combinations are still subject to the joint limitation of the benchmark fare and residual magnitude even when continuous samples are lacking. The quotation verification is extended from error evaluation to joint determination of interval positioning, residual boundaries, and risk levels, improving the interpretability of the estimation boundary and the stability of quotation risk identification. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the main steps of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] Example 1 Please see Figure 1 This invention provides a technical solution: a method for estimating ocean freight rates by fusing temporal convolutional networks, comprising the following steps: S1: Based on historical freight rate sequences, route numbers, ports of origin, ports of destination, carrying areas, trade directions, container type codes, container capacity levels, applicable fee types, and fee item amounts, the route embedding, container type embedding, and fee item embedding are cross-weighted according to carrying areas, container type families, and fee structures to generate a benchmark freight rate reference vector. S2: Based on the benchmark fare reference vector, the billing cycle coordinates are matched with soft dynamic time warping, offset paths are filtered out according to boundary thresholds, and mapping weights are selected according to jump magnitude to generate a cycle mapping weight tensor. S3: Based on the periodic mapping weight tensor and the benchmark freight rate reference vector, the mapping sequence is input into the causal dilation convolution, and the temporal convolutional network algorithm is adopted. The residual connection is superimposed, and the three-level quantile increment is accumulated according to the flexible positive value function to generate the ordered quantile freight rate range. S4: Based on the ordered quantile freight rate range and the benchmark freight rate reference vector, bilinear affinity is used for scoring. A multilayer perceptron algorithm is adopted to select the upper limit of the residual according to the number of quotations and truncate the offset to generate the affinity residual freight value. S5: Based on the affinity residual freight value and the ordered quantile freight rate range, the risk ratio is obtained by dividing the range width by the median quantile freight rate. The level is selected according to the landing area status and the threshold range to generate a quote risk level label.

[0021] The benchmark freight reference vector includes route reference value, container type reference value, cost item reference value, and billing cycle weight value. The cycle mapping weight tensor includes target cycle weight array, historical cycle weight array, origin boundary weight, destination boundary weight, and path jump indicator. The ordered quantile freight range includes low quantile freight rate, median quantile freight rate, high quantile freight rate, quantile interval width, and quantile skew direction. The affinity residual freight value includes combined affinity value, residual upper limit value, offset residual value, benchmark superimposed freight value, and quotation frequency tier indicator. The quotation risk level marker includes freight rate landing status, interval width ratio, risk level number, quantile deviation direction, and quotation verification marker.

[0022] The specific steps for generating the benchmark freight reference vector are as follows: Based on historical freight rate sequences, route numbers, ports of origin, ports of destination, carrying areas, trade directions, container type codes, container capacity classes, applicable fee types, and fee item amounts, an index-weighted moving average method is used. The grouping key is set as carrying area, container type family, and fee structure; the sorting key is set as billing cycle number; the amount key is set as freight rate amount; the attenuation coefficient is set to 0.85; and the number of valid cycles is set to 6. The quotations are divided into quotations based on carrying area, container type family, and fee structure. The freight rates of the 6 billing cycles within the same quotations are arranged from closest to furthest from the target billing cycle, and the cycle weights are calculated by decreasing by 0.85 period by period. The 6 cycle weights are normalized, and the freight rate amount is multiplied by the normalized cycle weights item by item. The product values ​​within the same quotations are summed to obtain the weighted quotations table. Based on the combined freight rate weighting table, the embedding lookup table method and the Hadamard product weighted summation method are adopted. The route embedding table is set to 32 dimensions, the container type embedding table to 16 dimensions, and the cost item embedding table to 16 dimensions. The route embedding is found by route number, port of origin, port of destination, carrying area, and trade direction. The container type embedding is found by container type code, container capacity class, and applicable cost type. The cost item embedding is found by cost structure. The container type embedding and cost item embedding are mapped to 32 dimensions respectively. The combined freight rate weight value in the combined freight rate weighting table is read. The route embedding, container type embedding, and cost item embedding are multiplied by the combined freight rate weight value respectively. The carrying area component, container type family component, and cost structure component are vector summed to generate the benchmark freight rate reference vector.

[0023] The specific steps for generating the periodic mapping weight tensor are as follows: Based on the benchmark fare reference vector, a soft dynamic time warping algorithm is adopted to read the target billing cycle location number, historical billing cycle location number, target cycle length value, and historical cycle length value. The temperature parameter is set to 0.1, the boundary threshold is set to 2, the row index of the distance matrix is ​​set to the historical billing cycle location number, and the column index of the distance matrix is ​​set to the target billing cycle location number. The difference operation is performed on the target billing cycle location number and the historical billing cycle location number. The squared term of the difference is multiplied by the reciprocal of the temperature parameter to obtain the location distance value. The location distance value is subjected to exponential negation to obtain the location matching weight. All location matching weights under the same target billing cycle location number are summed and normalized. The normalized location matching weights are compared with the boundary threshold. Weights whose historical billing cycle start point deviates from the first two locations of the target billing cycle are marked as boundary exceeding the limit. Weights whose historical billing cycle end point deviates from the last two locations of the target billing cycle are marked as boundary exceeding the limit. Boundary exceeding weights are filtered out to generate a cycle location matching table. Based on the periodic location matching table, a dynamic programming path backtracking method is adopted to read the target billing cycle location number, historical billing cycle location number, location matching weight, and boundary exceedance marker. The jump amplitude threshold is set to 1, and the continuous weight retention rule is set to the absolute value of the difference between the historical billing cycle location numbers corresponding to adjacent target billing cycle locations is no greater than 1. The target billing cycle location numbers are sorted in ascending order. The absolute value of the difference between the historical billing cycle location numbers corresponding to adjacent target billing cycle locations is calculated. The location weight with an absolute value of difference greater than 1 is marked as a jump exceedance. The location weight with both the boundary exceedance marker and the jump exceedance marker being 0 is written into the mapping candidate column. The mapping candidate column weights under the same target billing cycle location number are summed and normalized. The normalized historical location weights are written into the target cycle coordinate column. The historical billing cycle location numbers are written into the tensor row. The target cycle location numbers are written into the tensor column, generating a periodic mapping weight tensor.

[0024] The specific steps for generating ordered quantile freight rate ranges are as follows: Based on the periodic mapping weight tensor and the benchmark fare reference vector, the Einstein summation convention and tensor index mapping method are adopted. The target billing cycle coordinate sequence number is set as the column index, the historical billing cycle position sequence number is set as the row index, and the benchmark fare component sequence number and the fee item amount component sequence number are set as the channel index. The historical position weights at the row and column intersection positions are read according to the target billing cycle coordinates, and the weights of empty row and column cells are recorded as 0. The weights of all historical position weights under the same target billing cycle coordinates are summed and normalized. The historical position weights, benchmark fare components, and fee item amount components are multiplied item by item according to the same historical billing cycle position sequence number. The multiple product values ​​under the same target coordinates are summed to obtain the periodic weighted response sequence. Based on the periodically weighted response sequence, a temporal convolutional network algorithm is adopted. The target coordinates are set as the output position, the interval values ​​are set to 1, 2, 4, 8, and 16, the length of the convolutional parameter group is set to 3 and 5, and the padding value for out-of-bounds positions is set to 0. Only the target coordinates and the sequence values ​​before the target coordinates are selected. Under each interval value, 3 interval position values ​​and 5 interval position values ​​are extracted. The 3 interval position values ​​are multiplied and added to the convolutional parameter of length 3 item by item, and the 5 interval position values ​​are multiplied and added to the convolutional parameter of length 5 item by item. The two sets of multiplication and addition results under the same interval value are summed. The summed results under the 5 interval values ​​are arranged according to the coordinate index. The periodically weighted response sequence is linearly mapped to the same dimension to obtain the same dimension retained value. The same dimension retained value is added to the summed result according to the same coordinate index to obtain the freight rate time-series response vector. Based on the time-series response vector of freight rates, a quantile regression algorithm and a flexible positive value function are used. The low quantile is set to 0.1, the median quantile to 0.5, and the high quantile to 0.9. Low-quantile output weight vector, median increment weight vector, and high-quantile increment weight vector are set. The three weight vectors are then multiplied with the time-series response vector of freight rates and the corresponding bias values ​​are added to obtain the low-quantile output value, median increment value, and high-quantile increment value. The median increment value and the high-quantile increment value are then input into the flexible positive value function to obtain the non-negative median increment value and the non-negative high-quantile increment value, respectively. The non-negative median increment value is added to the low-quantile output value to obtain the median output value, and the non-negative high-quantile increment value is added to the median output value to obtain the high-quantile output value. The difference between the high-quantile output value and the low-quantile output value is calculated to generate the ordered quantile freight rate interval.

[0025] The temporal convolutional network algorithm first reads the periodically weighted response sequence, sets the target coordinates as the output coordinates, and establishes a sequence buffer window according to the coordinate order. The sequence buffer window only retains the target coordinates and the values ​​before the target coordinates. The interval values ​​are set sequentially to 1, 2, 4, 8, and 16. Under each interval value, a position group of length 3 and a position group of length 5 are selected respectively. The position group of length 3 consists of the target coordinates, the position one interval ahead, and the position two intervals ahead. The position group of length 5 consists of the target coordinates, the position one interval ahead, the position two intervals ahead, the position three intervals ahead, and the position four intervals ahead. The values ​​of the position group of length 3 are multiplied and added item by item with the convolution parameters of length 3, and the values ​​of the position group of length 5 are multiplied and added item by item with the convolution parameters of length 5. The two multiplication and addition results under the same interval value are summed to obtain the interval response value. Multiple interval response values ​​are accumulated in layer order. The residual retained value is called and the accumulated result is summed to output the freight rate temporal response vector.

[0026] Historical location weights are read based on the target billing cycle coordinates. The target billing cycle coordinate sequence number and the historical billing cycle location sequence number are determined. The target billing cycle coordinate sequence number is used as the tensor column index, and the historical billing cycle location sequence number is used as the tensor row index. The weight value of the row-column intersection cell is read to obtain the historical location weight. When the same target billing cycle coordinate corresponds to multiple historical billing cycle locations, multiple weight values ​​are read in ascending order of historical billing cycle location sequence number to form a historical location weight column vector. When the row-column intersection cell is empty, the weight value is recorded as 0. When the sum of multiple historical billing cycle location weight values ​​is not equal to 1, the multiple weight values ​​are normalized to obtain the historical location weight under the target billing cycle coordinate.

[0027] The specific steps for generating affinity residual transport value are as follows: Based on the ordered quantile fare range and the benchmark fare reference vector, the factor decomposition machine algorithm and bilinear mapping method are used to read the median quantile fare, range width, route component, box component, and cost structure component. The dimensions of the route component are set to 32, the box component to 16, the cost structure component to 16, and the second-order interaction latent vector to 16. The indexes of the route component, box component, and cost structure component are set. The route latent vector is read using the route component index, the box latent vector is read using the box component index, and the cost structure latent vector is read using the cost structure component index. The route latent vector and the box latent vector are multiplied in the same position, and the results of the same position multiplication are summed to obtain the route-box interaction value. The row values ​​of the weight matrix corresponding to the route component and the column values ​​of the weight matrix corresponding to the box component are read. The route component, the row and column values ​​of the weight matrix, and the box component are multiplied and added. The multiplied and added values, the route-box interaction value, and the cost structure component are summed to generate a combined affinity scoring table. Based on the combined affinity scoring table, a multilayer perceptron algorithm is used to read the combined affinity value, the number of bids, and the benchmark freight rate component. The residual input vector is arranged in the order of combined affinity value, number of bids, and benchmark freight rate component. The number of nodes in the first fully connected layer is set to 64, the number of nodes in the second fully connected layer is set to 32, and the number of nodes in the output layer is set to 1. The first layer weight matrix, the first layer bias vector, the second layer weight matrix, the second layer bias vector, the output layer weight vector, and the output layer bias value are set. The residual input vector is input into the first fully connected layer. The first layer weight matrix and the residual input vector are multiplied by matrix multiplication and then the first layer bias vector is added. The smaller of the resulting first layer node values ​​is then used to calculate the value. Replace 0 values ​​with 0 to obtain the first hidden layer vector, and input the first hidden layer vector into the second fully connected layer. Perform matrix multiplication between the second layer weight matrix and the first hidden layer vector, and add the second layer bias vector. Replace the values ​​less than 0 in the resulting second layer node values ​​with 0 to obtain the second hidden layer vector. Perform inner product between the output layer weight vector and the second hidden layer vector, and add the output layer bias value to obtain the offset candidate value. Compare the number of bids with 5 and 20. If the number of bids is less than 5, select 8% of the benchmark freight rate component. If the number of bids is not less than 5 and less than 20, select 15% of the benchmark freight rate component. If the number of bids is not less than 20, select the interval width ratio limit and generate the residual limit table. Based on the residual limit table, an interval truncation method is adopted. The offset candidate value, residual upper limit value, combined affinity value, and benchmark freight rate component are read. The lower truncation limit is set to the residual upper limit value multiplied by -1, and the upper truncation limit is set to the residual upper limit value. The offset candidate value sign and absolute value are read, and the absolute value of the offset candidate value is compared with the residual upper limit value. When the absolute value of the offset candidate value is greater than the residual upper limit value, the upper truncation limit is selected for positive offset candidate values, and the lower truncation limit is selected for negative offset candidate values. When the absolute value of the offset candidate value is not greater than the residual upper limit value, the offset candidate value is retained. The truncated offset value, combined affinity value, and benchmark freight rate component are then weighted and superimposed. When the combined affinity value is not less than 0.6, the full amount of the truncated offset value is selected; when the combined affinity value is less than 0.6, half of the truncated offset value is selected. The selected offset value is added to the benchmark freight rate component to generate the affinity residual freight value.

[0028] Employing a multilayer perceptron algorithm, the algorithm first reads the combined affinity value, the number of bids, and the benchmark freight rate component. These components are then arranged in the order of combined affinity value first, bids number identifier in the middle, and benchmark freight rate component last, yielding the residual input vector. The first-layer weight matrix and first-layer bias vector are then used to perform a weighted summation on the residual input vector, resulting in the first-layer node values. Values ​​less than 0 in the first-layer node values ​​are replaced with 0, yielding the first hidden layer vector. Finally, the second-layer weight matrix and second-layer bias vector are used to perform a weighted summation on the first hidden layer vector. The values ​​of the second-level nodes are obtained by summing the values ​​of the nodes. The values ​​less than 0 in the second-level nodes are replaced with 0 to obtain the second hidden layer vector. The output layer weight vector and the output layer bias value are called to perform a weighted summation on the second hidden layer vector to obtain the offset candidate value. The number of quotations is compared with 5 and 20. When the number of quotations is less than 5, 8% of the base freight rate component is recorded as the upper limit of the residual. When the number of quotations is not less than 5 and less than 20, 15% of the base freight rate component is recorded as the upper limit of the residual. When the number of quotations is not less than 20, the interval width ratio limit is recorded as the upper limit of the residual, thus obtaining the residual limit table.

[0029] The specific steps for generating a quote risk level label are as follows: Based on the affinity residual freight value and the ordered quantile freight rate range, a binary search algorithm and a ratio calculation method are used to read the affinity residual freight value, the low quantile freight rate, the median quantile freight rate, the high quantile freight rate, and the interval width. The ordered boundary sequence is set as the low quantile freight rate, the median quantile freight rate, and the high quantile freight rate. The search target value is set as the affinity residual freight value. The first comparison is performed between the search target value and the median quantile freight rate. When the search target value is less than the median quantile freight rate, the second comparison is performed between the search target value and the low quantile freight rate. When the search target value is not less than the median quantile freight rate, the second comparison is performed between the search target value and the high quantile freight rate. Values ​​less than the low quantile freight rate are marked as low outside, values ​​not less than the low quantile freight rate and less than the high quantile freight rate are marked as interval inside, and values ​​greater than the high quantile freight rate are marked as high outside. The interval width is divided by the median quantile freight rate to obtain the risk ratio, generating a quotation landing area ratio table. Based on the quotation landing area ratio table, a classification regression tree algorithm is used to read the risk ratio, low boundary, high boundary, landing area status, and affinity residual transport value. The first split node is set to a risk ratio of 3%, the second split node is set to a risk ratio of 8%, and the boundary nodes are set to low boundary and high boundary. When the risk ratio is not greater than 3% and the landing area status is within the interval, level number 1 is selected. When the risk ratio is greater than 3% but not greater than 8% and the landing area status is within the interval, level number 2 is selected. When the risk ratio is greater than 8%, level number 3 is selected. When the affinity residual transport value is less than the low boundary, a low deviation mark is selected. When the affinity residual transport value is greater than the high boundary, a high deviation mark is selected. The level number, low deviation mark, and high deviation mark are written into the same mark item to generate the quotation risk level mark.

[0030] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for estimating ocean freight rates by incorporating temporal convolutional networks, characterized in that, Includes the following steps: S1: Based on historical freight rate sequences, route numbers, ports of origin, ports of destination, carrying areas, trade directions, container type codes, container capacity levels, applicable fee types, and fee item amounts, the route embedding, container type embedding, and fee item embedding are cross-weighted according to carrying areas, container type families, and fee structures to generate a benchmark freight rate reference vector. S2: Based on the benchmark fare reference vector, the billing cycle coordinates are matched with soft dynamic time warping, offset paths are filtered out according to boundary thresholds, and mapping weights are selected according to jump amplitude to generate a cycle mapping weight tensor. S3: Based on the periodic mapping weight tensor and the benchmark freight rate reference vector, the mapping sequence is input into the causal dilation convolution, and the temporal convolutional network algorithm is used to superimpose residual connections and accumulate three quantile increments according to the flexible positive value function to generate an ordered quantile freight rate range. S4: Based on the ordered quantile freight rate range and the benchmark freight rate reference vector, the bilinear affinity score is used, and a multilayer perceptron algorithm is employed to select the upper limit of the residual based on the number of quotations and truncate the offset to generate the affinity residual freight rate. S5: Based on the affinity residual freight value and the ordered quantile freight rate range, the risk ratio is obtained by dividing the range width by the median quantile freight rate. The level is selected according to the landing area status and the threshold range, and a bid risk level label is generated.

2. The method for estimating ocean freight rates using fused temporal convolutional networks according to claim 1, characterized in that, The benchmark fare reference vector includes route reference value, container type reference value, cost item reference value, and billing cycle weight value. The cycle mapping weight tensor includes target cycle weight array, historical cycle weight array, origin boundary weight, destination boundary weight, and path jump identifier. The ordered quantile fare interval includes low quantile fare, median quantile fare, high quantile fare, quantile interval width, and quantile skew direction. The affinity residual fare value includes combined affinity value, residual upper limit value, offset residual value, benchmark superimposed fare value, and quotation frequency tier identifier. The quotation risk level marker includes fare landing status, interval width ratio, risk level number, quantile deviation direction, and quotation verification marker.

3. The method for estimating ocean freight rates using fused temporal convolutional networks according to claim 1, characterized in that, The specific steps for generating the benchmark freight reference vector are as follows: Based on historical freight rate sequences, route numbers, ports of origin, ports of destination, carrying areas, trade directions, container type codes, container capacity levels, applicable fee types, and fee item amounts, the quotations are divided into groups according to carrying areas, container families, and fee structures. The freight amount within the same quotation group is multiplied by the weight of the billing cycle and then summed to obtain a weighted table of combined freight rates. Based on the combined fare weighting table, the route embedding, container type embedding, and cost item embedding are multiplied by the combined fare weight value, and the regional component, container family component, and cost structure component are summed to generate a benchmark fare reference vector.

4. The method for estimating ocean freight rates using fused temporal convolutional networks according to claim 1, characterized in that, The specific steps for generating the periodic mapping weight tensor are as follows: Based on the benchmark fare reference vector, the distance value is calculated between the target billing cycle location and the historical billing cycle location, and locations exceeding the boundary threshold are filtered out to generate a cycle location matching table. Based on the periodic position matching table, the weights of continuous positions are retained according to the jump magnitude, and the historical position weights are assigned to the target periodic coordinates to generate a periodic mapping weight tensor.

5. The method for estimating ocean freight rates using fused temporal convolutional networks according to claim 1, characterized in that, The specific steps for generating the ordered quantile freight rate range are as follows: Based on the cycle mapping weight tensor and the benchmark fare reference vector, the historical location weight is read according to the target billing cycle coordinates. The historical location weight, the benchmark fare component, and the cost item amount component are multiplied one by one. The summation is performed on multiple product values ​​under the same target coordinates to obtain the cycle weighted response sequence. Based on the aforementioned periodically weighted response sequence, a temporal convolutional network algorithm is adopted. With the target coordinate as the output position, only the sequence values ​​of the target coordinate and the sequence values ​​before the target coordinate are selected. Three consecutive values ​​and five consecutive values ​​are extracted at intervals of one, two, four, eight, and sixteen, respectively. The extracted values ​​are multiplied and added to the corresponding convolution parameters one by one. The results of the multiplication and addition at multiple intervals are added to the retained values ​​of the same dimension to obtain the freight rate temporal response vector. Based on the aforementioned freight rate time-series response vector, the low-order output value, median increment value, and high-order increment value are read. The median increment value and the high-order increment value are restricted to non-negative values. The median increment value is added to the low-order output value, and the high-order increment value is added to the median output value. The difference between the high-order output value and the low-order output value is calculated to generate an ordered quantile freight rate interval.

6. The method for estimating ocean freight rates using fused temporal convolutional networks according to claim 5, characterized in that, The temporal convolutional network algorithm first reads the periodically weighted response sequence, sets the target coordinates as the output coordinates, and establishes a sequence buffer window according to the coordinate order. The sequence buffer window only retains the target coordinates and the values ​​before the target coordinates. The interval values ​​are set sequentially to 1, 2, 4, 8, and 16. Under each interval value, a position group of length 3 and a position group of length 5 are selected respectively. The position group of length 3 consists of the target coordinates, the position one interval ahead, and the position two intervals ahead. The position group of length 5 consists of the target coordinates, the position one interval ahead, the position two intervals ahead, the position three intervals ahead, and the position four intervals ahead. The values ​​of the position group of length 3 are multiplied and added item by item with the convolution parameters of length 3, and the values ​​of the position group of length 5 are multiplied and added item by item with the convolution parameters of length 5. The two multiplication and addition results under the same interval value are summed to obtain the interval response value. Multiple interval response values ​​are accumulated in layer order. The residual retained value is called and the accumulated result is summed to output the freight rate temporal response vector.

7. The method for estimating ocean freight rates using fused temporal convolutional networks according to claim 5, characterized in that, The process involves reading historical location weights based on the target billing cycle coordinates, determining the target billing cycle coordinate sequence number and the historical billing cycle location sequence number, using the target billing cycle coordinate sequence number as a tensor column index and the historical billing cycle location sequence number as a tensor row index, and reading the weight value of the row-column intersection cell to obtain the historical location weights. When the same target billing cycle coordinate corresponds to multiple historical billing cycle locations, multiple weight values ​​are read sequentially from smallest to largest according to the historical billing cycle location sequence number to form a historical location weight column vector. When the row-column intersection cell is empty, the weight value is recorded as 0. When the sum of multiple historical billing cycle location weight values ​​is not equal to 1, normalization processing is performed on multiple weight values ​​to obtain the historical location weights under the target billing cycle coordinates.

8. The method for estimating ocean freight rates using fused temporal convolutional networks according to claim 7, characterized in that, The specific steps for generating the affinity residual transport value are as follows: Based on the ordered quantile fare range and the benchmark fare reference vector, the median quantile fare, range width, route component, container type component, and cost structure component are read. The route component, the row and column values ​​of the weight matrix, and the container type component are multiplied and added. The multiplied and added values ​​are added to the cost structure component to generate a combined affinity scoring table. Based on the combined affinity scoring table, a multilayer perceptron algorithm is used to read the combined affinity value, the number of bids identifier, and the benchmark freight rate component. When the number of bids is less than five, 8% of the benchmark freight rate component is selected. When the number of bids is not less than five and less than twenty, 15% of the benchmark freight rate component is selected. When the number of bids is not less than twenty, the interval width ratio limit is selected to generate a residual limit table. Based on the residual limit table, the offset candidate value, residual upper limit value, and combined affinity value are read. The absolute value of the offset candidate value is compared with the residual upper limit value. If it exceeds the residual upper limit value, the residual upper limit value with the same sign is selected. If it does not exceed the residual upper limit value, the offset candidate value is retained. The selected offset value is added to the base freight rate component to generate the affinity residual freight value.

9. The method for estimating ocean freight rates using fused temporal convolutional networks according to claim 8, characterized in that, The algorithm employs a multilayer perceptron. First, it reads the combined affinity value, the number of quotes, and the benchmark freight rate component. These components are then arranged in the order of combined affinity value first, number of quotes in the middle, and benchmark freight rate component last, yielding a residual input vector. The first-layer weight matrix and the first-layer bias vector are then used to perform a weighted summation on the residual input vector, resulting in the first-layer node values. Values ​​less than 0 in the first-layer node values ​​are replaced with 0, yielding the first hidden layer vector. Finally, the second-layer weight matrix and the second-layer bias vector are used to perform a weighted summation on the first hidden layer vector. Summing yields the values ​​of the second-level nodes. Values ​​less than 0 in the second-level node values ​​are replaced with 0 to obtain the second hidden layer vector. The output layer weight vector and output layer bias value are called, and a weighted summation is performed on the second hidden layer vector to obtain the offset candidate values. The number of quotes is compared with 5 and 20. If the number of quotes is less than 5, 8% of the base freight rate component is recorded as the upper limit of the residual. If the number of quotes is not less than 5 and less than 20, 15% of the base freight rate component is recorded as the upper limit of the residual. If the number of quotes is not less than 20, the interval width ratio limit is recorded as the upper limit of the residual, thus obtaining the residual limit table.

10. The method for estimating ocean freight rates using fused temporal convolutional networks according to claim 1, characterized in that, The specific steps for generating the quoted risk level label are as follows: Based on the aforementioned affinity residual freight value and ordered quantile freight rate range, the range width is divided by the median quantile freight rate, and the residual freight value is compared with the quantile boundary to generate a table of bid landing ratios. Based on the aforementioned quotation landing area ratio table, a quotation risk level label is generated by selecting a level number according to the risk ratio threshold, low boundary, and high boundary.

Images