Infectious disease prediction method and system based on multi-time-sequence basic model fusion

Abstract

The invention discloses an infectious disease prediction method based on multi-time-sequence basic model fusion. The method comprises the following steps: acquiring time sequence data of an infectious disease and preprocessing the time sequence data; a plurality of time sequence basic models are selected for prediction and output; carrying out weighted splicing on output results of different time sequence basic models to obtain a fusion prediction combination yt; performing fine tuning processing on the fusion prediction combination yt to obtain a prediction model output; output is equal to KAN (yt); the weighted splicing part and the KAN represent the part of a forecast-head layer of the fine tuning model; and outputting the prediction model output to obtain a prediction result. According to the method, the capturing capability of a nonlinear time sequence mode in the prediction process can be improved, higher adaptability and accuracy are shown in an infectious disease prediction task, and the accuracy of a prediction result output by the prediction model output can be remarkably improved. The invention further provides an infectious disease prediction system based on multi-time-sequence basic model fusion.

Description

Technical Field

[0001] The present invention relates to a method for predicting infectious diseases, and in particular to a method and system for predicting infectious diseases based on the fusion of multiple time series basic models. Background Art

[0002] Against the backdrop of globalization and increasing population mobility, monitoring and early warning of infectious disease outbreaks and public health emergencies have become increasingly important. To effectively address this challenge, the Infectious Disease Monitoring and Early Warning Platform has emerged, aiming to comprehensively enhance the monitoring and early warning capabilities for major infectious disease outbreaks and public health emergencies.

[0003] The infectious disease prediction models used in existing infectious disease monitoring and early warning platforms have multiple limitations. First, in the early stages of a new or emerging infectious disease outbreak, the amount of available data samples is usually extremely limited. Traditional time series prediction models (such as ARIMA and Prophet) rely heavily on sufficient historical data sequences to construct robust predictions, making them difficult to effectively apply to such data-scarce scenarios. Second, although recently developed large-scale time series basic models (such as TimesFM and Moirai) have achieved zero-shot or small-shot prediction capabilities through pre-training on massive cross-domain time series data, the proportion of infectious disease-related time series data in the training corpus of such models is significantly low, resulting in insufficient capture of the dynamic characteristics of infectious diseases and difficulty adapting to the specific needs of epidemic prediction. In addition, existing model fusion methods generally adopt a static weight allocation strategy and fail to design a dynamic weight adjustment mechanism based on the spatiotemporal heterogeneity and sparsity characteristics of infectious disease data. At the same time, they lack a systematic quantitative assessment of the uncertainty of the prediction results, and cannot provide a reliable risk assessment basis for public health decision-making.

[0004] For example, Chinese patent document CN113658713A discloses a method, apparatus, device, and storage medium for predicting infectious trends, which fall within the field of artificial intelligence. The method comprises: obtaining historical infectious disease data; invoking at least one slave prediction model to perform predictions based on the historical infectious data, obtaining at least one prediction result for a target time period; invoking a master prediction model to perform predictions using the at least one prediction result and the historical infectious data as input features, obtaining a fused prediction result; and outputting the fused prediction result for the infectious trend for the target time period. This application can improve the accuracy of infectious trend predictions.

[0005] In this existing technology, the weight corresponding to each prediction result is determined by historical infectious disease data during the training process. Compared with the static weight allocation strategy with pre-defined weights, this existing technology can achieve dynamic adjustment of weight allocation to ensure the accuracy of the prediction.

[0006] Chinese patent document CN110136842A discloses an artificial intelligence technology and a method for predicting the onset of acute infectious diseases. The method includes: obtaining a time series data set that affects the onset of acute infectious diseases, and performing data processing on the time series data set according to a preset method to filter out principal component factor data that affects the onset of the acute infectious disease from the time series data set; based on the filtered principal component factor data that affects the onset of the acute infectious disease, combining the weighted average of multiple preset single prediction models, constructing a combined prediction model for the onset of acute infectious diseases; and using the calculation formula of the combined prediction model to automatically predict the outbreak period of the acute infectious disease. The present invention also proposes a device and a computer-readable storage medium. The present invention can have a good prediction effect on acute infectious diseases.

[0007] This prior art, in the process of constructing a combined prediction model for acute infectious disease incidence by combining the weighted average of multiple pre-set individual prediction models, calculates the sum of squared errors of the individual prediction models and assigns weights to the individual prediction models based on the principle of minimizing the overall sum of squared errors. This also addresses the low accuracy problem of static weight allocation strategies. However, all of the above prior art suffer from inaccurate prediction results. Summary of the Invention

[0008] In order to solve the technical problem of inaccurate prediction results in the prior art, the present invention provides an infectious disease prediction method based on the fusion of multiple time series basic models, comprising the following steps:

[0009] Collect and preprocess time series data of infectious diseases;

[0010] Select multiple time series basic models to make predictions and output based on preprocessed time series data;

[0011] The output results of different time series basic models are weighted and spliced to obtain the fusion prediction combination y t ;

[0012] For the fusion prediction combination y t Perform fine-tuning to obtain the prediction model output;

[0013] output=KAN(y t );

[0014] Among them, KAN represents a part of the forecast-head layer of the fine-tuning model;

[0015] The prediction model output outputs the prediction result.

[0016] Note that part of the forecast-head layer of the fine-tuning model in this solution refers to the fusion forecast combination yt The forecast-head layer is processed using the KAN neural network.

[0017] In the early stages of an infectious disease's spread, the number of infected people increases linearly. However, as the number of infected people increases, the spread accelerates, the number of infected people grows rapidly, and explosive growth occurs. Therefore, infectious disease data exhibits nonlinear time series patterns, making it difficult for models to capture the true patterns of the data. This makes it particularly difficult for models to capture the true patterns of the data. In long-term forecasts, the complexity of nonlinear dynamics leads to a significant decrease in the accuracy of forecast results. In this solution, after weighted concatenation of the outputs of different time series base models, the weighted concatenation is processed through a KAN neural network to produce the forecast model output. The introduced KAN neural network improves the ability to capture nonlinear time series patterns during the forecast process, making the overall framework of the forecast model output more adaptable and accurate for infectious disease forecasting tasks, significantly improving the accuracy of the forecast model output.

[0018] Preferably, when fine-tuning, weighted splicing is performed on the output results of the forecast-head layers of different time series basic models.

[0019] y t =Concat(w i ·f i (X t )), w i ≥0;

[0020] where f i (X t ) represents different time series basic models, N represents the number of time series basic models, i represents the i-th time series basic model, w i Represents the weight coefficient of the time series basic model; the weight coefficient is dynamically updated through backpropagation.

[0021] In this scheme, after dynamically updating the weight coefficients of the outputs of different time series basic models through back propagation, the contribution weights of each time series basic model at different prediction time scales can be dynamically adjusted, and finally a hybrid weighted prediction model output is generated, realizing the optimal combination of fusion predictions of multiple different time series basic models, which can effectively suppress the abnormal prediction behavior of a single model, thereby improving the accuracy of the output prediction results.

[0022] Preferably, during fine-tuning, for different time series base models, the embedding layer and decoder or encoder layer are frozen, and only the parameters of the forecast-head layer are fine-tuned. This adjustment method preserves the general time series patterns captured by the pre-trained model while also optimizing specifically for the characteristics of infectious disease transmission. This significantly reduces computational resource consumption, allowing the prediction model output to efficiently capture the specific prediction capabilities of infectious diseases while maintaining the original time series representation capabilities, achieving an optimal balance between computational efficiency and model performance.

[0023] Preferably, the preprocessed dataset is divided into a training set and a validation set, and a prediction interval of the prediction model output at a given confidence level is constructed based on the dataset and the validation set.

[0024] In this scheme, by constructing prediction intervals covering a specified confidence level and combining the prediction interval width with the actual coverage assessment, we can provide decision makers with an accurate and quantitative measure of prediction uncertainty. The actual coverage refers to the proportion of true values in the validation set that fall within the prediction interval.

[0025] Preferably, the construction of the prediction interval includes the following steps:

[0026] Based on the data set and validation set results, calculate the error of the prediction model output on each validation set, assuming that the error E has a mean of μ and a variance of σ 2 The prior normal distribution of :

[0027] E~N(μ,σ 2 );

[0028] Calculate the error and the corresponding mean and variance based on the validation set portion of the dataset:

[0029]

[0030] in Represents the error of the time series data validation set, y true Represents the true value of the predicted part of the dataset, y pred represents the predicted value of the predicted part, and Represents the mean and variance of the prediction result error, and n represents the number of data points in the validation set;

[0031] Assume that Y = (y1, y2, ..., y n ) is the true value of the predicted part of a new time series dataset, is the predicted value, and the mean and variance of the corresponding error are calculated, assuming that it obeys the posterior normal distribution:

[0032]

[0033] Use Bayesian inference to update the mean and variance of the error:

[0034]

[0035] where μ post and Represents the mean and variance of the updated error, and the confidence interval of the error at the corresponding level is obtained by giving a given confidence level:

[0036] [μ post -z α / 2 ·σ post , μ post +z α / 2 ·σ post ];

[0037] The prediction interval is obtained as:

[0038]

[0039] This solution can construct a prediction interval with theoretical guarantees, and the prediction interval will be dynamically adjusted as the data is updated.

[0040] Preferably, the preprocessing of the collected time series data includes interpolation of missing values of the time series data.

[0041] In this solution, the missing value interpolation method can ensure the continuity and integrity of time series data.

[0042] Preferably, the preprocessing further includes detecting the stationarity of the time series data and transforming the detected non-stationary time series data. This solution can meet the stationary data requirements of the modeling requirements.

[0043] Preferably, the preprocessing also includes normalizing the time series data. The purpose of normalization in this solution is to eliminate the dimensional differences of the time series data of different infectious diseases, so that the impact of all time series data on the prediction model is more balanced.

[0044] Preferably, the preprocessed time series data is sampled using an upsampling method, and the sampled time series data is masked. Because the time series data of different infectious diseases vary in length, this solution ensures that no redundant information is passed into the prediction model, thereby affecting the accuracy of the prediction results.

[0045] In a second aspect, the present invention provides an infectious disease prediction system based on the fusion of multiple time series basic models, which adopts the above-mentioned infectious disease prediction method based on the fusion of multiple time series basic models for prediction.

[0046] The present invention has the following beneficial effects:

[0047] 1. The present invention can enhance the ability to capture nonlinear time series patterns during the prediction process, thereby making the overall framework of the prediction model output more adaptable and accurate in infectious disease prediction tasks, and can significantly improve the accuracy of the prediction results output by the prediction model.

[0048] 2. The present invention can perform targeted fine-tuning of prediction model parameters, significantly improving the prediction model's ability to capture the temporal and spatial transmission patterns of infectious diseases and its domain adaptability.

[0049] 3. The present invention can automatically adjust the contribution weights of different time series basic models and effectively improve the prediction stability and robustness in small sample scenarios by integrating multi-source prediction results.

[0050] 4. The present invention can also generate prediction intervals with credibility levels, and combined with the knowledge constraints of epidemiological experts, provide accurate risk assessment indicators for epidemic prevention and control decisions at all levels, making the prediction results interpretable and operational, and effectively supporting scientific decision-making in public health emergency responses. DETAILED DESCRIPTION

[0051] The following is further described in detail through specific implementation methods:

[0052] 1. Definition

[0053] Time series basic model: A model established through curve fitting and parameter estimation based on the time series data obtained from system observations.

[0054] Time series data: refers to time series data, which is a series of observations recorded in sequence with time as the index, used to describe the state or law of the phenomenon changing over time.

[0055] KAN: refers to Kolmogorov–Arnold Networks, a type of neural network.

[0056] Forecast-head layer: refers to the forecast output layer.

[0057] Back Propagation: The Back Propagation (BP) algorithm is one of the most core and commonly used optimization algorithms in deep learning, and is widely used in the training process of neural networks. It updates the parameters by calculating the gradient of the loss function with respect to the network parameters, thereby minimizing the loss function and improving the model's prediction accuracy.

[0058] Embedding layer: embedding layer.

[0059] Decoder layer: decoding layer.

[0060] encoder layer: encoding layer.

[0061] Confidence level: It is an indicator to measure the reliability of the confidence interval, which refers to the degree to which a specific individual believes in the truth of a specific proposition. In other words, probability is a measure of the rationality of personal beliefs.

[0062] A confidence interval is an estimate of a population parameter constructed from a sample statistic, used to quantify the probability that the true value of the parameter falls within that interval. A 95% confidence interval means that there is a 95% probability that the true value of the population parameter falls within the interval calculated from the sample data.

[0063] A priori normal distribution: This refers to the prior distribution of the normal distribution. Based on general experience, a priori normal distribution is the distribution that random variables should satisfy. It is a type of probability distribution, in contrast to the "posterior distribution," and is independent of experimental results or random sampling. It reflects the distribution obtained before the statistical experiment is conducted based on knowledge of other relevant parameters, θ.

[0064] Posterior normal distribution: A type of probability distribution, as opposed to a prior distribution. It refers to the posterior distribution of a normal distribution, where the distribution of a random variable, modified by the current training data, is more consistent with the current data than the prior distribution.

[0065] Dimensional discrepancy: This refers to the situation where there are significant differences in the measurement scales or units of different features in the same set of data.

[0066] Upsampling: refers to the insertion of zeros or other values into a signal or data sequence to increase its sampling rate, typically to restore lost detail or prepare the data for subsequent processing steps.

[0067] Mask: It is a string of binary codes that performs a bitwise AND operation on the target field to mask the current input bit.

[0068] Mask operation: It is a common operation in deep learning, which is equivalent to covering the original tensor with a mask to shield or select some specific elements.

[0069] 2. The embodiment is basically as follows: The infectious disease prediction method based on the fusion of multiple time series basic models includes the following steps:

[0070] Collect time series data of infectious diseases and perform preprocessing. In this embodiment, multi-level data preprocessing is performed on the time series data. First, missing values are interpolated for the time series data. The missing values are filled using the linear interpolation method:

[0071]

[0072] in Represents the time series data values at time t0, t, and t1.

[0073] At the same time, the time series is smoothed to reduce the impact of random fluctuations on the stability of the forecast:

[0074] s t =α·x t +(1-α)·s t-1 ;

[0075] where s t , s t-1 Represents the smoothed value of the current moment and the previous moment, x t It represents the original value of the time series at the current moment, and α represents the smoothing factor.

[0076] Then, the min-max normalization technique is applied to map various indicators to the interval [0, 1] to eliminate the dimensional differences of different infectious disease time series data:

[0077]

[0078] where x min , x max , x′ represents the minimum value, maximum value and normalized corresponding value of the time series data.

[0079] Multiple time series basic models are selected for prediction and output based on preprocessed time series data. These include TimesFM, Moirai, and Chronos, which have been pre-trained on massive amounts of cross-domain time series data to achieve robust general representation capabilities. Upsampling is used to pad the time series data, followed by masking of the sampled data. Parameters of the embedding layer and decoder or encoder layer of each time series basic model are frozen.

[0080] The output results of different time series basic models are weighted and spliced to obtain the fusion prediction combination y t In this embodiment, a time sliding window strategy is used to divide the collected infectious disease time series data into a training set and a validation set. The validation set ratio is dynamically adjusted to adapt to the length of different infectious disease data sets. For the forecast-head layer output results of different time series basic models, and according to the mean absolute error of the time series basic model, the weight coefficient of each time series basic model output is dynamically updated through back propagation, and the forecast-head layer output results of different time series basic models are weighted and spliced to obtain the fusion prediction combination y t :

[0081] y t =Concat(w i ·f i (X t )), w i ≥0;

[0082] where f i (X t ) represents different time series basic models, N represents the number of time series basic models, i represents the i-th time series basic model, w i Represents the weight coefficient of the time series basic model.

[0083] For the fusion prediction combination y t Perform fine-tuning to obtain the prediction model output;

[0084] output=KAN(y t );

[0085] Among them, KAN represents a part of the forecast-head layer of the fine-tuning model.

[0086] The prediction model output outputs the prediction result.

[0087] The preprocessed dataset is divided into a training set and a validation set, and a prediction interval of the prediction model output at a given confidence level is constructed based on the dataset and the validation set. In this embodiment, the construction of the prediction interval includes the following steps:

[0088] Based on the data set and validation set results, calculate the error of the prediction model output on each validation set, assuming that the error E has a mean of μ and a variance of σ 2 The prior normal distribution of :

[0089] E~N(μ,σ 2 );

[0090] Calculate the error and the corresponding mean and variance based on the validation set portion of the dataset:

[0091]

[0092] in Represents the error of the time series data validation set, y true Represents the true value of the predicted part of the dataset, y pred represents the predicted value of the predicted part, and Represents the mean and variance of the prediction result error, and n represents the number of data points in the validation set;

[0093] Assume that Y = (y1, y2, ..., y n ) is the true value of the predicted part of a new time series dataset, is the predicted value, and the mean and variance of the corresponding error are calculated, assuming that it obeys the posterior normal distribution:

[0094]

[0095] Use Bayesian inference to update the mean and variance of the error:

[0096]

[0097] where μ post and Represents the mean and variance of the updated error, and the confidence interval of the error at the corresponding level is obtained by giving a given confidence level:

[0098] [μ post -z α / 2 ·σ post , μ post +z α / 2 ·σ post ];

[0099] The prediction interval is obtained as:

[0100]

[0101] Based on the above prediction method, this embodiment also discloses a prediction system for predicting infectious diseases based on a method for predicting infectious diseases that integrates multiple time series basic models.

[0102] The specific implementation process is as follows: 1. Time series data collection: Case data for various diseases from multiple countries around the world are collected. In this example, the diseases refer to infectious diseases, which include various pathogens such as dengue fever, influenza, and Mycoplasma pneumonia. A total of 289 infectious disease data points are collected, with a total number of time points exceeding 100,000. This example uses the number of influenza cases in China in 2023 as an example.

[0103] 2. Time series data preprocessing: First, implement a data cleaning strategy to address the missing value problem. For time points where there are no observations in the time series, use linear interpolation to fill in the missing data to ensure the continuity and integrity of the time series.

[0104] Subsequently, the stationary characteristics of the time series are evaluated by ADF (Augmented Dickey-Fuller test). For the detected non-stationary series, exponential smoothing technology is used to transform them to obtain stationary data that meets the modeling requirements.

[0105] Finally, the min-max normalization method is applied to standardize the processed data to construct a high-quality dataset suitable for model fine-tuning, thereby ensuring the accuracy and reliability of subsequent analysis.

[0106] In this example, the number of influenza cases in China in 2023 is pre-processed as follows:

[0107] [0.0031, 0.0571, 0.9046, 0.4073, 0.0511, 0.0152, 0.0112, 0.0140, 0.0404, 0.0823, 0.4525, 1.000].

[0108] 3. Model fine-tuning: Implement upsampling on the time series data, and then perform masking on the sampled data.

[0109] The results of upsampling the preprocessed number of influenza cases in China in 2023 are:

[0110] [0.2532, 0.2532, 0.2532, 0.2532, 0.0031, 0.0571, 0.9046, 0.4073, 0.0511, 0.0152, 0.0112, 0.0140, 0.0404, 0.0823, 0.4525, 1.000];

[0111] The result of the mask operation is:

[0112] [0,0,0,0,1,1,1,1,1,1,1,1,1,1,1,1,1].

[0113] For various time series basic models, the parameters of the embedding layer and the decoder or encoder layer are frozen, and only the parameters of the forecast-head layer are fine-tuned.

[0114] 4. Model fusion: In this embodiment, three basic time series models are used: TimesFM, Moirai and Chronos.

[0115] The outputs of the encoder layer or decoder layer of the three time series basic models are weighted and concatenated:

[0116] out1 = timesfm(X,X_mask);

[0117] out2 = moirai(X,X_mask);

[0118] out3 = chronos(X,X_mask);

[0119] out4=Concat(w1*out1,w2*out2,w3*out3);

[0120] out5 = KAN(out4);

[0121] where X, X mask represents the upsampling result and the mask result, out1, out2, and out3 represent the embedding operation results of TimesFM, Moirai, and Chronos respectively, out4 represents the result of weighted concatenation of the embedding operation results of the three models, and out5 represents the result obtained by processing the concatenated result through a single-layer KAN neural network.

[0122] 5. Construction of Prediction Intervals: In this example, the mean and variance of the prior normal distribution of the error calculated based on the validation set data in the fine-tuning dataset are 0.1381 and 0.1094, respectively. The mean and variance of the 23-year influenza data in China are 0.1657 and 0.1250, respectively. Assuming a 95% confidence level, then:

[0123]

[0124] Confidence interval of error at the 95% level:

[0125] [μ post -z α / 2 ·σ post ,μ post +z α / 2 ·σ post ]=[-0.0278,0.3542].

[0126] From the confidence interval, we can see that there is a 95% probability that the error of the prediction result is between -0.0278 and 0.3542.

[0127] The MAE and RMSE of various time series basic models are compared based on the validation set. The comparison results are shown in Table 1.

[0128] Table 1 Comparison results

[0129] Model AVER-MAE AVER-RMSE Our-model (prediction model of this example) 0.076953893 0.124000808 Chronos-small 0.078090275 0.126453582 Moirai-base 0.078096984 0.127283603 Chronos-mini 0.078534808 0.125982408 Chronos-base 0.078717377 0.127264008 TimesFM-200M 0.078925416 0.124417814 Moirai-large 0.079637555 0.129406353 Moirai-small 0.080038062 0.129033962

[0130] Based on the comparison of MAE and RMSE on the validation set, the smaller the two values, the better the effect. As can be seen from Table 1, the prediction model in this embodiment performs best in both MAE and RMSE.

[0131] The above is only an embodiment of the present invention. Common knowledge such as the known specific structures and characteristics in the scheme is not described in detail here. Ordinary technicians in the field are aware of all common technical knowledge in the technical field of the invention before the application date or priority date, can obtain all existing technologies in the field, and have the ability to apply conventional experimental means before that date. Ordinary technicians in the field can improve and implement this scheme in combination with their own abilities under the inspiration given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention. These will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. The infectious disease prediction method based on the fusion of multiple time series basic models includes the following steps: Collect and preprocess time series data of infectious diseases; Select multiple time series basic models to make predictions and output based on preprocessed time series data; The output results of different time series basic models are weighted and spliced to obtain the fusion prediction combination y t ; It is characterized by: the fusion prediction combination y t Perform fine-tuning to obtain the prediction model output; output=CAN(y t ); Among them, KAN represents a part of the forecast-head layer of the fine-tuning model; The prediction model output outputs the prediction result.

2. The infectious disease prediction method based on the fusion of multiple time series basic models according to claim 1 is characterized by: When fine-tuning, the forecast-head layer output results of different time series basic models are weighted and spliced. y t =Concat(w i ·f i (X t )), w i ≥0; where f i (X t ) represents different time series basic models, N represents the number of time series basic models, i represents the i-th time series basic model, w i Represents the weight coefficient of the time series basic model; The weight coefficients are dynamically updated through back-propagation.

3. The infectious disease prediction method based on the fusion of multiple time series basic models according to claim 2 is characterized by: During fine-tuning, for different time series base models, the embedding layer and decoder layer or encoder layer are frozen, and only the parameters of the forecast-head layer are fine-tuned.

4. The infectious disease prediction method based on the fusion of multiple time series basic models according to any one of claims 1 to 3, characterized in that: The preprocessed dataset is divided into a training set and a validation set, and the prediction interval of the prediction model output at a given confidence level is constructed based on the dataset and the validation set.

5. The infectious disease prediction method based on the fusion of multiple time series basic models according to claim 4 is characterized by: The construction of the prediction interval includes the following steps: Based on the data set and validation set results, calculate the error of the prediction model output on each validation set, assuming that the error E has a mean of μ and a variance of σ 2 The prior normal distribution of : E~N(μ,σ 2 ); Calculate the error and the corresponding mean and variance based on the validation set portion of the dataset: in Represents the error of the time series data validation set, y true Represents the true value of the predicted part of the dataset, y pred represents the predicted value of the predicted part, and Represents the mean and variance of the prediction result error, and n represents the number of data points in the validation set; Assume that Y = (y1, y2, ..., y n ) is the true value of the predicted part of a new time series dataset, is the predicted value, and the mean and variance of the corresponding error are calculated, assuming that it obeys the posterior normal distribution: Use Bayesian inference to update the mean and variance of the error: where μ post and Represents the mean and variance of the updated error, and the confidence interval of the error at the corresponding level is obtained by giving a given confidence level: [m post -z α / 2 ·s post ,m post +z α / 2 ·s post ]; The prediction interval is obtained as:

6. The infectious disease prediction method based on the fusion of multiple time series basic models according to claim 5 is characterized by: The preprocessing of the collected time series data includes interpolation of missing values in the time series data.

7. The infectious disease prediction method based on the fusion of multiple time series basic models according to claim 6 is characterized by: The preprocessing also includes detecting the stationarity of the time series data and performing transformation processing on the detected non-stationary time series data.

8. The infectious disease prediction method based on the fusion of multiple time series basic models according to any one of claims 5 to 7, characterized in that: The preprocessing also includes standardizing the time series data.

9. The infectious disease prediction method based on the fusion of multiple time series basic models according to claim 8 is characterized by: The pre-processed time series data is sampled using an upsampling method, and a mask operation is performed on the sampled time series data.

10. An infectious disease prediction system based on the fusion of multiple time series basic models, characterized by: Prediction is performed using the infectious disease prediction method based on the fusion of multiple time series basic models as described in any one of claims 1 to 9.

Images