Traffic injury emergency call volume prediction method and system based on deep learning model

Abstract

The invention discloses a traffic injury emergency call volume prediction method and system based on a deep learning model, and the method comprises the following steps: S1, obtaining historical call data, and carrying out the preprocessing of the historical call data; s2, dividing the preprocessed historical distress call data according to different time periods to obtain data sets corresponding to different time period division results; s3, establishing an LSTNFCL deep learning model, and respectively training the LSTNFCL deep learning model by adopting each data set; and S4, according to a training result, obtaining an optimal mode of time period division, and performing call volume prediction based on the correspondingly trained LSTNFCL deep learning model. The traffic injury first-aid call volume can be accurately predicted, and technical support is provided for pre-hospital first aid.

Description

Technical Field

[0001] The present invention relates to the technical field of call volume prediction, and particularly relates to a traffic injury first aid call volume prediction method and system based on a deep learning model. Background Art

[0002] In the past five years, with the rapid economic and social development of our country, whether it is the number of motor vehicles, the number of drivers, or the mileage of highways, they have all entered the fast lane of growth, and the traffic safety guarantee of the people faces greater challenges. Pre-hospital medical first aid is related to the life and health safety of the people. In the era of big data and artificial intelligence, monitoring and early warning based on digital technology to ensure the life safety of the people meet the current development needs.

[0003] In the field of pre-hospital first aid, current research mainly involves fields such as ambulance dispatching, construction of first aid information platforms, and early warning of sudden public events. There is little early warning research for specific types such as traffic injuries. The existing call volume prediction methods mainly have the following disadvantages: (1) The prediction model is relatively simple, and the historical call volume information for each time period is not fully utilized; (2) The time period division of the prediction model is too rough, and the utilization rate of call volume data is low; (3) The weight setting of the prediction model overly relies on historical data. The external influencing factors of call volume are dynamically changing, and their influence sizes are usually not fixed, and the prediction accuracy needs to be improved. Summary of the Invention

[0004] In view of the above problems, the present invention aims to provide a traffic injury first aid call volume prediction method and system based on a deep learning model.

[0005] The technical solution of the present invention is as follows:

[0006] On the one hand, a traffic injury first aid call volume prediction method based on a deep learning model is provided, including the following steps:

[0007] S1: Obtain historical call-for-help data and preprocess the historical call-for-help data;

[0008] S2: Divide the preprocessed historical call-for-help data according to different time periods to obtain data sets corresponding to the division results of each different time period;

[0009] S3: Establish an LSTNFCL deep learning model, and use each data set to train the LSTNFCL deep learning model respectively;

[0010] S4: According to the training results, obtain the optimal way of time period division, and perform call volume prediction with the corresponding trained LSTNFCL deep learning model.

[0011] Preferably, in step S1, the historical distress data includes the distress date and the call volume corresponding to the distress date.

[0012] Preferably, in step S1, the preprocessing includes outlier processing, null value processing, and data standardization processing.

[0013] Preferably, when performing outlier processing, a box plot is used to determine whether the data is an outlier, and the data determined to be an outlier is deleted.

[0014] Preferably, when performing null value processing, if the number of consecutive null values is greater than 2, the consecutive null values are deleted; if the number of consecutive null values is less than or equal to 2, the null values are filled with the average value.

[0015] Preferably, in step S2, the different time periods include 6-hour time intervals, 12-hour time intervals, and 24-hour time intervals.

[0016] Preferably, in step S3, the LSTNFCL deep learning model is based on the LSTNet model, extracts features through convolution combined with the Conv2Former self-attention mechanism, comprehensively fuses its features by combining GRU, CBAM, RNN, and LSTM time series modules, and finally outputs the features after combining AR autoregression.

[0017] Preferably, the LSTNFCL deep learning model includes an input layer, and the output end of the input layer is connected to the input end of the Conv layer;

[0018] One output end of the Conv layer is connected to the input end of the AR model;

[0019] Another output end of the Conv layer is connected to the input end of the Conv2Former model. One output end of the Conv2Former model is connected to the CBAM model and the RNN model in sequence, and another output end of the Conv2Former model is connected to the GRU model and the LSTM model in sequence;

[0020] The output ends of the RNN model and the LSTM model are both connected to the input end of the Concat layer, and the output end of the Concat layer is connected to the input end of the Linear layer;

[0021] The output ends of the AR model and the Linear layer are both connected to the input end of the Add layer, and the output end of the Add layer is connected to the input end of the output layer.

[0022] Preferably, in step S3, when training, the stochastic gradient descent method is used to optimize the loss function.

[0023] On the other hand, a traffic injury first aid call volume prediction system based on a deep learning model is also provided, including:

[0024] A data collection and preprocessing module, configured to obtain historical call-for-help data and preprocess the historical call-for-help data;

[0025] A data set establishment module, configured to divide the preprocessed historical call-for-help data according to different time periods to obtain data sets corresponding to the division results of each different time period;

[0026] A model establishment and training module, configured to establish an LSTNFCL deep learning model and train the LSTNFCL deep learning model respectively using each data set;

[0027] A screening and prediction module, configured to obtain the optimal way of time period division according to the training results, and perform call volume prediction using the trained LSTNFCL deep learning model corresponding thereto.

[0028] The beneficial effects of the present invention are:

[0029] By dividing according to different time periods, the present invention can obtain multiple data sets, with a large amount of data samples, can make full use of historical data, and is applicable to large-scale group call volume prediction scenarios; and by optimizing the optimal time period division method, the overall prediction accuracy can be improved.

[0030] Through the LSTNFCL deep learning model of the present invention, the characteristic information related to time series in the data can be fully extracted, and the prediction accuracy of the model can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0032] Figure 1 It is a schematic structural diagram of a specific embodiment of the LSTNFCL deep learning model of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0033] The present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, without conflict, the embodiments in the present application and the technical features in the embodiments may be combined with each other. It should be pointed out that unless otherwise specified, all technical and scientific terms used in the present application have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present application belongs. The terms "including" or "comprising" and the like used in the disclosure of the present invention mean that the elements or items appearing before the term cover the elements or items listed after the term and their equivalents, without excluding other elements or items.

[0034] On the one hand, the present invention provides a traffic injury first aid call volume prediction method based on a deep learning model, including the following steps:

[0035] S1: Obtain historical call-for-help data and preprocess the historical call-for-help data.

[0036] In a specific embodiment, the historical call-for-help data includes the call-for-help date and the call volume corresponding to the call-for-help date, and the preprocessing includes outlier processing, null value processing, and data standardization processing.

[0037] Optionally, when performing outlier processing, a box plot is used to determine whether the data is an outlier, and the data determined to be an outlier is deleted. It should be noted that the outlier judgment method in this embodiment is only a preferred judgment method of the present invention, and other outlier judgment methods in the prior art can also be applied to the present invention.

[0038] Optionally, when performing null value processing, if the number of consecutive null values is greater than 2, the consecutive null values are deleted; if the number of consecutive null values is less than or equal to 2, the null values are filled with the average value. Optionally, the average value is the average value of the past ten days.

[0039] S2: Divide the preprocessed historical call-for-help data according to different time periods to obtain data sets corresponding to the division results of each different time period.

[0040] In a specific embodiment, the different time periods include a 6-hour time interval, a 12-hour time interval, and a 24-hour time interval.

[0041] S3: Establish an LSTNFCL deep learning model, and use each data set to train the LSTNFCL deep learning model respectively.

[0042] In a specific embodiment, the LSTNFCL deep learning model is based on the LSTNet model, extracts features through convolution combined with the Conv2Former self-attention mechanism, comprehensively fuses its features by combining GRU, CBAM, RNN, and LSTM time series modules, and finally outputs the features after combining the AR autoregression.

[0043] The formula for convolution is:

[0044]

[0045] In the formula: z(u, v) is the output feature; i is the number of rows; j is the number of columns; x i,j is the element of the feature data at i, j; k u-i,v-j is the element of the convolution kernel at the i, j;

[0046] Conv2Former modulates V in the attention mechanism with convolution features:

[0047] Z = A · V (2)

[0048] A = DConv k×k (W 1 X)

[0049] V = W 2 X

[0050] In the formula: Z is the output attention matrix; V is the Value matrix, which contains the representation or features of the input sequence; DConv k×k is the Depth-wise convolution with a convolution kernel size of k; W 1 is the convolution kernel parameter; X is the input feature; W 2 is the weight matrix.

[0051] Using convolution can simplify the large computational amount and the number of parameters of the attention mechanism.

[0052] CBAM can effectively capture the complex spatio-temporal relationships between different time steps and different feature dimensions in time series data by combining channel attention and spatial attention mechanisms. This ability enables the model to better understand the temporal information in the data, thereby improving the modeling ability for time series data. At the same time, it can adaptively learn the important features in the data and improve the generalization performance of the model.

[0053] RNN is a type of neural network structure specifically designed for processing sequence data. Its characteristic lies in that it can introduce cyclic connections to process the temporal information of sequence data. The specific inference process is as follows:

[0054]

[0055] In the formula: a is the information input, and the superscript represents the corresponding time; is the information output; g() is the activation function; W aa , W ax , W y are all weight information; x <t>< / t> is the data feature at the t-th moment; b a is the bias corresponding to a; b y is the bias corresponding to y.

[0056] LSTM introduces three key gating structures: the input gate, the forget gate, and the output gate. Through these gating structures, the flow of information can be better controlled, and long-term dependencies can be effectively captured, thus alleviating the problems of gradient vanishing and gradient explosion, making LSTM perform better in processing sequence tasks that require long-term memory. Its calculation formula is:

[0057]

[0058] h t = u(W h c t + b y )

[0059] Γ u = σ(W u [c t-1 , x t + b u )

[0060] Γ r = σ(W r [c t-1 , x t + b r )

[0061] In the formula: is the memory matrix at the t-th moment; u() is the activation function; W c is the weight information corresponding to the memory matrix; Γ r is the reset gate matrix; c t-1 , c t are the memory state matrices at the (t - 1)-th moment and the t-th moment respectively; x t is the data at the t-th moment; b c is the bias corresponding to the memory state matrix; Γ u is the update gate matrix; h t is the past information data at the t-th moment; W h is the weight corresponding to h; b y is the output bias vector; σ() is the activation function; W u is the weight corresponding to the update gate matrix; b uTo update the bias corresponding to the gate matrix; W r To reset the weight corresponding to the gate matrix; b r To reset the bias corresponding to the gate matrix.

[0062] Through Γ u and Γ r The importance and weight information of c and x can be learned respectively.

[0063] GRU is more simplified compared to LSTM, only containing two gating structures, the update gate and the reset gate, reducing the number of parameters. The calculation formula is:

[0064]

[0065] In the formula: f and i are the forget gate and input gate functions respectively; O is the final output; g is the output value after the update gate, determined by the previous position output h t-1 and the input x at this position t decide; σ and tanh are both activation functions; W is the weight information.

[0066] In a specific embodiment, as Figure 1 shown, the LSTNFCL deep learning model includes an input layer, and the output end of the input layer is connected to the input end of the Conv layer;

[0067] One output end of the Conv layer is connected to the input end of the AR model;

[0068] Another output end of the Conv layer is connected to the input end of the Conv2Former model. One output end of the Conv2Former model is sequentially connected to the CBAM model and the RNN model, and another output end of the Conv2Former model is sequentially connected to the GRU model and the LSTM model;

[0069] The output ends of the RNN model and the LSTM model are both connected to the input end of the Concat layer, and the output end of the Concat layer is connected to the input end of the Linear layer;

[0070] The output ends of the AR model and the Linear layer are both connected to the input end of the Add layer, and the output end of the Add layer is connected to the input end of the output layer.

[0071] In the above embodiment, the convolved features are respectively input into the AR autoregressive module and the Conv2Former module. Time series data often contains information at multiple levels and scales, such as the minute-level and hour-level fluctuations in the data. The cross-layer connections in the module can better integrate and utilize information at different scales, improving the model's ability to extract features from time series data:

[0072] It should be noted that the LSTNFCL deep learning model in the above embodiments is only a preferred implementation model (with the smallest error) of the present invention. Other LSTNFCL deep learning models obtained by the inventor by changing the order of Conv2Former, GRU, CBAM, RNN, and LSTM (exchanging the positions of CBAM and Conv2Former, exchanging the positions of GRU and LSTM, and exchanging the positions of CBAM and Conv2Former and GRU and LSTM) can also accurately predict the call volume.

[0073] In a specific embodiment, during training, the stochastic gradient descent method is used to optimize the loss function. Optionally, the MSE algorithm is used as the loss function for weight optimization, and the output of the validation set is used for evaluation:

[0074]

[0075] In the formula: n is the number of a batch of training data; y is the predicted value, and Y is the true value.

[0076] In the above embodiments, by using the stochastic gradient descent method to optimize the loss function, the model weight parameters can be optimized, and the weight information can be optimized by backpropagation through the loss function L = MSE:

[0077]

[0078] In the formula: W l is the weight information; b l is the bias information; η is the parameter; B is the batch number; L is the loss function.

[0079] It should be noted that when using MSE for model effect evaluation, any one or more of Mae, Rmse, Mape, Mspe, RSE, RAE, and CORR can also be used for comprehensive evaluation.

[0080] S4: According to the training results, obtain the optimal way of time period division, and use the corresponding trained LSTNFCL deep learning model to predict the call volume.

[0081] On the other hand, the present invention also provides a traffic injury emergency call volume prediction system based on a deep learning model, including:

[0082] A data collection and preprocessing module, configured to obtain historical call-for-help data and preprocess the historical call-for-help data;

[0083] A dataset establishment module, which is used to divide the preprocessed historical emergency call data according to different time periods, and obtain datasets corresponding to the division results of each different time period;

[0084] A model establishment and training module, which is used to establish an LSTNFCL deep learning model, and use each dataset to train the LSTNFCL deep learning model respectively;

[0085] A screening and prediction module, which is used to obtain the optimal way of time period division according to the training results, and perform call volume prediction with the trained LSTNFCL deep learning model corresponding thereto.

[0086] In a specific embodiment, taking 12 districts in Chengdu as an example, the traffic injury emergency call volume prediction method based on the deep learning model of the present invention is used to predict the traffic injury emergency call volume thereof.

[0087] In this embodiment, the traffic injury call volume of the Chengdu Emergency Center is used as the feature data of the prediction model training dataset. The dataset is divided by 24 hours, 12 hours, and 6 hours as time periods respectively, and the datasets divided by different time periods are divided into training sets and test sets according to a ratio of 8:2 respectively. The data is input into the model in the way of a dynamic window. The data of the previous 24 days is input into the model, and the prediction result of the 25th day is output. The shape of the model input feature is 24×12, and the model output is 1×12. The sliding window 24 slides backward with a step size of 1, and then continuously predicts the subsequent values. The prediction results of this embodiment are shown in Table 1:

[0088] Table 1 Effects of the LSTNFCL deep learning model under different dataset divisions

[0089] Dataset division situation MSE MAE RMSE MAPE(%) MSPE(%) RES RAE CORR 6 hours 2.486 1.681 3.301 0.891 14.363 0.343 0.372 0.488 12 hours 3.683 2.881 5.002 0.581 12.812 0.3 0.327 0.266 24 hours 4.721 4.081 7.102 0.331 8.162 0.213 0.249 0.041

[0090] As can be seen from Table 1, the model trained with the data divided at 6-hour intervals has higher accuracy, and can improve the accuracy and model stability of the LSTNFCL deep learning model. Select this model as the final model for call volume prediction.

[0091] In addition, the datasets divided by 6 hours are used to train the traditional CNN model and the basic LSTNet model, and the results are shown in Table 2:

[0092] Table 2 Prediction effects of different models

[0093] Model MSE MAE RMSE MAPE(%) MSPE(%) RES RAE CORR LSTNFCL 2.486 1.681 3.301 0.891 14.363 0.343 0.372 0.488 CNN 2.931 2.944 3.573 2.937 19.366 0.369 0.394 0.473 LSTNet 2.673 1.899 3.365 1.028 16.531 0.361 0.382 0.496

[0094] As can be seen from Table 2, the LSTNFCL deep learning model adopted by the present invention can predict the call volume more accurately than the traditional CNN model and the basic LSTNet model.

[0095] The above are only the preferred embodiments of the present invention, and do not impose any form of limitation on the present invention. Although the present invention has been disclosed above with the preferred embodiments, it is not intended to limit the present invention. Any person skilled in the relevant art can make some changes or modifications to form equivalent embodiments by using the technical content disclosed above within the scope of the technical solution of the present invention. However, as long as it does not depart from the technical content of the technical solution of the present invention, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A method for predicting the number of traffic emergency calls based on a deep learning model, comprising the following steps: S1: Acquire historical distress call data and pre-process the historical distress call data; S2: Divide the pre-processed historical distress call data into different time periods to obtain data sets corresponding to the division results of different time periods; S3: establishing a LSTNFCL deep learning model, and using each data set to train the LSTNFCL deep learning model respectively; S4: According to the training results, the optimal way to divide the time periods is obtained, and the call volume is predicted using the corresponding trained LSTNFCL deep learning model.

2. The method for predicting the number of traffic emergency calls based on a deep learning model according to claim 1 is characterized in that: In step S1, the historical rescue data includes a rescue date and a call volume corresponding to the rescue date.

3. The method for predicting the number of traffic emergency calls based on a deep learning model according to claim 1, characterized in that: In step S1, the preprocessing includes outlier processing, null value processing and data standardization processing.

4. The method for predicting the number of traffic emergency calls based on a deep learning model according to claim 3 is characterized in that: When processing outliers, a box plot is used to determine whether the data is an outlier, and the data determined to be an outlier is deleted.

5. The method for predicting the number of traffic emergency calls based on a deep learning model according to claim 3 is characterized in that: When processing null values, if the number of consecutive null values ​​is greater than 2, the consecutive null values ​​are deleted; if the number of consecutive null values ​​is less than or equal to 2, the null values ​​are filled with the average value.

6. The method for predicting the number of traffic emergency calls based on a deep learning model according to claim 1, characterized in that: In step S2, the different time periods include 6-hour time intervals, 12-hour time intervals and 24-hour time intervals.

7. The method for predicting the number of traffic emergency calls based on a deep learning model according to any one of claims 1 to 6, characterized in that: In step S3, the LSTNFCL deep learning model is based on the LSTNet model, extracts features through convolution combined with the Conv2Former self-attention mechanism, integrates its features with the GRU, CBAM, RNN and LSTM time series modules, and finally outputs the features after AR autoregression.

8. The method for predicting the number of traffic emergency calls based on a deep learning model according to claim 7, characterized in that: The LSTNFCL deep learning model includes an input layer, and the output end of the input layer is connected to the input end of the Conv layer; One of the output ends of the Conv layer is connected to the input end of the AR model; Another output end of the Conv layer is connected to the input end of the Conv2Former model, one output end of the Conv2Former model is connected to the CBAM model and the RNN model in sequence, and another output end of the Conv2Former model is connected to the GRU model and the LSTM model in sequence; The output ends of the RNN model and the LSTM model are both connected to the input end of the Concat layer, and the output end of the Concat layer is connected to the input end of the Linear layer; The output ends of the AR model and the Linear layer are both connected to the input end of the Add layer, and the output end of the Add layer is connected to the input end of the output layer.

9. The method for predicting the number of traffic emergency calls based on a deep learning model according to claim 1, characterized in that: In step S3, during training, the loss function is optimized using the stochastic gradient descent method.

10. A traffic emergency call volume prediction system based on a deep learning model, characterized in that: include: A data collection and preprocessing module, used to obtain historical distress call data and preprocess the historical distress call data; The data set establishment module is used to divide the pre-processed historical distress call data into different time periods to obtain data sets corresponding to the division results of different time periods; A model building and training module, used to build a LSTNFCL deep learning model and train the LSTNFCL deep learning model using each data set; The screening and prediction module is used to obtain the optimal way to divide the time periods according to the training results, and use the corresponding trained LSTNFCL deep learning model to predict the call volume.

Images