Hybrid deep learning model time sequence prediction method for complex high-noise signals

Abstract

The invention discloses a hybrid deep learning model time sequence prediction method for a complex high-noise signal, and the method comprises the steps: dividing original time sequence data into a training set and a test set, generating a time sequence, carrying out the normalization preprocessing, and carrying out the VMD processing of the training set and the test set; modeling trend and seasonal characteristics in the data by using a Prophet model, outputting a preliminary prediction result, and extracting a residual error part; secondly, modeling is conducted on a residual error part generated by the Prophet model through an SARIMA model, modeling is conducted on a residual error part generated by the SARIMA model through an LSTM model, hyper-parameters of the LSTM model are optimized through a differential evolution algorithm, and an optimized prediction result is generated; and finally, fusing prediction results of the three models, and outputting a final prediction value. According to the method, the residual error is processed in a multi-stage progressive mode, the advantages of the three neural network models are fully utilized, the influence of noise on prediction can be reduced, different features in the signals are effectively captured, and the prediction accuracy and stability are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal data processing, and specifically relates to a hybrid deep learning model time series prediction method for complex high-noise signals. Background Art

[0002] With the development of big data and artificial intelligence, time series data is widely used in various fields (such as ocean, meteorology, finance, medicine, etc.), and has become a core task in many industries, including weather forecasting, financial market analysis, and traffic flow forecasting. However, in practical applications, many signal data are often affected by factors such as noise and periodic fluctuations, which makes prediction more difficult. High-noise signals and their complex characteristics pose great challenges to traditional prediction models. Traditional prediction models have the following limitations when processing complex and high-noise signals: 1. Mixed trends and cycles: Complex signals often contain multiple trends, seasonality, and cyclical components, and these components are intertwined. It is a challenge to effectively decompose these components and accurately predict future values; 2. A single model cannot capture the multi-level features of complex signals: Traditional time series models can usually only capture a certain type of data feature and are unable to simultaneously handle trends, seasonality, noise, and nonlinear relationships in the signal; 3. Insufficient residual modeling capabilities: Although SARIMA based on linear models can analyze residuals, it is not effective in dealing with residuals with strong nonlinearity and large fluctuations; 4. Overfitting and prediction stability issues: A single deep learning model (such as LSTM) is prone to overfitting in long-term predictions, resulting in insufficient model generalization ability and difficulty in stably coping with complex time series data.

[0003] Therefore, when dealing with complex and high-noise signals, the commonly used time series models such as Prophet, SARIMA and LSTM can solve the trend and seasonality problems of signals to a certain extent. However, facing the above-mentioned high noise, nonlinear and complex periodic mixed situations, their accuracy and robustness are insufficient, and the prediction effect is not ideal. Therefore, a multi-model fusion method is needed to improve the prediction accuracy and prediction effect of complex and high-noise signals. Summary of the invention

[0004] In view of the problems existing in the above-mentioned background technology, the purpose of the present invention is to provide a hybrid deep learning model time series prediction method for complex high-noise signals. The method combines the advantages of three neural network models, Prophet, SARIMA and LSTM, and adopts a multi-stage progressive modeling method to cope with the trends, seasonality, noise and nonlinear characteristics in complex signals, and can effectively process complex and high-noise time series signals.

[0005] In order to achieve the above object, the present invention adopts the following technical solution: A hybrid deep learning model time series prediction method for complex high-noise signals includes the following steps: S1. Data preprocessing The original time series data is divided into a training set and a test set to generate the time series of the training set and the test set; the time series data of the training set and the test set are normalized and preprocessed, and then the training set and the test set are processed by variational mode decomposition (VMD) respectively; S2. Modeling using the Prophet model Input the data of the training and test sets after VMD processing into the Prophet model, use the Prophet model to model the trend and seasonal characteristics in the data, capture the long-term trend and seasonal changes in the data, output preliminary prediction results, and extract the residual part; S3. Modeling residuals using the SARIMA model The SARIMA model is used to model the residual part generated by the Prophet model, capturing the linear characteristics and periodic noise in the residual and generating optimized forecast results. S4. Modeling residuals using LSTM models Use the LSTM model to model the residual part generated by the SARIMA model, capture the nonlinear characteristics in the data, and use the differential evolution algorithm to optimize the hyperparameters of the LSTM model to generate optimized prediction results; S5. Multi-model fusion The prediction results of the Prophet, SARIMA and LSTM models are added together, and then denormalized to output the final prediction value.

[0006] Furthermore, in step S1, the ratio of the training set and the test set divided by date and time is 70 / 30.

[0007] Furthermore, the parameters of the SARIMA model are selected to obtain the optimal combination through a grid search method.

[0008] Compared with the shortcomings and deficiencies of the prior art, the present invention has the following beneficial effects: (1) The present invention combines the advantages of three neural network models: Prophet, SARIMA and LSTM. By progressively processing residuals and optimizing layer by layer, the advantages of each model are fully utilized, which can reduce the impact of noise on prediction, effectively capture different features in the signal, achieve more efficient prediction, and significantly improve the accuracy and stability of prediction. (2) By combining traditional time series models with deep learning models, the prediction ability of complex nonlinear signals is improved and the risk of overfitting is reduced; (3) Parameter optimization: Use the differential evolution algorithm to optimize the hyperparameters of the LSTM model to determine the optimal hidden layer size, learning rate, and dropout probability to ensure the best performance of the model; (4) Compared with the traditional single model, the present invention can significantly improve the prediction accuracy in the prediction of complex high-noise signals, especially showing good stability in long-term prediction, overcoming the limitations of the single model in complex time series prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 It is a flowchart of a hybrid deep learning model time series prediction method for complex high-noise signals provided by an embodiment of the present invention; Figure 2 It is a comparison chart of the prediction results and actual results of the hybrid deep learning model for the training set data provided by the embodiment of the present invention; Figure 3 This is a comparison chart of the prediction results of the test set data by the hybrid deep learning model provided by an embodiment of the present invention and the actual results; Figure 4 This is a comparison chart of the prediction results of the Prophet model provided by the embodiment of the present invention on the test set data and the actual results; Figure 5 It is a comparison chart of the prediction results of the SARIMA model provided by the embodiment of the present invention on the test set data and the actual results; Figure 6 This is a comparison chart of the prediction results of the LSTM model provided by an embodiment of the present invention for the test set data and the actual results. DETAILED DESCRIPTION

[0010] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0011] The steps of the hybrid deep learning model time series prediction method for complex high-noise signals provided by the present invention are as follows, and the flowchart is as follows: Figure 1 shown.

[0012] S1. Data preprocessing Divide the original time series data into training set and test set, with the ratio of date and time being 70 / 30, and generate the time series of training set and test set; Data source: The daily brightness temperature data of a certain area from 2010 to 2019 were selected as the original time series data. The data contains the daily variation trend of the brightness temperature data and contains a large amount of high-frequency noise and irregular fluctuations; The time series data of the training set and the test set are normalized and preprocessed, and then the training set and the test set are processed by variational mode decomposition (VMD) respectively; Normalization: Use MinMaxScaler to normalize the selected temperature data. Normalization is to scale each eigenvalue in the data set to the range of [0,1] to ensure the stability and rapid convergence of model training.

[0013] S2. Modeling using the Prophet model The temperature data of the training set, validation set and test set after VMD processing are input into the Prophet model, and the Prophet model is used to model the trend and seasonal characteristics in the data, capture the long-term trend and seasonal changes in the data, output the preliminary prediction results, and extract the residual part, which mainly contains short-term fluctuations and components that cannot be explained by the trend.

[0014] S3. Modeling residuals using the SARIMA model The SARIMA model is used to model the residual part generated by the Prophet model, capturing the linear characteristics and periodic noise in the residual and generating optimized prediction results. The parameters of the SARIMA model are selected to achieve the optimal combination through the grid search method.

[0015] S4. Modeling residuals using LSTM models The LSTM model is used to model the residual part generated by the SARIMA model to capture the nonlinear characteristics in the data. The LSTM model further eliminates the noise and residual errors in the time series by learning the long-term and short-term dependencies of the sequence. In order to improve the prediction accuracy of the LSTM model, the differential evolution algorithm is used to optimize the hyperparameters of the LSTM model, such as the hidden layer size, learning rate, and Dropout probability, to generate optimized prediction results.

[0016] S5. Multi-model fusion The prediction results of Prophet, SARIMA and LSTM models are added together, and then denormalized to obtain the final prediction value. Figure 2-3 ,in, Figure 2 This is a comparison chart between the predicted results and the actual results of the training set data. Figure 3 This is a comparison chart of the predicted results and the actual test set data. It can be seen that the predicted curve of the present invention is nearly consistent with the actual curve, and the predicted results can still be stably output in a high noise environment.

[0017] The present invention combines the advantages of the three models of Prophet, SARIMA and LSTM, and adopts a multi-stage progressive modeling method to deal with the trend, seasonality, noise and nonlinear characteristics in complex signals. The prediction method of the present invention is evaluated. At the same time, the traditional Prophet model, the separate SARIMA model and the LSTM model are used as comparisons. The evaluation indicators are: by calculating the mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE), determination coefficient (R 2 ) and mean absolute percentage error (MAPE) to evaluate the performance of the model on the test set, where: Mean Absolute Error (MAE): The average of the absolute errors between the predicted values ​​and the true values; Mean Squared Error (MSE): The average of the sum of squares of the differences between the predicted values ​​and the true values; The coefficient of determination (R 2 ): It is the goodness-of-fit index of the regression model, indicating the explanatory power of the model; Root mean square error (RMSE): RMSE is the square root of MSE and directly reflects the difference between the predicted value and the actual value; Mean absolute percentage error (MAPE): the percentage error between the predicted value and the true value; The smaller the MAE, MSE, RMSE, and MAPE, the better. 2 The larger the better. The evaluation results of each indicator are shown in Table 1: Table 1 Evaluation index results of different models on the test set As can be seen from Table 1, the present invention has a better performance in capturing short-term fluctuations and trends of data. Compared with the traditional Prophet model, the single SARIMA model or the LSTM model, the multi-stage noise reduction and hybrid prediction model used in the present invention performs well on the test set. Figure 3-6 ,By comparing the original data and the prediction results, we can intuitively see the effect of the present invention in improving the prediction accuracy.

[0018] The present invention retains the main trend and seasonal characteristics of the data when processing time series data with high noise and irregular fluctuations, and provides an effective solution for the prediction of various time series data. It has broad application prospects and is applicable to multiple fields such as financial data, sensor data, and meteorological data. The present invention performs well in processing long-term trend prediction and short-term fluctuation prediction of temperature data, and can effectively reduce prediction errors and enhance the robustness of the model.

[0019] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A hybrid deep learning model time series prediction method for complex high-noise signals, characterized in that: The steps include: S1. Data preprocessing The original time series data is divided into a training set and a test set to generate the time series of the training set and the test set; the time series data of the training set and the test set are normalized and preprocessed, and then VMD is performed on the training set and the test set respectively; S2. Modeling using the Prophet model Input the data of the training set and test set after VMD processing into the Prophet model, use the Prophet model to model the trend and seasonal characteristics in the data, capture the long-term trend and seasonal changes in the data, output the preliminary prediction results, and extract the residual part; S3. Modeling residuals using the SARIMA model The SARIMA model is used to model the residual part generated by the Prophet model, capturing the linear characteristics and periodic noise in the residual and generating optimized forecast results. S4. Modeling residuals using LSTM models Use the LSTM model to model the residual part generated by the SARIMA model, capture the nonlinear characteristics in the data, and use the differential evolution algorithm to optimize the hyperparameters of the LSTM model to generate optimized prediction results; S5. Multi-model fusion The prediction results of the Prophet, SARIMA and LSTM models are added together, and then denormalized to output the final prediction value.

2. The hybrid deep learning model time series prediction method for complex high-noise signals according to claim 1, characterized in that: In step S1, the ratio of the training set and the test set divided by time is 70 / 30.

3. The hybrid deep learning model time series prediction method for complex high-noise signals according to claim 1, characterized in that: The parameters of the SARIMA model are optimally selected by a grid search method.

Images