CN110085327A - Multichannel LSTM neural network Influenza epidemic situation prediction technique based on attention mechanism - Google Patents

Abstract

The present invention provides the multichannel LSTM neural network Influenza epidemic situation prediction technique based on attention mechanism, belongs to epidemic disease monitoring technical field.The present invention first pre-processes data intensive data, standardizes, feature selecting, and the data of selection are divided into two class of weather dependent data and Influenza epidemic situation related data, generate training set；Then the multichannel LSTM neural network model including attention mechanism is established；Training set data is inputted the model to be trained, and carries out MAPE assessment, obtains trained multichannel LSTM neural network model；Test data is handled, test set is obtained；Test set data are inputted in trained LSTM neural network model and are tested；Inverse standardization finally is carried out to test output result, obtains Influenza epidemic situation predicted value.The present invention solves the problems, such as that existing Influenza epidemic situation Predicting Technique predictablity rate is lower.The present invention can be used for the influenza prediction of different zones.

Description

Attention mechanism-based multi-channel LSTM neural network influenza epidemic situation prediction method

Technical Field

The invention relates to an influenza epidemic situation prediction method, and belongs to the technical field of epidemic disease monitoring.

Background

Influenza is an acute respiratory infection caused by influenza virus. After the patient is infected with the disease, the primary disease is possibly aggravated, and secondary bacterial pneumonia, chronic heart and lung diseases and the like are caused. Outbreaks of influenza epidemics are seasonal and thus may trigger social panic, with major impact on human health and social stability (d.n.t.how, c.k.lo, and k.s.m. sahara. behavior recognition for human beings using short-term social-technical. international Journal of Advanced robotics Systems,13(6):1729881416663369,2016.). For example, influenza type H1N1, outbreak in 2009, caused death of 151,700 to 575,400 people worldwide in the first year of an outbreak (S.Yang, M. Santilana, and S.C.Kou.acquisition information of influenza using a genetic search data, visual. proceedings of the National Academy of Sciences,112(47): 14473-14478, 2015.). Therefore, the accurate and real-time monitoring and early warning of the influenza epidemic situation have important practical significance for public health epidemic prevention departments. The flu epidemic monitoring and early warning system can provide epidemiological information for public health departments, help the public health departments to do epidemic prevention work in advance, and then coordinate each Medical institution in the region to take corresponding countermeasures (J.S. Brown and K.D.Mandl.Regenerative real time outbreak detection systems for influenting Medical monitoring. in AMIA Annual Symposium Proceedings, volume 2006, page866.American Medical information Association, 2006.).

Ili is a criterion for acute respiratory infection established by the world health Organization (abbr. who). Symptoms in flu-like cases are fever above 38 ℃ within 10 days after infection and accompanying cough (W.H. Organization et al. Who intuitional local epidemic behaviour standards for influenca. geneva: world health Organization, pages 1-61,2012.). Our prediction target is ILI%, which is calculated as the ratio of the number of confirmed flu-like cases to the total number of patients. In the field of influenza epidemic monitoring, ILI% is often used as an indicator to determine whether an outbreak of influenza has occurred. When the ILI% exceeds a certain threshold value, the flu season comes, and the relevant department is reminded to do hygiene and epidemic prevention work in time.

In recent years, more and more scholars are focusing on the research of accurate real-time monitoring, early monitoring and epidemic situation early warning of influenza epidemic situations. Influenza epidemic prediction has become a research direction of great interest in academia by using information such as Twitter, Google Correlationship, etc. in web search or social networking sites ([ H.Achrekar, A.Gandhe, R.Lazarus, S. -H.Yu, and B.Liu. Predicting flash using and transmitting data. in computer communications works, (INFOCOM WKSHPS),2011IEEE Conference on, pages 702-707. IEEE,2011., [ D.A.Broniaywski, M.J. Paul 2006, and M.drive.national and localfluenza resonant transmission screw viewer: analysis of 2013 input, 3586.g.) (3527.35: 19. 33: 19. III.12. Hindu.) through the use of Internet, and the research of business discovery, III. Previous research methods are generally based on some commonly used linear models such as least absolute shrinkage and selection operator (LASSO LASSO algorithm) or normalized regression penalty regression, etc. (D.A. broadsword, M.J. Paul, and M.D. draft. national and local approximation screw: analysis of the 2012 and 2013 underfilling electronic plos, 8(12) e83672,2013.], [ M.Santillana, E.O. Nsorese, S.R. root, D.scales, and J.S. brown. using 'areas' discovery, analysis data, and M.S. broadcast of area, analysis of software, 10. export of software, 10. transform, M.D. export of software, see [ 10. export of software, 7. export of software, see [ 10. export of software, see, 7. export of software, 7. software, see, 7. export of software, see, 7. export, see, 2. export, see, 2. copy, see, section, see, section, see, section, 7, see, section, see, sample, see, sample, see. Still others have solved the influenza epidemic prediction problem using a deep learning approach (H.Hu, H.Wang, F.Wang, D.Langley, A.Avram, and M.Liu.Press of influenza-induced genetic basic on the improved anatomical tree, 8(1):4895,2018.], [ Q.Xu, Y.R.gel, L.R.Ramirez, K.Nezafati, Q.Zhang, and K. -L.Tsui.Foresting in fluidic search query and static fusion, P.P.355.: 0176690,2017). However, these methods do not predict the change in influenza epidemic (influenza-like case ratio ILI%) relatively accurately. First, the data collected in the internet is not accurate enough and lacks the features needed to accurately predict influenza epidemics. The result predicted by using the online data cannot accurately reflect the change trend of the influenza epidemic situation. Secondly, influenza epidemic situation data are complex in composition, strong in noise and data diversity, and information in multi-dimensional input data cannot be fully utilized by a traditional linear method. Thirdly, in the deep learning method proposed previously, the time sequence characteristics of the influenza epidemic situation data are not considered; therefore, a high-accuracy influenza epidemic situation prediction technology is urgently needed.

Disclosure of Invention

The invention provides a multichannel LSTM neural network influenza epidemic situation prediction method based on an attention mechanism, which aims to solve the problem of low prediction accuracy of the existing influenza epidemic situation prediction technology.

The invention discloses an attention mechanism-based multi-channel LSTM neural network influenza epidemic situation prediction method, which is realized by the following technical scheme:

step one, preprocessing and standardizing data in a data set; then, performing feature selection by using a mode based on model sorting, and dividing the selected data into weather related data and influenza epidemic situation related data to generate a training set;

step two, establishing a multi-channel LSTM neural network model comprising an attention mechanism; the input of the multichannel LSTM neural network model comprises influenza epidemic related data and weather related data;

inputting training set data into the multichannel LSTM neural network model for training, and performing MAPE (mapping adaptive mapping algorithm) evaluation to obtain a trained multichannel LSTM neural network model;

step four, processing the test data in the same way as the step one to obtain a test set;

inputting the test set data into the trained LSTM neural network model for testing;

and sixthly, carrying out inverse standardization processing on the test output result to obtain the influenza epidemic situation prediction value.

The most prominent characteristics and remarkable beneficial effects of the invention are as follows:

the invention relates to a multichannel LSTM neural network influenza epidemic situation prediction method based on an attention mechanism, which takes the time sequence characteristics of influenza epidemic situation data into consideration and uses a long-term and short-term memory neural network as a basis; by designing a multi-channel structure, the time sequence information in the data is better extracted, so that the data of different types are not influenced mutually in an underlying network, and the fusion of multi-dimensional data in a high-level network is also ensured; in addition, the accuracy of prediction is further improved by adding an Attention mechanism; thereby pertinently solving the influenza prediction problem. Simulation experiments show that the accuracy of the test result of the method is obviously higher than that of other traditional methods, and accurate and effective real-time prediction can be provided for ILI% (influenza sample case ratio).

Drawings

FIG. 1 is a flow chart of the method of the present invention;

FIG. 2 is a schematic diagram of an LSTM memory cell according to the present invention;

FIG. 3 is a schematic diagram of the Attention mechanism of the present invention;

FIG. 4 is a schematic diagram of an Attention-based Multi-channel LSTM model according to an embodiment of the present invention;

FIG. 5 is a graph comparing the real value and the predicted value of Att-MCLSTM in the embodiment of the present invention;

FIG. 6 is a graph comparing the real and predicted values of MCLSTM in an embodiment of the present invention;

FIG. 7 is a graph comparing the real and predicted values of LSTM in an embodiment of the present invention;

FIG. 8 is a graph comparing the real and predicted values of RNN in an embodiment of the present invention;

1. input gate, 2 output gate, 3 forget gate, 4 self-circulation neuron.

Detailed Description

The first embodiment is as follows: the embodiment is described with reference to fig. 1, and the method for predicting the influenza epidemic situation in the multichannel LSTM neural network based on the attention mechanism in the embodiment specifically includes the following steps:

step one, Preprocessing (Preprocessing) and standardizing (Normalization) data in a data set; then, a model-based ranking mode (model-based ranking) is used for feature selection, and the selected data are divided into weather related data and influenza epidemic situation related data to generate a training set;

establishing a multichannel LSTM neural network model including an Attention mechanism (Attention); this model is named Att-MCLSTM, namely Attention-based Multi-channel LSTM (Multi-channel LSTM neural network based on Attention-driven mechanism). Unlike the conventional recurrent Neural network RNN (Recurrent Neural network), the multi-channel LSTM Neural network model based on attention mechanism can solve the problem of gradient disappearance (S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural computation,9(8): 1735-1780, 1997); the memory module of the LSTM memory unit reserves the sequence information of the input context, and has better effect than the traditional RNN in the aspect of time sequence data processing. The input of the multichannel LSTM neural network model (Att-MCLSTM) comprises influenza epidemic related data and weather related data;

inputting the training set data processed in the step one into the multichannel LSTM neural network model for training, and performing MAPE (mapping adaptive mapping algorithm) evaluation to obtain a trained multichannel LSTM neural network model;

step four, processing the test data in the same way as the step one to obtain a test set;

inputting the test set data into the trained LSTM neural network model for testing;

and sixthly, carrying out inverse standardization processing on the test output result in order to reconstruct the original data to obtain the influenza epidemic situation predicted value.

The method adopts a deep neural network method to solve the problem of influenza epidemic situation prediction. In consideration of the time-series characteristics of the influenza epidemic situation data, the present embodiment uses a Long-short term memory (LSTM) neural network 0 as a basic prediction method. Because input data of different dimensions have different characteristics, the time sequence characteristics in multiple data dimensions cannot be fully extracted by using a single network structure. By designing a Multi-channel (Multi-channel) structure, timing information in data can be better extracted. The method not only ensures that different types of data do not influence each other in the underlying network, but also ensures the fusion of multi-dimensional data in a high-level network. The multi-level LSTM neural network model has strong fitting capability, and the accuracy of prediction is further improved by adding an Attention mechanism; in the Attention layer, the probability of occurrence of a value in the output sequence depends on the value in the input sequence. The Attention structure allows the model to better handle the relationships between the different regions of input data.

The second embodiment is as follows: the difference between the present embodiment and the first embodiment is that the normalization in the first step specifically includes the following steps:

and (3) carrying out Min-Max standardization (also called dispersion standardization) on the preprocessed data:

wherein x is a numerical value in the preprocessed data, and x_minIs the minimum value, x, in the pre-processed data_maxIs the maximum value in the preprocessed data, and y is the value of x after Min-Max standardization processing; after data normalization, the data values will scale between 0 and 1.

Other steps and parameters are the same as those in the first embodiment.

The third concrete implementation mode: the difference between this embodiment and the second embodiment is that, as shown in fig. 2, the LSTM memory unit in the multichannel LSTM neural network model in step two includes an input gate 1, an output gate 2, a forgetting gate 3, and a self-circulation neuron 4; the structure gate structure of the LSTM memory cell controls the transfer of data in the LSTM memory cell, including data transfer between different cells and data transfer within the cell. The input gate 1 controls the state updating process of the unit, the output gate 2 controls whether the output sequence of the unit changes the memory state of other units, and the forgetting gate 3 can selectively retain or forget the previous state.

The LSTM memory cell can be represented by the following system of equations:

wherein, sigma (-) is a logistic sigmoid function (mapping variables to be between 0-1), tanh (-) is a hyperbolic tangent function, and both sigma (-) and tanh (-) are activation functions; i.e. i_tInput Gate State at time t, f_tLeft door state at time t, o_tOutput gate state at time t, c_tRepresents the unit state (activation vector) at time t, h_tA hidden state (hidden vector) indicating time t; w_xiA weight matrix representing the weight between the input gate and the input data; w_hi(hidden-input gate) weight matrix representing input gate, hidden layer; w_ciRepresenting weight matrixes among input gates and units; w_xfA weight matrix representing the forgetting gate and the input data; w_hfRepresenting a weight matrix between a forgetting gate and a hidden layer; w_cfRepresenting a weight matrix between forgetting gates and units; w_xcA weight matrix between the unit and the input data is represented; w_hcA weight matrix between the presentation unit and the hidden layer; w_xoRepresenting a weight matrix between the output gate and the input data; w_hoRepresenting the weight matrix between the output gate and the hidden layer; w_coRepresenting the weight matrix between output gates and units; x is the number of_tInput value representing time t, b_iRepresenting an input gate bias term; b_fA representation forgetting gate bias term; b_cRepresenting a cell state bias term; b_oAn output gate bias term is indicated.

Other steps and parameters are the same as those in the first or second embodiment.

The fourth concrete implementation mode: the present embodiment is different from the first, second, or third embodiment in that the attention mechanism in the second step can be expressed as:

m_i＝tanh(W_cmc+W_ymy_i) (3)

a_i＝∝exp(<w_m,m_i>) (4)

z＝∑_ia_iy_i(6)

wherein the attention layer inputs n parameters y₁,…,y_nAnd context sequence c, output vector z; 1, …, n; w_cmIs a context sequence weight matrix, W_ymRepresenting an input vector weight matrix; vector z is given context sequence c, y_iThe weighted arithmetic mean of (a); sequence m_iDenotes c and y_iThe polymerization is calculated by the tanh layer; oc indicates a positive rate; w is a_mA weight matrix representing when the normalization index function softmax (i.e., equation (4)) is performed; a is_iRepresenting m given a context sequence c_iSoftmax results of (1).

Conventional codec structures typically encode an input sequence as a fixed-length vector. However, there are some drawbacks to this structure. When the input sequence is long, the structure is difficult to learn a proper vector characterization mode. The basic idea of the Attention mechanism is to break the traditional codec structure and to achieve selective learning of information in the input sequence by using the intermediate result training model of the LSTM encoder. There is therefore a correlation between the output sequence and the input sequence, i.e. the probability of each value in the output sequence appearing depends on the values in the input sequence.

FIG. 3 is a schematic diagram of the Attention mechanism. Attention layer computation y₁,…,y_nWeight distribution of (S)_tThe input to the LSTM layer at time t includes the output of the Attention layer. LSTM layer output sequence { …, x_t-1,x_t… } the probability of occurrence of the numerical value depends on the input sequence y₁,…,y_n}。

Other steps and parameters are the same as those in the first, second or third embodiment.

The fifth concrete implementation mode: the fourth difference between the present embodiment and the fourth embodiment is that, in the third step, the mape (mean absolute percentage) evaluation specifically includes:

wherein MAPE is the mean absolute percent error,representing the ith true value, p_iRepresenting the ith predicted value; the lower the MAPE number, the higher the accuracy of the model.

Other steps and parameters are the same as those in the fourth embodiment.

The sixth specific implementation mode: the difference between this embodiment and the fifth embodiment is that, in the sixth step, the inverse normalization processing on the test output result specifically includes:

wherein,the predicted value is the influenza epidemic situation, and q is the test output result; q. q.s_maxRepresents the maximum value, q, in the test output results_minRepresenting the minimum value in the test output.

The other steps and parameters are the same as those in the fifth embodiment.

Examples

The following examples were used to demonstrate the beneficial effects of the present invention:

this embodiment uses influenza epidemic data collected by the Guangzhou disease control epidemic prevention center as a data set. The data set comprises 6 modules, and each module has multiple dimensions. Data records are in weeks, including 52 weeks (week) of data each year. Data preprocessing, standardization and feature selection are carried out; selecting characteristics by adopting a model-based ranking (model-based ranking) mode; each time one dimension in the data set is removed, all the remaining dimensions are input into the same prediction model, and the output results of the models are compared. If the accuracy of the prediction result is lower, the removed data dimension is more relevant to the prediction target. And sorting all the prediction results according to the accuracy, thereby selecting 19 dimensions with higher correlation with the prediction target. Selected dimensions and their descriptions are listed in table 1, where the basic information modules (including time information, regions and demographics) are not listed.

TABLE 1 Module and selected dimension description

And dividing the selected dimension data into two types, namely weather related data and influenza epidemic related data. The weather related data comprises average air temperature, highest air temperature, lowest air temperature, rainfall, air pressure and relative humidity, and the rest dimensions are classified as flu epidemic situation related data. All regions have the same weather-related data and different influenza epidemic-related data every week; thus, the correspondence of the multi-channel LSTM neural network model includes two channels: a flu-related channel and a weather-related channel.

The overall structure of the Attention-based Multi-channel LSTM (abbreviated as Att-MCLSTM) model is shown in FIG. 4. Firstly, a network (LSTM 1, …, LSTM 9) consisting of a group of LSTM neural memory units is used for processing influenza epidemic situation related data, and the data of each area (District 1st, District 1nd … District 9th) are respectively input into one LSTM neural memory unit; while a LSTM neural network (LSTM 10) is used to process weather-related data. In order to combine the information extracted from the different regional influenza epidemic related data, the outputs of all LSTM neural networks in the first part are fused in its upper fusion layer (Merge 1). Although this set of LSTM neural networks extracts information about the influenza epidemic in each region, it is still necessary to weight the intermediate output sequences of the extracted information. Because the influenza epidemic information of different areas has different influences on the change trend of the whole influenza epidemic in Guangzhou city. Thus, the intermediate output sequence of the LSTM neural network passes through the Attention layer (Attention) and the fully connected layer (density 1) in sequence. Thereafter, the two pieces of data are fused in a higher layer (fusion layer Merge 2) network. And finally, extracting information fused with multi-dimensional input data after passing through two full connection layers (Dense 2 and Dense 3).

(1) Selection of input data length

I.e. how many consecutive weeks of data are used to optimize the prediction result for the next week. To validate the data from the experimental tests, we split the data into two parts, a training set and a test set. All experimental results are the average of 10 replicates.

The input data length was set to 6 weeks, 8 weeks, 10 weeks, 12 weeks, and 14 weeks, respectively. The parameter settings of each layer of the Attention-based Multi-channel LSTM neural network are shown in Table 2. The activation function is a linear activation function, the loss function is map, and the optimizer uses adam;

TABLE 2 neural network layer parameter settings

Name of neural network layer	Number of units
		LSTM 1,…,LSTM 9	32
LSTM 10	32
		Dense 1	16
Dense 2	10
		Dense 3	1

The data from the first 370 weeks were used for training and the remaining data were tested. Each data record includes weather related data and influenza epidemic related data for 9 regions. The weather related data has 6 dimensions, and the flu epidemic related data has 13 dimensions. The weather-related data (Climate data) is input into a weather-related channel (Climate-related channel), and the stream epidemic data of each area is input into a corresponding flu-related channel (flu-related channel). The prediction results of the model are shown in table 3:

TABLE 3 MAPE values for each prediction result

Number of time cycles	MAPE
		6	0.107
8	0.092
		10	0.086
12	0.106
		14	0.109

As can be seen from table 3, the prediction effect is best when the input data length is 10. This shows that the data of 10 consecutive weeks can sufficiently reflect the time sequence characteristics of the influenza epidemic data. If the length of the input data is too short, the time sequence characteristics of the data cannot be sufficiently reflected; if the input data length is too long, the noise of the data may increase. In the following experiments, we selected data for 10 consecutive weeks as input data.

(2) Evaluation of effects

The validity of the Attention mechanism is verified by comparing the predicted results of Att-MCLSTM and MCLSTM (multichannel LSTM neural network). The setting of neural network parameters and data entry methods are described in table 1.

Secondly, the effectiveness of the multi-channel structure is verified by comparing the prediction results of the MCLSTM and LSTM neural networks. The setting and data input method of the neural network parameters of the MCLSTM is the same as the step (1). The LSTM neural network receives all inputs using one LSTM layer. The influenza epidemic situation related data of the same week in all the areas are added correspondingly and used as input together with the weather related data, so that each data record comprises 19 dimensions. And the output result of the LSTM layer is output after passing through a full connection layer. The number of neurons in the LSTM and fully junctional layers were 32 and 1, respectively.

Finally, the LSTM neural network is proved to have better effect than the common RNN network. The setting of neural network parameters and data input methods are as described above.

MAPE values for the four methods described above are shown in table 4. As can be seen from the table, Att-MCLSTM can give the best results. The data in the first two rows of table 4 show that the MAPE value is increased from 0.105 to 0.086 by adding the Attention mechanism. This illustrates that the Attention structure allows the model to better learn to handle the relationships between the different regions of input data. The data in the second and third rows of table 4 show that the MAPE value was increased from 0.118 to 0.105 by designing a multichannel structure. This demonstrates that the multi-channel structure can better extract the timing characteristics of a variety of input data. The MAPE values for the LSTM neural network and the generic RNN neural network predictions were 0.118 and 0.132, respectively. This demonstrates that the LSTM neural network is better able to handle timing characteristics in the input data than the normal RNN neural network. Meanwhile, the influenza epidemic situation data is verified to have time sequence.

TABLE 4 MAPE values for each prediction result

Method of producing a composite material	MAPE
		The method of the invention, Att-MCLSTM	0.086
MCLSTM	0.105
		LSTM	0.118
RNN	0.132

The true values and predicted values of the four methods are shown in fig. 5, 6, 7, and 8, respectively. As can be seen from the figure, the predicted ILI% of the inventive method (Att-MCLSTM) is closest to the true ILI% (Actual ILI%). The predicted values and the true values of the other three methods have obvious differences. Experiments prove that the method can accurately and fully extract the implicit characteristics of the time sequence data and provide accurate flow epidemic situation prediction.

The present invention may be embodied in other specific forms without departing from the spirit or essential attributes thereof, and it is therefore intended that all such changes and modifications as fall within the true spirit and scope of the invention be considered as within the following claims.

Claims (6)

1. The attention mechanism-based multi-channel LSTM neural network influenza epidemic situation prediction method is characterized by comprising the following steps:

step one, preprocessing and standardizing data in a data set; then, selecting characteristics based on the ranking of the model, dividing the selected data into weather related data and influenza epidemic situation related data, and generating a training set;

step two, establishing a multi-channel LSTM neural network model comprising an attention mechanism; the input of the multichannel LSTM neural network model comprises influenza epidemic related data and weather related data;

inputting training set data into the multichannel LSTM neural network model for training, and performing MAPE (mapping adaptive mapping algorithm) evaluation to obtain a trained multichannel LSTM neural network model;

step four, processing the test data in the same way as the step one to obtain a test set;

inputting the test set data into the trained LSTM neural network model for testing;

and sixthly, carrying out inverse standardization processing on the test output result to obtain the influenza epidemic situation prediction value.

2. The attention mechanism-based multi-channel LSTM neural network influenza epidemic prediction method of claim 1, wherein the normalization in the first step specifically comprises the following processes:

carrying out Min-Max standardization treatment on the preprocessed data:

wherein x is a numerical value in the preprocessed data, and x_minIs the minimum value, x, in the pre-processed data_maxIs the maximum value in the preprocessed data, and y is the value of x after Min-Max standardization processing.

3. The attention mechanism-based multi-channel LSTM neural network influenza epidemic prediction method of claim 2, wherein in step two the LSTM memory unit in the multi-channel LSTM neural network model comprises an input gate, an output gate, a forgetting gate and a self-circulation neuron; the LSTM memory cell can be represented by the following system of equations:

where σ (·) is a logical sigmoid function (mapping variables between 0 and 1)Tanh (·) is a hyperbolic tangent function; i.e. i_tInput Gate State at time t, f_tLeft door state at time t, o_tOutput gate state at time t, c_tRepresents the cell state at time t, h_tRepresenting a hidden state at time t; w_xiA weight matrix representing the weight between the input gate and the input data; w_hiRepresenting the weight matrix between the input gate and the hidden layer; w_ciRepresenting weight matrixes among input gates and units; w_xfA weight matrix representing the forgetting gate and the input data; w_hfRepresenting a weight matrix between a forgetting gate and a hidden layer; w_cfRepresenting a weight matrix between forgetting gates and units; w_xcA weight matrix between the representation unit and the input data; w_hcA weight matrix between the presentation unit and the hidden layer; w_xoRepresenting a weight matrix between the output gate and the input data; w_hoRepresenting the weight matrix between the output gate and the hidden layer; w_coRepresenting the weight matrix between output gates and units; x is the number of_tInput value representing time t, b_iRepresenting an input gate bias term; b_fA representation forgetting gate bias term; b_cRepresenting a cell state bias term; b_oAn output gate bias term is indicated.

4. The method for predicting the influenza epidemic of the multichannel LSTM neural network based on the attention mechanism as claimed in claim 1,2 or 3, wherein the attention mechanism in the second step can be expressed as:

m_i＝tanh(W_cmc+W_ymy_i) (3)

a_i＝∝exp(<w_m，m_i>) (4)

z＝∑_ia_iy_i(6)

wherein the attention layer inputs n parameters y₁，…，y_nAnd context sequence c, output vector z; 1, …, n; w_cmIs a context sequence weight matrix, W_ymRepresenting an input vector weight matrix; vector z is given context sequence c, y_iWeighted arithmetic mean of (a); sequence m_iDenotes c and y_iPolymerization of (a); oc indicates a positive rate; w is a_mRepresenting a weight matrix when normalization is performed; a is_iRepresenting m given a context sequence c_iAnd (5) normalizing the result.

5. The attention mechanism-based multichannel LSTM neural network influenza epidemic prediction method according to claim 4, wherein the MAPE evaluation in step three specifically comprises:

wherein MAPE is the mean absolute percent error,representing the ith true value, p_iIndicates the ith prediction value.

6. The attention mechanism-based multi-channel LSTM neural network influenza epidemic prediction method according to claim 5, wherein the inverse standardization processing of the test output result in the sixth step is specifically:

wherein,the predicted value is the influenza epidemic situation, and q is the test output result; q. q.s_maxRepresents the maximum value, q, in the test output results_minRepresenting the minimum value in the test output.