Container terminal throughput prediction method, device and equipment based on deep learning

Abstract

The invention belongs to the field of throughput prediction, and provides a container terminal throughput prediction method, device and equipment based on deep learning, and the method comprises the steps: extracting key request content in a prediction request after the prediction request of the throughput of a container terminal is received; calling original characteristic variable data and throughput sample data from a plurality of data sources based on the key request content; determining a mutual dependency degree between each piece of original characteristic variable data and throughput sample data; taking the original characteristic variable data of which the mutual dependency degree is higher than a preset dependency degree threshold value as characteristic variable sample data; training, testing and evaluating the deep learning network model, and constructing to obtain a throughput prediction model; and preprocessing actual characteristic variable data in the key request content, and inputting the preprocessed actual characteristic variable data into the throughput prediction model to obtain a throughput prediction result. According to the scheme provided by the invention, the prediction accuracy and reliability of the throughput of the container terminal are improved.

Description

Technical Field

[0001] The present invention relates to the field of data prediction technology, and in particular to a method, device and equipment for predicting container terminal throughput based on deep learning. Background Art

[0002] Container terminal throughput is an important indicator for measuring port economic activities, and its forecast results are of great significance for port management and planning. Accurate forecasting of container terminal throughput helps port managers formulate reasonable operation plans, optimize resource allocation, and improve operational efficiency, thereby enhancing the competitiveness of ports.

[0003] In the related technology, traditional container terminal throughput prediction schemes are often implemented based on linear assumptions. Since the impact of nonlinear factors on the container terminal throughput is not considered, the accuracy and reliability of the throughput prediction results are low, which makes it difficult to meet the high-precision prediction requirements of container terminal throughput. Summary of the invention

[0004] The present invention provides a method, device and equipment for predicting container terminal throughput based on deep learning, so as to solve the defects that the traditional container terminal throughput prediction scheme does not consider the influence of nonlinear factors on the container terminal throughput, and the accuracy and reliability of the throughput prediction results are low.

[0005] In one aspect, the present invention provides a method for predicting container terminal throughput based on deep learning, comprising: After receiving a prediction request for container terminal throughput, extracting key request content in the prediction request; Based on the key request content, retrieve original feature variable data and throughput sample data corresponding to the original feature variable data from multiple data sources; Determining the mutual dependence between each of the original feature variable data and the throughput sample data; The original feature variable data whose mutual dependence is higher than a preset dependence threshold is used as feature variable sample data; Based on the characteristic variable sample data and the throughput sample data, a deep learning network model is trained, tested and evaluated to construct a throughput prediction model; The actual characteristic variable data in the key request content is pre-processed and then input into the throughput prediction model to obtain a throughput prediction result.

[0006] According to the container terminal throughput prediction method based on deep learning provided by the present invention, determining the mutual dependence between each of the original feature variable data and the throughput sample data includes: Convert each of the original feature variable data and the corresponding throughput sample data into time series data to obtain a plurality of feature variable sequences and throughput sample sequences; respectively determining a first probability density function of each characteristic element data in each characteristic variable sequence, a second probability density function of each throughput element data in each throughput sample sequence, and a joint density function between each characteristic element data and the corresponding throughput element data; Based on the first probability density function, the second probability density function and the joint density function, the mutual dependence between each of the original feature variable data and the throughput sample data is calculated.

[0007] According to the method for predicting container terminal throughput based on deep learning provided by the present invention, based on the characteristic variable sample data and the throughput sample data, the deep learning network model is trained, tested and evaluated to construct a throughput prediction model, including: Convert each of the characteristic variable sample data and the corresponding throughput sample data into time series data to obtain a plurality of characteristic variable sequences and throughput sample sequences; Establishing a sample data set based on the multiple feature variable sequences and throughput sample sequences; Dividing the sample data set into a training set, a test set and a validation set; The deep learning network model is trained by the training set, and the deep learning network model is tested by the test set to obtain a prediction model that passes the test; Inputting the characteristic variable sequence in the validation set into the prediction model that has passed the test to obtain a throughput prediction sequence; Based on the throughput prediction sequence and the throughput sample sequence in the validation set, an evaluation index value is calculated; When the evaluation index value is higher than a preset index threshold, a throughput prediction model is obtained.

[0008] According to the method for predicting throughput of a container terminal based on deep learning provided by the present invention, based on the throughput prediction sequence and the throughput sample sequence in the validation set, an evaluation index value is calculated, including: Determine the mean absolute percentage error, mean absolute error and mean square error between the throughput prediction sequence and the throughput sample sequence in the validation set respectively; Calculating the square values ​​corresponding to the mean absolute percentage error, the mean absolute error, and the mean square error, and performing weighted summation on the square values ​​to obtain a weighted sum of squares of errors; A square root operation is performed on the weighted sum of squares of the errors to obtain an evaluation index value.

[0009] According to the container terminal throughput prediction method based on deep learning provided by the present invention, the throughput prediction model includes: A convolutional neural network module is used to extract and reconstruct features from the preprocessed actual feature variable data to obtain a first output feature sequence; A bidirectional long short-term memory network module is used to perform forward data processing and reverse data processing on the first output feature sequence respectively to obtain a forward hidden state and a reverse hidden state, and to concatenate the forward hidden state and the reverse hidden state to obtain a second output feature sequence including a final hidden state; The self-attention module is used to assign a corresponding weight value to each value vector in the second output feature sequence, and perform feature enhancement and feature fusion processing on the second output feature sequence based on the weight value to obtain a throughput prediction result.

[0010] According to the container terminal throughput prediction method based on deep learning provided by the present invention, the convolutional neural network module includes: an input layer, a convolution layer, a pooling layer, a feature reconstruction layer and a priori knowledge layer; The convolution layer is used to extract features from the preprocessed actual feature variable data input into the input layer to obtain a first feature mapping sequence; The pooling layer is used to perform a maximum pooling operation on the first feature mapping sequence to obtain a second feature mapping sequence; The feature reconstruction layer is used to perform feature reconstruction on the second feature mapping sequence to sequentially connect the feature mapping elements in the second feature mapping sequence to obtain a preliminary feature sequence; The prior knowledge layer is used to combine the prior knowledge with the preliminary feature sequence to obtain a first output feature sequence.

[0011] According to the container terminal throughput prediction method based on deep learning provided by the present invention, the convolution layer includes: a data splitting sublayer, a standard convolution operator, a joint convolution operator and a splicing sublayer; The data splitting sublayer is used to split the preprocessed actual feature variable data into first variable sub-data and second variable sub-data; The standard convolution operator is used to perform a standard convolution operation on the first variable sub-data to obtain a first characteristic sub-sequence; The joint convolution operator is used to perform a PConv partial convolution operation and an MConv micro-point convolution operation on the first feature subsequence to obtain a second feature subsequence; The concatenation sublayer is used to concatenate the second variable sub-data with the second feature sub-sequence to obtain a first feature mapping sequence.

[0012] According to the container terminal throughput prediction method based on deep learning provided by the present invention, extracting key request content in the prediction request includes: Verifying the data format and data range of the request content in the prediction request to obtain a verification result; If the verification result is that the data format of the request content in the prediction request satisfies the verification reference format, and the data range satisfies the verification reference range, then the key request content is extracted from the request content.

[0013] On the other hand, the present invention also provides a container terminal throughput prediction device based on deep learning, comprising: An extraction module, configured to extract key request content in the prediction request after receiving the prediction request for the container terminal throughput; A calling module, used for calling original feature variable data and throughput sample data corresponding to the original feature variable data from multiple data sources based on the key request content; A determination module, configured to determine the mutual dependence between each of the original feature variable data and the throughput sample data; A processing module, used for taking the original feature variable data whose mutual dependence is higher than a preset dependence threshold as feature variable sample data; A construction module, used to train, test and evaluate the deep learning network model based on the characteristic variable sample data and the throughput sample data, and construct a throughput prediction model; The prediction module is used to pre-process the actual characteristic variable data in the key request content and input the data into the throughput prediction model to obtain a throughput prediction result.

[0014] On the other hand, the present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a container terminal throughput prediction method based on deep learning as described above is implemented.

[0015] The method, device and equipment for predicting throughput of container terminals based on deep learning provided by the present invention extract key request contents in prediction requests, retrieve original feature variable data and throughput sample data corresponding to the original feature variable data from multiple data sources based on the key request contents, determine the mutual dependence between each original feature variable data and the throughput sample data, take the original feature variable data with a mutual dependence higher than a preset dependence threshold as feature variable sample data, train, test and evaluate the deep learning network model based on the feature variable sample data and the throughput sample data, construct a throughput prediction model, and finally pre-process the actual feature variable data in the key request contents and input them into the throughput prediction model to obtain the throughput prediction result. Due to the sample data establishment link, data related to the key request contents can be retrieved from multiple data sources for prediction requests, and then relevant nonlinear factors are introduced in a targeted manner. Meanwhile, the mutual dependence between the original feature variable data and the throughput sample data is taken into consideration, and an effective data basis can be provided for the construction of the throughput prediction model, thereby improving the prediction accuracy and reliability of the throughput of container terminals. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0017] Figure 1 It is a flowchart of a method for predicting container terminal throughput based on deep learning provided by an embodiment of the present invention; Figure 2 It is a schematic diagram of the structure of the throughput prediction model; Figure 3 This is a schematic diagram of the data processing principle of the improved convolutional layer; Figure 4 is a schematic diagram of the structure of a container terminal throughput prediction device based on deep learning provided by an embodiment of the present invention; Figure 5 It is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] Combine the following Figures 1 to 5 Describe the detailed scheme of the method, device and equipment for predicting container terminal throughput based on deep learning provided by the embodiments of the present invention.

[0020] Figure 1 It is a flow chart of a method for predicting container terminal throughput based on deep learning provided in an embodiment of the present invention.

[0021] like Figure 1 As shown, the container terminal throughput prediction method based on deep learning provided by the embodiment of the present invention can be executed by a computer or server with data receiving and data processing functions. The method mainly includes the following steps: Step 110: After receiving the prediction request for container terminal throughput, extract key request content in the prediction request.

[0022] In this embodiment, the prediction request may be issued through a user client or a terminal supervision system, and key request content may include a prediction time range, a prediction granularity, and other specific prediction requirements.

[0023] Step 120: Based on the key request content, retrieve original feature variable data and throughput sample data corresponding to the original feature variable data from multiple data sources.

[0024] It is understandable that multiple data sources can be port management systems, customs databases, meteorological service stations, and other data storage media that store information that affects the throughput of container terminals. The original feature variable data mainly refers to relevant characteristic factors that can affect the throughput of container terminals, such as historical throughput data, time information, weather data, relevant economic indicators, yard area, and operational efficiency.

[0025] Step 130: Determine the mutual dependence between each original feature variable data and the throughput sample data.

[0026] It can be understood that the degree of mutual dependence can characterize the influence of each original characteristic variable data on the throughput of the container terminal. The greater the degree of mutual dependence, the greater the influence of the original characteristic variable data on the throughput of the container terminal, and the more important the original characteristic variable data is in the prediction link.

[0027] Step 140: The original feature variable data whose mutual dependence is higher than a preset dependence threshold is used as feature variable sample data.

[0028] In this embodiment, by screening some original characteristic variable data that have a greater impact on the throughput of the container terminal, characteristic variable sample data for subsequent model building can be obtained.

[0029] Step 150: Based on the feature variable sample data and the throughput sample data, the deep learning network model is trained, tested and evaluated to construct a throughput prediction model.

[0030] It can be understood that this embodiment adopts a deep learning approach to perform throughput prediction. By training, testing and evaluating the deep learning network model, a throughput prediction model that meets the prediction requirements can be obtained.

[0031] Step 160: Pre-process the actual feature variable data in the key request content and input it into the throughput prediction model to obtain a throughput prediction result.

[0032] The solution provided in this embodiment can retrieve data related to key request content from multiple data sources in response to prediction requests, and then introduce relevant nonlinear factors in a targeted manner. At the same time, it takes into account the interdependence between the original feature variable data and the throughput sample data, and can provide an effective data basis for the construction of the throughput prediction model, thereby improving the prediction accuracy and reliability of the container terminal throughput.

[0033] In one embodiment, extracting key request content in the prediction request specifically includes: First, the data format and data range of the request content in the prediction request are verified to obtain the verification result.

[0034] Then, if the verification result is that the data format of the request content in the prediction request satisfies the verification reference format, and the data range satisfies the verification reference range, the key request content is extracted from the request content.

[0035] In this embodiment, the request header information field of the prediction request can be checked to verify whether the data format of the request content meets the verification reference format, and the data range can be checked to determine whether the data of the request content is complete. If the data format is incorrect or the data is incomplete, an error handling mechanism will be triggered, and specifically, an error message can be sent to the request initiator, an error log can be recorded, the request can be discarded, or other appropriate error handling operations can be performed. The request content that passes the inspection will be stored in the internal storage unit.

[0036] In one embodiment, determining the mutual dependence between each original feature variable data and the throughput sample data specifically includes: In the first step, each original feature variable data and the corresponding throughput sample data are converted into time series data to obtain multiple feature variable sequences and throughput sample sequences.

[0037] In practical applications, since the original feature variable data involves multiple types of data, in order to avoid factors such as its value range and the correlation with throughput causing some features to have too great an impact on the prediction results, the original feature variable data needs to be normalized, as follows: (1) in, , are the original feature variable data before and after normalization, , are the mean and standard deviation of the original feature variable data respectively.

[0038] After normalizing the original feature variable data, each original feature variable data and throughput sample data can be further represented as corresponding time series data in chronological order, thereby obtaining multiple feature variable sequences and throughput sample sequences.

[0039] In the second step, the first probability density function of each feature element data in each feature variable sequence, the second probability density function of each throughput element data in each throughput sample sequence, and the joint density function between each feature element data and the corresponding throughput element data are determined respectively.

[0040] The third step is to calculate the mutual dependence between each original feature variable data and the throughput sample data based on the first probability density function, the second probability density function and the joint density function.

[0041] In this embodiment, the mutual dependence between the original feature variable data and the throughput sample data can be expressed as follows: (2) Among them, I(X,Y) represents the mutual dependence between the original feature variable data X and the throughput sample data Y, that is, the mutual information, n represents the number of element data in the feature variable sequence and the throughput sample sequence, and p(x i ) represents the first probability density function of the i-th feature element data, p(y i ) represents the second probability density function of the i-th throughput element data, p(x i ,y i ) represents the joint density function between the i-th feature element data and the i-th throughput element data.

[0042] After obtaining the interdependence between each original characteristic variable data and the throughput sample data, the original characteristic variable data with a smaller interdependence can be removed, and finally the characteristic variable sample data can be screened out. In this embodiment, the screened characteristic variable sample data can include eight original characteristic variable data such as relevant economic indicators (such as local GDP value), total retail sales of consumer goods, hinterland size, total imports and exports, seasonal time, etc.

[0043] In one embodiment, based on the feature variable sample data and the throughput sample data, the deep learning network model is trained, tested and evaluated to construct a throughput prediction model, including: In the first step, each characteristic variable sample data and the corresponding throughput sample data are converted into time series data to obtain multiple characteristic variable sequences and throughput sample sequences.

[0044] The second step is to establish a sample data set based on multiple feature variable sequences and throughput sample sequences.

[0045] In this embodiment, a sample data set X={x1,x2,…,x n}.

[0046] The third step is to divide the sample data set into training set, test set and validation set.

[0047] In practical applications, the training set, validation set, and test set can be divided in a ratio of 8:1:1. The training set is used to input into the deep learning network model for training, the validation set is used to adjust the model parameters, and the test set is used to evaluate the model performance.

[0048] The fourth step is to train the deep learning network model through the training set and test the deep learning network model through the test set to obtain a prediction model that passes the test.

[0049] The fifth step is to input the feature variable sequence in the validation set into the prediction model that has passed the test to obtain the throughput prediction sequence.

[0050] Step 6: Calculate the evaluation index value based on the throughput prediction sequence and the throughput sample sequence in the validation set.

[0051] In the seventh step, when the evaluation index value is higher than the preset index threshold, a throughput prediction model is obtained.

[0052] In one embodiment, based on the throughput prediction sequence and the throughput sample sequence in the validation set, the evaluation index value is calculated, specifically including: In the first step, the mean absolute percentage error, mean absolute error, and mean square error between the throughput prediction sequence and the throughput sample sequence in the validation set are determined respectively.

[0053] In this embodiment, the mean absolute percentage error, mean absolute error, and mean square error between the throughput prediction sequence and the throughput sample sequence in the validation set can be calculated as follows: (3) (4) (5) Where MSE represents mean square error, RMSE represents mean absolute percentage error, MAE represents mean absolute error, n represents the number of element data in the throughput prediction sequence and throughput sample sequence, and Y i represents the i-th predicted value in the throughput prediction sequence, y i Represents the i-th throughput element data in the throughput sample sequence, that is, the true value. The closer the predicted value is to the true value, the better the model is.

[0054] The second step is to calculate the square values ​​corresponding to the mean absolute percentage error, the mean absolute error, and the mean square error, and perform weighted summation of the square values ​​to obtain the weighted sum of squares of the error.

[0055] The third step is to perform a quadratic operation on the weighted sum of squares of the errors to obtain the evaluation index value.

[0056] In this embodiment, the evaluation index value can be calculated as follows: (6) in, represents the evaluation index value, Represents the weight value of the mean square error MSE, represents the weight value of the mean absolute percentage error RMSE, Represents the weight value of the mean absolute error MAE.

[0057] In one embodiment, if Figure 2 As shown in FIG. 1 , the throughput prediction model specifically includes: The convolutional neural network module 210 is used to perform feature extraction and feature reconstruction on the actual feature variable data after preprocessing to obtain a first output feature sequence.

[0058] The bidirectional long short-term memory network module 220 is used to perform forward data processing and reverse data processing on the first output feature sequence respectively to obtain a forward hidden state and a reverse hidden state, and to concatenate the forward hidden state and the reverse hidden state to obtain a second output feature sequence including a final hidden state.

[0059] The self-attention module 230 is used to assign a corresponding weight value to each value vector in the second output feature sequence, and perform feature enhancement and feature fusion processing on the second output feature sequence based on the weight value to obtain a throughput prediction result.

[0060] In one embodiment, if Figure 2 As shown, the convolutional neural network module specifically includes: an input layer 2101, a convolution layer 2102, a pooling layer 2103, a feature reconstruction layer 2104 and a priori knowledge layer 2105.

[0061] The convolution layer 2102 is used to perform feature extraction on the preprocessed actual feature variable data input into the input layer 2101 to obtain a first feature mapping sequence.

[0062] The pooling layer 2103 is used to perform a maximum pooling operation on the first feature mapping sequence to obtain a second feature mapping sequence.

[0063] The feature reconstruction layer 2104 is used to perform feature reconstruction on the second feature mapping sequence, so as to sequentially connect the feature mapping elements in the second feature mapping sequence to obtain a preliminary feature sequence.

[0064] The prior knowledge layer 2105 is used to combine the prior knowledge with the preliminary feature sequence to obtain a first output feature sequence.

[0065] It can be understood that the convolutional neural network module adopts the CNN convolutional neural network architecture. The CNN convolutional neural network is a feedforward neural network suitable for processing data with a grid topology structure. In this embodiment, it is mainly responsible for extracting local features in the throughput sample sequence and the feature variable sequence. After the data passes through the input layer and enters the convolution layer, it will be sorted according to the contribution of each relevant factor to the throughput, and the convolution kernel weight will be assigned to each element data in the feature variable sequence to obtain the feature representation of each element data in the feature variable sequence, where the feature representation of the i-th element data can be calculated as follows: (7) Among them, e i is the characteristic representation of the i-th element data, x i is the i-th element data, w i is the convolution kernel weight value of the i-th element data, and b is the bias value.

[0066] After performing the convolution operation on each element data, the first feature mapping sequence E can be obtained. Assuming that there are m convolution kernels in the convolution layer, the first feature mapping sequence E can be expressed as follows: (8) In the pooling layer, the maximum pooling operation can select the maximum value in each local area of ​​the feature map as the output, which can better retain the main features and suppress unimportant details. It can effectively reduce the size of the feature map and reduce the amount of calculation of the subsequent layers, thereby improving the computational efficiency of the entire throughput prediction model. The second feature map sequence F obtained after pooling compression can be expressed as follows: (9) In practical applications, the Sigmoid activation function can be used to introduce nonlinear factors into the convolutional neural network and add Parameters are optimized to accelerate model convergence, avoid gradient vanishing and explosion problems, and ensure that all features are considered on the same scale to make them better suitable for prediction tasks. The Sigmoid activation function S(x) can be expressed as follows: (10) Among them, x is the input data of the Sigmoid activation function, The parameter is used to control the center position of the activation function, so that the Sigmoid activation function can be offset according to the distribution characteristics of the input data, which helps to accelerate the propagation of gradients and the training of the model.

[0067] In this embodiment, in order to meet the input data format requirements of the bidirectional long short-term memory network module, a feature reconstruction layer is added after the pooling layer in the convolutional neural network module, and the feature mapping elements in the second feature mapping sequence can be sequentially connected to obtain a preliminary feature sequence Z, which can be specifically expressed as follows: Z=F1 F2 Fm (11) After flattening the feature reconstruction layer, the preliminary feature sequence also needs to input the prior knowledge layer. As an additional source of information, the prior knowledge layer can effectively combine the prior knowledge and the data in the preliminary feature sequence to help the model better understand and process the data, and further improve the feature representation based on the feature reconstruction layer to make it more in line with the needs.

[0068] In one embodiment, the convolution layer specifically includes: a data splitting sublayer, a standard convolution operator, a joint convolution operator, and a splicing sublayer.

[0069] The data splitting sublayer is used to split the preprocessed actual feature variable data into first variable sub-data and second variable sub-data.

[0070] The standard convolution operator is used to perform a standard convolution operation on the first variable sub-data to obtain a first characteristic sub-sequence.

[0071] The joint convolution operator is used to perform PConv partial convolution operation and MConv micro-point convolution operation on the first feature subsequence to obtain a second feature subsequence.

[0072] The concatenation sublayer is used to concatenate the second variable sub-data with the second feature sub-sequence to obtain a first feature mapping sequence.

[0073] This embodiment does not use all standard convolution operators in the convolution layer, but introduces a joint convolution operator. The joint convolution operator MPConv uses partial convolution PConv to connect with micro point convolution MConv, which can change the number of feature channels without changing the size of the feature map, and can effectively reduce the computational complexity.

[0074] like Figure 3 As shown, during the calculation process, the actual feature variable data is first split into the first variable sub-data X1 and the second variable sub-data X2. The number of channels of the first variable sub-data X1 is C P , the number of channels of the second variable sub-data X2 is CC P , a standard convolution operation is performed on the first variable sub-data X1 through a standard convolution operator with a convolution kernel size of k×k. No operation is performed on the data of the remaining channels to obtain a first feature sequence. Then, a convolution operation is performed through the partial convolution PConv and the micro-point convolution MConv in the joint convolution operator MPConv to output the second feature sequence X3. Finally, the second feature sequence X3 is concatenated with the second variable sub-data X2 to obtain a first feature mapping sequence.

[0075] It is understandable that the joint convolution operator can reduce the amount of memory access to a certain extent and solve the problem of weak information interaction ability of some convolution PConv. It not only reduces the computational complexity of the model, improves the model prediction speed, but also improves the actual prediction accuracy.

[0076] In this embodiment, the bidirectional long short-term memory network module can run two long short-term memory networks at each time step, one of which processes data in the forward order of the first output feature sequence, and processes it in sequence to the current time step to generate a series of forward hidden states, and the other processes data in the reverse order of the first output feature sequence, and processes it in sequence to the current time step to generate a series of reverse hidden states. At each time step, the forward hidden state and the reverse hidden state are concatenated to obtain the final hidden state representation of the time step, so as to obtain more context information at each time step, and thus can more comprehensively capture the context dependency in the first output feature sequence.

[0077] This embodiment uses a bidirectional long short-term memory network module to learn long-term dependencies. Two long short-term memory networks are run simultaneously in the bidirectional long short-term memory network module. The gate control mechanism is introduced to overcome the gradient vanishing and gradient exploding problems in long sequences. The main function of the forget gate is to decide when to discard the previous memory and selectively forget some of the previous throughput data information. The main function of the input gate is to decide when to add new information to the memory. The main function of the output gate is to control when to apply the memory to the current state, which determines which information in the memory unit will affect the current output. The forget gate, input gate, and output gate can be expressed as follows: (12) (13) (14) in, represents the forget gate, x t Represents input data, i t represents the input gate, o t represents the output gate, W f , W i as well as W o Both represent weight matrices, b f , b i , b o Both represent the offset. represents the Sigmoid activation function, h t-1 represents the hidden layer state at iteration t-1.

[0078] Below is the hidden layer state at time t h t For example, it can be expressed as follows: (15) (16) (17) in, h t represents the hidden layer state at the tth iteration, C t represents the memory cell state at the tth iteration, Indicates the current input unit status, represents the forget gate, it represents the input gate, o t represents the output gate, C t-1 represents the memory cell state at the t-1th iteration, W c represents the weight matrix, b c Indicates the offset, x t Represents input data.

[0079] The Adam optimizer is selected to allocate different learning rates to different parameters by calculating the first-order moment estimate and the second-order moment estimate. The main formula is as follows: (18) (19) In each iteration, it is necessary to calculate the gradient of the loss function with respect to each parameter. are model parameters, is the loss function, is the gradient, where m t is the first-order moment estimate of the t-th iteration, is the exponential decay rate of the first-order moment estimate, and this number is set to 0.9 in this embodiment. t is the second-order moment estimate of the t-th iteration, Represents the exponential decay rate of the second-order moment estimation. In this embodiment, the number is set to 0.999.

[0080] Due to m t and v t At the beginning, it is close to 0, so bias correction is needed. First, for the first-order moment estimate m t Correction, including refers to , the corrected first-order moment estimate It can be expressed as follows: (20) Then the second-order moment estimate v t Correction, including refers to , the corrected second-order moment estimate It can be expressed as follows: (twenty one) Finally, the parameters need to be updated. The update formula is: (twenty two) in, is the model parameter of the t+1th iteration, are the model parameters for the tth iteration, is a very small number to avoid division by 0. Represents the learning rate.

[0081] The learning rate is dynamically adjusted by calculating the momentum term of the gradient and the adaptive learning rate, so as to better adapt to the learning needs of different parameters according to each parameter, and avoid the problem of oscillation caused by too high learning rate or slow convergence caused by too low learning rate.

[0082] In the model training phase, the initial learning rate can be set to 0.0001 to facilitate adaptive adjustment of the learning rate. The number of training iterations can be set to 100 rounds, the batch size is 32, the number of BiLSTM neurons in the first layer of the bidirectional long short-term memory network module is 128, and the number of BiLSTM neurons in the second layer is 32. After the 93rd round of training, the model reaches a fitting state.

[0083] The second output feature sequence after the bidirectional long short-term memory network module contains the hidden state of each time step. These hidden states contain both past and future information and can be used as the input of the subsequent self-attention layer. The self-attention mechanism is an attention mechanism that associates different positions of a single sequence to calculate the representation of the same sequence. It allows the model to pay attention to the correlation between different positions within the sequence when processing the input sequence, and generate the output of the current position based on these correlations.

[0084] After the second output feature sequence enters the self-attention module, the similarity between each element in the second output feature sequence and other elements is calculated, and then the elements in the second output feature sequence are weighted and summed according to these similarities. Specifically, for each position in the second output feature sequence, three different linear transformations (i.e., three different weight matrices) can be used to generate query vectors, key vectors, and value vectors. Then, the dot product of each query vector and all key vectors is calculated, and normalized by the softmax function to obtain the attention weight. Finally, these attention weights are used to perform weighted summation on the value vector to obtain the output representation of each position, fuse the spatial features and temporal features, and input them into the fully connected layer for feature fusion to achieve accurate prediction of throughput.

[0085] The throughput prediction model provided in this embodiment can extract local features of input data through a convolutional neural network module, and use a bidirectional long short-term memory network module to extract global time dependencies. At the same time, a self-attention module is used to dynamically assign weights according to the importance of different parts of the input data to the prediction results, thereby enhancing the model's ability to capture key information. Finally, the captured features are converted into throughput prediction results through a fully connected layer, thereby achieving accurate prediction of container terminal throughput.

[0086] In practical applications, constraints can be added or modified according to actual forecasting needs. The throughput of container terminals in a certain period of time in the future can be predicted. For example, if the throughput of a certain operation route needs to be predicted, it is only necessary to add the operation set of the operation route and the time set consumed for loading and unloading as constraints.

[0087] The container terminal throughput prediction method based on deep learning provided by the embodiment of the present invention comprehensively uses a variety of prediction methods, such as time series analysis, regression analysis, and neural networks, etc., to significantly improve the accuracy and adaptability of throughput prediction. This not only enables port managers to grasp future operating trends more accurately, provides a solid basis for making scientific and reasonable decisions, but also effectively reduces the potential risks caused by inaccurate predictions. Faced with the complexity and uncertainty of container throughput data, whether it is the nonlinear relationship of data or complex features such as seasonal fluctuations, the throughput prediction model can accurately capture and effectively respond. This capability ensures the stability and reliability of throughput prediction results in different scenarios, and provides strong decision-making support for port managers. With the continuous advancement of technology, the container terminal throughput prediction process is gradually developing towards automation and intelligence. This can not only realize real-time monitoring and early warning of changes in container terminal throughput, but also automatically adjust the throughput prediction model to adapt to data changes. This intelligent prediction method not only improves the efficiency and accuracy of port operations, but also significantly enhances the competitiveness of ports.

[0088] Based on the same general inventive concept, the present invention also protects a container terminal throughput prediction device based on deep learning. The container terminal throughput prediction device based on deep learning provided by the present invention is described below. The container terminal throughput prediction device based on deep learning described below and the container terminal throughput prediction method based on deep learning described above can refer to each other.

[0089] like Figure 4 As shown, a container terminal throughput prediction device based on deep learning is characterized by comprising: The extraction module 310 is used to extract key request content in the prediction request after receiving the prediction request of the container terminal throughput; The retrieving module 320 is used to retrieve the original feature variable data and the throughput sample data corresponding to the original feature variable data from multiple data sources based on the key request content; A determination module 330, for determining the mutual dependence between each original feature variable data and the throughput sample data; A processing module 340 is used to use the original feature variable data with a mutual dependence higher than a preset dependence threshold as feature variable sample data; A construction module 350 is used to train, test and evaluate the deep learning network model based on the feature variable sample data and the throughput sample data to construct a throughput prediction model; The prediction module 360 ​​is used to pre-process the actual characteristic variable data in the key request content and input it into the throughput prediction model to obtain the throughput prediction result.

[0090] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated again here.

[0091] Figure 5 It is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention.

[0092] like Figure 5 As shown, the electronic device may include: a processor 410, a communication interface 420, a memory 430 and a communication bus 440, wherein the processor 410, the communication interface 420 and the memory 430 communicate with each other through the communication bus 440. The processor 410 may call the logic instructions in the memory 430 to execute the container terminal throughput prediction method based on deep learning, including: extracting the key request content in the prediction request after receiving the prediction request of the container terminal throughput; based on the key request content, retrieving the original feature variable data and the throughput sample data corresponding to the original feature variable data from multiple data sources; determining the mutual dependence between each original feature variable data and the throughput sample data; using the original feature variable data with a mutual dependence higher than the preset dependency threshold as the feature variable sample data; training, testing and evaluating the deep learning network model based on the feature variable sample data and the throughput sample data to construct a throughput prediction model; pre-processing the actual feature variable data in the key request content and inputting it into the throughput prediction model to obtain the throughput prediction result.

[0093] In addition, the logic instructions in the above-mentioned memory 430 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0094] On the other hand, the present invention also provides a computer program product, which includes a computer program, which can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute a container terminal throughput prediction method based on deep learning, including: after receiving a prediction request for container terminal throughput, extracting key request content in the prediction request; based on the key request content, retrieving original feature variable data and throughput sample data corresponding to the original feature variable data from multiple data sources; determining the mutual dependence between each original feature variable data and the throughput sample data; using the original feature variable data with a mutual dependence higher than a preset dependency threshold as feature variable sample data; based on the feature variable sample data and the throughput sample data, training, testing and evaluating the deep learning network model to construct a throughput prediction model; pre-processing the actual feature variable data in the key request content and inputting it into the throughput prediction model to obtain a throughput prediction result.

[0095] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a container terminal throughput prediction method based on deep learning, including: after receiving a prediction request for container terminal throughput, extracting key request content in the prediction request; based on the key request content, retrieving original feature variable data and throughput sample data corresponding to the original feature variable data from multiple data sources; determining the mutual dependence between each original feature variable data and the throughput sample data; using the original feature variable data with a mutual dependence higher than a preset dependency threshold as feature variable sample data; based on the feature variable sample data and the throughput sample data, training, testing and evaluating the deep learning network model to construct a throughput prediction model; pre-processing the actual feature variable data in the key request content and inputting it into the throughput prediction model to obtain a throughput prediction result.

[0096] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0097] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A container terminal throughput prediction method based on deep learning, characterized in that: include: After receiving a prediction request for container terminal throughput, extracting key request content in the prediction request; Based on the key request content, retrieve original feature variable data and throughput sample data corresponding to the original feature variable data from multiple data sources; Determining the mutual dependence between each of the original feature variable data and the throughput sample data; The original feature variable data whose mutual dependence is higher than a preset dependence threshold is used as feature variable sample data; Based on the characteristic variable sample data and the throughput sample data, a deep learning network model is trained, tested and evaluated to construct a throughput prediction model; The actual characteristic variable data in the key request content is pre-processed and then input into the throughput prediction model to obtain a throughput prediction result.

2. The method for predicting container terminal throughput based on deep learning according to claim 1, characterized in that: Determining the mutual dependence between each of the original feature variable data and the throughput sample data includes: Convert each of the original feature variable data and the corresponding throughput sample data into time series data to obtain a plurality of feature variable sequences and throughput sample sequences; respectively determining a first probability density function of each characteristic element data in each characteristic variable sequence, a second probability density function of each throughput element data in each throughput sample sequence, and a joint density function between each characteristic element data and the corresponding throughput element data; Based on the first probability density function, the second probability density function and the joint density function, the mutual dependence between each of the original feature variable data and the throughput sample data is calculated.

3. The method for predicting container terminal throughput based on deep learning according to claim 1, characterized in that: Based on the characteristic variable sample data and the throughput sample data, the deep learning network model is trained, tested and evaluated to construct a throughput prediction model, including: Convert each of the characteristic variable sample data and the corresponding throughput sample data into time series data to obtain a plurality of characteristic variable sequences and throughput sample sequences; Establishing a sample data set based on the multiple feature variable sequences and throughput sample sequences; Dividing the sample data set into a training set, a test set and a validation set; The deep learning network model is trained by the training set, and the deep learning network model is tested by the test set to obtain a prediction model that passes the test; Inputting the characteristic variable sequence in the validation set into the prediction model that has passed the test to obtain a throughput prediction sequence; Based on the throughput prediction sequence and the throughput sample sequence in the validation set, an evaluation index value is calculated; When the evaluation index value is higher than a preset index threshold, a throughput prediction model is obtained.

4. The method for predicting container terminal throughput based on deep learning according to claim 3, characterized in that: Based on the throughput prediction sequence and the throughput sample sequence in the validation set, an evaluation index value is calculated, including: Determine the mean absolute percentage error, mean absolute error and mean square error between the throughput prediction sequence and the throughput sample sequence in the validation set respectively; Calculating the square values ​​corresponding to the mean absolute percentage error, the mean absolute error, and the mean square error, and performing weighted summation on the square values ​​to obtain a weighted sum of squares of errors; A square root operation is performed on the weighted sum of squares of the errors to obtain an evaluation index value.

5. The method for predicting container terminal throughput based on deep learning according to claim 1, characterized in that: The throughput prediction model includes: A convolutional neural network module is used to extract and reconstruct features from the preprocessed actual feature variable data to obtain a first output feature sequence; A bidirectional long short-term memory network module is used to perform forward data processing and reverse data processing on the first output feature sequence respectively to obtain a forward hidden state and a reverse hidden state, and to concatenate the forward hidden state and the reverse hidden state to obtain a second output feature sequence including a final hidden state; The self-attention module is used to assign a corresponding weight value to each value vector in the second output feature sequence, and perform feature enhancement and feature fusion processing on the second output feature sequence based on the weight value to obtain a throughput prediction result.

6. The method for predicting container terminal throughput based on deep learning according to claim 5, characterized in that: The convolutional neural network module includes: an input layer, a convolution layer, a pooling layer, a feature reconstruction layer and a priori knowledge layer; The convolution layer is used to extract features from the preprocessed actual feature variable data input into the input layer to obtain a first feature mapping sequence; The pooling layer is used to perform a maximum pooling operation on the first feature mapping sequence to obtain a second feature mapping sequence; The feature reconstruction layer is used to perform feature reconstruction on the second feature mapping sequence to sequentially connect the feature mapping elements in the second feature mapping sequence to obtain a preliminary feature sequence; The prior knowledge layer is used to combine the prior knowledge with the preliminary feature sequence to obtain a first output feature sequence.

7. The method for predicting container terminal throughput based on deep learning according to claim 6, characterized in that: The convolution layer includes: a data splitting sublayer, a standard convolution operator, a joint convolution operator and a splicing sublayer; The data splitting sublayer is used to split the preprocessed actual feature variable data into first variable sub-data and second variable sub-data; The standard convolution operator is used to perform a standard convolution operation on the first variable sub-data to obtain a first characteristic sub-sequence; The joint convolution operator is used to perform a PConv partial convolution operation and an MConv micro-point convolution operation on the first feature subsequence to obtain a second feature subsequence; The concatenation sublayer is used to concatenate the second variable sub-data with the second feature sub-sequence to obtain a first feature mapping sequence.

8. The method for predicting container terminal throughput based on deep learning according to claim 1, characterized in that: Extract key request content from the prediction request, including: Verifying the data format and data range of the request content in the prediction request to obtain a verification result; If the verification result is that the data format of the request content in the prediction request satisfies the verification reference format, and the data range satisfies the verification reference range, then the key request content is extracted from the request content.

9. A container terminal throughput prediction device based on deep learning, characterized in that: include: An extraction module, configured to extract key request content in the prediction request after receiving the prediction request for the container terminal throughput; A calling module, used for calling original feature variable data and throughput sample data corresponding to the original feature variable data from multiple data sources based on the key request content; A determination module, configured to determine the mutual dependence between each of the original feature variable data and the throughput sample data; A processing module, used for taking the original feature variable data whose mutual dependence is higher than a preset dependence threshold as feature variable sample data; A construction module, used to train, test and evaluate the deep learning network model based on the characteristic variable sample data and the throughput sample data, and construct a throughput prediction model; The prediction module is used to pre-process the actual characteristic variable data in the key request content and input the data into the throughput prediction model to obtain a throughput prediction result.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the program, the container terminal throughput prediction method based on deep learning as described in any one of claims 1 to 8 is implemented.

Images