CN115936237A - Time series forecasting method, device, computer equipment and storage medium - Google Patents

Abstract

The embodiment of the invention provides a time series prediction method, a time series prediction device, computer equipment and a storage medium, and relates to the technical field of data processing. The method comprises the following steps: acquiring an initial season part and an initial trend part corresponding to the historical time sequence; inputting the historical time series into the encoder, and obtaining an encoding output result containing season information through the first cross attention unit; inputting the initial part of the season into the decoder, deriving a predicted part of the season via the second cross attention unit and the third cross attention unit, and separating a remaining trend part; and inputting the initial trend part into the decoder, and accumulating the initial trend part and the residual trend part to obtain a predicted trend part. The time series prediction method provided by the embodiment adopts a progressive decomposition and cross attention mechanism, so that the memory flow is reduced, and the time complexity is reduced. The method can accurately and efficiently predict the long-sequence time sequence.

Description

Time series forecasting method, device, computer equipment and storage medium

technical field

The present invention relates to the technical field of data processing, in particular to a time series prediction method, device, computer equipment and storage medium.

Background technique

Time series forecasting has been widely used in energy consumption, transportation and economic planning, weather and disease spread forecasting. In these practical engineering applications, it is urgent to extend the prediction time to the distant future, which is very meaningful for long-term planning and early warning.

Based on this, the existing time series forecasting methods generally build forecasting models based on historical data for predictive analysis of future data. An existing long-term time series prediction method extracts key information in historical data based on the self-attention mechanism to predict future data, but the self-attention mechanism has the following problems: it is not sensitive to local context, and it is easy to encounter Memory bottleneck, low computational efficiency.

Contents of the invention

The object of the present invention includes, for example, to provide a time series forecasting method, device, computer equipment and storage medium, capable of forecasting a long time series.

Embodiments of the present invention can be realized like this:

In the first aspect, an embodiment of the present invention provides a time series prediction method, which is applied to a time series prediction architecture, and the time series prediction architecture includes a multi-layer encoder and a multi-layer decoder, and the encoder includes a first cross attention A force unit, the decoder includes a second cross-attention unit and a third cross-attention unit, and the method includes:

Obtain the initial seasonal part and initial trend part corresponding to the historical time series;

Inputting the historical time series into the encoder, and obtaining an encoded output result containing seasonal information via the first cross-attention unit;

inputting the initial season part into the decoder, obtaining the predicted season part via the second cross-attention unit and the third cross-attention unit, and separating the remaining trend part;

The initial trend part is input into the decoder, and accumulated with the remaining trend part to obtain a predicted trend part.

In one embodiment, said obtaining the initial trend part corresponding to the historical time series includes:

performing moving average processing on the historical time series to obtain a historical trend sequence corresponding to the historical time series;

Obtain the last I/2 elements of the historical trend sequence to obtain a segmented trend sequence;

Obtain the average value of each element of the historical trend sequence;

After filling O average values for the segmented trend sequence, pass the concat function to obtain the initial trend part; wherein, I is the length of the historical time series, and O is the length of the time to be predicted.

In one embodiment, said obtaining the initial season part corresponding to the historical time series includes:

Subtracting the historical trend sequence from the historical time series to obtain a historical seasonal sequence corresponding to the historical time series;

Obtain the last I/2 elements of the historical season sequence to obtain the split season sequence;

After filling O zero values to the split season sequence, the concat function is used to obtain the initial season part.

In one embodiment, the encoder further includes a first sequence decomposing unit, a first feedforward unit, and a second sequence decomposing unit, the historical time series is input into the encoder, and through the first crossover The attention unit obtains encoded outputs containing seasonal information, including:

passing the historical time series through the first cross-attention unit to obtain the first attention sequence;

Adding the first attention sequence to the historical time series and passing through the first sequence decomposition unit to obtain a first decomposition result;

passing the first decomposition result through a first feedforward unit to obtain a first feedforward sequence;

The first feed-forward sequence is added to the first decomposition result and passed through a second sequence decomposition unit to obtain the encoding output result.

In one embodiment, the decoder further includes a third sequence decomposing unit, a fourth sequence decomposing unit, a second feedforward unit, and a fifth sequence decomposing unit; the initial season part is input into the decoder, and through the The second intersecting attention unit and the third intersecting attention unit obtain the prediction season part, and separate the remaining trend part, including:

passing the initial season part through the second cross-attention unit to obtain a second attention sequence;

After adding the second attention sequence and the historical time series, pass through the third sequence decomposition unit to obtain the second decomposition result and the first remainder;

passing the second decomposition result through the third cross-attention unit to obtain a third attention sequence;

After adding the third attention sequence and the second decomposition result, passing through the fourth sequence decomposition unit to obtain the third decomposition result and the second remainder;

passing the third decomposition result through the second feedforward unit to obtain a second feedforward sequence;

After adding the second feed-forward sequence and the third decomposition result, passing through the fifth sequence decomposition unit to obtain the predicted season part and the third remaining part;

The first remaining part, the second remaining part and the third remaining part are accumulated to obtain the remaining trend part.

In the second aspect, the embodiment of the present application provides a time series decomposition device, the device includes:

Decomposition module, for decomposing the historical time series, to obtain the seasonal part and the trend part of the historical time series;

The first filling module is used to divide and fill the seasonal part to obtain seasonal items;

The second filling module is used to divide and fill the trend part to obtain trend items;

an input module for inputting the seasonal item and the trend item into a decoder;

An acquisition module, configured to control the decoder to acquire season information from the season item through a cross-attention mechanism, and acquire trend information from the trend item through accumulation.

In a third aspect, an embodiment of the present application provides an electronic device, including a memory and a processor, the memory is used to store a computer program, and the computer program executes the time series decomposition provided in the first aspect when the processor is running method.

In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program, and the computer program executes the time series decomposition method provided in the first aspect when running on a processor.

The beneficial effects of the embodiments of the present invention include, for example: using a progressive decomposition method to extract the predicted season part and predicted trend part for complex time series, and use the cross-attention mechanism to use split and merge strategies across stages, effectively It reduces the possibility of repetition in the process of information integration and improves the learning ability of the network. The cross-attention mechanism also reduces memory traffic and reduces time complexity.

Description of drawings

In order to illustrate the technical solutions of the embodiments of the present invention more clearly, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention, and thus It should be regarded as a limitation on the scope, and those skilled in the art can also obtain other related drawings based on these drawings without creative work.

Fig. 1 shows a schematic flow chart of the time series prediction method provided by the embodiment of the present application;

Fig. 2 shows a schematic structural diagram of the time series prediction model provided by the embodiment of the present application;

FIG. 3 shows a schematic structural diagram of a cross-attention unit provided by an embodiment of the present application;

FIG. 4 shows a schematic structural diagram of a time series prediction device provided by an embodiment of the present application.

Detailed ways

In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments It is a part of embodiments of the present invention, but not all embodiments. The components of the embodiments of the invention generally described and illustrated in the figures herein may be arranged and designed in a variety of different configurations.

Accordingly, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

It should be noted that like numerals and letters denote similar items in the following figures, therefore, once an item is defined in one figure, it does not require further definition and explanation in subsequent figures.

In the description of the present invention, it should be noted that if the orientation or positional relationship indicated by the terms "upper", "lower", "inner" and "outer" appear, it is based on the orientation or positional relationship shown in the drawings, or It is the orientation or positional relationship that the invention product is usually placed in use, and it is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation , and therefore cannot be construed as a limitation of the present invention.

In addition, terms such as "first" and "second" are used only for distinguishing descriptions, and should not be understood as indicating or implying relative importance.

It should be noted that, in the case of no conflict, the features in the embodiments of the present invention may be combined with each other.

Example 1

Please refer to FIG. 1 , this embodiment provides a time series forecasting method.

Step S110, obtaining the initial seasonal part and initial trend part corresponding to the historical time series;

Time series decomposition is a method of time series analysis. The idea is to decompose data into different factors to achieve the purpose of explaining data, establishing mathematical models, and predicting data. Due to the unknowability of the future in the forecasting problem, the historical time series is usually decomposed first, and the decomposed residual series is used to replace the original series for forecasting, which helps to make better forecasts. A time series can be divided into three parts: a trend cycle part, a seasonal part, and a remaining part that contains everything else in the time series.

That is, the usual time series can be expressed as a product of three parts, as in Equation 1:

y _t =S _t *T _t *R _t

Among them, y _t is the original time series, S _t represents the seasonal part, T _t represents the trend period part, and R _t represents the remaining part.

However, how to decompose the time series is a problem that needs to be solved. Based on this, the embodiment of this application proposes a deep decomposition architecture based on the cross-attention mechanism, which uses the sequence decomposition as an internal unit of the framework and embeds it in the encoder and decoder. . During the forecasting process, the model alternately performs forecasting result optimization and sequence decomposition, that is, gradually separates trend items and periodic items from hidden variables to achieve progressive decomposition.

The series decomposition block is based on the idea of moving average, smoothing periodic items and highlighting trend items. Specifically, periodic fluctuations are smoothed out and long-term trends are highlighted by adjusting the moving average. For an input series X∈R ^Lxd of length L, the process is Equation 2:

X _t =AvgPool(Padding(X))

X _s =XX _t

In the above formula, X _t is the extracted trend item, which stores the mean value of each sliding window, that is, the short-term fluctuation of the sequence, and X _s is a smooth sequence that retains the seasonality after subtracting the short-term fluctuation.

Among them, padding is to ensure that the sequence length remains unchanged, and avgpool is a moving average. X is the latent variable to be decomposed, X _t and X _s are the trend item and the seasonal item respectively, and the above Series Decomp Block can be recorded as formula 3:

X _t ,X _s =SeriesDecomp(X)

The basic principles of time series decomposition and sequence decomposition unit work are described above. Based on this, the embodiment of the present application provides a time series decomposition method.

Specifically, in one embodiment, the acquiring the initial trend part corresponding to the historical time series includes: performing moving average processing on the historical time series to obtain the historical trend series corresponding to the historical time series; acquiring the historical time series The last 1/2 element of the trend sequence obtains the segmented trend sequence; obtains the average value of each element of the historical trend sequence; fills the 0 average values for the segmented trend sequence and obtains the initial trend part through the concat function; wherein , I is the length of the historical time series, O is the length of the time to be predicted.

The initial seasonal part corresponding to the acquisition of historical time series includes:

Subtract the historical trend sequence from the historical time series to obtain the corresponding historical season sequence of the historical time series; obtain the last 1/2 elements of the historical seasonal sequence to obtain a split seasonal sequence; for the split season After the sequence is filled with O zero values, it passes through the concat function to obtain the initial seasonal part.

First, the historical time series needs to be preprocessed to get the input to the encoder and decoder. Specifically, it is necessary to first split the historical trend series and historical season series obtained by decomposing the initial historical time series. Assuming that the length of the initial historical time series is I, then only the second half of the initial time series, that is, I /2 length.

Then, according to the purpose of prediction, the segmented trend sequence and season sequence after segmentation are filled, and the number of filled elements is the future length O that needs to be predicted. The filling method of the initial season part can be referred to in formula 4:

X _des ＝Concat(X _ens ,X ₀ )

Among them, X _en represents the historical time series, X _ens represents the split season series, X ₀ represents the zero-filled placeholders, and the number is O. X _des is all that is needed for the initial season part. X _ens and X ₀ are concatenated via the concat function to obtain the initial seasonal part.

Then the initial trend part can be obtained by formula 5:

X _det ＝Concat(X _ent ,X _Mean )

X _ent represents the segmented trend series, and X _mean represents the mean value of X _en .

Step S120, inputting the historical time series into the encoder, and obtaining an encoded output result containing seasonal information via the first cross-attention unit;

Specifically, please refer to FIG. 2 , which shows a schematic structural diagram of a time series decomposition model provided by an embodiment of the present application. The time series decomposition model includes an encoder 210 and a decoder 220 . N represents the number of layers in the encoder, and M represents the number of layers in the decoder. K, V, and Q represent key, query, and value, respectively, and are parameters of the attention mechanism.

The encoder 210 also includes a first sequence decomposition unit, a first feed-forward unit, and a second sequence decomposition unit, and the historical time series is input into the encoder 210, and obtained through the first cross-attention unit Encoded output with seasonal information, including:

passing the historical time series through the first cross-attention unit to obtain the first attention sequence; adding the first attention sequence to the historical time series and passing through the first sequence decomposition unit to obtain the first decomposition result ; passing the first decomposition result through a first feedforward unit to obtain a first feedforward sequence; adding the first feedforward sequence to the first decomposition result and passing through a second sequence decomposition unit to obtain the Encode the output result.

The input of the encoder 210 is the complete historical time series, and the output of the encoder 210 is a matrix containing seasonal information, which will be used as the input of the third cross-attention unit of the decoder to improve the decoder prediction result. The processing formula of encoder 210 is as formula 6:

表示经过第一交叉注意力单元得到的包含季节信息的分量，Sⁱ表示经过前馈单元，最终输出的编码输出结果。CrossAttention here represents the first cross-attention unit, "_" is the trend part eliminated after sequence decomposition;

Represents the component containing seasonal information obtained through the first cross-attention unit, and S ⁱ represents the final output encoding output after the feedforward unit.

Step S130, inputting the initial season part into the decoder 220, obtaining the predicted season part through the second cross-attention unit and the third cross-attention unit, and separating the remaining trend part;

Referring to FIG. 2 , the decoder 220 further includes a third sequence decomposing unit, a fourth sequence decomposing unit, a second feedforward unit, and a fifth sequence decomposing unit.

The initial season part is input into the decoder 220, the forecast season part is obtained through the second cross-attention unit and the third cross-attention unit, and the remaining trend part is separated, including: the initial season Partially pass through the second cross-attention unit to obtain a second attention sequence; add the second attention sequence and the historical time series and pass through the third sequence decomposition unit to obtain the second decomposition result and The first remaining part; pass the second decomposition result through the third cross-attention unit to obtain a third attention sequence; add the third attention sequence and the second decomposition result and pass the fourth a sequence decomposition unit to obtain a third decomposition result and a second remainder; pass the third decomposition result through the second feedforward unit to obtain a second feedforward sequence; combine the second feedforward sequence with the first After the three decomposition results are added, pass through the fifth sequence decomposition unit to obtain the predicted season part and the third remaining part; accumulate the first remaining part, the second remaining part and the third remaining part , to get the residual trend part.

The calculation process of the decoder 220 is a process of separating the trend information from the input information and retaining the seasonal information. Please refer to the structure of the dotted line box 221 and the dotted line box 222, the dotted line box 221 includes the second cross attention unit, the third sequence decomposition unit, the third cross attention unit, the fourth sequence decomposition unit, the second feedforward unit and the first The five-sequence decomposition unit is mainly used to gradually separate trend information from input information.

Specifically, the previously obtained initial season part is input into the decoder 220, and the operation of each layer of the decoder 220 is shown in formula 7:

表示经过第二交叉注意力单元和第三序列分解单元得到的含季节信息的序列，也就是第二分解结果；

表示经过第二交叉注意力单元和第三序列分解单元分离出的趋势信息，也就是第一剩余部分。Equation 7 corresponds to part 221 of the decoder, which progressively extracts seasonal information from the initial seasonal sequence containing seasonal information and the encoded output, further separating trend and seasonal information. Among them, X _de represents the input of the decoder,

Represents the sequence containing seasonal information obtained through the second cross-attention unit and the third sequence decomposition unit, that is, the second decomposition result;

Represents the trend information separated by the second cross-attention unit and the third sequence decomposition unit, that is, the first remaining part.

表示第三分解结果，

表示第二剩余部分；

表示最终得到的预测季节部分，

表示第三剩余部分。

会在222进行累加。In the same way,

represents the third decomposition result,

represents the second remainder;

Denotes the resulting predicted seasonal component,

Indicates the third remainder.

Will be accumulated at 222.

Step S140, input the initial trend part into the decoder, and accumulate with the remaining trend part to obtain a predicted trend part.

Specifically, the dotted line box 222 contains a plurality of accumulation units, which are used to accumulate the separated trend information and the initial trend part input to the decoder 220, that is, the initial trend part and the first remaining part, the second remaining part and The third remaining part is gradually accumulated to obtain the final predicted trend part.

The resulting trend information can be found in Equation 8:

Wherein, P _i , P _i+1 and P _i+2 represent weights corresponding to each remaining part.

The cross-attention mechanism uses the periodic nature of seasonal information to aggregate subsequences with similar processes in different cycles; for trend information, it uses an accumulation method to gradually extract trend information from predicted latent variables, alternately, and promote each other.

Based on the model architecture shown in Figure 2, the model can gradually decompose hidden variables during the forecasting process, and obtain the forecast results of season and trend information respectively through the cross-attention mechanism and accumulation method, realizing the alternate and mutual promotion of decomposition and forecast result optimization .

In addition, please refer to FIG. 3 , which shows a schematic structural diagram of a cross-attention unit provided by an embodiment of the present application, that is, a CrossAttention unit. A CrossAttention unit divides the input into two parts. The first part is propagated through layer 310, passing through a 1×1 convolutional layer, while the second part is propagated through layer 320, passing through a self-attention = block. Finally, the output of the two parts is connected together through the concat function as the final output of the entire CrossAttention unit.

Using the CrossAttention unit has the following advantages, for example:

分为两部分。X₁在经过一个1×1卷积层后连接到CrossAttention单元的末端，而X₂充当自注意块的输入。A和B的输出通过维度连接起来作为整个CrossAttention单元的输出。Input R ^L*d to the CrossAttention unit, where L is the input length, d is the input dimension, and pass the dimension

Divided into two parts. _X1 is connected to the end of the CrossAttention unit after going through a 1×1 convolutional layer, while _X2 serves as the input of the self-attention block. The outputs of A and B are concatenated by dimensions as the output of the whole CrossAttention unit.

The output matrix of a stage of CrossAttention is shown in Formula 9:

where A(X _2h ) is the scaled dot product of the hth self-attention block, and W _h is a d _h * d _h linear projection matrix. H is the number of heads in the attention mechanism, d _h is the size of each head, assuming that each head has the same size, W _c is a value weight matrix of a 1*1 convolutional layer.

Compared with the self-attention (self-attention) mechanism, CrossAttention (cross-attention) can alleviate the memory bottleneck and computational efficiency problems of the self-attention mechanism. CrossAttention also reduces the memory traffic and time complexity of the self-attention mechanism.

For example, if two matrices A∈R ^a×b and B∈R ^b×c are given, and only the calculation of multiplication is considered, the time complexity of AxB is axbxc, expressed by T(A×B)=a×b×c . Let X ∈ R ^L×d be the input matrix consisting of L tokens, and the self-attention block has H heads. Excluding bias, the output of the hth self-attention head can be written as Equation 10:

A _h (X) = P _h X W _{V, h}

Calculate the scaled dot product _Ph by Equation 11:

In this way, ignoring the calculation of softmax, the time complexity of a single self-attention head can be written as Equation 12:

where d represents the size of a single self-attention head.

And the full time complexity of a typical self-attention block is Equation 13:

T(SA(X))＝H×T(A _h (X))+Ld ²

=4d ² L+2HdL ²

However, when calculating the time complexity of CrossAttention, assuming that CrossAttention divides the input dimension in half, the first part of CrossAttention has only one linear projection layer, which means that the time complexity of the first part can be written as Equation 14:

T(CroA ₁ (X ₁ ))=L×(d/2) ²

The time complexity of the second part is given by Equation 15:

T(CroA ₂ (X ₂ ))＝4L(d/2) ² +2L ² (d/2)

＝Ld ² +HL ² d

So the total time complexity of CrossAttention is formula 16:

T(CroA(X))＝1.25d ² L+HdL ²

Obviously, as L grows, the time complexity is approximately related to the coefficient of ^L2 . So when the network propagates forward, the time complexity of CrossAttention is almost 2Hd/Hd=50% of canonical selfattention (standardized self-attention mechanism). In addition, the L coefficient of the time complexity of CrossAttention is 1.25d ² /4d ² =31.25% of the time complexity of canonical self-attention. Therefore, CrossAttention reduces the time complexity by at least 50% compared to the canonical self-attention mechanism.

Therefore, CrossAttention reduces the memory traffic and time complexity of the self-attention mechanism.

A time series forecasting method provided in this embodiment uses a progressive decomposition method to extract the predicted season part and forecasted trend part for complex time series, and uses the cross-attention mechanism to use split and merge strategies across stages, It effectively reduces the possibility of repetition in the process of information integration and improves the learning ability of the network. The cross-attention mechanism also reduces memory traffic and reduces time complexity.

Example 2

This embodiment also provides a time series forecasting device, and the device 400 includes:

An acquisition module 410, configured to acquire an initial seasonal part and an initial trend part corresponding to a historical time series;

An encoding module 420, configured to input the historical time series into the encoder, and obtain an encoded output result containing seasonal information via the first cross-attention unit;

The decoding module 430 is used to input the initial season part into the decoder, obtain the predicted season part through the second cross-attention unit and the third cross-attention unit, and separate the remaining trend part;

The accumulation module 440 is configured to input the initial trend part into the decoder, and accumulate it with the remaining trend part to obtain a predicted trend part.

A time series forecasting device provided in this embodiment adopts a progressive decomposition method to extract the predicted season part and predicted trend part from complex time series, and uses the cross-attention mechanism to use split and merge strategies across stages, It effectively reduces the possibility of repetition in the process of information integration and improves the learning ability of the network. The cross-attention mechanism also reduces memory traffic and reduces time complexity.

Example 3

In addition, an embodiment of the present disclosure provides an electronic device, including a memory and a processor, the memory stores a computer program, and the computer program executes the time series decomposition method provided in Embodiment 1 when running on the processor .

The electronic device provided in the embodiment of the present invention can implement the time series decomposition method provided in Embodiment 1, and to avoid repetition, details are not repeated here.

The electronic device provided by the embodiment of the present invention adopts the method of progressive decomposition to extract the predicted season part and predicted trend part from the complex time series, and uses the cross-attention mechanism to use split and merge strategies across stages to effectively reduce information The possibility of repetition during the ensemble improves the learning ability of the network. The cross-attention mechanism also reduces memory traffic and reduces time complexity.

Example 4

The present application also provides a computer-readable storage medium, where a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the time series decomposition method provided in Embodiment 1 is implemented.

In this embodiment, the computer-readable storage medium may be a read-only memory (Read-Only Memory, ROM for short), a random access memory (Random Access Memory, RAM for short), a magnetic disk or an optical disk, and the like.

The computer-readable storage medium provided in this embodiment can implement the time series decomposition method provided in Embodiment 1, and details are not repeated here to avoid repetition.

It should be noted that, in this document, the term "comprising", "comprising" or any other variation thereof is intended to cover a non-exclusive inclusion such that a process, method, article or terminal comprising a set of elements includes not only those elements, It also includes other elements not expressly listed, or elements inherent in the process, method, article, or terminal. Without further limitations, an element defined by the phrase "comprising a ..." does not preclude the presence of additional identical elements in the process, method, article or terminal comprising the element.

Through the description of the above embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus a necessary general-purpose hardware platform, and of course also by hardware, but in many cases the former is better implementation. Based on such an understanding, the technical solution of the present application can be embodied in the form of a software product in essence or the part that contributes to the prior art, and the computer software product is stored in a storage medium (such as ROM/RAM, disk, CD) contains several instructions to enable a terminal (which may be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in various embodiments of the present application.

The embodiments of the present application have been described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementations. The above-mentioned specific implementations are only illustrative and not restrictive. Those of ordinary skill in the art will Under the inspiration of this application, without departing from the purpose of this application and the scope of protection of the claims, many forms can also be made, all of which belong to the protection of this application.

The above is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Anyone skilled in the art can easily think of changes or substitutions within the technical scope disclosed in the present invention. All should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the protection scope of the claims.

Claims (10)

1. A time series forecasting method, characterized in that it is applied to a time series forecasting framework, the time series forecasting framework comprising a multi-layer encoder and a multi-layer decoder, the encoder comprising a first cross-attention unit, the Said decoder comprises a second cross-attention unit and a third cross-attention unit, said method comprising: Obtain the initial seasonal part and initial trend part corresponding to the historical time series; Inputting the historical time series into the encoder, and obtaining an encoded output result containing seasonal information via the first cross-attention unit; inputting the initial season part into the decoder, obtaining the predicted season part via the second cross-attention unit and the third cross-attention unit, and separating the remaining trend part; The initial trend part is input into the decoder, and accumulated with the remaining trend part to obtain a predicted trend part. 2. The time series forecasting method according to claim 1, wherein said obtaining the initial trend part corresponding to the historical time series comprises: performing moving average processing on the historical time series to obtain a historical trend sequence corresponding to the historical time series; Obtain the last I/2 elements of the historical trend sequence to obtain a segmented trend sequence; Obtain the average value of each element of the historical trend sequence; After filling O average values for the segmented trend sequence, pass the concat function to obtain the initial trend part; wherein, I is the length of the historical time series, and O is the length of the time to be predicted. 3. The time series prediction method according to claim 2, wherein the initial seasonal part corresponding to the acquisition of the historical time series comprises: Subtracting the historical trend sequence from the historical time series to obtain a historical seasonal sequence corresponding to the historical time series; Obtain the last I/2 elements of the historical season sequence to obtain the split season sequence; After filling O zero values to the split season sequence, the concat function is used to obtain the initial season part. 4. The time series prediction method according to claim 1, wherein the encoder further comprises a first sequence decomposing unit, a first feedforward unit and a second sequence decomposing unit, the historical time series Input the coder, obtain the coding output result containing seasonal information via the first cross-attention unit, including: passing the historical time series through the first cross-attention unit to obtain the first attention sequence; Adding the first attention sequence to the historical time series and passing through the first sequence decomposition unit to obtain a first decomposition result; passing the first decomposition result through a first feedforward unit to obtain a first feedforward sequence; The first feed-forward sequence is added to the first decomposition result and passed through a second sequence decomposition unit to obtain the encoding output result. 5. time series prediction method according to claim 1, is characterized in that, described decoder also comprises the 3rd sequence decomposing unit, the 4th sequence decomposing unit, the 2nd feed-forward unit, the 5th sequence decomposing unit; The initial season part is input into the decoder, the predicted season part is obtained through the second cross-attention unit and the third cross-attention unit, and the remaining trend part is separated, including: passing the initial season part through the second cross-attention unit to obtain a second attention sequence; After adding the second attention sequence and the historical time series, pass through the third sequence decomposition unit to obtain the second decomposition result and the first remainder; passing the second decomposition result through the third cross-attention unit to obtain a third attention sequence; After adding the third attention sequence and the second decomposition result, passing through the fourth sequence decomposition unit to obtain the third decomposition result and the second remainder; passing the third decomposition result through the second feedforward unit to obtain a second feedforward sequence; After adding the second feed-forward sequence and the third decomposition result, passing through the fifth sequence decomposition unit to obtain the predicted season part and the third remaining part; The first remaining part, the second remaining part and the third remaining part are accumulated to obtain the remaining trend part. 6. A time series forecasting device, characterized in that the device comprises: The obtaining module is used to obtain the initial seasonal part and the initial trend part corresponding to the historical time series; An encoding module, configured to input the historical time series into an encoder, and obtain an encoded output result containing seasonal information via the first cross-attention unit; The decoding module is used to input the initial season part into the decoder, obtain the predicted season part through the second cross-attention unit and the third cross-attention unit, and separate the remaining trend part; An accumulating module, configured to input the initial trend part into the decoder, and accumulate it with the remaining trend part to obtain a predicted trend part. 7. The time series prediction device according to claim 6, wherein the encoding module is also used for: passing the historical time series through the first cross-attention unit to obtain the first attention sequence; Adding the first attention sequence to the historical time series and passing through the first sequence decomposition unit to obtain a first decomposition result; passing the first decomposition result through a first feedforward unit to obtain a first feedforward sequence; The first feed-forward sequence is added to the first decomposition result and passed through a second sequence decomposition unit to obtain the encoding output result. 8. The time series prediction device according to claim 6, wherein the decoding module is also used for: passing the initial season part through the second cross-attention unit to obtain a second attention sequence; After adding the second attention sequence and the historical time series, pass through the third sequence decomposition unit to obtain the second decomposition result and the first remainder; passing the second decomposition result through the third cross-attention unit to obtain a third attention sequence; After adding the third attention sequence and the second decomposition result, passing through the fourth sequence decomposition unit to obtain the third decomposition result and the second remainder; passing the third decomposition result through a second feedforward unit to obtain a second feedforward sequence; Adding the second feed-forward sequence to the third decomposition result and passing through a fifth sequence decomposition unit to obtain the predicted season part; The first remaining part, the second remaining part and the third remaining part are accumulated to obtain the remaining trend part. 9. An electronic device, characterized in that it comprises a memory and a processor, the memory stores a computer program, and the computer program executes the time described in any one of claims 1 to 5 when the processor is running. Sequence prediction methods. 10. A computer-readable storage medium, characterized in that it stores a computer program, and the computer program executes the time series prediction method according to any one of claims 1 to 5 when running on a processor.

Images