阅读说明：本技术 一种面向covid-19的经验模态分解模糊预测方法 (COVID-19-oriented empirical mode decomposition fuzzy prediction method ) 是由 陈伯伦 沈怡芸 姜文心 纪敏 于永涛 朱国畅 戚梓凡 于 2021-08-16 设计创作，主要内容包括：发明公开了一种面向COVID-19的经验模态分解模糊预测方法,适于疫情预测领域,步骤包括：新冠疫情原始数据使用经验模态分解算法,得到不同时间尺度下数据的变化趋势；将得到的数据用大小为1*(c+1)的窗口,以步长为1,使用极限学习机进行滑动训练,得到不同时间尺度下的预测值；使用自适应模糊推理系统对训练结果进行拟合,得到最终预测值。该方法在保证学习精度的前提下,简化了训练过程,提升了算法运行速度和泛化能力,同时解决了过拟合和局部最小等问题,能够很好地满足了用户对预测系统中的高准确率的要求,具有切实可行的应用前景和实用价值。(The invention discloses a COVID-19-oriented empirical mode decomposition fuzzy prediction method, which is suitable for the field of epidemic situation prediction and comprises the following steps: obtaining the change trend of the new crown epidemic situation original data under different time scales by using an empirical mode decomposition algorithm; using a window with the size of 1 x (c +1) to perform sliding training by using an extreme learning machine with the step length of 1 to obtain predicted values under different time scales; and fitting the training result by using an adaptive fuzzy reasoning system to obtain a final predicted value. On the premise of ensuring the learning precision, the method simplifies the training process, improves the algorithm running speed and generalization capability, solves the problems of overfitting, local minimum and the like, can well meet the requirement of a user on high accuracy in a prediction system, and has practical and feasible application prospect and practical value.)

1. A COVID-19-oriented empirical mode decomposition fuzzy prediction method is characterized by comprising the following steps:

step 1: obtaining the change trend of the new crown epidemic situation original data under different time scales by using an empirical mode decomposition algorithm;

step 2: using a window with the size of 1 x (c +1) to perform sliding training by using an extreme learning machine with the step length of 1 to obtain predicted values under different time scales;

and step 3: and fitting the training result by using an adaptive fuzzy reasoning system to obtain a final predicted value.

2. The COVID-19-oriented empirical mode decomposition fuzzy prediction method of claim 1, wherein the empirical mode decomposition algorithm in the step 1 comprises the following specific steps:

step 1.1: finding out all maximum value points and minimum value points of the original signal, fitting the maximum value points and minimum value points to obtain an upper envelope line and a lower envelope line, and taking the mean value of the maximum value points and the minimum value points to obtain a mean envelope line;

step 1.2: removing the mean value from the original signal to obtain a new signal;

step 1.3: judging whether the new signal meets the IMF condition:

if not, the new signal replaces the original signal, and the steps are repeated until the IMF condition is met;

if yes, recording the signal as a first IMF;

step 1.4: repeating the steps on the IMF to obtain more IMFs;

step 1.5: and when the n-order IMF component meets the empirical mode decomposition end condition, completing the decomposition.

3. The COVID-19-oriented empirical mode decomposition fuzzy prediction method of claim 1, wherein the specific steps of the extreme learning machine algorithm in the step 2 are as follows:

assuming that N samples and L hidden layer nodes are provided; for a single hidden layer neural network, input x_iThe corresponding desired output is t_iThe activation function is g (x), the input layer to hidden layer input weight w_iAnd bias b_iHidden layer output weight beta_iThen, it can be:

the matrix can be expressed as:

Hβ＝T

where H is the hidden layer output matrix, β is the output weight, T is the desired output:

inputting weights w in extreme learning machines_iAnd bias b_iThe hidden layer output matrix H is uniquely determined, the training process of the extreme learning machine is converted into the solving output weight beta, namely:

β＝H^-T。

4. the COVID-19-oriented empirical mode decomposition fuzzy prediction method of claim 1, wherein in the step 3, the adaptive fuzzy inference system is assumed to have two inputs x₁、x₂And one output y, the rule base has the following rules:

if x is input₁、x₂More satisfy A₁、B₁If y is equal to p₁x₁+q₁x₂+r₁；

If x is input₁、x₂More satisfy A₂、B₂If y is equal to p₂x₁+q₂x₂+r₂；

Wherein A is₁、A₂、B₁、B₂For values of language variables, p₁、p₂、q₁、q₂、r₁、r₂Is an adjustable parameter; the network is divided into 5 layers, the first 3 layers are regular front parts, and the second 2 layers are regular back parts.

5. The COVID-19-oriented empirical mode decomposition fuzzy prediction method of claim 4, wherein the layer 5 network comprises:

a first layer: and the fuzzy layer converts the input variable into the membership degree of each fuzzy set, and outputs the membership degree as follows:

wherein x is_iIs input;the membership function of the ith node under the jth rule is called a front part parameter;

a second layer: and the rule applicability layer calculates the product of the input signals, and the result is the applicability of the rule, and the output is as follows:

wherein, ω is_jSuitability for the jth rule;is the membership function of the ith node;

and a third layer: and normalizing the applicability layer, calculating the ratio of the applicability of the jth rule to the sum of the applicability of all the rules for the ith node, and outputting:

wherein the content of the first and second substances,normalized fitness for the jth rule;

a fourth layer: and the fuzzy rule output layer calculates the output of each rule, and the output is as follows:

wherein p is_i、q_i、r_iA parameter set for the node, called a back-piece parameter;is that the third layer is standardizedThe suitability of (2); f. of_jIs the output of the jth rule;

and a fifth layer: a summation layer for summing each output to obtain a total output, outputting:

and obtaining a final prediction result.

Technical Field

The invention belongs to prediction of the number of confirmed people for new crown epidemic situation applied to a prediction system, and particularly relates to a COVID-19-oriented empirical mode decomposition fuzzy prediction method.

Background

The application of the prediction system is very wide, and the significance is very far. For example, forecasting the financial status of a business is of great practical significance to protect the interests of investors and creditors, to protect operators from financial crisis, and to monitor the quality of listed companies and the risk of securities markets by government regulators. For example, historical meteorological data is used for predicting factors such as future soil moisture and the like, and defense preparation is made in advance to improve the yield of agricultural products. For example, real-time traffic passenger flow data is analyzed and studied to predict it, and coordinated scheduling is performed to select the best route to mitigate traffic flow. Therefore, the method has extremely important significance for effectively controlling the epidemic situation by utilizing a reasonable prediction method.

Many scholars have been engaged in predicting new cases of coronary pneumonia in recent years. The COVID-19 elastic net predictor (EN-CoF for short) proposed by ohnsen et al aims to provide an intuitive, versatile and easy-to-use predictor. EN-CoF is a multiple linear regressor trained on time series data to predict the number of new daily coronary pneumonia cases. Compared with more complex models such as ARIMA and Bi-LSTM, the EN-CoF maintains higher precision and has the advantages of transparency, generalization and accessibility. Mouswavi et al propose a novel covi-19 confirmed case prediction system platform based on factors such as spreading rate, temperature, humidity, etc. The prediction platform can systematically derive a group of characteristics suitable for training a Recurrent Neural Network (RNN), and respectively improve the prediction of the trend of stationarity and non-stationarity of the diagnosed case number by using the stationarity and the non-stationarity of the characteristics. Shahid et al used prediction models such as auto-regressive integrated moving average (ARIMA), Support Vector Regression (SVR), long term memory (LSTM), Bi-LSTM, and time series predictions of confirmed cases, deaths, and recovery in 10 countries affected primarily by 2019 coronavirus disease. The performance of the model is measured in terms of mean absolute error, root mean square error, and r2_ score index. In the whole scene, the models from good performance to the lowest performance are Bi-LSTM, GRU, SVR and ARIMA in sequence. Masum et al, when predicting daily cumulative confirmed cases based on ARIMA model and LSTM-based recurrent neural network, proposed a repeatable LSTM framework (r-LSTM) for the randomness of neural network optimization and weight random initialization, and the result repeatability generated by the LSTM-based model was poor, and tested the repeatability and robustness of the framework using z-score outliers. Based on the ARIMA model and the lstm-based recurrent neural network, Assimakis et al researches the applicability of Kalman filtering as decision support in the infectious disease early warning and emergency response systems such as COVID-19 and the like, and predicts daily accumulated confirmed cases. Friji et al proposed a generalized mechanism model containing eight states to describe the progression of COVID-19 from a susceptible state to a discharged state through isolated and hospitalized states. The parameters of the model are determined by solving a fitting optimization problem solved using the Levenberg-Marquardt algorithm. The problem has three observations input, the number of infected, dead and reported cases. The target function of the model is weighted within the training days so as to guide the fitting algorithm to approach the latest pandemic period, so that the trend prediction is more accurate and the forecasting capability is stronger. Gaglione et al apply the algorithms for tracking and predicting targets such as missiles and ships, i.e., Bayesian sequences and adaptive dynamic estimation, to epidemiology COVID-19 to reliably estimate and predict the evolution of infection. Beche et al attempted to address the prediction of the spread of new coronaviruses by exploiting the ability of cyclic autoencoders on time series and the implicit semi-supervised training process, and introduced the notion of neighborhoods for better estimation when the cumulative number of cases diagnosed in any country. The results show that the method can make reliable predictions for a period of 30 days. Kumar et al summarized a time series prediction model and analyzed the occurrence of new coronary pneumonia epidemics to examine whether these numbers increased or decreased in the near future. Statistical pattern analysis and data visualization employ widely accepted time series methods such as autoregressive integrated moving average (ARIMA) and its components Moving Average (MA) and Autoregressive (AR). Finally, time-dependent parameters may provide an indication of the trend of indian COVID-19 outbreaks. Iqbal et al propose a long-short term memory (LSTM) model based on the Recurrent Neural Network (RNN). RNN and LSTM training predictions were made for the number of patients with Pakistan COVID-19, respectively, and the Mean Absolute Percentage Error (MAPE) was calculated to determine the predicted effect of the model at different LSTM units, batch sizes and time points.

Therefore, much time is consumed for completing prediction in mass data, and how to design a prediction method with high accuracy and small calculation amount is a main problem at present.

Disclosure of Invention

The purpose of the invention is as follows: aiming at the problems in the prior art, the invention provides an empirical mode decomposition fuzzy prediction method which is suitable for various countries and has better accuracy in long-term and short-term prediction.

The technical scheme is as follows: a COVID-19-oriented empirical mode decomposition fuzzy prediction method comprises the following steps:

step 1: obtaining the change trend of the new crown epidemic situation original data under different time scales by using an empirical mode decomposition algorithm;

step 2: using a window with the size of 1 x (c +1) to perform sliding training by using an extreme learning machine with the step length of 1 to obtain predicted values under different time scales;

and step 3: and fitting the training result by using an adaptive fuzzy reasoning system to obtain a final predicted value.

Further, the experimental mode decomposition algorithm in step 1 comprises the following specific steps:

step 1.1: finding out all maximum value points and minimum value points of the original signal, fitting the maximum value points and minimum value points to obtain an upper envelope line and a lower envelope line, and taking the mean value of the maximum value points and the minimum value points to obtain a mean envelope line;

step 1.2: removing the mean value from the original signal to obtain a new signal;

step 1.3: judging whether the new signal meets the IMF condition:

if not, the new signal replaces the original signal, and the steps are repeated until the IMF condition is met;

if yes, recording the signal as a first IMF;

step 1.4: repeating the steps on the IMF to obtain more IMFs;

step 1.5: and when the n-order IMF component meets the empirical mode decomposition end condition, completing the decomposition.

Further, the specific steps of the limit learning machine algorithm in step 2 are as follows:

assuming that N samples and L hidden layer nodes are provided; for a single hidden layer neural network, input x_iThe corresponding desired output is t_iThe activation function is g (x), the input layer to hidden layer input weight w_iAnd bias b_iHidden layer output weight beta_iThen, it can be:

the matrix can be expressed as:

Hβ＝T

where H is the hidden layer output matrix, β is the output weight, T is the desired output:

inputting weights w in extreme learning machines_iAnd bias b_iThe hidden layer output matrix H is uniquely determined, the training process of the extreme learning machine is converted into the solving output weight beta, namely:

β＝H^-T。

further, in step 3, assume that the adaptive fuzzy inference system has two inputs x₁、x₂And one output y, the rule base has the following rules:

if x is input₁、x₂More satisfy A₁、B₁If y is equal to p₁x₁+q₁x₂+r₁；

If x is input₁、x₂More satisfy A₂、B₂If y is equal to p₂x₁+q₂x₂+r₂；

Wherein A is₁、A₂、B₁、B₂For values of language variables, p₁、p₂、q₁、q₂、r₁、r₂Is an adjustable parameter; the network is divided into 5 layers, the first 3 layers are regular front parts, and the second 2 layers are regular back parts.

Further, a layer 5 network includes:

a first layer: and the fuzzy layer converts the input variable into the membership degree of each fuzzy set, and outputs the membership degree as follows:

wherein x is_iIs input;the membership function of the ith node under the jth rule is called a front-part parameter.

A second layer: and the rule applicability layer calculates the product of the input signals, and the result is the applicability of the rule, and the output is as follows:

wherein, ω is_jSuitability for the jth rule;is the membership function of the ith node.

And a third layer: and normalizing the applicability layer, calculating the ratio of the applicability of the jth rule to the sum of the applicability of all the rules for the ith node, and outputting:

wherein the content of the first and second substances,normalized suitability for the jth rule.

A fourth layer: and the fuzzy rule output layer calculates the output of each rule, and the output is as follows:

wherein p is_i、q_i、r_iA parameter set for the node, called a back-piece parameter;is the third layer's normalized suitability; f. of_jIs the output of the jth rule.

And a fifth layer: a summation layer for summing each output to obtain a total output, outputting:

and obtaining a final prediction result.

Compared with the prior art, the method has the advantages and effects of short training time, strong generalization capability and high prediction accuracy, and is specifically represented as follows:

(1) the empirical mode decomposition algorithm is used for carrying out stabilization processing on the original data, the change trend of the data under different time scales is mined, and result prediction can be better carried out;

(2) the connection weight between the hidden layer and the output layer in the extreme learning machine can obtain the optimal solution at one time in a way of solving an equation set, thereby reducing the operation amount and improving the operation speed;

(3) the self-adaptive fuzzy inference system can change system parameters according to the prior knowledge, so that the final output accuracy is higher;

the invention provides a fuzzy prediction method based on empirical mode decomposition aiming at a prediction problem. Decomposing original new crown epidemic situation data into data under different time scales by using an empirical mode algorithm; respectively carrying out sliding training on data under different time scales by using a sliding block with the size of 1 x (c +1) by using an extreme learning machine to obtain predicted values under different time scales; and finally, organically fusing the predicted values under different time scales by using a self-adaptive fuzzy reasoning system to obtain the final predicted value. On the premise of ensuring the learning precision, the method simplifies the training process, improves the operation speed and generalization capability of the algorithm, solves the problems of overfitting, local minimum and the like, and can well meet the requirement of a user on high accuracy in a prediction system.

Drawings

FIG. 1 is a general flow diagram of the present invention.

Detailed Description

The invention is further elucidated with reference to the drawings and the detailed description.

The invention realizes prediction based on empirical mode decomposition, and solves the problem that other prediction methods cannot well analyze and process nonlinear non-stationary signals. When training data, the connection weight of the input layer and the hidden layer and the threshold value of the hidden layer are set randomly without adjustment, and the connection weight can obtain an optimal solution at one time by solving an equation set, so that the calculation amount is reduced, the running speed is improved, and the problems of long time consumption, high calculation cost and the like of other training methods are solved. The invention can obtain a higher-quality prediction result on the basis of reduced calculation overhead.

The method comprises the following steps:

take the newly-increased number of confirmed diagnosis persons in the United states every day as an example:

analyzing data using empirical mode decomposition

(1) Normalizing the newly-added number of confirmed American patients to obtain NEW (t), wherein t is days

(2) Decomposing the data NEW (t) by empirical mode decomposition, which comprises the following steps:

firstly, the methodFinding out all maximum value points in NEW (t), and obtaining an upper envelope line max (t) through function fitting; finding out all minimum value points, and fitting by using the same method to obtain a lower envelope line min (t); the mean value of the upper and lower envelope lines is denoted as m₁(t)：

Then, m is subtracted from the original signal₁(t) obtaining a new signal with low frequencies removed

The first calculation is based on the raw data NEW (t) being chaotic and irregularGenerally, the IMF condition can not be satisfied, and the steps are repeated for k times until the IMF condition is satisfiedIMF satisfied condition, namely the first-order IMF component of NEW (t), noted as IMF₁(t)：

Subtracting imf1 from NEW (t)₁(t) obtaining a new signal r with the high frequency components removed₁(t)：

r₁(t)＝NEW(t)-imf1₁(t)(4)

To r₁(t) repeating the above process to obtain a second order IMF component IMF₂(t) of (d). And so on, up to the IMF component IMF of order n_n(t) stopping the decomposition if the residual component is less than a predetermined value or is a monotonic function or a constant. Obtaining IMF after N decompositions of NEW (t), wherein IMF Is (IMF)₁,imf₂,···,imf_n)。

Two, using extreme learning machine and combining 1 x (c +1) sliding window training data

(1) Pair imf_iSliding by using windows with the size of 1 x (c +1) and the step length of 1, wherein the first c values of each sliding window are input of the extreme learning machine, and the c +1 th value is output of the extreme learning machine;

(2) training the obtained input and output by using an extreme learning machine to obtain elm_iIf the new number of IMFs for the confirmed diagnosis is equal to (ELM)₁,elm₂,···,elm_n)。

Thirdly, carrying out organic fitting on data by using self-adaptive fuzzy reasoning system

Using self-adaptive fuzzy reasoning system to process the elm obtained in the second step_iFitting, the first layer output is:

wherein the content of the first and second substances,the membership function of the ith node under the jth rule is determined by some parameters, which are called as front-part parameters; output ofIs the degree of membership of the fuzzy set.

Output of the second layer:

wherein, ω is_jFor the applicability of the jth rule,is the membership function of the ith node.

Output of the third layer:

wherein the content of the first and second substances,normalized suitability for the jth rule.

Output of the fourth layer:

wherein p is_i、q_iA parameter set for the node, called a back-piece parameter; f. of_jIs the output of the jth rule.

Outputting by the fifth layer:

and obtaining a final prediction result.

Details not described herein are well within the skill of those in the art.