A cargo volume prediction model training method, cargo volume prediction method and system
By introducing the routing mapping matrix and line segment error loss function in the cargo volume prediction model, the problem of superposition of cargo volume prediction errors is solved and more accurate allocation of capacity resources is achieved, and resource waste and shortage of logistics companies is reduced.
Patent Information
- Application Number
- CN202111238840.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-10-25
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2041-10-25
AI Technical Summary
The existing logistics cargo forecasting methods have error superposition problems in the mapping process from flow to cargo volume to line segment cargo volume, and the flow accuracy has no direct relationship with the line segment accuracy, resulting in inaccurate allocation of transportation resources and waste of resources and shortage of resources.
A routing mapping matrix is introduced, and the line segment error is added to the loss function. During the training of machine learning model, the flow direction and line segment error are comprehensively considered, and the model is optimized to improve the prediction accuracy of line segment cargo volume.
By optimizing the model, the accuracy of line segment cargo volume prediction is improved, more accurate capacity resource scheduling is provided, resource waste and shortage are reduced, and the operational efficiency of logistics companies is improved.
Smart Images

Figure CN114169564B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of logistics, and more specifically, to a cargo volume prediction model training method, a cargo volume prediction method and system, a storage medium and a computer device. Background Art
[0002] Cargo volume is the core data of logistics companies. For logistics companies, cargo volume can be regarded as the demand for transportation resources, and transportation resources can be regarded as the supply. The accurate matching relationship between transportation resources (supply) and cargo volume (demand) is of great significance to optimizing the timeliness of cargo transportation, improving service quality and reducing operating costs. In the logistics industry, since the allocation of transportation resources has a lag, it is necessary to predict the demand for cargo volume in a certain period of time in the future in advance and allocate transportation resources accordingly. Summary of the invention
[0003] In view of this, a first aspect of the present invention provides a cargo volume prediction model training method, comprising:
[0004] S1. Obtain the characteristic data X and actual cargo volume data Y of all flows in the past N days from the database;
[0005] S2. Preprocess the feature data X and the actual cargo volume data Y to obtain the ideal feature data X' and cargo volume data Y', and divide the training set X train and Y train , validation set X valid and Y valid ;
[0006] S3, constructing a routing mapping matrix A according to the mapping relationship between the flow directions and the line segments in the planned routing;
[0007] S4. Select a machine learning model and customize the loss function of the machine learning model to integrate the loss function of the line segment error:
[0008] (1-λ)*L[f(X train ),Y train ]+λ*L[A T f(X train ),A T Y train ]
[0009] Among them, L[f(X train ),Y train ] represents the flow loss, L[A T f(X train ),A T Y train ] represents the line segment loss, and λ is a positive number less than 1;
[0010] S5. Using training set Xtrain and Y train Train the machine learning model to obtain the model f * ;
[0011] S6. X valid Input the model f * , output Calculate flow direction error and segment error
[0012] S7. Select different machine learning models and λ to train S4-S6, compare the obtained line segment errors, and select the model with the smallest line segment error as the best model.
[0013] Optionally, X is a matrix with P rows and F columns, each row is called a sample point, representing all features of a certain flow on a certain day, and each column is called a set of eigenvalues, representing the values of all sample points in this feature; Y is a vector with a length of P, and each element represents the actual cargo volume of a certain flow on a certain day.
[0014] Optionally, the preprocessing includes: filling missing values in the data by using a zero-filling method or a mean filling method; trimming outliers that deviate greatly from actual values, or discarding the sample points where they are located.
[0015] Optionally, the part of X' and Y' that belongs to the latest N' days is used as the validation set, denoted as X valid and Y valid ; The remaining earlier N-N' days are used as training sets, denoted as X train and Y train .
[0016] Optionally, in the routing mapping matrix, the row index of the matrix is the flow direction, and the column index is the line segment. If the flow direction i is mapped to the line segment j according to the routing, the element of the i-th row and j-th column of the matrix is 1, otherwise the element is 0.
[0017] A second aspect of the present invention provides a cargo volume forecasting method, comprising:
[0018] Get the feature data X of the day 1 ;
[0019] The feature data X of the day 1 By selecting the best model, we can get the forecast value of cargo flow
[0020] According to the planned route, the flow of cargo is mapped to the corresponding line segment, and the cargo volume on each line segment is summed up to obtain the line segment cargo volume forecast value.
[0021] A third aspect of the present invention provides a cargo volume forecasting system, comprising:
[0022] Feature data acquisition module, used to obtain feature data X of the day 1 ;
[0023] The flow volume forecasting module is used to convert the feature data X 1 By selecting the best model, we can get the forecast value of cargo flow
[0024] The line segment cargo volume prediction module is used to map the flow direction cargo volume prediction value to the corresponding line segment according to the route planning, and sum up the cargo volume prediction values on each line segment to obtain the line segment cargo volume prediction value.
[0025] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method provided in the first aspect of the present invention or the method provided in the second aspect of the present invention.
[0026] A fifth aspect of the present invention provides a computer device, comprising a processor and a memory storing a program, wherein when the processor executes the program, the method provided in the first aspect of the present invention or the method provided in the second aspect of the present invention is implemented.
[0027] The beneficial effects of the present invention are as follows:
[0028] The present invention introduces a routing mapping matrix, adds the line segment aggregation error to the loss function, and can simultaneously optimize the flow prediction accuracy and the line segment prediction accuracy when training the model, rather than improving the line segment accuracy by improving the flow accuracy alone, so that the model finally trained has good line segment cargo volume prediction accuracy. Since the capacity allocation is directly based on the line segment cargo volume rather than the flow cargo volume, the model obtained by the method provided by the present invention can provide a good basis for logistics companies to dispatch capacity resources and reduce the waste and shortage of capacity resources. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] The specific implementation modes of the present invention are further described in detail below in conjunction with the accompanying drawings.
[0030] Figure 1 Shows the flow direction and line segment schematic of logistics.
[0031] Figure 2 A flow chart of a cargo volume prediction model training method according to an embodiment of the present invention is shown.
[0032] Figure 3 A flow chart of a cargo volume prediction method according to an embodiment of the present invention is shown.
[0033] Figure 4 A diagram showing the structure of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION
[0034] In order to more clearly illustrate the present invention, the present invention is further described below in conjunction with preferred embodiments and accompanying drawings. Similar components in the accompanying drawings are represented by the same reference numerals. It should be understood by those skilled in the art that the content specifically described below is illustrative rather than restrictive, and should not be used to limit the scope of protection of the present invention.
[0035] The logistics network can be viewed as a network consisting of distribution nodes and lines connecting different distributions. Specifically, a distribution is a node where goods are transferred. Each order of goods has an originating distribution and a destination distribution. (Original distribution, destination distribution) is the flow direction of goods. In the actual operation process, in order to improve resource utilization, some goods need to be transferred to distributions other than the originating distribution and the destination distribution. Therefore, the actual route of the goods may be: originating distribution, first transfer distribution, second transfer distribution... third transfer distribution, destination distribution. For example Figure 1 As shown in the figure, the origin of the goods is Ningbo distribution, the destination is Foshan distribution, and the flow direction is (Ningbo distribution, Foshan distribution). In the actual transportation process, the goods are transferred in Hangzhou distribution and Guangzhou distribution in turn, and the actual route is Ningbo distribution, Hangzhou distribution, Guangzhou distribution, and Foshan distribution.
[0036] Goods are transported between different distribution rooms using transport resources such as shuttle buses. Line segments are an attribute of transport resources and can also be represented by a tuple. To distinguish them from flow directions, they are recorded as (starting distribution, ending distribution). The difference between flow directions and line segments is that flow directions describe the starting and ending points of goods, while line segments focus on the description of transport resources such as shuttle buses. Figure 1 In the example, there are three different line segments: (Ningbo distribution, Hangzhou distribution), (Hangzhou distribution, Guangzhou distribution) and (Guangzhou distribution, Foshan distribution).
[0037] In the logistics network cargo volume forecast, according to the statistical dimension, it can generally be divided into three types: full network cargo volume forecast, line segment cargo volume forecast, and flow direction cargo volume forecast, and the granularity of these three is gradually refined. Full network cargo volume forecast is to predict the total amount of cargo entering the network; line segment cargo volume forecast is to predict the total amount of cargo on a certain line segment; flow direction cargo volume forecast is to predict the amount of cargo in a certain flow direction. Among them, line segment cargo volume forecast is directly related to the allocation of transportation resources.
[0038] The existing methods mainly predict from the flow dimension, and then map the cargo volume in the flow direction to the corresponding line segment according to the route planning, and then add up the cargo volume on each line segment to obtain the capacity demand on each line segment, and then allocate the capacity resources according to the capacity demand. If the capacity demand on a certain day is small, some capacity resources can be idle to save costs. If the capacity demand on a certain day is high and exceeds the available capacity resources, it is necessary to temporarily add capacity resources to avoid cargo backlogs. The specific flow volume prediction method is as follows:
[0039] 1. Collect the characteristic data X of each flow direction on different days in history, including weather conditions, holidays, etc., and the corresponding cargo volume data, recorded as Y.
[0040] 2. Use machine learning algorithms to learn the mapping from X to Y, that is, to find f * =argmin f∈F L[f(X),Y], where L is the loss function, which measures the difference between two vectors. Commonly used methods in practice include mean absolute error, mean square error, etc. * The final machine learning model.
[0041] 3. In the prediction stage, we collect the characteristic data x of each flow direction on that day and use the model f obtained in step 2 * ,calculate Will As the flow prediction value. Then use route planning to map the flow volume to the corresponding line segment, and finally add up the volume on each line segment to get the line segment volume prediction value, and formulate the transport resource allocation plan based on it.
[0042] In the process of implementing the present invention, the inventors found that there are at least the following problems in the prior art:
[0043] 1. The method of summing up the forecasted flow volume faces the problem of error superposition. Since the allocation of transport resources is based on the predicted volume on the line segment, the results of the flow volume forecast need to be mapped to the corresponding line segment through routing planning and then summed. In this process, the prediction errors may be accumulated. For example, for the two directions (Ningbo, Guangzhou) and (Ningbo, Foshan), (Ningbo, Guangzhou) is mapped to the two line segments (Ningbo, Hangzhou) and (Hangzhou, Guangzhou), and (Ningbo, Foshan) is mapped to the two line segments (Ningbo, Hangzhou) and (Hangzhou, Foshan). It can be seen that these two directions have a common line segment (Ningbo, Hangzhou). If the forecast values of the two directions are both too large, the error superposition makes the volume on the line segment (Ningbo, Hangzhou) seriously too large, thereby misleading the subsequent allocation of transport resources. If a line segment corresponds to many directions, this situation is even more serious.
[0044] 2. There is no direct relationship between flow accuracy and line segment accuracy. A high flow accuracy does not strictly correspond to a high line segment accuracy. Therefore, the existing method of training by optimizing the flow error may not necessarily produce an excellent model.
[0045] In view of this, if Figure 2 As shown, the first embodiment of the present invention provides a cargo volume prediction model training method, comprising:
[0046] S1. Obtain the characteristic data X and actual cargo volume data Y of all flows in the past N days from the database;
[0047] X is a matrix with P rows and F columns. Each row is called a sample point, which represents all the features of a certain flow on a certain day. Each column is called a set of eigenvalues, which represents the values of all sample points in this feature. In practical applications, features often have practical significance, such as weather conditions, whether it is a holiday, etc., which can be used as the basis for model training. Y is a vector of length P, each element of which represents the actual volume of a certain flow on a certain day. In practice, a larger N includes more information, which can reduce the overfitting of the model and improve the generalization ability of the model. For example, N=800.
[0048] S2. Preprocess the feature data X and the actual cargo volume data Y to obtain the ideal feature data X' and cargo volume data Y', and divide the training set X train and Y train , validation set X valid and Y valid ;
[0049] In a possible implementation, the preprocessing includes: filling missing values in the data by using the zero-filling method or the mean filling method; trimming outliers that deviate seriously from the actual values, or discarding the sample points where they are located. In this way, relatively ideal feature data X and cargo volume data Y can be obtained.
[0050] In a possible implementation, the part of X' and Y' that belongs to the latest N' days is used as a validation set, denoted as X valid and Y valid ; The remaining earlier N-N' days are used as training sets, denoted as X train and Y train ; For example, 0.05N≤N'≤0.1N.
[0051] S3, constructing a routing mapping matrix A according to the mapping relationship between the flow directions and the line segments in the planned routing;
[0052] In a possible implementation, the routing mapping matrix has a row index for flow directions and a column index for line segments. If flow direction i is mapped to line segment j according to the routing, the element in the i-th row and j-th column of the matrix is 1, otherwise the element is 0. Since each flow direction can be mapped to multiple line segments, there may be more than one element 1 in each row of the matrix. Similarly, since each line segment can carry multiple flow directions, there may also be more than one element 1 in each column of the matrix. The matrix is denoted as A.
[0053] S4. Select a machine learning model and customize the loss function of the machine learning model to integrate the loss function of the line segment error:
[0054] (1-λ)*L[f(X train ),Y train ]+λ*L[A T f(X train ),A T Y train ]
[0055] Among them, L[f(X train ),Y train ] represents the flow loss, L[A T f(X train ),A T Y train ] represents the line segment loss, and λ is a positive number less than 1; the two are weightedly summed using a positive number λ less than 1, so that the flow loss and line segment loss can be comprehensively considered during the model training process.
[0056] S5. Using training set X train and Y train Train the machine learning model to obtain the model f * ;
[0057] S6. X valid Input the model f * , output Calculate flow direction error and segment error
[0058] S7. Select different machine learning models and λ to train S4-S6, compare the obtained line segment errors, and select the model with the smallest line segment error as the best model.
[0059] Since the segment error is directly related to the subsequent resource allocation, this error can be used to compare different models. In practice, different machine learning models can be used with different hyperparameters to train different models, and finally the segment error is selected. The smallest model.
[0060] The embodiment of the present invention introduces a routing mapping matrix, adds the line segment aggregation error to the loss function, and can optimize the flow prediction accuracy and line segment prediction accuracy at the same time when training the model, rather than improving the line segment accuracy by improving the flow accuracy alone, so that the model finally trained has good line segment cargo volume prediction accuracy. Since the capacity allocation is directly based on the line segment cargo volume rather than the flow cargo volume, the model obtained by the method described in this embodiment can provide a good basis for logistics companies to dispatch capacity resources and reduce the waste and shortage of capacity resources.
[0061] like Figure 3 As shown, the second embodiment of the present invention provides a cargo volume prediction method, comprising:
[0062] Get the feature data X of the day 1 ;
[0063] The feature data X of the day 1 By selecting the best model, we can get the forecast value of cargo flow
[0064] According to the planned route, the flow volume forecast value is mapped to the corresponding line segment, and the volume forecast values on each line segment are summed up to obtain the line segment volume forecast value.
[0065] According to the forecast value of cargo volume of the line segment, a transportation resource allocation plan can be formulated.
[0066] This embodiment can provide a good basis for logistics companies to dispatch transportation resources and reduce the waste and shortage of transportation resources.
[0067] A third embodiment of the present invention provides a cargo volume forecasting system, comprising:
[0068] Feature data acquisition module, used to obtain feature data X of the day 1 ;
[0069] The flow volume forecasting module is used to convert the feature data X 1 By selecting the best model, we can get the forecast value of cargo flow
[0070] The line segment cargo volume prediction module is used to map the flow cargo volume to the corresponding line segment according to the route planning, and sum up the cargo volume on each line segment to obtain the line segment cargo volume prediction value.
[0071] A fourth embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the method described in the first or second embodiment of the present invention is implemented.
[0072] In practical applications, the computer-readable storage medium can adopt any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, a system, device or device of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination of the above. More specific examples (non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present embodiment, the computer-readable storage medium can be any tangible medium containing or storing a program, which can be used by an instruction execution system, device or device or used in combination with it.
[0073] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, which carry computer-readable program code. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium other than a computer-readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device.
[0074] The program code embodied on the computer readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.
[0075] Computer program code for performing the operations of the present invention may be written in one or more programming languages or a combination thereof, including object-oriented programming languages such as Java, Smalltalk, C++, and conventional procedural programming languages such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).
[0076] like Figure 4As shown, a fifth embodiment of the present invention provides a schematic diagram of the structure of a computer device. Figure 4 The computer device 12 shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.
[0077] like Figure 4 As shown, the computer device 12 is in the form of a general-purpose computing device. The components of the computer device 12 may include, but are not limited to: one or more processors or processing units 16, a system memory 28, and a bus 18 that connects various system components (including the system memory 28 and the processing unit 16).
[0078] Bus 18 represents one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, a processor or a local bus using any of a variety of bus architectures. By way of example, these architectures include, but are not limited to, an Industry Standard Architecture (ISA) bus, a Micro Channel Architecture (MAC) bus, an Enhanced ISA bus, a Video Electronics Standards Association (VESA) local bus, and a Peripheral Component Interconnect (PCI) bus.
[0079] The computer device 12 typically includes a variety of computer system readable media. These media can be any available media that can be accessed by the computer device 12, including volatile and non-volatile media, removable and non-removable media.
[0080] The system memory 28 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) 30 and / or cache memory 32. The computer device 12 may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, the storage system 34 may be used to read and write non-removable, non-volatile magnetic media ( Figure 4 not shown, usually called a "hard drive"). Although Figure 4 Not shown in the figure, a disk drive for reading and writing to a removable non-volatile disk (e.g., a "floppy disk"), and an optical disk drive for reading and writing to a removable non-volatile optical disk (e.g., a CD-ROM, a DVD-ROM, or other optical media) may be provided. In these cases, each drive may be connected to the bus 18 via one or more data medium interfaces. The memory 28 may include at least one program product having a set (e.g., at least one) of program modules that are configured to perform the functions of the various embodiments of the present invention.
[0081] A program / utility 40 having a set (at least one) of program modules 42 may be stored, for example, in the memory 28, such program modules 42 including, but not limited to, an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment. The program modules 42 generally perform the functions and / or methods of the embodiments described herein.
[0082] The computer device 12 may also communicate with one or more external devices 14 (e.g., keyboards, pointing devices, displays 24, etc.), one or more devices that enable a user to interact with the computer device 12, and / or any device that enables the computer device 12 to communicate with one or more other computing devices (e.g., network cards, modems, etc.). Such communication may be performed through an input / output (I / O) interface 22. Furthermore, the computer device 12 may also communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) through a network adapter 20. Figure 4 As shown, the network adapter 20 communicates with other modules of the computer device 12 via the bus 18. It should be understood that although Figure 4 Not shown, other hardware and / or software modules may be used in conjunction with computer device 12, including but not limited to microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.
[0083] The processor unit 16 executes the programs stored in the system memory 28 to perform various functional applications and data processing, such as implementing the method described in the first or second embodiment of the present invention.
[0084] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not limitations on the implementation methods of the present invention. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made based on the above description. It is impossible to list all the implementation methods here. All obvious changes or modifications derived from the technical solution of the present invention are still within the protection scope of the present invention.
Claims
1. A cargo volume prediction model training method, characterized in that: include: S1. Obtain the characteristic data X and actual cargo volume data Y of all flows in the past N days from the database; S2. Preprocess the feature data X and the actual cargo volume data Y to obtain the ideal feature data X' and cargo volume data Y', and divide the training set X train and Y train , validation set X valid and Y valid ; X is a matrix with P rows and F columns. Each row is called a sample point, which represents all the features of a certain flow on a certain day. Each column is called a set of eigenvalues, which represents the values of all sample points in this feature. Y is a vector with a length of P. Each element represents the actual volume of a certain flow on a certain day. The pretreatment includes If there are missing values in the data, fill them with zero or mean values; if the values are seriously deviated from the actual values, trim them or discard the sample points where they are located; The part of X' and Y' that belongs to the latest N' days is used as the validation set, and is denoted as X valid and Y valid ; The remaining earlier N-N' days are used as training sets, denoted as X train and Y train ; S3, constructing a routing mapping matrix A according to the mapping relationship between the flow directions and the line segments in the planned routing; The row index of the routing mapping matrix is the flow direction, and the column index is the line segment. If the flow direction i is mapped to the line segment j according to the routing, the element of the i-th row and j-th column of the matrix is 1, otherwise the element is 0; S4. Select a machine learning model and customize the loss function of the machine learning model to integrate the loss function of the line segment error: (1-λ)*L[f(X train ),Y train ]+λ*L[A T f(X train ),A T Y train ], Among them, L[f(X train ),Y train ] represents the flow loss, L[A T f(X train ),A T Y train ] represents the line segment loss, and λ is a positive number less than 1; S5. Using training set X train and Y train Train the machine learning model to obtain the model f * ; S6. X valid Input the model f * , output Calculate flow direction error and segment error S7. Select different machine learning models and λ to train S4-S6, compare the obtained line segment errors, and select the model with the smallest line segment error as the best model.
2. A method for predicting cargo volume using the model obtained by the training method according to claim 1, characterized in that: include: Get the feature data X1 of the day; Pass the characteristic data X1 of the day through the selected best model to obtain the predicted value of the flow volume The predicted value of the flow volume is mapped to the corresponding line segment according to the planned route, and the predicted value of the volume on each line segment is summed up to obtain the predicted value of the line segment volume.
3. A cargo volume prediction system according to the cargo volume prediction model training method according to claim 1, characterized in that: include: The feature data acquisition module is used to obtain the feature data X1 of the day; The flow volume prediction module is used to obtain the flow volume prediction value by passing the feature data X1 of the day through the selected best model The line segment cargo volume prediction module is used to map the flow cargo volume to the corresponding line segment according to the route planning, and sum up the cargo volume on each line segment to obtain the line segment cargo volume prediction value.
4. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to claim 1 or the method according to claim 2 is implemented.
5. A computer device comprising a processor and a memory storing a program, characterized in that: When the processor executes the program, the method according to claim 1 is implemented or the method according to claim 2 is implemented.
Citation Information
Patent Citations
sales volume forecasting method and a training method, a device and an electronic system of a model thereof
CN109509030A
Traffic flow prediction method based on bidirectionally nested LSTM neural network
CN110070713A
