An influenza seasonal trend prediction method and system based on an ARIMA-LSTM hybrid model and a storage medium

Abstract

The application belongs to the technical field of public health and artificial intelligence, and particularly relates to an influenza seasonal trend prediction method based on an ARIMA-LSTM hybrid model, which comprises: preprocessing and stationary conversion of influenza time series data; decomposition of the sequence into a trend item, a seasonal item and a holiday item by using a Prophet model, directional input of the trend item into an ARIMA model, directional input of the seasonal and holiday items into an LSTM model, and realization of feature decoupling; division of the data set by using a year-by-year rolling strategy, parallel training of the double models; dynamic updating of the fusion weight by using a batch offline optimization algorithm based on the error of the verification set, realization of self-adaptive weighting; construction of a dynamic prior by combining historical errors, updating of a posterior distribution by using Bayesian inference to generate a prediction interval; and triggering of an early warning by comparing an actual value with an upper limit of the interval, so that the application effectively overcomes the limitations of a single model and the rigidity of a static weight, and significantly improves the influenza prediction accuracy and uncertainty quantification capability under a complex epidemic background.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of public health and artificial intelligence, specifically relating to a method, system, and storage medium for predicting the seasonal trend of influenza based on an ARIMA-LSTM hybrid model. Background Technology

[0002] Influenza, a respiratory infectious disease with a significant winter-spring peak, exhibits complex temporal characteristics in its transmission dynamics, often displaying a mixture of linear drift and nonlinear fluctuations. For a long time, the industry has primarily relied on statistical time series models and machine learning algorithms to predict epidemic trends. Among these, the autoregressive integral moving average model, with its robust ability to capture linear autocorrelation structures, is widely used for long-term trend fitting; while long short-term memory networks, relying on gating mechanisms, demonstrate unique advantages in characterizing nonlinear temporal dependencies and sudden fluctuations. With the digital upgrade of public health monitoring systems, single models are gradually evolving towards multi-model fusion, aiming to improve the robustness and foresight of predictions in complex epidemic contexts.

[0003] While existing prediction technologies have achieved some success within regular epidemic cycles, several technical bottlenecks remain to be overcome when facing sudden public health events and the intertwining of multi-source time-series signals. First, single models have inherent limitations in feature capture. Traditional statistical models, based on the assumption of linear stationarity, struggle to effectively respond to nonlinear jumps caused by interventions or environmental abrupt changes. While deep learning models possess nonlinear fitting capabilities, they are prone to overfitting in scenarios with non-stationary sequences or limited sample sizes, leading to drift in long-term trend projections. Second, existing fusion strategies lack the ability to dynamically adapt to changes in epidemic patterns. Although some publicly available solutions attempt to integrate results using multi-model weighted averaging or principal component analysis, their weight allocation largely relies on the static principle of minimizing the sum of squared historical errors, failing to adaptively adjust to the seasonal patterns reconstructed by flu peaks, troughs, or sudden public health events. Third, feature input mechanisms suffer from redundant interference. When attempting to introduce external time series decomposition terms to enhance model representation, existing techniques generally employ full feature input or random allocation strategies, which cause gradient interference between trend components and periodic components within the model. This not only wastes model capacity but also weakens the sensitivity to key fluctuation features.

[0004] The aforementioned technical deficiencies directly lead to a sharp decline in the accuracy of existing prediction tools when dealing with fundamental changes in influenza seasonality caused by large-scale non-pharmaceutical interventions such as the COVID-19 pandemic. How to effectively isolate periodic noise interference while ensuring stable tracking of linear trends, and how to construct an adaptive fusion architecture that can automatically evolve weights according to the epidemic stage, have become urgent technical challenges for those skilled in the art. Therefore, there is a pressing need for a hybrid prediction method that can achieve feature-oriented decoupling based on temporal decomposition, support rolling validation and batch offline dynamic weighting, and possess uncertainty quantification capabilities. This would compensate for the cognitive blind spots of single models and improve the accuracy and engineering feasibility of influenza seasonality trend prediction in the context of complex public health events. Summary of the Invention

[0005] To address the aforementioned shortcomings in existing technologies, this invention provides a method, system, and storage medium for predicting seasonal influenza trends based on an ARIMA-LSTM hybrid model. This addresses the problems of single prediction models struggling to balance linear long-term trends with nonlinear sudden fluctuations, existing fusion strategies using static weights failing to adapt to the characteristics of different stages of influenza epidemics, and the tendency for multi-source temporal feature coupling inputs to lead to model interference and overfitting.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] One of the objectives of this invention is to provide a method for predicting the seasonal trend of influenza based on an ARIMA-LSTM hybrid model. By decoupling time-series decomposition features, constructing dual models in parallel, and batch offline dynamic weighted fusion, combined with Bayesian dynamic prior interval construction, the method achieves a steady improvement in prediction accuracy and a scientific quantification of prediction uncertainty at different epidemic stages.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows:

[0009] A method for predicting the seasonal trend of influenza based on an ARIMA-LSTM hybrid model includes the following steps:

[0010] Acquire and preprocess time-series influenza surveillance data for the target region;

[0011] The preprocessed sequence was decomposed using a time-series decomposition algorithm to obtain trend, seasonal, and holiday effect components.

[0012] The trend term is then directed into the linear prediction sub-model as an external regression variable, while the seasonal term and the holiday effect term are directed into the nonlinear prediction sub-model as specific features.

[0013] A rolling validation strategy is adopted to divide the training set and the validation set;

[0014] Based on the training set, construct the linear prediction sub-model and the nonlinear prediction sub-model respectively, and output their respective prediction values;

[0015] Based on the validation set, the fusion weights of the linear prediction sub-model and the nonlinear prediction sub-model are dynamically updated using a batch offline optimization algorithm, and the weighted fusion prediction value is calculated.

[0016] A dynamic prior distribution is constructed based on the prediction error statistics of the historical validation set. The prediction error likelihood function of the current validation set is combined with Bayesian inference to update the posterior distribution and construct the prediction interval.

[0017] The actual monitored value is compared with the upper limit of the predicted range, and an early warning signal is triggered when the actual value exceeds the upper limit.

[0018] To further optimize the design, the preprocessing includes:

[0019] The positive rate of influenza virus nucleic acid testing was calculated as time series data; missing time points were filled using linear interpolation; the extended Dickey-Fuller test was used to assess the stationarity of the sequence, and if the sequence was non-stationary, difference transformation was performed; and the time series data were mapped to a preset interval using min-max normalization.

[0020] Further design optimization involves inputting the trend term into the linear prediction sub-model as an external regression variable, and inputting the seasonal term and the holiday effect term into the nonlinear prediction sub-model as specific features. Specifically, this includes:

[0021] The influenza positivity rate sequence was decomposed into a superposition of trend term, annual cycle seasonal term, and public holiday effect term using the Prophet model;

[0022] The linear prediction sub-model only accepts the trend term as an external regression variable to focus on fitting the linear drift component;

[0023] The nonlinear prediction sub-model only accepts the seasonal and holiday effect terms as input features to centrally model periodic fluctuations and sudden holiday effects.

[0024] Further design optimization involves adopting a rolling annual validation strategy to divide the training and validation sets, specifically including:

[0025] Using data from year m to year n as the training set, predict the influenza positivity rate in year (n+1), where n is a preset sequence of years, and generate multiple independent validation windows to evaluate the model's predictive performance at different stages of the epidemic.

[0026] Further optimization of the design: the linear prediction sub-model is an ARIMA model. The optimal autoregression order, difference order, and moving average order are determined through grid search. The trend term is used as an external regression variable input to output a linear prediction value.

[0027] Further optimization of the design: the nonlinear prediction sub-model is an LSTM model. Its input layer is composed of the standardized influenza positivity rate, the Fourier feature of the seasonal term, and the holiday effect indicator. The hidden layer contains multiple LSTM units and a Dropout layer. The output layer is a fully connected layer to output nonlinear prediction values.

[0028] Further design optimization involves using a batch offline optimization algorithm to dynamically update the fusion weights, specifically including:

[0029] Initialize the weights of the linear prediction sub-model and the nonlinear prediction sub-model, such that the sum of the weights is 1 and each weight is non-negative;

[0030] Define a weighted fusion prediction value and calculate the mean absolute error on the validation set as the objective function;

[0031] The weights are iteratively updated using gradient descent. The update process includes a subgradient term of the sign function and an L2 regularization penalty term. An early stopping mechanism and a learning rate search strategy are set to prevent extreme weight allocation.

[0032] Further design optimization includes the following specific steps for constructing the prediction interval:

[0033] The prediction error sequence is tested for normality. If it follows a normal distribution, the prediction error statistics of the historical verification window are taken as the Gaussian prior distribution, and the prediction error of the current verification window is taken as the Gaussian likelihood function. The mean and variance of the posterior distribution are calculated under the conjugate condition.

[0034] An exponentially weighted moving average mechanism is used to dynamically update the prior distribution parameters, enabling the prior distribution to adaptively adjust with changes in popular patterns. At a given confidence level, the prediction interval is calculated based on the posterior distribution parameters.

[0035] The second objective of this invention is to provide an influenza seasonality trend prediction system based on an ARIMA-LSTM hybrid model, for running the aforementioned influenza seasonality trend prediction method based on an ARIMA-LSTM hybrid model on a computer.

[0036] To achieve the above objectives, the technical solution of the present invention is as follows:

[0037] A seasonal influenza trend prediction system based on an ARIMA-LSTM hybrid model includes:

[0038] The data acquisition and preprocessing module is used to acquire and preprocess time-series influenza surveillance data from the target area.

[0039] The feature decoupling input module is used to decompose the preprocessed sequence using a time series decomposition algorithm to obtain trend terms, seasonal terms and holiday effect terms. The trend terms are then directed to the linear prediction sub-model as external regression variables, and the seasonal terms and holiday effect terms are directed to the nonlinear prediction sub-model as exclusive features.

[0040] The rolling validation partitioning module is used to partition the training set and validation set using a yearly rolling validation strategy.

[0041] A dual-model construction module is used to construct the linear prediction sub-model and the nonlinear prediction sub-model based on the training set, and output their respective prediction values.

[0042] An adaptive fusion module is used to dynamically update the fusion weights of the linear prediction sub-model and the nonlinear prediction sub-model based on the validation set using a batch offline optimization algorithm, and to calculate the weighted fusion prediction value.

[0043] The interval construction module is used to construct a dynamic prior distribution based on the prediction error statistics of the historical validation set, and combined with the prediction error likelihood function of the current validation set, update the posterior distribution through Bayesian inference and construct the prediction interval.

[0044] The early warning output module is used to compare the actual monitored value with the upper limit of the prediction range, and trigger an early warning signal when the actual value exceeds the upper limit.

[0045] The third objective of this invention is to provide a computer-readable storage medium for storing program instructions that are executed on a computer to implement a method and optimization design for predicting the seasonal trend of influenza based on an ARIMA-LSTM hybrid model.

[0046] To achieve the above objectives, the technical solution of the present invention is as follows:

[0047] A computer-readable storage medium storing program instructions that, when executed by a processor, are used to implement an influenza seasonal trend prediction method and optimization design based on an ARIMA-LSTM hybrid model, as described in one of the objectives of this invention.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] This invention achieves significant technical results in the field of influenza prediction through architectural innovation and algorithm-level optimization. First, the feature decoupling and directional input mechanism fundamentally resolves the modeling bottleneck caused by the mutual interference between linear trends and periodic signals. This allows the linear sub-model to focus on macroscopic drift, while the nonlinear sub-model focuses on microscopic fluctuations. The complementary advantages of the two models result in fusion prediction accuracy superior to a single model across multiple epidemic stages. Second, the batch offline dynamic weighting strategy abandons traditional static weight allocation, enabling the model to keenly perceive the evolution of epidemic characteristics at different stages before, during, and after the epidemic, automatically adjusting the contribution ratio of linear and nonlinear components, significantly improving prediction robustness. Third, the prediction interval construction method based on Bayesian inference and exponentially weighted moving average not only provides a statistically guaranteed measure of uncertainty but also achieves adaptive scaling of the interval width, providing public health institutions with a scientific and quantifiable decision-making basis for formulating tiered prevention and control strategies and allocating medical resources. Attached Figure Description

[0050] Figure 1 A flowchart illustrating an influenza seasonal trend prediction method based on an ARIMA-LSTM hybrid model is provided in this embodiment of the invention.

[0051] Figure 2 This invention provides a method for predicting the seasonal trend of influenza based on an ARIMA-LSTM hybrid model, specifically a method for predicting the seasonal trend of influenza cases based on the Prophet model (2011-2019).

[0052] Figure 3 This invention provides a method for predicting the seasonal trend of influenza based on an ARIMA-LSTM hybrid model, specifically a method for predicting the trend and seasonality of influenza cases based on the Prophet model (2020-2022).

[0053] Figure 4 This invention provides a method for predicting the seasonal trend of influenza based on an ARIMA-LSTM hybrid model, specifically a method for predicting the seasonal trend of influenza cases based on the Prophet model (2023-2024).

[0054] Figure 5 This is a comparison chart of the predictions and actual values ​​of the ARIMA sub-model in 2017-2018.

[0055] Figure 6 This is a comparison chart of the predictions and actual values ​​of the ARIMA sub-model in 2018-2019.

[0056] Figure 7 This is a comparison chart of the predictions and actual values ​​of the ARIMA sub-model in 2019-2020.

[0057] Figure 8 This is a comparison chart of the ARIMA sub-model's predictions and actual values ​​for 2020-2021.

[0058] Figure 9 This is a comparison chart of the ARIMA sub-model's predictions and actual values ​​for 2021-2022.

[0059] Figure 10 This is a comparison chart of the ARIMA sub-model's predictions and actual values ​​for 2022-2023.

[0060] Figure 11 This is a comparison chart of the ARIMA sub-model's predictions and actual values ​​for 2023-2024.

[0061] Figure 12 This is a comparison chart of the prediction results and actual values ​​of the LSTM sub-model in 2017-2018.

[0062] Figure 13 This is a comparison chart of the prediction results and actual values ​​of the LSTM sub-model in 2018-2019.

[0063] Figure 14 This is a comparison chart of the prediction results and actual values ​​of the LSTM sub-model in 2019-2020.

[0064] Figure 15 This is a comparison chart of the LSTM sub-model's predictions and actual values ​​for 2020-2021.

[0065] Figure 16 This is a comparison chart of the LSTM sub-model's predictions and actual values ​​for 2021-2022.

[0066] Figure 17 This is a comparison chart of the LSTM sub-model's predictions and actual values ​​for 2022-2023.

[0067] Figure 18 This is a comparison chart of the LSTM sub-model's predictions and actual values ​​for 2023-2024.

[0068] Figure 19 This is a comparison chart of the prediction results and actual values ​​of the ARIMA-LSTM hybrid model in 2017-2018.

[0069] Figure 20 This is a comparison chart of the prediction results and actual values ​​of the ARIMA-LSTM hybrid model in 2018-2019.

[0070] Figure 21 This is a comparison chart of the prediction results and actual values ​​of the ARIMA-LSTM hybrid model in 2019-2020.

[0071] Figure 22 This is a comparison chart of the prediction results and actual values ​​of the ARIMA-LSTM hybrid model in 2020-2021.

[0072] Figure 23 This is a comparison chart of the prediction results and actual values ​​of the ARIMA-LSTM hybrid model in 2021-2022.

[0073] Figure 24 This is a comparison chart of the predictions and actual values ​​of the ARIMA-LSTM hybrid model in 2022-2023.

[0074] Figure 25 This is a comparison chart of the predictions and actual values ​​of the ARIMA-LSTM hybrid model in 2023-2024. Detailed Implementation

[0075] The present invention will be further described below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0076] This application provides a method for predicting the seasonal trend of influenza based on an ARIMA-LSTM hybrid model, including the following steps:

[0077] Acquire and preprocess time-series influenza surveillance data for the target region;

[0078] The preprocessed sequence was decomposed using a time-series decomposition algorithm to obtain trend, seasonal, and holiday effect components.

[0079] The trend term is then directed into the linear prediction sub-model as an external regression variable, while the seasonal term and the holiday effect term are directed into the nonlinear prediction sub-model as specific features.

[0080] A rolling validation strategy is adopted to divide the training set and the validation set;

[0081] Based on the training set, construct the linear prediction sub-model and the nonlinear prediction sub-model respectively, and output their respective prediction values;

[0082] Based on the validation set, the fusion weights of the linear prediction sub-model and the nonlinear prediction sub-model are dynamically updated using a batch offline optimization algorithm, and the weighted fusion prediction value is calculated.

[0083] A dynamic prior distribution is constructed based on the prediction error statistics of the historical validation set. The prediction error likelihood function of the current validation set is combined with Bayesian inference to update the posterior distribution and construct the prediction interval.

[0084] The actual monitored value is compared with the upper limit of the predicted range, and an early warning signal is triggered when the actual value exceeds the upper limit.

[0085] Specifically, the implementation includes the following:

[0086] 1. Data Sources and Preprocessing

[0087] This embodiment uses data from the Chongqing area's childhood influenza surveillance from January 2011 to August 2024 as an example to specifically illustrate the implementation of the present invention.

[0088] Data source: A retrospective collection of all respiratory virus nucleic acid testing case records from January 2011 to August 2024 at the Children's Hospital Affiliated to Chongqing Medical University, totaling 168,255 cases. Nucleic acid test results for influenza A and influenza B viruses were extracted simultaneously, and the weekly influenza positivity rate was calculated. The positivity rate for influenza A was 2.8% (4712 / 168255), the positivity rate for influenza B was 0.64% (1078 / 168255), and the overall positivity rate was 3.4% (5760 / 168255).

[0089] The specific steps of data preprocessing are as follows:

[0090] Step 1: Missing value imputation. For individual weeks with no test data, linear interpolation is used to imput missing values ​​based on the positivity rates of adjacent weeks.

[0091] Step 2: Stationarity test. The ADF test was used to verify the stationarity of the weekly positivity rate sequence. The test results showed that the original sequence was non-stationary (p > 0.05), therefore, the first-order differencing method was used for stationarization.

[0092] Step 3: Min-Max Normalization. Map the processed positivity rate data to the [0,1] interval. An example of the normalized data is shown below:

[0093] The normalized positive rates for weeks 1 to 12 were: 0.0031, 0.0571, 0.9046, 0.4073, 0.0511, 0.0152, 0.0112, 0.0140, 0.0404, 0.0823, 0.4525, and 1.000, respectively.

[0094] 2. Seasonal decomposition of Prophet

[0095] The Prophet model was used to perform seasonal decomposition on the normalized influenza positivity rate series from 2011 to 2024. The trend term was set as a piecewise linear growth function, the seasonal term was Fourier order 10, and the holiday effect included a ±7-day window of the Spring Festival holiday. The decomposition yielded:

[0096] (1) Trend term g(t): shows that the influenza positivity rate was stable and fluctuated seasonally from 2011 to 2019; it dropped sharply to a low plateau from 2020 to 2022; and it gradually rebounded from 2023 to 2024, but the peak was delayed.

[0097] (2) Seasonal term s(t): It is clear that before the epidemic (2011-2019), the peak of influenza was concentrated in December to March of the following year; during the epidemic (2020-2022), the seasonal fluctuations disappeared; in the post-epidemic era (2023-2024), the seasonal pattern began to recover but the pattern changed.

[0098] (3) Holiday effect h(t): The positive rate of influenza during the Spring Festival holiday window showed a significant negative effect, which was consistent with the decrease in the number of patients and the decrease in the number of tests during this period.

[0099] 3. ARIMA Sub-model Construction

[0100] The trend term g(t) obtained from Prophet decomposition is used as an external regression variable, and the ARIMA model is trained using data from 2011 to the nth year. Optimal parameters are determined through grid search: for the scenario of predicting 2018 using the 2011-2017 training set, the optimal parameters are ARIMA(2,1,2). After model training is complete, the linear prediction values ​​ARIMA_out for each week of 2018 are output.

[0101] Similarly, ARIMA models with a total of 7 prediction windows are constructed sequentially: 2011-2018 for training and predicting 2019, 2011-2019 for training and predicting 2020, and so on, up to 2011-2023 for training and predicting 2024.

[0102] 4. LSTM Sub-model Construction

[0103] LSTM model architecture: Input layer dimension is 1 (positive rate) + 12 (Fourier features for seasonality) + 1 (holiday effect indicator) = 14 dimensions; hidden layer is a 2-layer LSTM, each with 64 neurons, and a dropout rate of 0.2; output layer is a fully connected layer, outputting a 1-dimensional prediction value. The optimizer is Adam, with a learning rate of 0.001, the loss function is mean squared error (MSE), the training epochs are 100, and the batch size is 32.

[0104] The seasonal term s(t) and holiday effect term h(t) from the Prophet decomposition are concatenated with the normalized positivity rate sequence and then input into the LSTM model. The LSTM model for each prediction window is trained independently, and the nonlinear prediction value LSTM_out is output.

[0105] 5. Adaptive weighted fusion

[0106] Taking the training and prediction of 2018 from 2011 to 2017 as an example, the adaptive weighted fusion process is explained in detail:

[0107] Step 1: Initialize weights w1 = 0.5 (ARIMA weights), w2 = 0.5 (LSTM weights).

[0108] Step 2: Calculate the initial fused prediction value y_fused = 0.5 × ARIMA_out + 0.5 × LSTM_out.

[0109] Step 3: Calculate MAE on the validation set (2018 data) and update the weights using gradient descent. Set the learning rate η = 0.01, the regularization coefficient λ = 0.01, the maximum number of iterations to 20, and the MAE improvement to <10 for 5 consecutive iterations. -4 The iteration terminates prematurely. After iteration, the weights converge to w1 = 0.4 and w2 = 0.6, at which point the MAE of the validation set reaches its minimum.

[0110] Step 4: Output the final fused prediction value and use the convergence weights to calculate the final prediction result.

[0111] Repeating the above process for each prediction window, the resulting weight change trends are as follows: In the pre-pandemic stable period (2018), the LSTM weight was relatively high (w2 = 0.6) because although the fluctuations were low, the LSTM was more sensitive to residual nonlinearity; during the trend cliff phase during the pandemic (2020-2022), the ARIMA weight was relatively high (w1 = 0.6) because the trend dominated the prediction; in the post-pandemic fluctuating period (2023), the LSTM weight increased (w2 = 0.6), verifying the effectiveness of the adaptive weight mechanism.

[0112] 6. Comparison Experiment of Feature Allocation Schemes

[0113] To verify the non-obviousness of the decoupled input strategy of "trend item → ARIMA, seasonal item + holiday item → LSTM" in this invention, three feature allocation schemes were set up for comparative experiments:

[0114] Option A (Full Feature Input to All Models): The trend term g(t), seasonal term s(t), and holiday effect term h(t) are all used as input features for ARIMA and LSTM. Under this option, ARIMA's difference stabilization effect decreases due to seasonal noise interference, and LSTM overfits due to feature redundancy, resulting in limited prediction accuracy after fusion.

[0115] Option B (random assignment): The trend term g(t) is assigned to the LSTM, and the seasonal term s(t) is assigned to the ARIMA. Under this option, the linear model (ARIMA) is forced to fit the periodic signal, resulting in a significant decrease in modeling efficiency; the LSTM is forced to focus on trend prediction, and its nonlinear modeling capabilities are wasted.

[0116] Scheme C (Decoupled Allocation of the Invention): Trend term g(t) → ARIMA, Seasonal term s(t) + Holiday effect h(t) → LSTM.

[0117] The following table compares the MAE of the three schemes across seven prediction windows (data comes from actual model training and validation results on the same dataset):

[0118]

[0119] Experimental results show that Scheme C achieves the lowest MAE across all seven prediction windows, validating the inventive advantage of the decoupled input strategy of this invention. Scheme A suffers from overfitting due to feature redundancy, and Scheme B suffers from low modeling efficiency due to feature-model mismatch; both are inferior to Scheme C of this invention.

[0120] 7. Predictive Performance Evaluation

[0121] The MAE and RMSE of the ARIMA, LSTM, and ARIMA-LSTM hybrid models were calculated for seven prediction windows (2018-2024), and the results are shown in the table below:

[0122]

[0123] Analysis conclusion:

[0124] (1) 2018 (before the pandemic): The hybrid model MAE=3.45, which is lower than ARIMA (4.19) and LSTM (4.26), verifying the effectiveness of the fusion strategy.

[0125] (2) 2021 (during the pandemic): The errors of all three models were at their lowest levels (MAE approximately 1.1-1.15), because the influenza positivity rate remained low during the pandemic, making prediction easier. The mixed model still achieved the best MAE of 1.11 and RMSE of 1.36.

[0126] (3) 2023 (post-pandemic era): The errors of all models increased significantly (MAE approximately 6.26-6.36), reflecting the predictive challenges brought about by changes in the intensity of influenza epidemics after the pandemic. The mixed model had the lowest MAE of 6.26.

[0127] (4) Overall, the hybrid model achieved the lowest MAE in all 7 prediction windows and the lowest RMSE in 5 of them, verifying the robustness and superiority of the method of the present invention in different popularization stages.

[0128] 8. Forecast Range and Early Warning

[0129] Using 2023 influenza data from China as an example, a prediction interval with a 95% confidence level is constructed:

[0130] Step 1: Based on the prediction error statistics of the historical validation window (2018-2022), the prior distribution parameters are obtained: mean μ0=0.1381, variance σ0²=0.1094.

[0131] Step 2: Calculate the mean prediction error for the 2023 validation window: μ_hat = 0.1657, variance σ_hat² = 0.1250.

[0132] Step 3: Perform the Shapiro-Wilk normality test on the 2023 prediction error, and get W=0.9832, p=0.6675 > 0.05, accepting the normal distribution hypothesis.

[0133] Step 4: Update the posterior parameters using Bayesian algorithm to obtain the posterior error interval at a 95% confidence level as [-0.0278, 0.3542].

[0134] Step 5: The prediction interval is [y_fused - 0.3542, y_fused + 0.3542]. When the actual influenza positivity rate exceeds the upper limit of this interval, the system will automatically trigger an alert.

[0135] Step 6: After each validation window ends, the prior parameters are updated using the EWMA mechanism: μ0^(new) = 0.3×0.1657 + 0.7×0.1381 = 0.1462, σ0²^(new) = 0.3×0.1250 + 0.7×0.1094 = 0.1143, which will be used as the prior distribution for the next round (2024).

[0136] This application provides an influenza seasonality trend prediction system based on an ARIMA-LSTM hybrid model, including:

[0137] The data acquisition and preprocessing module is used to acquire and preprocess time-series influenza surveillance data from the target area.

[0138] The feature decoupling input module is used to decompose the preprocessed sequence using a time series decomposition algorithm to obtain trend, seasonal and holiday effect terms. The trend terms are then directed into the linear prediction sub-model as external regression variables, while the seasonal and holiday effect terms are directed into the nonlinear prediction sub-model as exclusive features.

[0139] The rolling validation partitioning module is used to partition the training set and validation set using a yearly rolling validation strategy.

[0140] The dual-model construction module is used to build a linear prediction sub-model and a non-linear prediction sub-model based on the training set, and output their respective prediction values.

[0141] An adaptive fusion module is used to dynamically update the fusion weights of the linear prediction sub-model and the nonlinear prediction sub-model based on the validation set using a batch offline optimization algorithm, and to calculate the weighted fusion prediction value.

[0142] The interval construction module is used to construct a dynamic prior distribution based on the prediction error statistics of the historical validation set, and combined with the prediction error likelihood function of the current validation set, update the posterior distribution through Bayesian inference and construct the prediction interval.

[0143] The early warning output module is used to compare the actual monitored value with the upper limit of the prediction range, and trigger an early warning signal when the actual value exceeds the upper limit.

[0144] This embodiment provides a computer-readable storage medium storing program instructions. When executed by a processor, the program instructions are used to implement the above-described influenza seasonal trend prediction method based on the ARIMA-LSTM hybrid model.

[0145] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications should be covered within the scope of the claims of the present invention. The ARIMA-LSTM hybrid model of the present invention is not limited to influenza prediction. Its core idea, based on the adaptive weighted fusion framework of Prophet seasonal decomposition, can be extended to other infectious disease prediction scenarios with seasonal characteristics. In the present invention, the LSTM sub-model can be replaced with other deep learning time series models (such as GRU, Transformer, etc.), and the ARIMA sub-model can be replaced with other linear time series models (such as Prophet, ETS, etc.), all of which fall within the protection scope of the present invention. The adaptive weight update algorithm in the present invention is not limited to the gradient descent method. Other optimization algorithms (such as Newton's method, Adam, etc.) are used to achieve dynamic weight updates, all of which fall within the protection scope of the present invention. The prediction interval construction method involved in the present invention, in addition to Bayesian inference, uses quantile regression, Bootstrap, and other methods to construct dynamic prediction intervals, all of which fall within the protection scope of the present invention.

Claims

1. A method for predicting the seasonal trend of influenza based on an ARIMA-LSTM hybrid model, characterized in that, Includes the following steps: Acquire and preprocess time-series influenza surveillance data for the target region; The preprocessed sequence is decomposed using a time series decomposition algorithm to obtain a trend term, a seasonal term, and a holiday effect term. The trend term is then directed into a linear prediction sub-model as an external regression variable, while the seasonal term and the holiday effect term are directed into a nonlinear prediction sub-model as specific features. A rolling annual validation strategy is adopted to divide the training set and the validation set; based on the training set, the linear prediction sub-model and the nonlinear prediction sub-model are constructed respectively, and their respective prediction values ​​are output; Based on the validation set, the fusion weights of the linear prediction sub-model and the nonlinear prediction sub-model are dynamically updated using a batch offline optimization algorithm, and the weighted fusion prediction value is calculated. A dynamic prior distribution is constructed based on the prediction error statistics of the historical validation set. The prediction error likelihood function of the current validation set is combined with Bayesian inference to update the posterior distribution and construct the prediction interval. The actual monitored value is compared with the upper limit of the predicted range, and an early warning signal is triggered when the actual value exceeds the upper limit.

2. The influenza seasonality trend prediction method based on the ARIMA-LSTM hybrid model as described in claim 1, characterized in that, The preprocessing includes: The positive rate of influenza virus nucleic acid testing was calculated as time series data; missing time points were filled using linear interpolation; the extended Dickey-Fuller test was used to assess the stationarity of the sequence, and if the sequence was non-stationary, difference transformation was performed; and the time series data were mapped to a preset interval using min-max normalization.

3. The influenza seasonality trend prediction method based on the ARIMA-LSTM hybrid model as described in claim 2, characterized in that, The step of inputting the trend term into the linear prediction sub-model as an external regression variable, and inputting the seasonal term and the holiday effect term into the nonlinear prediction sub-model as specific features, specifically includes: The influenza positivity rate sequence was decomposed into a superposition of trend term, annual cycle seasonal term, and public holiday effect term using the Prophet model; The linear prediction sub-model only accepts the trend term as an external regression variable to focus on fitting the linear drift component; The nonlinear prediction sub-model only accepts the seasonal and holiday effect terms as input features to centrally model periodic fluctuations and sudden holiday effects.

4. The influenza seasonality trend prediction method based on the ARIMA-LSTM hybrid model as described in claim 3, characterized in that, The method of dividing the training set and validation set using a rolling annual validation strategy specifically includes: Using data from year m to year n as the training set, predict the influenza positivity rate in year (n+1), where n is a preset sequence of years, and generate multiple independent validation windows to evaluate the model's predictive performance at different stages of the epidemic.

5. The influenza seasonality trend prediction method based on the ARIMA-LSTM hybrid model as described in claim 4, characterized in that: The linear prediction sub-model is an ARIMA model, which determines the optimal autoregression order, difference order, and moving average order through grid search, and uses the trend term as an external regression variable to output a linear prediction value.

6. The influenza seasonality trend prediction method based on the ARIMA-LSTM hybrid model as described in claim 5, characterized in that: The nonlinear prediction sub-model is an LSTM model. Its input layer is composed of the standardized influenza positivity rate, the Fourier feature of the seasonal term, and the holiday effect indicator. The hidden layer contains multiple LSTM units and a Dropout layer. The output layer is a fully connected layer to output nonlinear prediction values.

7. The influenza seasonality trend prediction method based on the ARIMA-LSTM hybrid model as described in claim 6, characterized in that, The method of dynamically updating the fusion weights using a batch offline optimization algorithm specifically includes: Initialize the weights of the linear prediction sub-model and the nonlinear prediction sub-model, such that the sum of the weights is 1 and each weight is non-negative; Define a weighted fusion prediction value and calculate the mean absolute error on the validation set as the objective function; The weights are iteratively updated using gradient descent. The update process includes a subgradient term of the sign function and an L2 regularization penalty term. An early stopping mechanism and a learning rate search strategy are set to prevent extreme weight allocation.

8. The influenza seasonality trend prediction method based on the ARIMA-LSTM hybrid model as described in claim 7, characterized in that, The construction of the prediction interval specifically includes: The prediction error sequence is tested for normality. If it follows a normal distribution, the prediction error statistics of the historical verification window are taken as the Gaussian prior distribution, and the prediction error of the current verification window is taken as the Gaussian likelihood function. The mean and variance of the posterior distribution are calculated under the conjugate condition. The prior distribution parameters are dynamically updated using an exponentially weighted moving average mechanism, enabling the prior distribution to adapt adaptively to changes in popular patterns; at a given confidence level, the prediction interval is calculated based on the posterior distribution parameters.

9. A seasonal influenza trend prediction system based on an ARIMA-LSTM hybrid model, characterized in that, include: The data acquisition and preprocessing module is used to acquire and preprocess time-series influenza surveillance data from the target area. The feature decoupling input module is used to decompose the preprocessed sequence using a time series decomposition algorithm to obtain trend terms, seasonal terms and holiday effect terms. The trend terms are then directed to the linear prediction sub-model as external regression variables, and the seasonal terms and holiday effect terms are directed to the nonlinear prediction sub-model as exclusive features. The rolling validation partitioning module is used to partition the training set and validation set using a yearly rolling validation strategy. A dual-model construction module is used to construct the linear prediction sub-model and the nonlinear prediction sub-model based on the training set, and output their respective prediction values. An adaptive fusion module is used to dynamically update the fusion weights of the linear prediction sub-model and the nonlinear prediction sub-model based on the validation set using a batch offline optimization algorithm, and to calculate the weighted fusion prediction value. The interval construction module is used to construct a dynamic prior distribution based on the prediction error statistics of the historical validation set, and combined with the prediction error likelihood function of the current validation set, update the posterior distribution through Bayesian inference and construct the prediction interval. The early warning output module is used to compare the actual monitored value with the upper limit of the prediction range, and trigger an early warning signal when the actual value exceeds the upper limit.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 8.

Images