Influenza incidence prediction method and system based on multi-scale multivariate time sequence feature fusion

Abstract

The invention discloses an influenza incidence prediction method and system based on multi-scale multivariate time sequence feature fusion. The method comprises the following steps: (1) collecting multivariate time sequence data daily; (2) preprocessing the collected multivariate time series data to obtain multivariate time series data with the scales of day, week, month and quarter; (3) performing deep fusion and characterization on the multivariate time series data of different scales to obtain time series data characterization of different scales, and performing fusion by using a pyramid feature fusion network to obtain time series feature coding sequences under different scales; time sequence feature coding sequences under different scales are optimized by calculating similarity; (4) constructing an influenza incidence prediction model, and training the prediction model by using the daily time sequence feature coding sequence as input; and (5) after the training is completed, predicting the influenza incidence at the next time point according to the input multi-scale multivariate time sequence data. According to the invention, the accuracy of influenza incidence prediction can be improved.

Description

Technical Field

[0001] The present invention belongs to the field of medical information technology, and in particular relates to a method and system for predicting influenza incidence based on multi-scale multivariate time series feature fusion. Background Art

[0002] The incidence of influenza is affected by many factors, including virus mutation, seasonal and climatic conditions, population immunity, vaccination rate and effectiveness, population characteristics such as the susceptibility of the elderly and children, social behavior and population mobility, and the implementation of public health measures. These factors work together to affect the spread and scale of influenza. The incidence of influenza usually shows obvious seasonal fluctuations, with the autumn and winter being the peak seasons and the incidence rate being higher in susceptible populations. The incidence decreases when the vaccination rate is high or public health measures are strict, and virus mutations may lead to influenza pandemics or more severe epidemic waves.

[0003] For example, a Chinese patent document with publication number CN109545386A discloses a spatiotemporal prediction method and device for influenza based on deep learning. The method and device first collect outpatient influenza-like case data in the region; construct an influenza spatiotemporal feature matrix based on the collected outpatient influenza-like case data in the region; generate a data set based on the influenza spatiotemporal feature matrix; construct a convolutional neural network model; use the generated data set to train the constructed convolutional neural network model to obtain parameter information of the convolutional neural network model; use the trained convolutional neural network model to predict the currently input outpatient influenza-like case data. The method and device mine the spatiotemporal characteristics of influenza from two dimensions, time and space, and effectively improve the prediction performance of influenza. The invention uses the superior performance of convolutional neural networks to extract the spatial correlation and temporal correlation between influenza in each sub-region under the region, and is used to predict future influenza trends in each sub-region.

[0004] For example, a Chinese patent document with publication number CN113257426A discloses a clustered group influenza prediction system, storage medium and device, including an LSTM module, a LightGBM module and a prediction module: the LSTM module is configured to use LSTM to obtain a predicted value n of the number of influenza infections in the clustered group; the LightGBM module is configured to use the LightGBM algorithm to predict the probability of the ith person in the clustered group being infected with influenza, and to obtain an infection probability sequence by sorting from large to small; the prediction module is configured to calculate the predicted value n of the number of influenza infections according to the probability correction obtained by the LightGBM module, obtain the predicted correction value of the number of influenza infections, and determine the individuals with the highest probability of influenza infection as the high-risk group. This invention combines the influenza outbreak trend in the clustered group with the influenza infection probability of each individual in the group, and can accurately predict the susceptible people in the clustered group, wherein the correction calculation step improves the accuracy of the prediction results while ensuring the calculation speed.

[0005] The influenza incidence time series forecast aims to predict future time series changes based on past time series observations. However, due to the complexity and non-stationarity of real factors, the observed influenza incidence time series usually presents a complex time pattern, in which multiple change trends (such as rising, falling and fluctuating) are deeply mixed, which brings great challenges to the prediction task. The influenza incidence time series presents different time series changes at different sampling scales. For example, the influenza incidence recorded daily shows the trend of influenza incidence in the short term, and the impact of a certain virus epidemic or climate change on the influenza incidence can be observed. In the time series sampled on a weekly basis, fine-grained time series information is integrated, and the impact of different viruses, seasonal changes, public health measures and other factors on the influenza incidence can be observed. Traditional time series methods usually only use fixed time windows to extract information, which is difficult to adapt to pattern changes on different time scales. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention provides a method and system for predicting influenza incidence based on the fusion of multi-scale multivariate time series features. By adjusting the time resolution and distance, it can not only capture the local details of the time series, but also grasp its long-term trends and periodic changes; at the same time, the complementary prediction capabilities of multi-scale multivariate sequences are utilized to model complex time series change trends, thereby improving the accuracy of influenza incidence prediction.

[0007] A method for predicting influenza incidence based on multi-scale multivariate time series feature fusion, comprising the following steps:

[0008] (1) Collect multivariate time series data on a daily basis, including meteorological information, influenza incidence, influenza type distribution ratio, gender ratio of the patient population, and age structure of the patient population;

[0009] (2) Fill missing values ​​and process outliers in the multivariate time series data collected daily, and then calculate the time series data X based on the processed multivariate time series data X. D Generate multivariate time series data X with weekly, monthly, and quarterly scales W , X M , X S ;

[0010] (3) Deeply fuse and represent multivariate time series data of different scales to obtain time series data representations of different scales; use the pyramid feature fusion network to fuse time series data representations of different scales to obtain time series feature coding sequences at different scales; optimize time series feature coding sequences at different scales by calculating the similarity between time series feature coding sequences at different scales;

[0011] (4) Construct an influenza incidence prediction model and use the daily time series feature coding sequence with the finest time granularity as input to train the prediction model;

[0012] (5) After training is completed, the influenza incidence at the next time point is predicted based on the input multi-scale multivariate time series data.

[0013] In step (2), the missing values ​​of the multivariate time series data are filled using the linear interpolation method; the outliers of the multivariate time series data are identified using the isolation forest algorithm, and then the replacement values ​​of the outliers are calculated and replaced using the linear interpolation method.

[0014] In step (3), the multivariate feature fusion component is used to perform deep fusion and characterization on multivariate time series data of different scales. The working process of the multivariate feature fusion component is as follows:

[0015] For multivariate time series data of each scale, the time series features of each dimension are encoded by its exclusive LSTM feature encoder to obtain the corresponding exclusive coding sequence; at the same time, the time series features of all dimensions are encoded by a shared LSTM feature encoder with shared weights to obtain a shared coding sequence; then, the exclusive coding sequence and the shared coding sequence are added together to obtain the time series data representation that integrates multivariate information;

[0016] Multivariate time series data X of different scales D , X W , X M , X S After the multi-feature fusion component, the time series data representation Z at the corresponding time scale is obtained. D , Z W , Z M , Z S .

[0017] In step (3), the pyramid feature fusion network is used to fuse the time series data representations of different scales to obtain the time series feature coding sequences at different scales. The specific process is as follows:

[0018] Seasonal time series data representation Z S Through the seasonal LSTM encoding network, the seasonal time series feature encoding sequence L is obtained S ; Monthly corresponding time series data representation Z M and seasonal time series feature encoding sequence L S Combined as the input of the monthly LSTM encoding network, the monthly time series feature encoding sequence L is obtained M ; The time series data corresponding to the week represents Z W and monthly time series feature coding sequence L M Combined as the input of the weekly LSTM encoding network, the weekly time series feature encoding sequence L is obtained. W ; The time series data corresponding to the day represents Z D and weekly time series feature coding sequence L W The combined data is used as the input of the daily LSTM encoding network to obtain the daily time series feature encoding sequence L D .

[0019] Due to the seasonal time series feature encoding sequence L S The sequence length is less than the corresponding monthly time series data representation Z M The sequence length is L through linear interpolation upsampling technology. S Expanded to time series length and Z M Consistent upsampled seasonal time series feature encoding sequence Then, the monthly time series data is represented by Z M And upsampled seasonal time series feature encoding sequence After adding each element, it is used as the input of the monthly LSTM feature encoding network to obtain the monthly time series feature encoding sequence L M Similarly, we need to change L M Expanded to time series length and Z W Consistent upsampled monthly time series feature encoding sequence L W Expanded to time series length and Z D Consistent upsampled weekly time series feature encoding sequence

[0020] In step (3), the similarity between time series feature coding sequences at different scales is calculated by dynamic time warping distance. Specifically, dynamic time warping distance allows the time series to be stretched or compressed at certain positions to match similar patterns in another sequence. The sequence contrast loss calculation formula is as follows:

[0021] Losss =-dtw(L S ,L M )-dtw(L M ,L W )-dtw(L W ,L D )

[0022] Among them, dtw(L S ,L M ) represents the seasonal time series feature coding sequence L S and monthly time series feature coding sequence L M The dynamic time warping distance between M ,L W ) represents the monthly time series feature coding sequence L M and weekly time series feature coding sequence L W The dynamic time warping distance between W ,L D ) represents the weekly time series feature coding sequence L W and daily time series feature coding sequence L D The dynamic time warping distance between them.

[0023] In step (4), the daily time series feature encoding sequence L D As input, according to the sequence contrast loss Loss s And mean square error loss Loss MSE Construct the overall loss function L and train the model by minimizing the overall loss:

[0024] L=Loss s +Loss MSE

[0025] Among them, the mean square error loss Loss MSe It is used to measure the difference between the model prediction value and the true value.

[0026] Based on the same inventive principle, the present invention also provides an influenza incidence prediction system based on multi-scale multivariate time series feature fusion, including:

[0027] The time series data collection module is used to collect multivariate time series data on a daily basis, including meteorological information, influenza incidence, influenza type distribution ratio, gender ratio of the patient population, and age structure of the patient population;

[0028] The time series data preprocessing module is used to fill missing values ​​and process outliers in the multivariate time series data collected daily, and to calculate the time series data based on the processed multivariate time series data X D Generate multivariate time series data X with weekly, monthly, and quarterly scales W , XM , X S ;

[0029] The multi-scale multivariate time series data modeling module is used to perform deep fusion and representation of multivariate time series data of different scales to obtain time series data representations of different scales; the pyramid feature fusion network is used to fuse time series data representations of different scales to obtain time series feature coding sequences at different scales; the time series feature coding sequences at different scales are optimized by calculating the similarity between the time series feature coding sequences at different scales;

[0030] The time series data prediction module uses the model built by the multiscale multivariate time series data modeling module to predict the influenza incidence at the next time point based on the input multiscale multivariate time series data.

[0031] Based on the same inventive principle, the present invention also provides an influenza incidence prediction system based on multi-scale multivariate time series feature fusion, including a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the above-mentioned influenza incidence prediction method.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] The present invention firstly uses sampling technology to efficiently align time series data of different granularities to ensure that various types of data can be seamlessly integrated and processed at different time scales; based on the LSTM network architecture, the deep integration of multivariate time series data is realized, and comprehensive feature extraction and characterization of multiscale multivariate time series data are performed to capture complex time dependencies and dynamic patterns; further, by calculating the similarity between time series data of different scales, the embedded representation is optimized, the accuracy and consistency of data representation are improved, and the generalization ability of the model is enhanced; in the model training process, the mean square error loss function is used as the core indicator to maximize the fitting effect, so that the accuracy and stability of influenza incidence prediction are significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A schematic diagram of an influenza incidence prediction system based on multi-scale multivariate time series feature fusion provided by an embodiment of the present invention;

[0035] Figure 2 This is a schematic diagram of the structure of a multi-scale multivariate time series data modeling module in an embodiment of the present invention;

[0036] Figure 3 Schematic diagram of a multi-feature fusion component in an embodiment of the present invention. DETAILED DESCRIPTION

[0037] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be pointed out that the embodiments described below are intended to facilitate the understanding of the present invention and do not have any limiting effect on the present invention.

[0038] like Figure 1 As shown, the influenza incidence prediction system based on multi-scale multivariate time series feature fusion includes a time series data acquisition module, a time series data preprocessing module, a multi-scale multivariate time series data modeling module, and an influenza incidence prediction module.

[0039] The time series data collection module is used to collect multivariate time series data on a daily basis to support the prediction of influenza incidence. Specifically, this module collects meteorological information including temperature, humidity, wind speed, weather, influenza incidence, influenza type distribution ratio, gender ratio and age structure of the sick population. Through the collection of these data, the changes in seasonal and environmental factors in the process of influenza transmission can be captured, and the epidemic trends of different influenza types in different periods can be analyzed. Data on gender and age structure can help identify high-risk groups and reveal the trend of influenza transmission in different groups. These multivariate time series data support the construction of multiscale multivariate time series data prediction models to help timely predict the spread of influenza. The formal definition of multivariate time series data is:

[0040]

[0041] Among them, X D The dimension is N×T, where N is the dimension of the multivariate time series data. If the multivariate time series data includes information in three dimensions, such as temperature, humidity, and influenza incidence, then N = 3. T is the length of the time series data in one dimension. The length of the time series data in different dimensions is T.

[0042] The time series data preprocessing module preprocesses the collected multivariate time series data, including filling missing values ​​in time series data, outlier processing, and multi-scale time series data construction.

[0043] First, the linear interpolation method is used to fill missing values ​​in time series data. For each missing value, find the valid data points before and after it; calculate the values ​​of the previous valid point and the next valid point, as well as their time interval; use the linear interpolation formula to perform linear interpolation to ensure data integrity.

[0044] Next, the isolation forest algorithm is used to identify outliers in time series data. The number of trees in the model and the sample size of each tree are specified, and the model is trained using multivariate time series data to establish an isolation forest model. After training, the isolation forest model is used to predict each data point and calculate its outlier score. The lower the score, the more likely it is an outlier. Next, a suitable threshold is selected based on the score distribution, and data points with scores below the threshold are marked as outliers. Finally, the linear interpolation method is used to calculate the replacement value of the outlier point and replace it. By processing outliers in time series data, it can ensure that subsequent feature extraction, modeling, and prediction can be more stable and reliable.

[0045] Finally, time series data of different scales are generated according to different time granularities. For data collected daily, coarse-grained time series information is generated according to time granularities such as weeks, months, and quarters, so that the data can show trends and characteristics over a wider time range. In addition, constructing multi-scale time series features can capture the time series characteristics presented by time series data at different time granularities, thereby helping the model to more comprehensively model the patterns in time series data. D Generate coarse-grained multivariate time series data according to the time granularity of week, month, quarter, etc., denoted as X W , X M , X S .

[0046] The multi-scale multivariate time series data modeling module fuses multi-variate time series data based on the multi-feature fusion component, deeply fuses and characterizes the multi-scale time series data; uses the pyramid feature fusion network to fuse multi-scale time series features; optimizes the embedding representation of multi-scale multivariate time series data by calculating the similarity between time series of different scales, and uses the square loss to optimize the fitting effect of influenza incidence. The specific structure is as follows: Figure 2 shown.

[0047] (1) Use the multi-feature fusion component to fuse time series features of different dimensions to obtain a time series data representation that integrates multi-feature information. The structure of the multi-feature fusion component is as follows: Figure 3 As shown. The time series features of each dimension are encoded by its own LSTM feature encoder. At the same time, the time series features of all dimensions are encoded by a shared LSTM feature encoder with shared weights to obtain a coded sequence. Then, the output of the exclusive LSTM feature encoder corresponding to the time series features of different dimensions and the output of the shared LSTM feature encoder are added to obtain a time series data representation that integrates multivariate information. The LSTM feature encoder dedicated to time series features takes the time series features of a specific dimension as input and can learn the time series patterns in the time series features of a specific dimension. The shared LSTM feature encoder takes the time series features of all dimensions as input to simultaneously learn the correlation and complementarity of the time series features of all dimensions.

[0048] Multivariate time series data X of different scales D , X W , X M , X S After the multi-feature fusion component, the time series data representation Z at the corresponding time scale is obtained. D , Z W , Z M , Z S .

[0049] (2) The pyramid feature fusion network is used to fuse multi-scale time series data representations to obtain time series feature coding sequences at different scales.

[0050] Specifically, seasonal time series data characterizes Z S Through the seasonal LSTM encoding network, the seasonal time series feature encoding sequence L is obtained S . Monthly time series data characterization Z M and seasonal time series feature encoding sequence L S Combined as the input of the monthly LSTM encoding network, the monthly time series feature encoding sequence L is obtained M . Weekly time series data characterization Z W and monthly time series feature coding sequence L M Combined as the input of the weekly LSTM encoding network, the weekly time series feature encoding sequence L is obtained. W . Daily time series data characterization Z D and weekly time series feature coding sequence L W The combined data is used as the input of the daily LSTM encoding network to obtain the daily time series feature encoding sequence L D Among them, due to the seasonal time series feature encoding sequence L S The sequence length is less than the monthly time series data representation Z M The sequence length is L through linear interpolation upsampling technology. S Expanded to time series length and Z M Consistent upsampled seasonal time series feature encoding sequence Then, the monthly time series data is represented by Z M And upsampled seasonal time series feature encoding sequence After adding each element, it is used as the input of the monthly LSTM feature encoding network. M Expanded to time series length and Z W Consistent upsampled monthly time series feature encoding sequence L W Expanded to time series length and Z D Consistent upsampled weekly time series feature encoding sequence The pyramid feature fusion network fuses the time series data features at different scales to ensure that the time series feature information at each scale can be fully utilized and improve the quality of time series data representation.

[0051] (3) Using dynamic time warping distance, the similarity between temporal feature coding sequences at different scales is calculated, and the temporal feature coding sequences at different scales are optimized.

[0052] Multi-scale time series data fusion can capture and integrate information at different time scales, and improve the model's ability to understand and predict complex time series patterns. By fusing short-term and long-term time series features, the model can not only capture local detail changes, but also pay attention to global trends, thereby improving the ability to fully characterize time series. In addition, multi-scale fusion can also enhance the robustness of the model, effectively deal with noise, outliers, and changes at different time frequencies, making the model more robust and accurate. However, the multi-scale time series data fusion method is not effective, which may lead to the loss of fine-grained time series features or noise amplification, weakening the model's ability to capture short-term changes; at the same time, it may also lead to the inability to effectively extract coarse-grained features, ignoring long-term trends or global dynamics. In order to ensure effective multi-scale time series feature fusion, both fine-grained and coarse-grained time series features are fully extracted and retained, and the complementary role of multi-scale time series features is fully utilized. Consider using the consistency between the overall trends of time series data at different scales to optimize the encoding of multi-scale time series data.

[0053] Specifically, the sequence contrast loss is defined as the similarity of time series data at different time scales within a similar time range. The greater the similarity, the smaller the corresponding loss. The dynamic time warping distance is used to calculate the similarity between sequences. Dynamic time warping distance is a method for measuring the similarity between two time series, even if they have nonlinear changes on the time axis. The dynamic time warping distance uses dynamic programming to find the optimal alignment path between two sequences so that the distance between the aligned time series is minimized. Specifically, the dynamic time warping distance allows the time series to be "stretched" or "compressed" at certain positions to match similar patterns in another sequence. The sequence contrast loss calculation formula is as follows:

[0054] Loss s =-dtw(L S ,L M )-dtw(L M ,L W )-dtw(L W ,L D )

[0055] Among them, dtw(L S ,L M ) represents the seasonal time series feature coding sequence LS and monthly time series feature coding sequence L M The dynamic time warping distance between M ,L W ) represents the monthly time series feature coding sequence L M and weekly time series feature coding sequence L W The dynamic time warping distance between W ,L D ) represents the weekly time series feature coding sequence L W and daily time series feature coding sequence L D The dynamic time warping distance between them.

[0056] (4) Encode the daily time series feature sequence L with the finest time granularity D Used for influenza incidence prediction, supervised learning using mean square error loss.

[0057] The mean square error loss is used to measure the difference between the model prediction value and the true value. The formula is expressed as:

[0058]

[0059] Among them, n is the sequence length of the test time series data, y i is the true value of the sequence at the i-th time step, is the sequence prediction value at the i-th time step. When training the model, the goal is to minimize Loss through an optimization algorithm (such as gradient descent). MSE , so that the model's prediction results are as close to the true value as possible.

[0060] (5) Finally, the overall loss function is constructed based on the sequence contrast loss and the mean square error loss, and the model is trained by minimizing the overall loss:

[0061] L=Loss s +Loss MSE

[0062] The time series data prediction module uses the influenza incidence prediction model constructed by the multi-scale multivariate time series data modeling module to predict the influenza incidence at the next time point based on the input multi-scale multivariate time series data.

[0063] Based on the same inventive principle, this embodiment also provides a method for predicting influenza incidence based on multi-scale multivariate time series feature fusion, comprising the following steps:

[0064] S01, collects multivariate time series data on a daily basis, including meteorological information, influenza incidence, influenza type distribution ratio, gender ratio of the sick population, and age structure of the sick population.

[0065] S02, fill missing values ​​and process outliers in the multivariate time series data collected daily, and generate multivariate time series data X_W, X_M, and X_S with weekly, monthly, and quarterly scales based on the processed multivariate time series data X_D with daily scale.

[0066] S03, perform deep fusion and representation on multivariate time series data of different scales respectively to obtain time series data representations of different scales; use the pyramid feature fusion network to fuse the time series data representations of different scales to obtain time series feature coding sequences at different scales; optimize the time series feature coding sequences at different scales by calculating the similarity between the time series feature coding sequences at different scales.

[0067] S04, construct an influenza incidence prediction model, use the daily time series feature coding sequence with the finest time granularity as input, and train the prediction model.

[0068] S05, after the training is completed, the influenza incidence at the next time point is predicted based on the input multi-scale multivariate time series data.

[0069] The embodiments described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting influenza incidence based on multi-scale multivariate time series feature fusion, characterized in that: The following steps are involved: (1) Collect multivariate time series data on a daily basis, including meteorological information, influenza incidence, influenza type distribution ratio, gender ratio of the patient population, and age structure of the patient population; (2) Fill missing values ​​and process outliers in the multivariate time series data collected daily, and then calculate the time series data X based on the processed multivariate time series data X. D Generate multivariate time series data X with weekly, monthly, and quarterly scales W , X M , X S ; (3) Deeply fuse and represent multivariate time series data of different scales to obtain representations of time series data of different scales; The pyramid feature fusion network is used to fuse the time series data representations of different scales to obtain the time series feature coding sequences at different scales. By calculating the similarity between the temporal feature coding sequences at different scales, the temporal feature coding sequences at different scales are optimized; (4) Construct an influenza incidence prediction model and use the daily time series feature coding sequence with the finest time granularity as input to train the prediction model; (5) After training is completed, the influenza incidence at the next time point is predicted based on the input multi-scale multivariate time series data.

2. The influenza incidence prediction method based on multi-scale multivariate time series feature fusion according to claim 1 is characterized in that: In step (2), the missing values ​​of the multivariate time series data are filled using the linear interpolation method; the outliers of the multivariate time series data are identified using the isolation forest algorithm, and then the replacement values ​​of the outliers are calculated and replaced using the linear interpolation method.

3. The influenza incidence prediction method based on multi-scale multivariate time series feature fusion according to claim 1 is characterized in that: In step (3), the multivariate feature fusion component is used to perform deep fusion and characterization on multivariate time series data of different scales. The working process of the multivariate feature fusion component is as follows: For multivariate time series data of each scale, the time series features of each dimension are encoded by its exclusive LSTM feature encoder to obtain the corresponding exclusive coding sequence; at the same time, the time series features of all dimensions are encoded by a shared LSTM feature encoder with shared weights to obtain a shared coding sequence; then, the exclusive coding sequence and the shared coding sequence are added together to obtain the time series data representation that integrates multivariate information; Multivariate time series data X of different scales D , X W , X M , X S After the multi-feature fusion component, the time series data representation Z at the corresponding time scale is obtained. D , Z W , Z M , Z S .

4. The influenza incidence prediction method based on multi-scale multivariate time series feature fusion according to claim 3 is characterized in that: In step (3), the pyramid feature fusion network is used to fuse the time series data representations of different scales to obtain the time series feature coding sequences at different scales. The specific process is as follows: Seasonal time series data representation Z S Through the seasonal LSTM encoding network, the seasonal time series feature encoding sequence L is obtained S ; Monthly corresponding time series data representation Z M and seasonal time series feature encoding sequence L S Combined as the input of the monthly LSTM encoding network, the monthly time series feature encoding sequence L is obtained M ; The time series data corresponding to the week represents Z W and monthly time series feature coding sequence L M Combined as the input of the weekly LSTM encoding network, the weekly time series feature encoding sequence L is obtained. W ; The time series data corresponding to the day represents Z D and weekly time series feature coding sequence L W The combined data is used as the input of the daily LSTM encoding network to obtain the daily time series feature encoding sequence L D .

5. The influenza incidence prediction method based on multi-scale multivariate time series feature fusion according to claim 4 is characterized in that: Due to the seasonal time series feature encoding sequence L S The sequence length is less than the corresponding monthly time series data representation Z M The sequence length is L through linear interpolation upsampling technology. S Expanded to time series length and Z M Consistent upsampled seasonal time series feature encoding sequence Then, the monthly time series data is represented by Z M And upsampled seasonal time series feature encoding sequence After adding each element, it is used as the input of the monthly LSTM feature encoding network to obtain the monthly time series feature encoding sequence L M Similarly, we need to change L M Expanded to time series length and Z W Consistent upsampled monthly time series feature encoding sequence L W Expanded to time series length and Z D Consistent upsampled weekly time series feature encoding sequence 6. The influenza incidence prediction method based on multi-scale multivariate time series feature fusion according to claim 1 is characterized in that: In step (3), the similarity between time series feature coding sequences at different scales is calculated by dynamic time warping distance. Specifically, dynamic time warping distance allows the time series to be stretched or compressed at certain positions to match similar patterns in another sequence. The sequence contrast loss calculation formula is as follows: Loss s =-dtw(L S ,L M )-dtw(L M ,L W )-dtw(L W ,L D ) Among them, dtw(L S ,L M ) represents the seasonal time series feature coding sequence L S and monthly time series feature coding sequence L M The dynamic time warping distance between M ,L W ) represents the monthly time series feature coding sequence L M and weekly time series feature coding sequence L W The dynamic time warping distance between W ,L D ) represents the weekly time series feature coding sequence L W and daily time series feature coding sequence L D The dynamic time warping distance between them.

7. The influenza incidence prediction method based on multi-scale multivariate time series feature fusion according to claim 6 is characterized in that: In step (4), the daily time series feature encoding sequence L D As input, according to the sequence contrast loss Loss s And mean square error loss Loss MSE Construct the overall loss function L and train the model by minimizing the overall loss: L=Loss s +Loss MSE Among them, the mean square error loss Loss MSE It is used to measure the difference between the model prediction value and the true value.

8. A influenza incidence prediction system based on multi-scale multivariate time series feature fusion, characterized in that: include: The time series data collection module is used to collect multivariate time series data on a daily basis, including meteorological information, influenza incidence, influenza type distribution ratio, gender ratio of the patient population, and age structure of the patient population; The time series data preprocessing module is used to fill missing values ​​and process outliers in the multivariate time series data collected daily, and to calculate the time series data based on the processed multivariate time series data X D Generate multivariate time series data X with weekly, monthly, and quarterly scales W , X M , X S ; The multi-scale multivariate time series data modeling module is used to deeply fuse and characterize multivariate time series data of different scales to obtain time series data representations of different scales; The pyramid feature fusion network is used to fuse the time series data representations of different scales to obtain the time series feature coding sequences at different scales. By calculating the similarity between the temporal feature coding sequences at different scales, the temporal feature coding sequences at different scales are optimized; The time series data prediction module uses the model built by the multiscale multivariate time series data modeling module to predict the influenza incidence at the next time point based on the input multiscale multivariate time series data.

9. A influenza incidence prediction system based on multi-scale multivariate time series feature fusion, characterized in that: The method comprises a memory and one or more processors, wherein the memory stores executable codes, and when the one or more processors execute the executable codes, the method is used to implement the influenza incidence prediction method according to any one of claims 1 to 7.

Images