A channel prediction method for Internet of Vehicles based on decomposition prediction and error correction

Abstract

The present invention discloses a vehicle network channel prediction method based on decomposition prediction and error correction. First, variational mode decomposition is used to decompose channel data into several sub-components, and then a prediction model based on a bidirectional gated recurrent neural network is established for each sub-component. The prediction results of each sub-component are superimposed to obtain a preliminary prediction result. Next, the prediction difference of the sub-component is calculated, and the prediction error of the sub-component is processed by an error correction method to obtain an error prediction result. Finally, the preliminary prediction result and the error prediction result are superimposed to complete the prediction. This vehicle network channel prediction method based on decomposition prediction and error correction can solve the problems of insufficient data in channel modeling and long-term error accumulation in the prediction process in existing vehicle network channel prediction technologies, achieve higher vehicle network channel prediction accuracy, and thus facilitate better optimization of the performance of vehicle network communication systems.

Description

Technical Field

[0001] The present invention relates to a vehicle networking channel prediction method, in particular to a vehicle networking channel prediction method based on decomposition prediction and error correction, and belongs to the field of vehicle networking communication technology. Background Art

[0002] As an important part of the intelligent transportation system, the Internet of Vehicles is gradually changing people's transportation methods. Through information exchange between vehicles and interconnection with infrastructure, the Internet of Vehicles communication technology can realize real-time traffic information monitoring, intelligent navigation, vehicle remote control and other functions, greatly improving traffic safety and traffic efficiency. However, factors such as Doppler frequency shift and multipath fading in the rapid movement of vehicles will make the Internet of Vehicles channel data lose timeliness, which will lead to serious deterioration of the performance of the Internet of Vehicles communication system. Channel prediction technology can provide reliable channel data, so as to better adjust the transmission mode of the Internet of Vehicles communication system to adapt to the current channel status. Therefore, accurately predicting the channel data of the Internet of Vehicles communication system has become the key to optimizing the performance of the Internet of Vehicles communication and improving the overall operation efficiency of the transportation system.

[0003] In the study of channel prediction for IoV communication systems, traditional channel prediction methods based on statistical methods need to meet certain statistical assumptions and perform well in linear stationary data, but the prediction accuracy of such methods is not ideal for large-scale networks and IoV channels with nonlinear changes. Deep learning is a promising tool that has been proven to have good results in processing data with strong randomness and non-stationarity. In IoV channel prediction, deep learning has competitive performance and affordable and reasonable computational costs in processing complex channel data. However, as the prediction step increases, the long-term error accumulation of IoV channel prediction methods based on deep learning will increase sharply during the prediction process, which will lead to a decrease in prediction accuracy. In addition, when processing small sample data, IoV channel prediction methods based on deep learning may not be able to fully learn the complex features in the channel data, making it difficult to achieve high prediction accuracy. Summary of the invention

[0004] In view of the problems existing in the above-mentioned prior art, the present invention provides a vehicle network channel prediction method based on decomposition prediction and error correction, which can solve the problems of insufficient data in channel modeling and long-term error accumulation in the prediction process in the existing vehicle network channel prediction technology, and can achieve higher vehicle network channel prediction accuracy, thereby facilitating better optimization of the performance of the vehicle network communication system.

[0005] To achieve the above purpose, the vehicle network channel prediction method based on decomposition prediction and error correction specifically includes the following steps:

[0006] Step 1, after collecting the IoV channel data, normalize the collected IoV channel data;

[0007] Step 2, decompose the normalized IoV channel data into multiple sub-components with intrinsic modal components of different center frequencies through VMD;

[0008] Step 3, for the vehicle network channel sub-component data y=[f1,f2,...,f n ], where f i For the data at the i-th moment, each subcomponent is divided into a training set and a test set. A channel prediction model based on the BiGRU model is established for the channel prediction optimization problem, and the prediction results of each subcomponent are completed. The prediction results of each subcomponent are superimposed to obtain a preliminary prediction result.

[0009] Step 4: Calculate the prediction difference of the sub-components, process the prediction error of the sub-components by the error correction method, and obtain the error prediction result

[0010] Step 5: Preliminary prediction results And the error prediction results The final vehicle network channel prediction results obtained by superposition are as follows:

[0011]

[0012] Where: This is the final prediction result of the Internet of Vehicles channel.

[0013] Furthermore, Step 2 is as follows:

[0014] Step 2-1, construct a variational problem: obtain the analytical signal of each subcomponent through Hilbert transform, and multiply the analytical signal by the estimated center frequency It is modulated to the corresponding base spectrum, and the square of the second norm of the gradient of the analytical signal is estimated by Gaussian smoothing to obtain the signal bandwidth of each modal component, thereby obtaining a constrained variational model:

[0015]

[0016] Where: u k is the subcomponent, represents partial derivative, δ(t) represents Dirac distribution, {ω k} is the set of center frequencies, K is the sub-component data, and f is the original signal;

[0017] Step 2-2, variational problem solution: introduce penalty factor α and Lagrange multiplier λ(t), transform the constrained variational problem into an unconstrained variational problem, and use the alternating direction operator method to solve it, and get the center frequency and each sub-component as follows:

[0018]

[0019] Where: is the center frequency, For subcomponents.

[0020] Furthermore, when each subcomponent is divided into a training set and a test set in Step 3, for the vehicle network channel subcomponent data y=[f1,f2,...,f n ], where f i is the data at the i-th moment, and the sliding prediction method is used for prediction. The window size is m, where m ≥ 2. Divide y into nm windows. Then, the first window w1 = [f1, f2, ..., f m-1 ,f m ], the i-th window w i =[f i ,f i+1 ,...,f i+m-2 ,f i+m-1 ], the last window is w n-m =[f n-m ,f n-m+1 ,...,f n-2 ,f n-1 ], the output is multiple for multi-step prediction.

[0021] Furthermore, for the optimization problem of channel prediction in Step 3, the channel prediction function can be expressed as:

[0022]

[0023] Where: is the predicted channel data at time t+1; f(·) is a channel prediction function; R is the history length used for prediction; h(t) is the actual channel data at time t;

[0024] The optimization problem of channel prediction is defined as:

[0025]

[0026] Where: h(t+1) is the channel data at time t+1, Predict channel data for time t+1.

[0027] Furthermore, when establishing a channel prediction model based on the BiGRU model in Step 3, the number of BiGRU layers is set to 1, the regularization parameter is set to 0.001, the activation function is set to sigmoid, and the training data is input into the BiGRU. The model calculates the output value through forward propagation. The BiGRU combines the previous and next information of the sequence for calculation, generates the output of each time step, calculates the loss function between the predicted value and the true value, calculates the gradient through the back-propagation algorithm, and uses the Adam optimizer to update the model parameters. Multiple rounds of training are performed, and the model parameters are updated in each round of training so that the loss function gradually converges.

[0028] Furthermore, in Step 4, the prediction error of the sub-component is processed by the error correction method as follows:

[0029] For a channel data y=[f1,f2,...,f n ], the prediction step length is m, and the test set length is l, where m<<l<n;

[0030] Using channel data y = [f1, f2, ..., f n ] and the preliminary prediction results obtained by decomposition prediction Get the first layer error data e1 = [e m+1 ,e m+2 ,...,e n ], specifically:

[0031] Take e1 as experimental data to predict and get the prediction result of the first layer error data

[0032] By e1 and Get the second layer error data e2 = [e 2m+1 ,e 2m+2 ,...,e n ], specifically:

[0033] The prediction result of the second layer error data is obtained by e2

[0034] By analogy, the prediction result of the error data of the i-th layer is

[0035] Furthermore, the termination condition of the error calculation is set as:

[0036]

[0037] Where: represents the expected sum of squares, and Represent the prediction results of the error data of the i-th layer and the i-1-th layer respectively.

[0038] Furthermore, after completing the iteration, the final error prediction result is:

[0039]

[0040] In the formula, is the final error prediction result, i is the number of error layers, l is the length of the test set, and n is the length of the channel data.

[0041] Furthermore, in Step 2, the number of subcomponents of VMD is set to 6.

[0042] Compared with the prior art, the vehicle network channel prediction method based on decomposition prediction and error correction first uses VMD to decompose the channel data into several sub-components, reduces the complexity of the channel data, and can better capture the frequency characteristics in the signal. Then, a BiGRU prediction model is established for each sub-component, and the prediction results of each sub-component are superimposed to obtain a preliminary prediction result. Next, the prediction difference of the sub-component is calculated, and the prediction error of the sub-component is processed by the error correction method to obtain the error prediction result. Finally, the preliminary prediction result and the error prediction result are superimposed to complete the prediction. Experiments have shown that the vehicle network channel prediction method based on decomposition prediction and error correction can achieve a high prediction accuracy when the number of sub-components is 6. Further correction of the vehicle network channel prediction error can improve the prediction accuracy. It can solve the problems of insufficient data in channel modeling and long-term error accumulation in the prediction process in the existing vehicle network channel prediction technology, achieve a high vehicle network channel prediction accuracy, and thus facilitate better optimization of the performance of the vehicle network communication system. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a flow chart of the present invention;

[0044] Figure 2 is a schematic diagram of error correction of the present invention;

[0045] Figure 3 It is a diagram of prediction results of the present invention with different numbers of sub-components and iteration times;

[0046] Figure 4 It is the prediction error QQ (sample data - standard normal) diagram of the vehicle networking channel of the present invention;

[0047] Figure 5 It is a diagram of the prediction results of the present invention at different step sizes. DETAILED DESCRIPTION

[0048] The present invention is further described below by taking vehicle network channel data with a communication frequency band of sub-6GHz (5.9GHz) and containing information of multiple vehicles and multiple road terminals as an example in conjunction with the accompanying drawings.

[0049] like Figure 1 As shown, the vehicle network channel prediction method based on decomposed prediction (Decomposed Prediction) and error correction (ErrorCorrection) is a DE-BiGRU (bidirectional gated recurrent neural network model with decomposed prediction and error correction) method, which specifically includes the following steps:

[0050] Step 1, after collecting the Internet of Vehicles channel data, the collected Internet of Vehicles channel data is normalized. In this embodiment, channel impulse response data with a length of 3000 is selected as the channel data, and the channel data is normalized.

[0051] Step 2, decompose the normalized IoV channel data into multiple sub-components with intrinsic modal components of different center frequencies through variational mode decomposition (VMD), as follows:

[0052] Step 2-1, construct a variational problem: obtain the analytical signal of each subcomponent through Hilbert transform, and multiply the analytical signal by the estimated center frequency e -jωkt , modulate it to the corresponding base spectrum, estimate the square of the second norm of the gradient of the analytical signal through Gaussian smoothing, and obtain the signal bandwidth of each modal component, thereby obtaining a constrained variational model:

[0053]

[0054] Where: u k is the subcomponent, represents partial derivative, δ(t) represents Dirac distribution, {ω k} is the set of center frequencies, K is the sub-component data, and f is the original signal.

[0055] Step 2-2, variational problem solution: introduce penalty factor α and Lagrange multiplier λ(t), transform the constrained variational problem into an unconstrained variational problem, and use the alternating direction operator method to solve it, and get the center frequency and each sub-component as follows:

[0056]

[0057] Where: is the center frequency, For subcomponents.

[0058] Step 3, for the vehicle network channel sub-component data y=[f1,f2,...,f n ], where f i For the data at the i-th moment, each subcomponent is divided into a training set and a test set. In this embodiment, the first 80% is used as a training set and the last 20% is used as a test set. A channel prediction model based on a bidirectional gated recurrent neural network model (BidirectionalGated Recurrent Unit, BiGRU) is established for the optimization problem of channel prediction, and the prediction result of each subcomponent is completed. The prediction results of each subcomponent are superimposed to obtain a preliminary prediction result

[0059] When each subcomponent is divided into a training set and a test set, for the vehicle network channel subcomponent data y=[f1,f2,...,f n ], where f i is the data at the i-th moment, and the sliding prediction method is used for prediction. The window size is m, where m ≥ 2. Divide y into nm windows. Then, the first window w1 = [f1, f2, ..., f m-1 ,f m ], the i-th window w i =[f i ,f i+1 ,...,f i+m-2 ,f i+m-1 ], the last window is w n-m =[f n-m ,f n-m+1 ,...,f n-2 ,f n-1 ], the output is multiple for multi-step prediction.

[0060] For the optimization problem of channel prediction, the channel prediction function can be expressed as:

[0061]

[0062] Where: is the predicted channel data at time t+1; f(·) is a channel prediction function; R is the history length used for prediction; h(t) is the actual channel data at time t.

[0063] The optimization problem of channel prediction is defined as:

[0064]

[0065] Where: h(t+1) is the channel data at time t+1, Predict channel data for time t+1.

[0066] When establishing a channel prediction model based on the BiGRU model, set the number of BiGRU layers to 1, set the regularization parameter to 0.001, set the activation function to sigmoid, input the training data into the BiGRU, and the model calculates the output value through forward propagation. The BiGRU combines the previous and next information of the sequence for calculation, generates the output of each time step, calculates the loss function between the predicted value and the true value, calculates the gradient through the back propagation algorithm, and uses the Adam optimizer to update the model parameters. Perform multiple rounds of training, and set the epoch to 100. The model parameters are updated in each round of training to gradually converge the loss function.

[0067] Step 4, calculate the prediction difference of the sub-component, and process the prediction error of the sub-component by the error correction method, such as Figure 2 As shown, the error prediction result is obtained

[0068] For a channel data y=[f1,f2,...,f n ], the prediction step length is m, and the test set length is l, where m<<l<n.

[0069] Using channel data y = [f1, f2, ..., f n ] and the preliminary prediction results obtained by decomposition prediction Get the first layer error data e1 = [e m+1 ,e m+2 ,...,e n ], specifically:

[0070] Take e1 as experimental data to predict and get the prediction result of the first layer error data

[0071]

[0072] By e1 and The second layer error data e2 = [e 2m+1 ,e 2m+2 ,...,e n ], specifically:

[0073]

[0074] The prediction result of the second layer error data can be obtained from e2

[0075] Similarly, the prediction result of the error data of the i-th layer is

[0076] The prediction accuracy of the IoV channel data can be improved by error correction. However, when the number of corrections is too many, the prediction results will diverge and the accuracy of the prediction results will be reduced. Therefore, the termination condition of the error calculation is set as:

[0077]

[0078] Where: represents the expected sum of squares, and Represent the prediction results of the error data of the i-th layer and the i-1-th layer respectively.

[0079] After completing the iteration, the final error prediction result is:

[0080]

[0081] In the formula, is the final error prediction result, i is the number of error layers, l is the length of the test set, and n is the length of the channel data.

[0082] Step 5: Preliminary prediction results And the error prediction results The final vehicle network channel prediction results obtained by superposition are as follows:

[0083]

[0084] Where: This is the final prediction result of the Internet of Vehicles channel.

[0085] In order to verify the prediction accuracy of the vehicle network channel prediction method based on decomposition prediction and error correction, the DE-BiGRU (bidirectional gated recurrent neural network model with decomposition prediction and error correction) method of the present invention is used as a benchmark model, and the BiGRU (bidirectional gated recurrent neural network model) method and the D-BiGRU (bidirectional gated recurrent neural network model with decomposition prediction) method are added as comparisons, wherein the BiGRU method and the D-BiGRU method do not consider error correction.

[0086] The prediction results of different numbers of subcomponents when the number of iterations is 50, 100, and 200 are as follows: Figure 3As shown. When the number of sub-components is less than 6, the NMSE of the D-BiGRU method and BiGRU is much larger than that of the DE-BiGRU method, indicating that the DE-BiGRU method can effectively suppress the prediction error. When the number of sub-components of VMD exceeds 6, the NMSE of the DE-BiGRU method tends to converge, so the number of sub-components of VMD is set to 6 in the simulation. The prediction accuracy of the D-BiGRU method and the DE-BiGRU method increases with the increase of the number of iterations. When the number of iterations increases from 50 to 100, the prediction accuracy is greatly improved. When the number of iterations increases from 100 to 200, the prediction accuracy is further improved, but the improvement is limited, so the number of iterations is set to 200. As shown Figure 3 The prediction results shown show that the prediction accuracy of the DE-BiGRU method is better than that of the D-BiGRU method, and it is more suitable for the prediction of Internet of Vehicles channel data.

[0087] Considering that the prediction error of the IoV channel data may have certain regularities, the prediction error sequence QQ (sample data-standard normal) from step 1 to step 10 is analyzed as shown in the figure Figure 4 As shown in the figure, it can be seen that the relationship between the sample data in the QQ graph and the standard normal distribution is approximately a straight line, and only a small number of points are far away from the straight line, indicating that the distribution of the prediction error sequence can be approximated as a normal distribution, that is, the mean of the prediction error sequence is 0 and the variance is a constant. The approximate normal distribution of the prediction error of the Internet of Vehicles channel shows that it is predictable.

[0088] The performance of multi-step methods for different prediction methods is as follows Figure 5 As shown. The NMSE of each method increases with the increase of the prediction step size, indicating that the prediction accuracy of each method is inversely proportional to the prediction step size. This is because the information loss and uncertainty accumulation will increase with the increase of the step size, thereby increasing the prediction error. The prediction accuracy of the D-BiGRU method and the DE-BiGRU method is much higher than that of the BiGRU method. This is because after they decompose the channel data into multiple sub-components of different frequencies, the single sub-components have strong correlation and regularity, which can reduce the prediction difficulty of the prediction module. Compared with the D-BiGRU method, the DE-BiGRU method has higher prediction accuracy, but with the increase of the step size, the correlation within the prediction error is greatly reduced, which leads to the gradual weakening of the error compensation ability of the DE-BiGRU method. Overall, the DE-BiGRU method has the best performance in predicting the channel data of the Internet of Vehicles.

[0089] This IoV channel prediction method based on decomposition prediction and error correction has the ability to decompose prediction and suppress long-term error accumulation. First, the channel data is decomposed into multiple sub-components using variational mode decomposition, which can better capture the frequency characteristics in the signal. The prediction accuracy improvement effect for small sample channel data is more obvious, and the recommended number of sub-components is given; then the prediction error of the IoV channel is analyzed, and it is observed that its error distribution approximately conforms to the normal distribution and is predictable; finally, the prediction results of the sub-components and the prediction results of the prediction error are superimposed to complete the IoV channel prediction. Through the above experimental simulation verification, this IoV channel prediction method based on decomposition prediction and error correction can solve the problems of insufficient data in channel modeling and long-term error accumulation in the prediction process in the existing vehicle network channel prediction technology, achieve higher IoV channel prediction accuracy, and thus facilitate better optimization of the performance of the IoV communication system.

Claims

1. A vehicle network channel prediction method based on decomposition prediction and error correction, characterized in that: The specific steps include: Step 1, after collecting the IoV channel data, normalize the collected IoV channel data; Step 2, decompose the normalized IoV channel data into multiple sub-components with intrinsic modal components of different center frequencies through VMD; Step 2-1, construct a variational problem: obtain the analytical signal of each subcomponent through Hilbert transform, and multiply the analytical signal by the estimated center frequency It is modulated to the corresponding base spectrum, and the square of the second norm of the gradient of the analytical signal is estimated by Gaussian smoothing to obtain the signal bandwidth of each modal component, thereby obtaining a constrained variational model: Where: u k is the subcomponent, represents partial derivative, δ(t) represents Dirac distribution, {ω k } is the set of center frequencies, K is the sub-component data, and f is the original signal; Step 2-2, variational problem solution: introduce penalty factor α and Lagrange multiplier λ(t), transform the constrained variational problem into an unconstrained variational problem, and use the alternating direction operator method to solve it, and get the center frequency and each sub-component as follows: Where: is the center frequency, is the subcomponent; Step 3, for the vehicle network channel sub-component data y=[f1,f2,...,f n ], where f i For the data at the i-th moment, each subcomponent is divided into a training set and a test set. A channel prediction model based on the BiGRU model is established for the channel prediction optimization problem, and the prediction results of each subcomponent are completed. The prediction results of each subcomponent are superimposed to obtain a preliminary prediction result. When each subcomponent is divided into a training set and a test set, for the vehicle network channel subcomponent data y=[f1,f2,...,f n ], where f i is the data at the i-th moment, and the sliding prediction method is used for prediction. The window size is m, where m ≥ 2. Divide y into nm windows, then the first window w1 = [f1, f2, ..., f m-1 ,f m ], the i-th window w i =[f i ,f i+1 ,...,f i+m-2 ,f i+m-1 ], the last window is w n-m =[f n-m ,f n-m+1 ,...,f n-2 ,f n-1 ], the output is multiple when multi-step prediction; Step 4: Calculate the prediction difference of the sub-components, process the prediction error of the sub-components by the error correction method, and obtain the error prediction result The prediction error of the sub-component is processed by the error correction method as follows: For a channel data y=[f1,f2,...,f n ], the prediction step length is m, and the test set length is l, where m<<l<n; Using channel data y = [f1, f2, ..., f n ] and the preliminary prediction results obtained by decomposition prediction Get the first layer error data e1 = [e m+1 ,e m+2 ,...,e n ], specifically: Take e1 as experimental data to predict and get the prediction result of the first layer error data By e1 and Get the second layer error data e2 = [e 2m+1 ,e 2m+2 ,...,e n ], specifically: The prediction result of the second layer error data is obtained by e2 By analogy, the prediction result of the error data of the i-th layer is Step 5: Preliminary prediction results And the error prediction results The final vehicle network channel prediction results obtained by superposition are as follows: Where: This is the final prediction result of the Internet of Vehicles channel.

2. The vehicle network channel prediction method based on decomposition prediction and error correction according to claim 1 is characterized in that: In Step 3, for the optimization problem of channel prediction, the channel prediction function can be expressed as: Where: is the predicted channel data at time t+1; f(·) is a channel prediction function; R is the history length used for prediction; h(t) is the actual channel data at time t; The optimization problem of channel prediction is defined as: Where: h(t+1) is the channel data at time t+1, Predict channel data for time t+1.

3. The vehicle network channel prediction method based on decomposition prediction and error correction according to claim 1, characterized in that: When establishing a channel prediction model based on the BiGRU model in Step 3, set the number of BiGRU layers to 1, set the regularization parameter to 0.001, set the activation function to sigmoid, input the training data into the BiGRU, and the model calculates the output value through forward propagation. The BiGRU combines the previous and next information of the sequence for calculation, generates the output of each time step, calculates the loss function between the predicted value and the true value, calculates the gradient through the back-propagation algorithm, and uses the Adam optimizer to update the model parameters. Perform multiple rounds of training, and update the model parameters in each round of training to gradually converge the loss function.

4. The vehicle network channel prediction method based on decomposition prediction and error correction according to claim 1, characterized in that: Set the termination condition of error calculation to: Where: represents the expected sum of squares, and Represent the prediction results of the error data of the i-th layer and the i-1-th layer respectively.

5. The vehicle network channel prediction method based on decomposition prediction and error correction according to claim 4 is characterized in that: After completing the iteration, the final error prediction result is: In the formula, is the final error prediction result, i is the number of error layers, l is the length of the test set, and n is the length of the channel data.

6. The vehicle network channel prediction method based on decomposition prediction and error correction according to claim 1, characterized in that: In Step 2, the number of subcomponents of VMD is set to 6.

Images