Frequency spectrum prediction method for channel separation

Abstract

The invention discloses a frequency spectrum prediction method for channel separation, and belongs to the technical field of frequency spectrum prediction. According to the invention, the problems of low prediction efficiency, low prediction precision and large storage resource occupation of the existing spectrum prediction method are solved. According to the method, the channel dimension and the batch dimension are fused, the channel separation is realized, the focus point of the spectrum prediction model is put on the extraction of time correlation, and the influence of channel interference and large distribution difference of channel data on the model feature extraction process is avoided. After the model outputs the prediction result, the batch dimension is split, the result of each channel is restored to the original dimension, the future information of all channels is predicted at the same time, the method combines the advantages of the channel independence and channel joint prediction method, and the precision and efficiency of spectrum data prediction can be improved. And through sharing of the multi-channel model, the generalization ability of the model is improved, and resource occupation is reduced. The method can be applied to the technical field of spectrum prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of spectrum prediction, and in particular relates to a spectrum prediction method for channel separation. Background Art

[0002] In recent years, with the rapid development of mobile Internet, Internet of Things (IoT) and Internet of Vehicles (IoV), the demand for limited spectrum resources has been growing. This surging demand urgently requires smarter and more complete spectrum utilization strategies. To achieve this goal, it is first necessary to be able to obtain the state of the electromagnetic spectrum in advance. There are two main ways to obtain this information: spectrum sensing and spectrum prediction. Spectrum sensing uses various signal processing methods to passively detect the current state of the radio spectrum. In contrast, spectrum prediction actively analyzes historical spectrum usage patterns to predict future usage. Accurate spectrum prediction (low prediction loss) can help new users identify idle or underutilized spectrum, thereby achieving dynamic spectrum access and improving overall spectrum utilization. However, due to the development of wireless communication technology and the deployment of numerous heterogeneous wireless communication systems, spectrum prediction still faces many challenges brought by complex electromagnetic environments.

[0003] Spectrum prediction is to predict the future spectrum usage of a channel by using temporal correlation and inter-channel correlation. At present, the development of neural networks has enabled single-channel prediction to achieve a high prediction accuracy. As we all know, multi-channel prediction is to simultaneously predict the spectrum status of multiple channels in multiple steps in the future, but there are still many problems in multi-channel prediction. There are two main multi-channel prediction methods: channel-independent prediction and channel-joint prediction. The channel-joint prediction method is to train a model for multiple channels and use the model to predict the future results of all channels. This method has high prediction efficiency. The similarity of channels can provide inter-channel correlation when channel data is missing or insufficient. This correlation can assist low-quality channels in prediction. When there is sufficient data between channels, temporal correlation dominates the prediction process. At this time, the similarity between channels is not enough to provide additional information. Instead, the prediction accuracy will be reduced due to interference between channels and large differences in channel distribution. For this, we should adopt a channel-independent method. The channel-independent method trains a model for each channel for prediction, which reduces interference between channels. Large differences in channel distribution will not affect the feature extraction process of the model. However, in this method, each model focuses more on its own channel, which inevitably leads to a reduction in its generalization ability. Predicting each channel in turn will also bring about a huge time delay, reducing the prediction efficiency. In addition, the resource usage of this method is closely related to the number of channels, and a large number of channels will occupy huge storage resources. Summary of the invention

[0004] The purpose of the present invention is to solve the problems of low prediction efficiency, low prediction accuracy and large storage resource occupation in the existing spectrum prediction methods, and to propose a spectrum prediction method with channel separation.

[0005] The technical solution adopted by the present invention to solve the above technical problems is: a spectrum prediction method for channel separation, the method specifically comprising the following steps:

[0006] Step 1: Define the spectrum data that needs to be predicted for the next P time slots, and build a corresponding spectrum prediction model based on the value of P;

[0007] And define the number of historical time slots H∈O corresponding to the historical spectrum data sequence input by the spectrum prediction model, where O represents the set of all values ​​of H;

[0008] Step 2: Record the historical spectrum data sequence input into the spectrum prediction model as in, represents a real number, M represents the number of channels that need to be predicted for spectrum data in each batch, and B represents the number of batches;

[0009] The spectrum data of each channel in the historical spectrum data sequence X is normalized respectively, and the channel dimension of the normalized input data is fused to the batch dimension to obtain the processed historical spectrum data sequence

[0010] Step 3: Use the processed historical spectrum data sequence X′ to train the spectrum prediction model to obtain the trained spectrum prediction model corresponding to the current H value;

[0011] Step 4: For each value of H in the set O, steps 2 and 3 are executed, that is, the trained spectrum prediction model corresponding to each H value is obtained respectively;

[0012] Then, the spectrum prediction model with the smallest mean square error loss is selected from the trained spectrum prediction models corresponding to each H value;

[0013] Step 5: After processing the historical spectrum data sequence of the channel to be predicted, the processed data is input into the selected spectrum prediction model to obtain the spectrum data prediction result of the channel to be predicted;

[0014] The batch dimension of the spectrum data prediction result is then split and restored to the original channel dimension to obtain the final prediction result.

[0015] Furthermore, the corresponding spectrum prediction model is constructed according to the value of P, specifically:

[0016] When P=1, the constructed spectrum prediction model is a LSTM-based Seq2Seq model or an improved Autoformer model;

[0017] When P>1, the constructed spectrum prediction model is an improved Autoformer model.

[0018] Furthermore, the improved Autoformer model includes an encoder and a decoder, and the encoder includes N en The decoder includes N de A decoding unit;

[0019] In the improved Autoformer model, the input of the improved Autoformer model is passed through the linear mapping layer, the output of the linear mapping layer is used as the input of the first encoding unit, and the seasonal part of the output of the first encoding unit is used as the input of the second encoding unit, and the seasonal part of the output of the second encoding unit is used as the input of the third encoding unit, and so on. en The seasonal part of the output of the -1 coding unit is taken as the Nth en The input of the encoding unit is obtained en The seasonal component of the coding unit output;

[0020] The input of the improved Autoformer model is decomposed into the cumulative structure of the trend cycle component and the cumulative structure of the seasonal component, and the cumulative structure of the decomposed trend cycle component, the cumulative structure of the seasonal component and the Nth en The seasonal part output by the encoding unit is used as the input of the first decoding unit, and the trend cycle part, seasonal part and the Nth en The seasonal part of the output of the encoding unit is used as the input of the second decoding unit, and so on. de -1 The trend cycle part, seasonal part and Nth part of the decoding unit output en The seasonal part of the encoding unit output is used as the Nth de The input of the decoding unit is based on the Nth de The output of the decoding unit obtains the spectrum data prediction result.

[0021] Furthermore, the working process of the encoder is:

[0022] First, the processed historical spectrum data sequence X′ is dimensionally expanded to d dimensions, and the dimension expansion result X en for:

[0023]

[0024] Among them, Embedding(·) represents the linear mapping layer;

[0025] Then the output X of the linear mapping layer en As the input of the first coding unit, in the first coding unit, the input of the first coding unit is used as the input of the first correlation extraction module, then the output X of the first correlation extraction module is (1) for:

[0026] X (1) =Auto-Correlation1(X en )

[0027] Wherein, Auto-Correlation1(·) represents the operation of the first correlation extraction module;

[0028] The output of the first correlation extraction module is compared with X en Add, and the result is X (2) for:

[0029] X (2) =Auto-Correlation1(X en )+X en

[0030] Add the result X (2) As the input of the first sequence decomposition module, the trend cycle part X is output by the first sequence decomposition module t(3) and seasonal part X s(3) ;

[0031] The seasonal part X s(3) As the input of the first forward propagation module, the output X of the first forward propagation module (4) for:

[0032] X (4) =FFN1(X s(3) )

[0033] Where FFN1(·) represents the operation of the first forward propagation module;

[0034] Combine the output of the first forward propagation module with the seasonal part X s(3) Add them together to get the result X (5) :

[0035] X (5) =X (4) +X s(3)

[0036] Add the result X (5) As the input of the second sequence decomposition module, the trend cycle part of the second sequence decomposition module output is recorded as Xt(6) , the seasonal part of the output of the second sequence decomposition module is recorded as X s(6) , and X s(6) The seasonal part as output of the first encoding unit;

[0037] Then the seasonal part X output by the first encoding unit is s(6) as the input of the second coding unit; and the seasonal part output by the second coding unit is further used as the input of the third coding unit;

[0038] And so on, the Nth en The seasonal part of the output of the -1 coding unit is taken as the Nth en The input of the encoding unit is obtained en The seasonal part of the encoding unit output

[0039] Furthermore, the cumulative structure of the trend cycle component and the cumulative structure of the seasonal component are obtained by:

[0040]

[0041] Among them, the processed historical spectrum data sequence X′ is passed through the sequence decomposition module to obtain the historical trend cycle sequence and the historical seasonal sequence. It is the last H / 2 row of the historical trend cycle sequence. represents the last H / 2 rows of the historical seasonal series, mean(X′ H ) P represents a vector of P rows and 1 column consisting of the mean values ​​of each row in the historical spectrum data sequence X′, O P represents a 0 vector with P rows and 1 columns, represents the cumulative structure of the trend period component, Represents the cumulative structure of the seasonal component.

[0042] Furthermore, the working process of the decoder is:

[0043] Will And Nth en The seasonal part of the encoding unit output As the input of the first decoding unit, in the first decoding unit, As the input of the second correlation extraction module, the output of the second correlation extraction module is then combined with Add, and use the addition result as the input of the third sequence decomposition module, and output X through the third sequence decomposition module s(7) and X t(7) :

[0044]

[0045] Wherein, Auto-Correlation2(·) represents the operation of the second correlation extraction module, and SeriesDecomp3(·) represents the operation of the third series decomposition module;

[0046] X s(7) and As the input of the third correlation extraction module, the output of the third correlation extraction module and X s(7) Add, and use the addition result as the input of the fourth sequence decomposition module to obtain the output X of the fourth sequence decomposition module s(8) and X t(8) :

[0047]

[0048] Wherein, Auto-Correlation3(·) represents the operation of the third correlation extraction module, and SeriesDecomp4(·) represents the operation of the fourth series decomposition module;

[0049] X s(8) As the input of the second forward propagation module, the output of the second forward propagation module is combined with X s(8) Add, and use the added result as the input of the fifth sequence decomposition module to obtain the output X of the fifth sequence decomposition module s(9) and X t(9) :

[0050] X s(9) ,X t(9) = SeriesDecomp5(FFN2(X s(8) )+X s(8) )

[0051] Wherein, FFN2(·) represents the operation of the second forward propagation module; SeriesDecomp5(·) represents the operation of the fifth series decomposition module;

[0052] Will X t(7) , X t(8) and X t(9) Add together to get the sum of the trend cycle components

[0053]

[0054] X s(9) As the seasonal part of the output of the first decoding unit, As the trend cycle part of the first decoding unit output, X s(9) and as an input of a second decoding unit;

[0055] And so on, we get the Nth de The seasonal part and trend cycle part of the decoding unit output; de The seasonal part of the decoding unit output is denoted by The Nth de The trend cycle part of the decoding unit output is recorded as Then the Nth de Trend cycle part of the decoding unit output and seasonal part merge:

[0056]

[0057] Among them, Y de Indicates the merge result;

[0058] Then merge the result Y de Mapping is performed to obtain the spectrum data prediction result Y′:

[0059] Y′=W o Y de +b o

[0060] Among them, W o and b o is the model parameter.

[0061] Furthermore, the working process of the first correlation extraction module is:

[0062] Step 1: X en After linear mapping, we get the sequences q, k and v;

[0063] Step 2, initialize τ=1;

[0064] Step 3: cyclically shift the sequence v to the right according to the time delay τ to obtain the result Roll(v,τ) after the cyclic shift to the right;

[0065] Calculate X again en The autocorrelation function

[0066]

[0067] Among them, q t represents the tth element in the sequence q, k t-τ represents the t-τth element in sequence k;

[0068] Step 4: Determine whether τ=H is satisfied;

[0069] If satisfied, proceed to step 5;

[0070] If not satisfied, set τ = τ + 1 and return to step 3;

[0071] Step 5: According to Select the k most correlated delays:

[0072]

[0073] Among them, τ1, τ2, ..., τ k represents the selected k delays;

[0074] For the selected k delays corresponding to To process:

[0075]

[0076] in, Represent k delays τ1, τ2, ..., τ k The corresponding weights, softmax(·) is the normalized exponential function;

[0077] Then the output of the first correlation extraction module is Auto-Correlation(q,k,v):

[0078]

[0079] Furthermore, the first correlation extraction module performs calculations based on FFT and IFFT, and the specific calculation method is:

[0080] Step 1: Use sequence q as the input of the first linear layer, sequence k as the input of the second linear layer, and sequence v as the input of the third linear layer;

[0081] Step 2: Pass the output of the first linear layer through the first FFT module, pass the output of the second linear layer through the second FFT module, multiply the output of the first FFT module by the output of the second FFT module, and use the multiplication result as the input of the first IFFT module. Output the autocorrelation function value corresponding to each delay through the first IFFT module.

[0082] Then select the delays corresponding to the first k largest autocorrelation function values, and record the selected k delays as τ1, τ2, ..., τ k ;

[0083] Step 3: Delay the output of the third linear layer by τ1, τ2, ..., τ k The delay processing is to obtain delays τ1, τ2, ..., τk The corresponding delayed processing results;

[0084] Step 4: multiply the delay processing result corresponding to delay τ1 by the autocorrelation function value corresponding to delay τ1, multiply the delay processing result corresponding to delay τ2 by the autocorrelation function value corresponding to delay τ2, and so on. k Corresponding delay processing results and delay τ k The corresponding autocorrelation function values ​​are multiplied to obtain the delays τ1, τ2, ..., τ k The corresponding multiplication results are then accumulated to obtain an accumulated result;

[0085] The accumulated result is passed through a fourth linear layer, and the output of the fourth linear layer is used as the output of the first correlation extraction module.

[0086] Furthermore, the working process of the first sequence decomposition module is:

[0087] X t(3) =AvgPool(Padding(X (2) ))

[0088] X s(3) =X (2) -X t(3)

[0089] Among them, Padding(·) represents the data padding operation, AvgPool(·) represents the moving average operation, Represents the trend cycle part of the first sequence decomposition module output, Represents the seasonal part of the output of the first series decomposition module.

[0090] Furthermore, the working process of the first forward propagation module is:

[0091] FFN1(X s(3) )=w2relu(w1X s(3) +b1)+b2

[0092] Among them, FFN1(·) represents the operation of the first forward propagation module, w1, w2, b1 and b2 are model parameters, and relu(·) is the activation function.

[0093] The beneficial effects of the present invention are:

[0094] The present invention integrates the channel dimension with the batch dimension to achieve channel separation. The spectrum prediction model focuses on the extraction of time correlation, avoiding the impact of channel interference and large differences in data distribution of each channel on the model feature extraction process. After the model outputs the prediction results, the batch dimension is split, the results of each channel are restored to the original dimension, and the future information of all channels is predicted at the same time. The method of the present invention combines the advantages of channel independence and channel joint prediction methods, and can use a single model to predict the future results of multiple channels at the same time, thereby improving the accuracy and efficiency of spectrum data prediction. The sharing of multi-channel models also prevents the model from paying too much attention to its own channels, improves the generalization ability of the model, and reduces resource usage. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] Figure 1 is the two-dimensional power spectral density plot of the data set;

[0096] Figure 2 is the three-dimensional power spectral density plot of the data set;

[0097] Figure 3 It is a structural diagram of the spectrum prediction model;

[0098] Figure 4 is the spectrum occupancy status prediction diagram;

[0099] Figure 5 It is the structural diagram of the improved Autoformer model;

[0100] Figure 6 is a schematic diagram of the calculation process of the correlation extraction module;

[0101] Figure 7 This is a comparison chart of the average MSE under different prediction methods;

[0102] Figure 8 This is a comparison chart of the average MAE under different prediction methods. DETAILED DESCRIPTION

[0103] Specific implementation method 1: Combination Figure 3 The present embodiment describes a spectrum prediction method for channel separation, and the method specifically comprises the following steps:

[0104] Step 1: define the spectrum data that needs to be predicted for the next P time slots (i.e., the power spectrum density values ​​of the next P time slots need to be predicted), and construct a corresponding spectrum prediction model according to the value of P;

[0105] And define the number of historical time slots H∈O corresponding to the historical spectrum data sequence input by the spectrum prediction model, where O represents a set consisting of all values ​​of H (here some values ​​can be set based on experience to form a set, and H time slots should be selected continuously);

[0106] Step 2: When the value of H is any value in the set O, the historical spectrum data sequence input into the spectrum prediction model is recorded as in, represents a real number, M represents the number of channels that need to be predicted for spectrum data in each batch, and B represents the number of batches;

[0107] The spectrum data of each channel in the historical spectrum data sequence X is normalized respectively, and then the channel dimension of the normalized input data is fused to the batch dimension according to the channel separation method to obtain the processed historical spectrum data sequence

[0108] Step 3: Use the processed historical spectrum data sequence X′ to train the spectrum prediction model, with the training label being the actual spectrum data of the prediction time slot, to obtain the trained spectrum prediction model corresponding to the current H value;

[0109] Step 4: For each value of H in the set O, steps 2 and 3 are executed, that is, the trained spectrum prediction model corresponding to each H value is obtained respectively;

[0110] Then, the spectrum prediction model with the smallest mean square error loss is selected from the trained spectrum prediction models corresponding to each H value;

[0111] Step 5: After processing the historical spectrum data sequence of the channel to be predicted (the processing method is the same as the processing method of the historical spectrum data sequence during training), the processed data is input into the selected spectrum prediction model to obtain the spectrum data prediction result of the channel to be predicted;

[0112] The batch dimension of the spectrum data prediction result is then split and restored to the original channel dimension to obtain the final prediction result.

[0113] In the traditional method, each batch of data input to the model includes spectrum data of multiple channels. Therefore, during the model training process, it will be affected by the mutual interference between channels and the large difference in channel data distribution. The method of the present invention first separates the channels and merges them into the batch dimension, so that each batch contains only the data of one channel, and uses multiple channels to train a common prediction model, thereby reducing the mutual interference between channels and the large difference in channel data distribution, and improving the accuracy of prediction. Then, the improved Autoformer model is used for prediction. In the encoder part, the trend cycle component is gradually separated, and the complex seasonal fluctuation component is extracted by the correlation extraction module. In the decoder, the seasonal component of the part to be predicted is gradually established by the autocorrelation and cross-autocorrelation extraction modules, and the trend cycle component is gradually added to the final prediction result. Finally, the batch dimension of the decoder prediction result is split and restored to the original channel. In the training stage, the spectrum data of the training set is divided by changing the historical time slot length and the prediction time slot length, and the model of the corresponding prediction scenario is trained. The trained model is applied to the actual multi-channel multi-step joint prediction, and the known spectrum power spectral density is used to predict the spectrum data in the unknown future time slot, and the next step of spectrum sensing, spectrum decision-making, spectrum sharing and spectrum switching operations are carried out based on this.

[0114] Specific implementation method 2: This implementation method is different from the specific implementation method 1 in that the corresponding spectrum prediction model is constructed according to the value of P, specifically:

[0115] When P=1, the constructed spectrum prediction model is a LSTM-based Seq2Seq model or an improved Autoformer model;

[0116] When P>1, the constructed spectrum prediction model is an improved Autoformer model.

[0117] The other steps and parameters are the same as those in the first embodiment.

[0118] In the actual training process, we can first determine the value of P according to the actual prediction needs, then build a spectrum prediction model based on the value of P, and then continuously change the value of H to obtain the trained spectrum prediction model under each H value, and then determine the best spectrum prediction model and the H value corresponding to the best spectrum prediction model based on the training results. In the actual prediction process, for each channel, we first obtain the spectrum data of H consecutive time slots with the current time slot as the end point, and then fuse the channel dimension to the batch dimension to obtain the input data. After normalizing the data of each channel in the input data, the normalized result is used as the input of the best spectrum prediction model, and then the batch dimension of the output result of the best spectrum prediction model is split and restored to the original channel dimension to obtain the final prediction result.

[0119] Specific implementation method three: Combination Figure 5 The present embodiment is different from the first or second embodiment in that the improved Autoformer model includes an encoder and a decoder, and the encoder includes N en The decoder includes N de A decoding unit;

[0120] In the improved Autoformer model, the input of the improved Autoformer model is passed through the linear mapping layer, the output of the linear mapping layer is used as the input of the first encoding unit, and the seasonal part of the output of the first encoding unit is used as the input of the second encoding unit, and the seasonal part of the output of the second encoding unit is used as the input of the third encoding unit, and so on. en The seasonal part of the output of the -1 coding unit is taken as the Nth en The input of the encoding unit is obtained en The seasonal component of the coding unit output;

[0121] The input of the improved Autoformer model is decomposed into the cumulative structure of the trend cycle component and the cumulative structure of the seasonal component, and the cumulative structure of the decomposed trend cycle component, the cumulative structure of the seasonal component and the Nth en The seasonal part output by the encoding unit is used as the input of the first decoding unit, and the trend cycle part, seasonal part and the Nth en The seasonal part of the output of the encoding unit is used as the input of the second decoding unit, and so on. de -1 The trend cycle part, seasonal part and Nth part of the decoding unit output en The seasonal part of the encoding unit output is used as the Nth de The input of the decoding unit is based on the Nth de The output of the decoding unit obtains the spectrum data prediction result.

[0122] The other steps and parameters are the same as those in the first or second embodiment.

[0123] In order to ensure that the model can stably perform multi-step predictions, the present invention adopts an improved Autoformer model structure as a spectrum prediction model. The improved Autoformer model is an improved model based on Transformer, which aims to solve the limitations of the traditional Transformer model in time series processing. The improved Autoformer model decomposes the time series into a trend cycle part and a seasonal part, and captures long-term trends and short-term fluctuations through explicit modeling, making it more suitable for processing complex time series data. The number of encoding units and decoding units in the model of the present invention can be any integer.

[0124] Specific implementation method 4: This implementation method is different from any one of the specific implementation methods 1 to 3 in that the working process of the encoder is:

[0125] First, the processed historical spectrum data sequence X′ is dimensionally expanded to d dimensions, and the dimension expansion result X en for:

[0126]

[0127] Among them, Embedding(·) represents the linear mapping layer; the dimension expansion process is equivalent to the process before input encoder;

[0128] Then the output X of the linear mapping layer en As the input of the first coding unit, in the first coding unit, the input of the first coding unit is used as the input of the first correlation extraction module, then the output X of the first correlation extraction module is (1) for:

[0129] X (1) =Auto-Correlation1(X en )

[0130] Wherein, Auto-Correlation1(·) represents the operation of the first correlation extraction module; seasonal information can be extracted through the correlation extraction module;

[0131] The output of the first correlation extraction module is compared with X en Add, and the result is X (2) for:

[0132] X (2) =Auto-Correlation1(X en )+X en

[0133] Residual connections can alleviate the problems of gradient vanishing and gradient exploding, and promote information flow;

[0134] Add the result X (2) As the input of the first sequence decomposition module, the trend cycle part X is output by the first sequence decomposition module t(3) and seasonal part X s(3) ;

[0135] The encoder part mainly models the seasonal part. Therefore, the input of the forward propagation module only contains the seasonal part. s(3) As the input of the first forward propagation module, the output X of the first forward propagation module (4) for:

[0136] X (4) =FFN1(X s(3) )

[0137] Where FFN1(·) represents the operation of the first forward propagation module;

[0138] Combine the output of the first forward propagation module with the seasonal part X s(3) Add them together to get the result X (5) :

[0139] X (5) =X (4) +X s(3)

[0140] Further separate the trend cycle part and the seasonal part, and add the result X (5) As the input of the second sequence decomposition module, the trend cycle part of the second sequence decomposition module output is recorded as X t(6) , the seasonal part of the output of the second sequence decomposition module is recorded as X s(6) , and X s(6) The seasonal part (X) output by the first encoding unit s(6) That is, the output of the first encoding unit);

[0141] Then the seasonal part X output by the first encoding unit is s(6) As the input of the second encoding unit (X s(6) For the second coding unit, it is equivalent to X en For the first coding unit); the seasonal part output by the second coding unit is used as the input of the third coding unit;

[0142] And so on, the Nth en The seasonal part of the output of the -1 coding unit is taken as the Nth en The input of the encoding unit is obtained en The seasonal part of the encoding unit output

[0143] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.

[0144] The present invention adopts N en By stacking coding units with the same structure, we can learn deeper seasonal fluctuations and separate deeper trend cycle components.

[0145] Specific implementation method 5: This implementation method is different from any one of specific implementation methods 1 to 4 in that the cumulative structure of the trend period component and the cumulative structure of the seasonal component are obtained by:

[0146]

[0147] Among them, the processed historical spectrum data sequence X′ is passed through the sequence decomposition module to obtain the historical trend cycle sequence and the historical seasonal sequence. It is the last H / 2 row of the historical trend cycle sequence. represents the last H / 2 rows of the historical seasonal series, mean(X′ H ) P represents a vector of P rows and 1 column composed of the mean values ​​of each row in the historical spectrum data sequence X′ (for any channel in the processed historical spectrum data sequence, the historical trend period data corresponding to the channel is equivalent to a vector of H rows and 1 column. First, select the data of the last H / 2 rows in the vector, then calculate the mean value of the H rows of data corresponding to the channel, and use the selected H / 2 rows of data as the first H / 2 rows of the column vector, that is, and in A vector with P rows and 1 column is added after the supplementary vector, and the value of each element in the supplementary vector is the calculated mean. Similarly, each channel is processed separately, and each channel is processed in parallel). P Represents a 0 vector with P rows and 1 columns (the processing process is similar to the historical trend cycle data, the difference is The elements of the vector added after are all 0), and we get and After that, the linear mapping layer Embedding(·) is used to expand the dimension. represents the cumulative structure of the trend period component, Represents the cumulative structure of the seasonal component.

[0148] The other steps and parameters are the same as those in Specific Embodiments 1 to 4.

[0149] Specific implementation method 6: This implementation method is different from any one of the specific implementation methods 1 to 5 in that the working process of the decoder is as follows:

[0150] Will And Nth enThe seasonal part of the encoding unit output As the input of the first decoding unit, in the first decoding unit, As the input of the second correlation extraction module, the output of the second correlation extraction module is then combined with Add, and use the addition result as the input of the third sequence decomposition module, and output X through the third sequence decomposition module s(7) and X t(7) :

[0151]

[0152] Wherein, Auto-Correlation2(·) represents the operation of the second correlation extraction module, and SeriesDecomp3(·) represents the operation of the third series decomposition module;

[0153] X s(7) and As the input of the third correlation extraction module, the output of the third correlation extraction module and X s(7) Add, and use the addition result as the input of the fourth sequence decomposition module to obtain the output X of the fourth sequence decomposition module s(8) and X t(8) :

[0154]

[0155] Wherein, Auto-Correlation3(·) represents the operation of the third correlation extraction module, and SeriesDecomp4(·) represents the operation of the fourth series decomposition module;

[0156] X s(8) As the input of the second forward propagation module, the output of the second forward propagation module is combined with X s(8) Add, and use the added result as the input of the fifth sequence decomposition module to obtain the output X of the fifth sequence decomposition module s(9) and X t(9) :

[0157] X s(9) ,X t(9) = SeriesDecomp5(FFN2(X s(8) )+X s(8) )

[0158] Wherein, FFN2(·) represents the operation of the second forward propagation module; SeriesDecomp5(·) represents the operation of the fifth series decomposition module;

[0159] Will X t(7) , Xt(8) and X t(9) Add together to get the sum of the trend cycle components

[0160]

[0161] X s(9) As the seasonal part of the output of the first decoding unit, As the trend cycle part of the first decoding unit output, X s(9) and As the input of the second decoding unit (X s(9) Similar to the input of the first decoding unit Similar to the input of the first decoding unit );

[0162] And so on, we get the Nth de The seasonal part and trend cycle part of the decoding unit output; de The seasonal part of the decoding unit output is denoted by The Nth de The trend cycle part of the decoding unit output is recorded as Then the Nth de Trend cycle part of the decoding unit output and seasonal part merge:

[0163]

[0164] Among them, Y de Indicates the merge result;

[0165] Then merge the result Y de Mapping is performed to obtain the spectrum data prediction result Y′:

[0166] Y′=W o Y de +b o

[0167] Among them, W o and b o are the learnable parameters of the model.

[0168] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 5.

[0169] Restore the spectrum data prediction result Y′ to the original channel and get Prediction method Figure 4 As shown. By deBy stacking up decoding units with the same structure, more complex internal relationships can be modeled. Each decoding unit contains two correlation extraction modules, one for modeling future seasonal components and the other for extracting the relationship with past seasonal components. By combining the two, the time dependency of future prediction time steps can be mined by using the periodic properties of the sequence. The trend part information is accumulated on the trend input of the decoding unit to obtain the future trend cycle part.

[0170] The residual connection (i.e., adding the output of the current layer to the value input to the layer) is applied in the improved Autoformer model, which can ensure that the effect of the deep network is better than that of the shallow network.

[0171] Specific implementation method 7: This implementation method is different from any one of specific implementation methods 1 to 6 in that the working process of the first correlation extraction module (Auto-Correlation) is as follows:

[0172] Step 1: X en After linear mapping, we get the sequences q, k and v, which can represent X en More information in

[0173] Step 2, initialize τ=1;

[0174] Step 3: cyclically shift the sequence v to the right according to the time delay τ to obtain the result Roll(v,τ) after the cyclic shift to the right;

[0175] Calculate X again en The autocorrelation function

[0176]

[0177] Among them, q t represents the tth element in the sequence q, k t-τ represents the t-τth element in sequence k;

[0178] Step 4: Determine whether τ=H is satisfied;

[0179] If satisfied, proceed to step 5;

[0180] If not satisfied, set τ = τ + 1 and return to step 3;

[0181] Step 5: According to Select the k most correlated delays:

[0182]

[0183] Among them, τ1, τ2, ..., τ k represents the selected k delays; select The first k delays with larger values ​​of , k delays represent the sequence X en The most likely cycle length, the autocorrelation function corresponding to each delay is the probability of each cycle;

[0184] For the selected k delays corresponding to To process:

[0185]

[0186] in, Represent k delays τ1, τ2, …, τ k The corresponding weights, softmax(·) is the normalized exponential function;

[0187] Then the output of the first correlation extraction module is Auto-Correlation(q,k,v):

[0188]

[0189] The other steps and parameters are the same as those in Specific Embodiments 1 to 6.

[0190] Specific implementation method eight: Combination Figure 6 The present embodiment is different from the first to seventh embodiments in that the first correlation extraction module performs calculations based on FFT and IFFT, and the specific calculation method is:

[0191] Step 1: Use sequence q as the input of the first linear layer, sequence k as the input of the second linear layer, and sequence v as the input of the third linear layer;

[0192] Step 2: Pass the output of the first linear layer through the first FFT module, pass the output of the second linear layer through the second FFT module, multiply the output of the first FFT module by the output of the second FFT module, and use the multiplication result as the input of the first IFFT module. Output the autocorrelation function value corresponding to each delay through the first IFFT module.

[0193] Then select the delays corresponding to the first k largest autocorrelation function values ​​(sort the obtained autocorrelation function values ​​in descending order, and take the delays corresponding to the first k autocorrelation function values ​​as the selected k delays), and record the selected k delays as τ1, τ2, ..., τ k ;

[0194] Step 3: Delay the output of the third linear layer by τ1, τ2, …, τ k Delay processing (i.e., the output of the third linear layer is cyclically shifted to the right by τ1 bits, the output of the third linear layer is cyclically shifted to the right by τ2 bits, ... ...1 bits, k The rightward circular shift of the bits is equivalent to Roll(v,τ1), Roll(v,τ2),…, Roll(v,τ k )), and we get the delays τ1,τ2,...,τ k The corresponding delayed processing results;

[0195] Step 4: multiply the delay processing result corresponding to delay τ1 by the autocorrelation function value corresponding to delay τ1, multiply the delay processing result corresponding to delay τ2 by the autocorrelation function value corresponding to delay τ2, and so on. k Corresponding delay processing results and delay τ k The corresponding autocorrelation function values ​​are multiplied to obtain the delays τ1, τ2, …, τ k The corresponding multiplication results are then accumulated to obtain an accumulated result;

[0196] The accumulated result is passed through a fourth linear layer, and the output of the fourth linear layer is used as the output of the first correlation extraction module.

[0197] The other steps and parameters are the same as those in Specific Embodiments 1 to 7.

[0198] For cross autocorrelation The calculation is to convert X s(7) After linear mapping, the sequence q is obtained as Figure 6 In q, After linear mapping, the sequences k and v are obtained as Figure 6 The k and v in .

[0199] Specific implementation method 9: This implementation method is different from any one of specific implementation methods 1 to 8 in that the working process of the first sequence decomposition module is as follows:

[0200] X t(3) =AvgPool(Padding(X (2) ))

[0201] X s(3) =X (2) -X t(3)

[0202] Among them, Padding(·) represents the data padding operation (padding is performed to keep the sequence length unchanged), AvgPool(·) represents the moving average operation, Represents the trend cycle part of the first sequence decomposition module output, Represents the seasonal part of the output of the first series decomposition module.

[0203] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0204] The first sequence decomposition module can adjust the moving average to smooth periodic fluctuations and highlight long-term trends. The working process of each sequence decomposition module in the present invention is the same.

[0205] Specific implementation method 10: This implementation method is different from any one of specific implementation methods 1 to 9 in that the working process of the first forward propagation module is as follows:

[0206] FFN1(X s(3) )=w2relu(w1X s(3) +b1)+b2

[0207] Among them, FFN1(·) represents the operation of the first forward propagation module, w1, w2, b1 and b2 are all learnable parameters of the model, and relu(·) is the activation function.

[0208] The other steps and parameters are the same as those in Specific Embodiments 1 to 9.

[0209] The working process of each forward propagation module in the present invention is the same.

[0210] Experimental Section

[0211] The following is an example of an open source spectrum dataset from RWTH Aachen University in Germany, which contains data from different locations and frequency bands. We selected data collected in a residential area in Maastricht, the Netherlands, with a center frequency of 770MHz and a bandwidth of 1500MHz. The data is divided into 8192 channels with an interval of 200kHz, and there is a slight overlap between adjacent channels. In order to facilitate analysis and not affect problems in the actual environment, we selected 150 channels for description. The two-dimensional and three-dimensional power spectra of the data are shown as follows: Figure 1 and Figure 2 shown.

[0212] The selected 150 channels are divided into three parts: training set, validation set and test set in a ratio of 7:2:1. To verify the effectiveness of the channel prediction method of the present invention, a Seq-to-Seq model based on LSTM, a bidirectional LSTM model based on attention mechanism, a fusion model of CNN and LSTM, and an improved Autoformer model of the present invention are respectively used. en =1,N de=1) Multi-channel joint prediction, single channel prediction and prediction of the method of the present invention are performed, and the results are compared. In the four models, the widely used and excellent Adam is used as the optimizer, and the initial learning rate is set to 10 -4 , the loss function adopts mean square error loss, and MAE and MSE are selected as evaluation indicators. In all models, the number of hidden units is set to 128, the learning rate of the optimizer is reduced by half each round, and an early stopping mechanism is set, and patience is set to 10. In each model, the window length corresponding to the lowest MSE is selected, that is, the required historical time length. Since the bidirectional LSTM model based on the attention mechanism and the fusion model of CNN and LSTM can only perform single-step prediction, the present invention takes single-step prediction as an example to verify the effectiveness of the prediction method of the present invention.

[0213] Compare the average MSE and MAE of all channels of the above models on the test set, as shown in Figure 7 and Figure 8 As shown. When the LSTM-based Seq-to-Seq model is used for multi-channel joint prediction, the average prediction MSE of all channels is 0.4007 and the MAE is 0.4750; when single channel prediction is performed, the average prediction MSE of all channels is 0.0912 and the MAE is 0.2023; and when the prediction method of the present invention is used, the average prediction MSE of all channels is 0.0895 and the MAE is 0.2015. When the bidirectional LSTM model based on the attention mechanism is used for multi-channel joint prediction, the average prediction MSE of all channels is 0.3918 and the MAE is 0.4721; when single channel prediction is performed, the average prediction MSE of all channels is 0.0918 and the MAE is 0.2033; and when the prediction method of the present invention is used, the average prediction MSE of all channels is 0.0894 and the MAE is 0.2020. When the fusion model of CNN and LSTM is used for multi-channel joint prediction, the average prediction MSE of all channels is 0.3975 and the MAE is 0.4766; when single channel prediction is performed, the average prediction MSE of all channels is 0.0959 and the MAE is 0.2089; and when the prediction method of the present invention is used, the average prediction MSE of all channels is 0.0923 and the MAE is 0.2076. When the improved Autoformer model is used for multi-channel joint prediction, the average prediction MSE of all channels is 0.1221 and the MAE is 0.2514; when single channel prediction is performed, the average prediction MSE of all channels is 0.0921 and the MAE is 0.2087; and when the prediction method of the present invention is used, the average prediction MSE of all channels is 0.0898 and the MAE is 0.2037.

[0214] It can be seen that the four models have much lower losses than the multi-channel joint prediction when using the prediction method of the present invention, indicating that the prediction method of the present invention can avoid the mutual interference between channels and the problem of large differences in signal distribution of each channel. In addition, the loss of the method proposed by the present invention is further reduced compared with the single-channel prediction, and the future information of all channels can be predicted at the same time, which shows that the prediction method of the present invention has stronger generalization ability and higher efficiency. Since the improved Autoformer model is more suitable for long-term prediction and is not prone to the problems of gradient disappearance and gradient explosion, the improved Autoformer is selected as the prediction model to verify the effectiveness of the method of the present invention.

[0215] In summary, the present invention proposes a multi-channel joint prediction method that can avoid interference between channels and large differences in channel data distribution. First, the channel dimension is split and fused into the batch dimension, and then any spectrum prediction model is used for prediction. The method of the present invention can be regarded as training a common model for all channels. The model will learn the most core conditional distribution P(Y|X) in each channel, avoiding the model from being more biased towards any channel, thereby improving the generalization ability of the model. At the same time, the method of the present invention can predict the future information of all channels at once, thereby solving the problem that single-channel prediction can only predict one channel at a time, and the number of models that need to be trained and stored for single-channel prediction is proportional to the number of channels. When the number of channels is large, time and storage resources will be seriously wasted. Compared with multi-channel joint prediction, due to the obvious difference in data distribution of each channel, direct multi-channel joint prediction of unprocessed data will result in a significant increase in prediction loss. In addition, when there is a problem of misalignment of correlation between channels, the multi-channel joint prediction model will incorrectly learn these correlations, thereby interfering with the predicted channel. The method proposed in the present invention will avoid these problems well, and make full use of time correlation to obtain prediction results with lower loss when the spectrum data is sufficient. The present invention uses an improved Autoformer that can perform stable long-term predictions as a spectrum prediction model to avoid the gradient vanishing problem of LSTM in multi-step predictions and the inefficiency of multi-step recursive predictions. Low-loss prediction results can help new users identify idle or underutilized spectrum, thereby achieving dynamic spectrum access and improving overall spectrum utilization.

[0216] The above calculation examples of the present invention are only used to explain the calculation model and calculation process of the present invention in detail, and are not intended to limit the implementation methods of the present invention. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made based on the above description. It is impossible to list all the implementation methods here. All obvious changes or modifications derived from the technical solution of the present invention are still within the scope of protection of the present invention.

Claims

1. A spectrum prediction method for channel separation, characterized in that: The method specifically comprises the following steps: Step 1: Define the spectrum data that needs to be predicted for the next P time slots, and build a corresponding spectrum prediction model based on the value of P; And define the number of historical time slots H∈O corresponding to the historical spectrum data sequence input by the spectrum prediction model, where O represents the set of all values ​​of H; Step 2: Record the historical spectrum data sequence input into the spectrum prediction model as in, represents a real number, M represents the number of channels that need to be predicted for spectrum data in each batch, and B represents the number of batches; The spectrum data of each channel in the historical spectrum data sequence X is normalized respectively, and the channel dimension of the normalized input data is fused to the batch dimension to obtain the processed historical spectrum data sequence Step 3: Use the processed historical spectrum data sequence X′ to train the spectrum prediction model to obtain the trained spectrum prediction model corresponding to the current H value; Step 4: For each value of H in the set O, steps 2 and 3 are executed, that is, the trained spectrum prediction model corresponding to each H value is obtained respectively; Then, the spectrum prediction model with the smallest mean square error loss is selected from the trained spectrum prediction models corresponding to each H value; Step 5: After processing the historical spectrum data sequence of the channel to be predicted, the processed data is input into the selected spectrum prediction model to obtain the spectrum data prediction result of the channel to be predicted; The batch dimension of the spectrum data prediction result is then split and restored to the original channel dimension to obtain the final prediction result.

2. The spectrum prediction method for channel separation according to claim 1, characterized in that: The corresponding spectrum prediction model is constructed according to the value of P, specifically: When P=1, the constructed spectrum prediction model is a LSTM-based Seq2Seq model or an improved Autoformer model; When P>1, the constructed spectrum prediction model is an improved Autoformer model.

3. The spectrum prediction method for channel separation according to claim 2, characterized in that: The improved Autoformer model includes an encoder and a decoder, and the encoder includes N en The decoder includes N de A decoding unit; In the improved Autoformer model, the input of the improved Autoformer model is passed through the linear mapping layer, the output of the linear mapping layer is used as the input of the first encoding unit, and the seasonal part of the output of the first encoding unit is used as the input of the second encoding unit, and the seasonal part of the output of the second encoding unit is used as the input of the third encoding unit, and so on. en The seasonal part of the output of the -1 coding unit is taken as the Nth en The input of the encoding unit is obtained en The seasonal component of the coding unit output; The input of the improved Autoformer model is decomposed into the cumulative structure of the trend cycle component and the cumulative structure of the seasonal component, and the cumulative structure of the decomposed trend cycle component, the cumulative structure of the seasonal component and the Nth en The seasonal part output by the encoding unit is used as the input of the first decoding unit, and the trend cycle part, seasonal part and the Nth en The seasonal part of the output of the encoding unit is used as the input of the second decoding unit, and so on. de -1 The trend cycle part, seasonal part and Nth part of the decoding unit output en The seasonal part of the encoding unit output is used as the Nth de The input of the decoding unit is based on the Nth de The output of the decoding unit obtains the spectrum data prediction result.

4. The spectrum prediction method for channel separation according to claim 3, characterized in that: The working process of the encoder is: First, the processed historical spectrum data sequence X′ is dimensionally expanded to d dimensions, and the dimension expansion result X en for: Among them, Embedding(·) represents the linear mapping layer; Then the output X of the linear mapping layer en As the input of the first coding unit, in the first coding unit, the input of the first coding unit is used as the input of the first correlation extraction module, then the output X of the first correlation extraction module is (1) for: X (1) =Auto-Correlation1(X en ) Wherein, Auto-Correlation1(·) represents the operation of the first correlation extraction module; The output of the first correlation extraction module is compared with X en Add, and the result is X (2) for: X (2) =Auto-Correlation1(X en )+X en Add the result X (2) As the input of the first sequence decomposition module, the trend cycle part X is output by the first sequence decomposition module t(3) and seasonal part X s(3) ; The seasonal part X s(3) As the input of the first forward propagation module, the output X of the first forward propagation module (4) for: X (4) =FFN1(X s(3) ) Where FFN1(·) represents the operation of the first forward propagation module; Combine the output of the first forward propagation module with the seasonal part X s(3) Add them together to get the result X (5) : X (5) =X (4) +X s(3) Add the result X (5) As the input of the second sequence decomposition module, the trend cycle part of the second sequence decomposition module output is recorded as X t(6) , the seasonal part of the output of the second sequence decomposition module is recorded as X s(6) , and X s(6) The seasonal part as output of the first coding unit; Then the seasonal part X output by the first encoding unit is s(6) as the input of the second coding unit; and the seasonal part output by the second coding unit is further used as the input of the third coding unit; And so on, the Nth en The seasonal part of the output of the -1 coding unit is taken as the Nth en The input of the encoding unit is obtained en The seasonal part of the encoding unit output 5. The spectrum prediction method for channel separation according to claim 3, characterized in that: The cumulative structure of the trend cycle component and the cumulative structure of the seasonal component are obtained by: Among them, the processed historical spectrum data sequence X′ is passed through the sequence decomposition module to obtain the historical trend cycle sequence and the historical seasonal sequence. It is the last H / 2 row of the historical trend cycle sequence. represents the last H / 2 rows of the historical seasonal series, mean(X′ H ) P represents a vector of P rows and 1 column consisting of the mean values ​​of each row in the historical spectrum data sequence X′, O P represents a 0 vector with P rows and 1 columns, represents the cumulative structure of the trend period component, Represents the cumulative structure of the seasonal component.

6. A spectrum prediction method for channel separation according to claim 4, characterized in that: The working process of the decoder is: Will And Nth en The seasonal part of the encoding unit output As the input of the first decoding unit, in the first decoding unit, As the input of the second correlation extraction module, the output of the second correlation extraction module is then combined with Add, and use the addition result as the input of the third sequence decomposition module, and output X through the third sequence decomposition module s (7) and X t(7) : Wherein, Auto-Correlation2(·) represents the operation of the second correlation extraction module, and SeriesDecomp3(·) represents the operation of the third series decomposition module; X s(7) and As the input of the third correlation extraction module, the output of the third correlation extraction module and X s(7) Add, and use the addition result as the input of the fourth sequence decomposition module to obtain the output X of the fourth sequence decomposition module s(8) and X t(8) : Wherein, Auto-Correlation3(·) represents the operation of the third correlation extraction module, and SeriesDecomp4(·) represents the operation of the fourth series decomposition module; X s(8) As the input of the second forward propagation module, the output of the second forward propagation module is combined with X s(8) Add, and use the added result as the input of the fifth sequence decomposition module to obtain the output X of the fifth sequence decomposition module s(9) and X t(9) : X s(9) ,X t(9) =SeriesDecomp5(FFN2(X s(8) )+X s(8) ) Wherein, FFN2(·) represents the operation of the second forward propagation module; SeriesDecomp5(·) represents the operation of the fifth series decomposition module; Will X t(7) , X t(8) and X t(9) Add together to get the sum of the trend cycle components X s(9) As the seasonal part of the output of the first decoding unit, As the trend cycle part of the first decoding unit output, X s(9) and as an input of a second decoding unit; And so on, we get the Nth de The seasonal part and trend cycle part of the decoding unit output; de The seasonal part of the decoding unit output is denoted by The Nth de The trend cycle part of the decoding unit output is recorded as Then the Nth de Trend cycle part of the decoding unit output and seasonal part merge: Among them, Y de Indicates the merge result; Then merge the result Y de Mapping is performed to obtain the spectrum data prediction result Y′: Y′=W o Y de +b o Among them, W o and b o is the model parameter.

7. The spectrum prediction method for channel separation according to claim 4, characterized in that: The working process of the first correlation extraction module is as follows: Step 1: X en After linear mapping, we get the sequences q, k and v; Step 2, initialize τ=1; Step 3: cyclically shift the sequence v to the right according to the time delay τ to obtain the result Roll(v,τ) after the cyclic shift to the right; Calculate X again en The autocorrelation function Among them, q t represents the tth element in the sequence q, k t-τ represents the t-τth element in sequence k; Step 4: Determine whether τ=H is satisfied; If satisfied, proceed to step 5; If not satisfied, set τ = τ + 1 and return to step 3; Step 5: According to Select the k most correlated delays: Among them, τ1, τ2, ..., τ k represents the selected k delays; For the selected k delays corresponding to To process: in, Represent k delays τ1, τ2, ..., τ k The corresponding weights, softmax(·) is the normalized exponential function; Then the output of the first correlation extraction module is Auto-Correlation(q,k,v):

8. A spectrum prediction method for channel separation according to claim 7, characterized in that: The first correlation extraction module performs calculations based on FFT and IFFT, and the specific calculation method is: Step 1: Use sequence q as the input of the first linear layer, sequence k as the input of the second linear layer, and sequence v as the input of the third linear layer; Step 2: Pass the output of the first linear layer through the first FFT module, pass the output of the second linear layer through the second FFT module, multiply the output of the first FFT module by the output of the second FFT module, and use the multiplication result as the input of the first IFFT module. Output the autocorrelation function value corresponding to each delay through the first IFFT module. Then select the delays corresponding to the first k largest autocorrelation function values, and record the selected k delays as τ1, τ2, ..., τ k ; Step 3: Delay the output of the third linear layer by τ1, τ2, ..., τ k The delay processing is to obtain delays τ1, τ2, ..., τ k The corresponding delayed processing results; Step 4: multiply the delay processing result corresponding to delay τ1 by the autocorrelation function value corresponding to delay τ1, multiply the delay processing result corresponding to delay τ2 by the autocorrelation function value corresponding to delay τ2, and so on. k Corresponding delay processing results and delay τ k The corresponding autocorrelation function values ​​are multiplied to obtain the delays τ1, τ2, ..., τ k The corresponding multiplication results are then accumulated to obtain an accumulated result; The accumulated result is passed through a fourth linear layer, and the output of the fourth linear layer is used as the output of the first correlation extraction module.

9. The spectrum prediction method for channel separation according to claim 4, characterized in that: The working process of the first sequence decomposition module is as follows: X t(3) =AvgPool(Padding(X (2) )) X s(3) =X (2) -X t(3) Among them, Padding(·) represents the data padding operation, AvgPool(·) represents the moving average operation, Represents the trend cycle part of the first sequence decomposition module output, Represents the seasonal part of the output of the first series decomposition module.

10. The spectrum prediction method for channel separation according to claim 4, characterized in that: The working process of the first forward propagation module is: <h2 style=";text-align:left;direction:ltr">FFN1(X<h2 style=";text-align:left;direction:ltr"> s(3) <h2 style=";text-align:left;direction:ltr"> )=w2relu(w1X<h2 style=";text-align:left;direction:ltr"> s(3) <h2 style=";text-align:left;direction:ltr"> +b1)+b2 Among them, FFN1(·) represents the operation of the first forward propagation module, w1, w2, b1 and b2 are model parameters, and relu(·) is the activation function.

Images