CN108564326A - Prediction technique and device, computer-readable medium, the logistics system of order - Google Patents

Abstract

A kind of prediction technique and device, computer-readable medium, logistics system of order, the prediction technique of the order include：Obtain the corresponding time series of History Order；Based on the time series, characteristic sequence is extracted；Based on the characteristic sequence, 2D signal characteristic pattern is generated；Based on the 2D signal characteristic pattern, neural network model is built, and order forecasting is carried out according to constructed neural network model.Using said program, the accuracy of order forecasting can be improved.

Description

Order forecasting method and device, computer readable medium, logistics system

technical field

The embodiments of the present invention relate to the field of logistics, and in particular, to an order forecasting method and device, a computer readable medium, and a logistics system.

Background technique

Because the order forecast of vehicle logistics can make dispatchers prepare in advance and plan ahead, so that the scheduling of transportation resources is more reasonable, so accurate order forecast of vehicle logistics plays a very important role in vehicle logistics.

Time series can describe the discrete values of variables over time, such as temperature, humidity, stock prices, car sales trends and other variables. Time series prediction refers to determining an unknown time point (t+ 1) or the value of an unknown time period (t+1, t+n). Often, a time series forecast may depend on one or more other time series in addition to past time. Therefore, a time series forecasting problem refers to a regression problem in a high-dimensional space. In a high-dimensional space, the predicted value is a highly non-linear function of its past values and other related time series. The order forecasting problem of vehicle logistics can be regarded as a time series forecasting problem in essence. The vehicle logistics order forecasting problem can be regarded as a time series forecasting problem in essence, so it can be solved by a time series forecasting model, such as the traditional Auto Regressive Moving Average model (Auto Regressive Moving Average, ARMA), feedforward neural network model , Recurrent Neural Network (RNN), variant Long Short-Term Memory neural network (Long Short-Term Memory, LSTM), etc. When these models solve the problem of vehicle logistics order forecasting, the logistics order data is combined into one-dimensional data as input, and then the time correlation on the data is extracted.

In the existing forecasting method of vehicle logistics orders, when there are multiple time series or multiple feature sequences as input, through a simple method of merging into one sequence, the synthesized one-dimensional sequence is used as model input for learning and predict.

The accuracy of the forecasting method for the above-mentioned vehicle logistics orders is poor.

Contents of the invention

The technical problem solved by the embodiments of the present invention is how to improve the accuracy of order forecasting.

In order to solve the above technical problems, an embodiment of the present invention provides an order prediction method, the method comprising: obtaining a time series corresponding to historical orders; extracting a feature sequence based on the time series; and generating a two-dimensional order based on the feature sequence A signal feature map; constructing a neural network model based on the two-dimensional signal feature map, and performing order prediction according to the constructed neural network model.

Optionally, the acquiring the time series corresponding to the order includes: acquiring raw data corresponding to historical orders; preprocessing the raw data to acquire the time series corresponding to the order.

Optionally, the preprocessing includes at least one of the following: outlier processing and missing value processing.

Optionally, the extracting the feature sequence includes: extracting the feature sequence based on a wavelet transform algorithm.

Optionally, the generating a two-dimensional signal feature map includes: dividing each feature sequence into a plurality of sequence segments with a length of n, where n is a positive integer; copying sequence segments corresponding to different feature sequences line by line to generate m rows of sequence fragments, wherein the m rows of sequence fragments satisfy different feature sequence rows adjacent to each other, wherein m is a positive integer; based on the m rows of sequence fragments, an m*n two-dimensional signal feature map is generated.

Optionally, the dividing each feature sequence into a plurality of sequence fragments with a length of n includes: dividing each characteristic sequence into a plurality of sequence fragments with a length of n based on a shift operation.

Optionally, the constructing a neural network model includes: constructing a neural network model; based on the two-dimensional signal feature map, training the neural network model and acquiring parameters of the neural network model.

Optionally, after the neural network model is constructed, the order prediction method further includes: acquiring online data of the order; training and updating the neural network model based on the online data.

Optionally, the neural network is: a convolutional neural network.

An embodiment of the present invention provides an order prediction device, including: a first acquisition unit, adapted to acquire a time series corresponding to historical orders; an extraction unit, adapted to extract a feature sequence based on the time series; a generation unit, adapted to The feature sequence generates a two-dimensional signal feature map; the construction unit is adapted to construct a neural network model based on the two-dimensional signal feature map, and perform order prediction according to the constructed neural network model.

Optionally, the first acquisition unit includes: a first acquisition subunit adapted to acquire raw data corresponding to historical orders; a second acquisition subunit adapted to preprocess the raw data and acquire time series corresponding to orders.

Optionally, the preprocessing includes at least one of the following: outlier processing and missing value processing.

Optionally, the extraction unit is adapted to extract the feature sequence based on a wavelet transform algorithm.

Optionally, the generation unit includes: a segmentation subunit, adapted to divide each feature sequence into a plurality of sequence fragments with a length of n, where n is a positive integer; a copy subunit, adapted to divide the corresponding The sequence fragments are copied line by line to generate m-line sequence fragments, and the m-line sequence fragments satisfy different characteristic sequence lines adjacent to each other, wherein m is a positive integer; generate subunits, which are suitable for generating sub-units based on the m-line sequence fragments , to generate a two-dimensional signal feature map of m*n.

Optionally, the segmentation subunit is adapted to segment each feature sequence into a plurality of sequence fragments with a length of n based on a shift operation.

Optionally, the construction unit includes: a construction subunit adapted to construct a neural network model based on the neural network model parameters; a training subunit adapted to train the neural network model based on the two-dimensional signal feature map , to obtain the parameters of the neural network model; the prediction subunit is adapted to perform order prediction according to the trained neural network model.

Optionally, the device for predicting orders further includes: a second acquiring unit adapted to acquire online data of orders; an updating unit adapted to train and update the neural network model based on the online data.

Optionally, the neural network is: a convolutional neural network.

An embodiment of the present invention provides a computer-readable storage medium, on which computer instructions are stored, and the steps of any one of the above-mentioned methods are executed when the computer instructions are executed.

An embodiment of the present invention provides a logistics system, including a memory and a processor. The memory stores computer instructions that can run on the processor. When the processor runs the computer instructions, any one of the above rights is executed. The steps of the method.

Compared with the prior art, the technical solutions of the embodiments of the present invention have the following beneficial effects:

In the embodiment of the present invention, a two-dimensional signal feature map is generated based on the feature sequence, and then a neural network model is constructed based on the two-dimensional signal feature map, and order prediction is performed according to the constructed neural network model. Since the one-dimensional related sequences are merged into two-dimensional feature maps, and the neural network is used to learn the correlation between the time adjacent points of each sequence and each adjacent sequence point at the same time, the maximum learning of the signal feature map is achieved. Containing features can effectively improve the accuracy of order forecasting.

Further, by performing outlier processing on the raw data corresponding to historical orders, outliers can be eliminated, so as to prevent outliers from affecting the accuracy of subsequent predictions.

Furthermore, the accuracy of order forecasting is improved by performing missing value processing on the raw data corresponding to historical orders.

Description of drawings

FIG. 1 is a flow chart of an order forecasting method provided by an embodiment of the present invention;

Fig. 2 is a flowchart of another order forecasting method provided by an embodiment of the present invention;

Fig. 3 is a flow chart of another order forecasting method provided by an embodiment of the present invention;

Fig. 4 is a schematic structural diagram of an order forecasting device provided by an embodiment of the present invention.

Detailed ways

The existing methods cannot learn the correlation between multiple sequences or multiple features well, resulting in the features contained in the input data not being well learned by the machine learning model. Therefore, the accuracy of order forecasting is poor.

In the embodiment of the present invention, a two-dimensional signal feature map is generated based on the feature sequence, and then a neural network model is constructed based on the two-dimensional signal feature map, and order prediction is performed according to the constructed neural network model. Since the one-dimensional related sequences are merged into two-dimensional feature maps, and the neural network is used to learn the correlation between the time adjacent points of each sequence and each adjacent sequence point at the same time, the maximum learning of the signal feature map is achieved. Containing features can effectively improve the accuracy of order forecasting.

In order to make the above objects, features and beneficial effects of the present invention more comprehensible, specific embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings.

Referring to Fig. 1, an embodiment of the present invention provides a method for forecasting an order, which may include the following steps:

Step S101, obtaining the time series corresponding to historical orders.

In the specific implementation, in order to predict the order, it is first necessary to obtain the time series corresponding to the historical order, that is, the relevant data of the historical order, and then predict the future data of the order based on the relevant data of the historical order, that is, the regression problem of the high-dimensional space. In a high-dimensional space, the predicted value is a highly non-linear function of its past values and other related time series.

In specific implementation, it may not be possible to directly obtain the desired time series, but only raw data such as vehicle logistics order time series data and inventory sequence data corresponding to historical orders can be obtained. Therefore, the original data corresponding to historical orders can be obtained, and then the original data can be preprocessed through certain algorithms and logic, for example, data cleaning, and the time series corresponding to orders can be obtained.

In an embodiment of the present invention, the preprocessing includes: abnormal value processing. For example, check and check the original data corresponding to historical orders, and eliminate outliers, so as to prevent outliers from affecting the accuracy of subsequent forecasts.

In another embodiment of the present invention, the preprocessing includes: missing value processing. For example, check and check the original data corresponding to historical orders, and fill in corresponding missing values to improve the accuracy of order forecasting.

In yet another embodiment of the present invention, missing value processing and outlier value processing exist simultaneously.

It can be understood that other preprocessing can also be performed, which will not be repeated here.

Step S102, extracting feature sequences based on the time series.

In a specific implementation, the feature sequence corresponding to the time series may be extracted based on a wavelet transform algorithm.

In the specific implementation, for the preprocessed time series, all the features can be extracted first, for example, M different features are extracted, and then correlation analysis and judgment are performed, and the N most relevant feature sequences are selected from the M feature sequences , where N and M are both positive integers, and N≤M.

Step S103, generating a two-dimensional signal feature map based on the feature sequence.

In a specific implementation, each feature sequence can be divided into a plurality of sequence fragments with a length of n, where n is a positive integer; then the sequence fragments corresponding to different feature sequences are copied row by row to generate m rows of sequence fragments, the m rows of sequence fragments satisfy that different feature sequence rows are adjacent to each other in pairs, where m is a positive integer, that is, copy the sequence fragments corresponding to the same index for each feature sequence to generate m rows of sequence fragments, and the m rows of sequences Among the fragments, the sequence fragments corresponding to the same index of any two feature sequences are located in two adjacent rows, where m is a positive integer, and each row is a sequence fragment corresponding to the same index of the feature sequence; finally, based on the m row sequence fragments, generate Two-dimensional signal feature map of m*n.

In an embodiment of the present invention, based on a shift operation, each feature sequence is divided into a plurality of sequence fragments with a length of n.

In a specific implementation, the shift operation may be one bit at a time, or more than two bits at a time, which can be flexibly selected according to actual conditions, and is not limited in this embodiment of the present invention.

For example, based on step S102, 5 feature sequences are extracted, respectively a, b, c, d, e, and the length of each feature sequence is L, that is, each feature sequence contains L bits of information. First, through shift slice segmentation, the five characteristic sequences of length L are divided into L-n sequence fragments of length n, respectively: sequence fragment 1 corresponds to the first to nth bits of the characteristic sequence, and sequence fragment 2 corresponds to The 2nd to the n+1th bit of the characteristic sequence, ..., the sequence fragment L-n corresponds to the L-n to the Lth bit of the characteristic sequence; then generate m row sequence fragments that satisfy different characteristic sequence rows adjacent to each other, For the five feature sequences a, b, c, d, e, generate m-row sequence fragments satisfying that the rows are adjacent to each other in pairs: abcdeacebda, m=11, satisfying that any two different feature sequences are adjacent to each other. Finally, based on the m-row sequence fragments, an m*n two-dimensional signal feature map can be generated.

Step S104, constructing a neural network model based on the two-dimensional signal feature map, and performing order prediction according to the constructed neural network model.

In a specific implementation, the parameters of the neural network model may be calculated based on the two-dimensional signal feature map; and then a neural network model may be constructed based on the parameters of the neural network model.

In a specific implementation, the neural network model can be constructed by using a convolutional neural network. A convolutional neural network (Convolutional Neural Network, CNN) is a feed-forward neural network, and its artificial neurons can respond to surrounding areas within a part of the coverage. Unit, which has excellent performance for large image processing. It includes alternating convolutional layers and pooling layers. It has good characteristics such as displacement invariance and weight sharing, and can extract abstract features well.

In a specific implementation, the characteristic map can be divided into several parts of training, verification and testing, and input to the neural network model to complete the process of training and adjusting parameters. The trained neural network model can be used as a model for order prediction.

In specific implementation, as time goes by, new logistics order data will be continuously input into the neural network system, so it is also possible to train and update the neural network model based on real-time online data. The relearning of data updates the convolutional neural network model that was previously trained offline so that it can also keep pace with the times.

Applying the above scheme, since the one-dimensional related sequences are merged into two-dimensional feature maps, and the neural network is used to learn the correlation between the time adjacent points of each sequence and the adjacent sequence points at the same time, the maximum learning The implied features of the signal feature map can effectively improve the accuracy of order forecasting.

In order to enable those skilled in the art to better understand and implement the present invention, the embodiment of the present invention provides the accuracy comparison results of order forecasting using the scheme provided by the embodiment of the present invention and the existing scheme, as shown in the chart.

Table 1

Referring to Table 1, based on the vehicle logistics order data during the four years from 2013 to 2016, verify the accuracy of the order forecast of the solution provided by the embodiment of the present invention and the existing solution. The three-year data from 2013 to 2015 is used as training data to generate a neural network model, and then based on the generated neural network, the order data in 2016 is predicted and compared with the real order data in 2016, and the error rate is given. The specific calculation formula is: : Error rate=(actual value-forecast value)/actual value.

It can be seen from Table 1 that the error rate of order forecasting can be effectively reduced by adopting the solution provided by the embodiment of the present invention, thereby providing higher accuracy of order forecasting.

In order to enable those skilled in the art to better understand and implement the present invention, the embodiment of the present invention provides another order forecasting method, as shown in FIG. 2 .

Referring to Fig. 2, the forecasting method of the order may include the following steps:

Step S201, acquiring offline order data.

Step S202, preprocessing the order data.

Step S203, extracting a feature sequence based on the order data.

Step S204, constructing a two-dimensional signal feature map and dividing it into a training set, a verification set and a test set.

Step S205, constructing a convolutional neural network.

In a specific implementation, the convolutional neural network is a convolutional neural network model for order prediction.

Step S206, training the convolutional neural network based on the two-dimensional signal feature map.

In a specific implementation, the convolutional neural network may be trained based on the two-dimensional signal feature map, so as to modify parameters of the convolutional neural network.

Step S207, judging whether the convolutional neural network satisfies the stop condition, when the convolutional neural network satisfies the stop condition, execute step S208, otherwise execute step S206.

In a specific implementation, the stop condition may be: based on the verification set, the calculated error of the convolutional neural network no longer decreases, that is, the convolutional neural network converges.

Step S208, judging whether the error of the convolutional neural network satisfies a preset threshold, and when the error of the convolutional neural network meets the preset threshold, execute step S209, otherwise execute step S210.

Step S209, saving the convolutional neural network model for order prediction.

Step S210, adjust the convolutional neural network structure and parameters, and repeat step S205.

In order to enable those skilled in the art to better understand and implement the present invention, the embodiment of the present invention provides another order forecasting method, as shown in FIG. 3 .

Step S301, acquiring online order data.

Step S302, preprocessing the order data.

Step S303, extracting a feature sequence based on the order data.

Step S304, constructing a two-dimensional signal feature map.

Step S305, updating the convolutional neural network model based on online learning.

Step S306, predicting orders based on the convolutional neural network model.

Step S307, outputting the prediction result.

In order to enable those skilled in the art to better understand and implement the present invention, an embodiment of the present invention further provides a forecasting device capable of implementing the above order forecasting method, as shown in FIG. 4 .

Referring to Fig. 4, the forecasting device of the order includes: a first acquisition unit 41, an extraction unit 42, a generation unit 43 and a construction unit 44, wherein:

The first acquisition unit 41 is adapted to acquire time series corresponding to historical orders.

The extraction unit 42 is adapted to extract a feature sequence based on the time sequence.

The generating unit 43 is adapted to generate a two-dimensional signal feature map based on the feature sequence.

The construction unit 44 is adapted to construct a neural network model based on the two-dimensional signal feature map, and perform order prediction according to the constructed neural network model.

In a specific implementation, the first acquisition unit 41 may include: a first acquisition subunit 411 and a second acquisition subunit 412, wherein:

The first obtaining subunit 411 is adapted to obtain raw data corresponding to historical orders.

The second acquiring subunit 412 is adapted to preprocess the raw data and acquire the time series corresponding to the order.

In an embodiment of the present invention, the preprocessing includes at least one of the following: abnormal value processing and missing value processing.

In an embodiment of the present invention, the extraction unit 42 is adapted to extract a feature sequence based on a wavelet transform algorithm.

In a specific implementation, the generation unit 43 includes: a division subunit 431, a replication subunit 432, and a generation subunit 433, wherein:

The segmentation subunit 431 is adapted to segment each feature sequence into a plurality of sequence fragments with a length of n, where n is a positive integer.

The copying subunit 432 is adapted to copy sequence fragments corresponding to different characteristic sequences row by row to generate m rows of sequence fragments, and the m rows of sequence fragments satisfy that two rows of different characteristic sequences are adjacent to each other.

The generation subunit 433 is adapted to generate m*n two-dimensional signal feature maps based on the m-row sequence segments.

In an embodiment of the present invention, the segmentation subunit 431 is adapted to segment each feature sequence into a plurality of sequence segments with a length of n based on a shift operation.

In a specific implementation, the construction unit 44 includes: a construction subunit 441, a training subunit 442 and a prediction subunit 443, wherein:

The construction subunit 441 is adapted to construct a neural network model based on the neural network model parameters.

The training subunit 442 is adapted to train the neural network model based on the two-dimensional signal feature map, and acquire parameters of the neural network model.

The prediction subunit 443 is adapted to perform order prediction according to the trained neural network model.

In a specific implementation, the order forecasting device 40 may also include: a second acquiring unit (not shown) and an updating unit (not shown), wherein:

The second acquiring unit is adapted to acquire online data of the order.

The update unit is adapted to train and update the neural network model based on the online data.

In an embodiment of the present invention, the neural network is a convolutional neural network.

In a specific implementation, the workflow and principle of the order forecasting device 40 can refer to the description in the method provided in the above embodiment, and will not be repeated here.

An embodiment of the present invention provides a computer-readable storage medium. The computer-readable storage medium is a non-volatile storage medium or a non-transitory storage medium, and computer instructions are stored thereon. When the computer instructions are executed, any one of the above-mentioned The steps corresponding to the method will not be repeated here.

An embodiment of the present invention provides a logistics system, including a memory and a processor, the memory stores computer instructions that can run on the processor, and the processor executes any of the above-mentioned computer instructions when running the computer instructions. The steps corresponding to the above method will not be repeated here.

Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above-mentioned embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, and the storage medium can include: ROM, RAM, disk or CD, etc.

Although the present invention is disclosed above, the present invention is not limited thereto. Any person skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, so the protection scope of the present invention should be based on the scope defined in the claims.

Claims (20)

1. A method for forecasting an order, comprising: Obtain the time series corresponding to historical orders; extracting a feature sequence based on the time series; Generate a two-dimensional signal feature map based on the feature sequence; A neural network model is constructed based on the two-dimensional signal feature map, and order prediction is performed according to the constructed neural network model. 2. The method for forecasting an order according to claim 1, wherein said obtaining the time series corresponding to the order comprises: Obtain raw data corresponding to historical orders; Preprocess the raw data to obtain the time series corresponding to the order. 3 . The order prediction method according to claim 2 , wherein the preprocessing includes at least one of the following: abnormal value processing and missing value processing. 4 . 4. The order prediction method according to claim 1, wherein said extracting the feature sequence comprises: extracting the feature sequence based on a wavelet transform algorithm. 5. The method for predicting an order according to claim 1, wherein said generating a two-dimensional signal feature map comprises: Divide each feature sequence into multiple sequence fragments of length n, where n is a positive integer; The sequence fragments corresponding to different characteristic sequences are copied line by line to generate m-row sequence fragments, and the m-row sequence fragments satisfy that different characteristic sequence lines are adjacent to each other, wherein m is a positive integer; An m*n two-dimensional signal feature map is generated based on the m-row sequence fragments. 6. The method for predicting an order according to claim 5, wherein said dividing each feature sequence into a plurality of length n sequence fragments comprises: Based on the shift operation, each feature sequence is divided into multiple sequence fragments of length n. 7. The prediction method of order according to claim 1, is characterized in that, described construction neural network model comprises: Build a neural network model; Based on the two-dimensional signal feature map, train the neural network model, and acquire parameters of the neural network model. 8. The prediction method of order according to claim 7, is characterized in that, after constructing neural network model, also comprises: Obtain online data of orders; Training and updating the neural network model based on the online data. 9. The order prediction method according to claim 7 or 8, wherein the neural network is a convolutional neural network. 10. An order forecasting device, comprising: The first acquisition unit is adapted to acquire time series corresponding to historical orders; an extraction unit adapted to extract a feature sequence based on the time series; a generating unit adapted to generate a two-dimensional signal feature map based on the feature sequence; The construction unit is adapted to construct a neural network model based on the two-dimensional signal feature map, and perform order prediction according to the constructed neural network model. 11. The order forecasting device according to claim 10, characterized in that the first acquisition unit comprises: The first acquisition subunit is suitable for acquiring raw data corresponding to historical orders; The second acquisition subunit is suitable for preprocessing the raw data and acquiring the time series corresponding to the order. 12 . The order prediction device according to claim 11 , wherein the preprocessing includes at least one of the following: outlier value processing and missing value processing. 13 . 13. The order prediction device according to claim 10, wherein the extraction unit is adapted to extract the feature sequence based on a wavelet transform algorithm. 14. The order forecasting device according to claim 10, wherein the generating unit comprises: The segmentation subunit is suitable for dividing each feature sequence into a plurality of sequence fragments with a length of n, where n is a positive integer; The copying subunit is adapted to copy sequence fragments corresponding to different characteristic sequences line by line to generate m-row sequence fragments, and the m-line sequence fragments satisfy that different characteristic sequence lines are adjacent to each other, wherein m is a positive integer; The generation subunit is adapted to generate m*n two-dimensional signal feature maps based on the m-line sequence fragments. 15. The order prediction device according to claim 14, wherein the segmentation subunit is adapted to segment each feature sequence into a plurality of sequence fragments with a length of n based on a shift operation. 16. The order forecasting device according to claim 10, wherein the construction unit comprises: Constructing a subunit, adapted to construct a neural network model based on the neural network model parameters; The training subunit is adapted to train the neural network model based on the two-dimensional signal feature map, and acquire parameters of the neural network model; The prediction subunit is adapted to perform order prediction according to the trained neural network model. 17. The order forecasting device according to claim 16, further comprising: The second acquisition unit is adapted to acquire online data of the order; An update unit, adapted to train and update the neural network model based on the online data. 18. The order prediction device according to claim 16 or 17, wherein the neural network is a convolutional neural network. 19. A computer-readable storage medium on which computer instructions are stored, wherein the steps of the method according to any one of claims 1 to 9 are executed when the computer instructions are run. 20. A logistics system, comprising a memory and a processor, the memory is stored with computer instructions that can run on the processor, characterized in that, when the processor runs the computer instructions, it executes claims 1 to 20. The steps of the method described in any one of 9.