Prompt information fused multivariate energy data long-time prediction method and system

Abstract

The invention relates to a prompt information fused multivariate energy data long-time prediction method and system. The method comprises the steps of collecting multivariate energy data, performing missing filling and normalization, and determining target data and multi-source factors; a time sequence prompt module is constructed in the encoder, modeling analysis, average pooling and decomposition are performed on an original input sequence, and a trend sequence and a seasonal input sequence are obtained; sequence updating is carried out through a feedforward module and a sequence decomposition module to obtain cross information, and the cross information is input into a decoder; carrying out modeling analysis on the seasonal input sequence through a time sequence prompt module in the decoder, and fusing seasonal data and trend data; carrying out modeling and sequence updating on the seasonal input sequence and the cross information through a multi-head prompting attention module; and updating the sequence, and inputting an output result to a full connection layer to obtain a prediction result. According to the scheme of the invention, accurate long-time prediction of energy data is realized.

Description

Technical Field

[0001] The present invention belongs to the field of computer technology, and particularly relates to a long-term prediction method and system for multi-source energy data integrating prompt information. Background Art

[0002] With the rapid development of Internet of Things technology and deep learning technology, a variety of sensors are applied to various different fields, such as energy prediction, traffic flow monitoring, ancient building monitoring, etc. All kinds of monitoring are to better obtain and analyze the situation to make adjustments. However, the existing monitoring technologies are difficult to predict future data situations and cannot accurately make meaningful decisions. The time series prediction technology of deep learning can learn data and predict future data. Therefore, by combining the time series prediction technology, the monitoring can play a greater role.

[0003] In the field of deep learning, various time series prediction methods have been relatively mature in terms of applications. The traditional Transformer-based efficient structure suitable for long-term time series prediction halves the input by using a stacked structure, makes the extraction of self-attention more concentrated, and effectively processes extremely long input sequences. The traditional sequence decoupling Transformer framework learns the seasonal information of the time series through internal autocorrelation and encoder autocorrelation modules and combines the trend sequence to obtain the prediction effect.

[0004] The above methods have realized the prediction of multi-source energy data sequences to a certain extent, but there are still certain defects, including the lack of a global perspective in long-term prediction of energy data and the difficulty in capturing long-term dependencies to achieve accurate prediction. It should be particularly noted that the multi-source energy data in the real world is input in the form of a data stream, so the prediction model can only obtain the semantic information within the current time period. And the long-term time series contains more data information, and a global perspective is needed to obtain more semantic information to realize relationship construction. Summary of the Invention

[0005] In order to solve the deficiencies existing in the prior art, the present invention provides a long-term prediction method and system for multi-source energy data integrating prompt information to solve the technical problems of the lack of a global perspective in long-term prediction of energy data and the difficulty in capturing long-term dependencies to achieve accurate prediction.

[0006] To solve the above technical problems, the present invention adopts the following technical solutions.

[0007] The present invention first discloses a long-term prediction method for multi-source energy data integrating prompt information, and the method includes the following steps: Collect multi-source energy data, and perform missing value filling and normalization processing on the collected data; Build a temporal prompt module in the encoder to model and analyze the original input sequence of the multi-source energy data. Perform average pooling through the sequence decomposition module to decompose the original input sequence and obtain a trend sequence and a seasonal input sequence; Perform sequence update on the original input sequence through the feed-forward module and the sequence decomposition module with residual connection to obtain cross information, and input the cross information into the decoder; In the decoder, model and analyze the seasonal input sequence through the temporal prompt module, and fuse the calculated seasonal data and trend data through the sequence decomposition module; Build a multi-head prompt attention module to model and analyze the cross information between the seasonal input sequence and the encoder input; Perform sequence update through the feed-forward module and the sequence decomposition module with residual connection, and input the output result into the fully connected layer to obtain the prediction result.

[0008] The present invention further includes the following preferred solutions: Normalize the collected data, which further includes: Perform normalization using the following formula:

[0009] where represents each data, represents the minimum value of the entire data, represents the maximum value of the entire data.

[0010] The construction of the temporal prompt module further includes: Build a temporal prompt pool , represents the size of the temporal prompt pool, i.e., the prompt key; Build Prompt Key , is an adjustable hyperparameter representing the dimension of Prompt Key; Build the relationship between and Prompt Key:

[0011] where is obtained after projection by ; By selecting the most similar prompt information, supplement the sequence:

[0012] wherein represents the value of the first most similar hint messages, represents the index value of the first most similar messages, represents the information of the first data with the largest selected value; Update the time-series hint pool information and index value to the time-series hint pool:

[0013] wherein, is a splicing operation, represents the data of the hint pool after updating the index value of the first k most similar messages after matching to the hint pool.

[0014] The construction of the multi-head hint attention module, through which the cross information between the seasonal input sequence and the encoder input is modeled and analyzed, further includes: Construct a hint pool , represents the size of the hint pool, represents the dimension of the hint pool, and is consistent with the dimension, and construct a Prompt Key , represents the size of the Prompt Key, which is the same as the size of the hint pool, represents the dimension of the Prompt Key; Construct the relationship with the Prompt Key, relationship with the Prompt Key:

[0015]

[0016]

[0017] wherein is obtained after projection, is obtained through projection; By selecting the most similar hint messages, supplement the and sequences:

[0018]

[0019]

[0020] wherein represents the value of the top most similar hint messages, represents the index value of the top most similar messages, represents the information of the top data with the largest selected values; Update the hint pool and the most similar hint messages to the hint pool:

[0021]

[0022]

[0023] Concatenate with to obtain sequence, and perform dual-channel attention processing on the multi-head channel and the data dimension respectively:

[0024]

[0025] wherein, is a learnable parameter, and the final weight sequence is obtained through concatenation:

[0026] wherein, is a projection operation.

[0027] The present invention also discloses a long-term prediction system for multi-source energy data using the fused hint information as described above, including: A multi-source energy data acquisition module for acquiring multi-source energy data and performing missing value filling and normalization processing on the acquired data; A time series hint module construction module for constructing a time series hint module in the encoder, performing modeling analysis on the original input sequence of the multi-source energy data, and performing average pooling through a sequence decomposition module to decompose the original input sequence to obtain a trend sequence and a seasonal input sequence; An input module for updating the original input sequence through a feed-forward module and a sequence decomposition module with residual connections to obtain cross information, and inputting the cross information into the decoder; A fusion module, which is used to perform modeling analysis on the seasonal input sequence in the decoder through a timing prompt module, and fuse the calculated seasonal data and trend data through a sequence decomposition module; A multi-head prompt attention module construction module, which is used to construct a multi-head prompt attention module, and perform modeling analysis on the cross information between the seasonal input sequence and the encoder input through the multi-head prompt attention module; A prediction result output module, which is used to update the sequence through a feed-forward module and a sequence decomposition module with residual connection, and input the output result into a fully connected layer to obtain a prediction result.

[0028] Correspondingly, the present application also discloses a terminal, including a processor and a storage medium; The storage medium is used to store instructions; The processor is used to operate according to the instructions to execute the steps of the multi-source energy data long-term prediction method according to the foregoing fused prompt information.

[0029] Correspondingly, the present application also discloses a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the multi-source energy data long-term prediction method according to the foregoing fused prompt information are implemented.

[0030] The beneficial effect of the present invention is that, compared with the prior art, the present invention provides a multi-source energy data long-term prediction method and system that fuses prompt information. Based on Transformer, a timing prompt module, a multi-head prompt attention module, and a sequence decomposition module, etc., it solves the problem of the lack of a global view in the long-term prediction of energy data and the difficulty in capturing long-term dependencies to achieve accurate prediction. Description of the Drawings

[0031] Figure 1 is a flowchart of the multi-source energy data long-term prediction method that fuses prompt information in the present invention.

[0032] Figure 2 is a schematic diagram of the overall framework of the algorithm in the present invention.

[0033] Figure 3 is a schematic diagram of the multi-head prompt attention module included. Detailed Embodiments

[0034] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0035] The embodiments described in this application are only a part of the embodiments of the present invention, not all of them. Based on the spirit of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work belong to the protection scope of the present invention.

[0036] Aiming at the deficiencies of the prior art, the present invention proposes a long-term prediction method and system for multi-source energy data integrated with prompt information, which collects multi-source energy data, fills in the missing values and normalizes the collected multi-source energy data, and uses the temporal prompt method to learn the decoupled seasonal data information, stores these information as knowledge parameters and updates them to the prompt pool, and shares the seasonal knowledge parameters at multiple time steps in the form of multi-channel multi-head prompt attention to capture the overall trend details of the time series seasonality from a global perspective. Based on Transformer, temporal prompt module, multi-head prompt attention module and sequence decomposition module, etc., the present invention solves the problem that the global view is missing during the long-term prediction of energy data and it is difficult to capture long-term dependencies to achieve accurate prediction.

[0037] The long short-term memory network adds three gating units on the basis of the recurrent neural network, which can effectively solve the problem of gradient disappearance or explosion in long sequences. The attention mechanism is a model that simulates the human brain's attention. It draws on the characteristics of the human brain that at a certain specific moment, the attention to things will be concentrated on specific places, while reducing or even ignoring the attention to other parts. The attention mechanism mainly obtains the key information by allocating probabilities, so as to assist the long short-term memory network to better capture long-term dependencies, thereby improving the accuracy of model prediction. It is more intelligent than traditional mathematical statistical methods and can adaptively adjust the weights of different key information.

[0038] Long-term sequence prediction has very broad application scenarios in the real world, such as practical problems like power resource estimation, disease spread and diffusion, economic development prediction, and mountain change trend prediction.

[0039] See Figure 1 As shown, the long-term prediction method for multi-source energy data integrated with prompt information disclosed by the present invention includes the following steps: Step 1: Collect multi-source energy data, and perform missing value filling and normalization processing on the collected data.

[0040] The collected multi-source energy data is filled in and complemented for missing values by the interpolation method, and the data is normalized, which is expressed as follows:

[0041] Among them, represents each data, represents the minimum value of the entire data, represents the maximum value of the entire data.

[0042] Step 2: Build a temporal prompt module in the encoder to perform modeling analysis on the original input sequence of the multi-source energy data. Through the sequence decomposition module, average pooling is performed to decompose the original input sequence to obtain a trend sequence and a seasonal input sequence.

[0043] Specifically, in combination with Figure 2 and Figure 3 as shown, the multi-source energy data input sequence is copied into , which means that the input energy data is copied into three copies of the original data through self-attention.

[0044] And it is formed into a multi-head form through projection , which means performing a dimensional transformation on Q, K, V and converting their dimensions to n×d so that the data can be input into the multi-head attention of n heads. Among them, . represents the dimension of the temporal prompt pool and is consistent with the dimension of . The construction process of the temporal prompt module is as follows: Step 2.1: Build a temporal prompt pool , which represents the size of the temporal prompt pool (prompt key), and build the prompt key Prompt Key , is an adjustable hyperparameter representing the dimension of Prompt Key.

[0045] Step 2.2: Build the relationship between and Prompt Key:

[0046] Among them is obtained after projection through .

[0047] Step 2.3: By selecting the most similar prompt messages, supplement the sequence:

[0048] Among them represents the value of the first most similar prompt messages, Representing the previous index values of the k most similar pieces of information, representing the information of the top data with the largest selected values.

[0049] Step 2.4: Update the temporal prompt pool information and index values to the temporal prompt pool:

[0050] Among them, is the concatenation operation. represents the data in the prompt pool after updating the index values of the top k most similar pieces of information after matching to the prompt pool.

[0051] Step 2.5: Concatenate with to obtain sequence, and based on the sequence, obtain the seasonal sequence with prompt information through the attention mechanism .

[0052] Step 3: Update the sequence of the original input sequence through a feed-forward module and a sequence decomposition module with residual connections to obtain cross information, and input the cross information into the decoder.

[0053] The input in the encoder stage is the data of the past time steps , and it is assumed that there are encoder layers. The overall equation of the th encoder layer is . Its expression is:

[0054]

[0055] Among them is the eliminated trend part. represents the output of the th encoder layer, and is the embedded . respectively represent the seasonal components after the th sequence decomposition module in the st layer. SeriesDecomp is the sequence decomposition operation.

[0056] Step 4: In the decoder, model and analyze the seasonal input sequence through the temporal prompt module, and fuse the calculated seasonal data and trend data through the sequence decomposition module.

[0057] The data input is decomposed into two parts, which respectively reflect the long-term and seasonal characteristics of the sequence. However, direct decomposition is not achievable for future sequences because the future is unknown. The long-term stationary trend is gradually extracted from the predicted intermediate latent variables through the sequence decomposition module. The moving average is adjusted to eliminate periodic fluctuations and highlight the long-term trend. For a sequence with an input length of the process is as follows: For a sequence with an input length of

[0058]

[0059] where represent seasonal data and the extracted trend data respectively. Padding represents the padding operation. represents the moving average operation and performs a padding operation to keep the sequence length unchanged.

[0060] Step 5: Construct a multi-head prompt attention module to model and analyze the cross information between the seasonal input sequence and the encoder input through the multi-head prompt attention module.

[0061] Specifically, the construction of the multi-head prompt attention module includes constructing a prompt pool and a Prompt Key, and the construction method is as follows: Step 5.1: Construct the prompt pool , represents the size of the prompt pool, represents the dimension of the prompt pool, which is the same as the dimension of . And construct the Prompt Key , represents the size of the Prompt Key, which is the same as the size of the prompt pool, represents the dimension of the Prompt Key, which is also an adjustable hyperparameter.

[0062] Step 5.2: Construct the relationship between and the Prompt Key, and the Prompt Key:

[0063]

[0064]

[0065] where is obtained after projection, is obtained through projection; O represents the length of the predicted sequence given each time for the model prediction task; Step 5.3: By selecting the most similar prompt messages, and the sequence is supplemented:

[0066]

[0067]

[0068] Among them, represents the value of the most similar prompt messages, represents the index value of the most similar messages, represents the information of the largest

[0069] Step 5.4: Update the prompt pool and the most similar prompt messages to the prompt pool:

[0070]

[0071]

[0072] Step 5.5: Concatenate with to obtain the sequence, and perform dual-channel attention processing on the multi-head channel and the data dimension respectively:

[0073]

[0074] Among them, is a learnable parameter. And the final weight sequence is obtained through concatenation:

[0075] Among them, is a projection operation.

[0076] Step 6: Update the sequence through the feed-forward module and the sequence decomposition module with residual connections, and input the output result into the fully connected layer to obtain the prediction result.

[0077] Assume there are decoder layers. Using the latent variable from the encoder, the th decoder layer is , the decoder expression is shown as follows:

[0078]

[0079]

[0080]

[0081] where , represents the output of the -th decoder layer. is the embedded , used for depth transformation, used for accumulation. respectively represent the seasonal component and the trend periodic component after the -th layer and the -th sequence decomposition module. represents the projection of the -th extracted trend . PromptAtt represents the multi-head prompt attention operation; FeedForward represents the feed-forward network layer in the Transformer framework.

[0082] To quantitatively evaluate the performance of the model proposed by the present invention, the present invention uses the root mean square error (MSE) index and the mean absolute error (MAE) index as error evaluation indexes. As follows:

[0083]

[0084] where: is the number of prediction points; represents the true value of the i-th point; represents the predicted value of the i-th point. The smaller the values of the above evaluation indexes, the higher the accuracy of the prediction task.

[0085] The multi-step long-term prediction of the prediction method of the present invention is carried out on (1) the ETT dataset: It contains power-related data collected from two sites within two years. To explore the performance of the model on data with different granularities, different sampling frequencies are used to obtain hourly data {ETTh 1, ETTh 2} and 15-minute data {ETTm 1, ETTm 2}. (2) The Electricity dataset: It contains the hourly electricity consumption of 321 customers within 2 years. And it is compared with Atuoformer, Informer, LogTrans, Reformer, LSTNet, LSTM, and TCN, and better results are obtained. The model is evaluated by the evaluation metrics MSE and MAE. The smaller the values of these two evaluation metrics, the higher the prediction accuracy of the model algorithm. The results are shown in Table 1 below.

[0086] It can be seen that the present invention is superior to all baseline models in all cases, demonstrating powerful performance. Under the input96-predict-192 setting, compared with the previous state-of-the-art results, the performance of the present invention has the highest relative improvement of 12.9% (0.255 → 0.222) in the MSE evaluation metric of the ETTm2 dataset, and the highest relative improvement of 6.4% (0.222 → 0.208) in the MSE evaluation metric of the Electricity dataset.

[0087]

[0088] Table 1 The present invention conducts experimental comparisons of relevant long / short-term time series prediction methods on the ETT dataset by setting the same input length and prediction length. The experimental results of the unit data are shown in Table 2. Among them, Atuoformer, Informer, LogTrans, N-BEATS, DeepAR, Prophet, and ARMIA are all mainstream methods for unit time series prediction. The model is evaluated by the evaluation metrics MSE and MAE. The smaller the values of these two evaluation metrics, the higher the prediction accuracy of the model algorithm. The results are shown in Table 2.

[0089]

[0090] Table 2 The beneficial effect of the present invention is that, compared with the prior art, the present invention provides a multi-source energy data long-term prediction method and system that integrates hint information. Based on Transformer, a time series hint module, a multi-head hint attention module, and a sequence decomposition module, etc., it solves the problem of the lack of a global view in the long-term prediction of energy data and the difficulty in capturing long-term dependencies to achieve accurate prediction.

[0091] The present invention can be a system, a method, and / or a computer program product. The present invention also discloses a long-term prediction system for multi-source energy data with fused prompt information based on the aforementioned long-term prediction method for multi-source energy data with fused prompt information, including: A multi-source energy data acquisition module, configured to acquire multi-source energy data and perform missing value filling and normalization processing on the acquired data; A temporal prompt module construction module, configured to construct a temporal prompt module in an encoder, perform modeling analysis on the original input sequence of the multi-source energy data, perform average pooling through a sequence decomposition module, decompose the original input sequence, and obtain a trend sequence and a seasonal input sequence; An input module, configured to perform sequence update on the original input sequence through a feed-forward module and a sequence decomposition module with residual connections to obtain cross information, and input the cross information into a decoder; A fusion module, configured to perform modeling analysis on the seasonal input sequence through a temporal prompt module in the decoder, and fuse the calculated seasonal data and trend data through a sequence decomposition module; A multi-head prompt attention module construction module, configured to construct a multi-head prompt attention module, and perform modeling analysis on the cross information between the seasonal input sequence and the encoder input through the multi-head prompt attention module; A prediction result output module, configured to perform sequence update through a feed-forward module and a sequence decomposition module with residual connections, and input the output result into a fully connected layer to obtain a prediction result.

[0092] Based on the spirit of the present invention, those skilled in the art can easily conceive that a computer program product can be obtained based on the aforementioned long-term prediction method for multi-source energy data with fused prompt information. The computer program product may include a computer-readable storage medium having computer-readable program instructions thereon for causing a processor to implement various aspects of the present disclosure. That is, the present application also includes a terminal, including a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the long-term prediction method for multi-source energy data with fused prompt information according to the foregoing.

[0093] A computer-readable storage medium can be a tangible device that can hold and store instructions for use by an instruction execution device. A computer-readable storage medium can be, for example - but not limited to - an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer-readable storage medium include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disc (DVD), a memory stick, a floppy disk, a mechanically encoded device, such as a punched card or raised structures in grooves storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage medium used herein is not construed as an instantaneous signal itself, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagated through a waveguide or other transmission medium (e.g., an optical pulse through an optical fiber cable), or an electrical signal transmitted through a wire.

[0094] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to various computing / processing devices, or downloaded to an external computer or external storage device through a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include a copper transmission cable, an optical fiber transmission, a wireless transmission, a router, a firewall, a switch, a gateway computer, and / or an edge server. A network adapter or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in a computer-readable storage medium in each computing / processing device.

[0095] The computer program instructions for performing the operations of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state-setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages such as Smalltalk, C++, etc., and conventional procedural programming languages such as the "C" language or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, executed as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., through the Internet using an Internet service provider). In some embodiments, by using the status information of the computer-readable program instructions to customize an electronic circuit, such as a programmable logic circuit, a field-programmable gate array (FPGA), or a programmable logic array (PLA), the electronic circuit can execute the computer-readable program instructions to implement various aspects of the present disclosure.

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: modifications or equivalent substitutions can still be made to the specific embodiments of the present invention, and any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered by the protection scope of the claims of the present invention.

Claims

1. A multivariate energy data long-term prediction method integrating prompt information, characterized in that: The following steps are involved: Collect multi-energy data, and perform missing fill and normalization processing on the collected data; A time series prompt module is constructed in the encoder to model and analyze the original input sequence of the multi-energy data, and the original input sequence is decomposed by average pooling through the sequence decomposition module to obtain a trend sequence and a seasonal input sequence; Performing sequence updating on the original input sequence through a feedforward module and a sequence decomposition module with a residual connection to obtain cross information, and inputting the cross information into a decoder; In the decoder, the seasonal input sequence is modeled and analyzed by a timing prompt module, and the calculated seasonal data is integrated with the trend data by a sequence decomposition module; Constructing a multi-head prompt attention module, and modeling and analyzing the cross information of the seasonal input sequence and the encoder input through the multi-head prompt attention module; The sequence is updated through the feedforward module and the sequence decomposition module with residual connection, and the output result is input into the fully connected layer to obtain the prediction result.

2. The multivariate energy data long-term prediction method integrating prompt information according to claim 1 is characterized in that: The collected data is normalized, further comprising: The following formula is used for normalization: in, Represents each data, Represents the minimum value of the entire data, Indicates the maximum value of the entire data.

3. The multivariate energy data long-term prediction method integrating prompt information according to claim 2 is characterized in that: The construction timing prompt module further comprises: Build a Timing Hint Pool , Represents the size of the timing prompt pool, i.e., the prompt key; Constructing the Prompt Key , is an adjustable hyperparameter, representing the dimension of PromptKey; Build Relationship with Prompt Key: in It is through Obtained after projection; By selecting the most similar A reminder message, The sequence is supplemented: in Before the Representative The value of the most similar prompt information, Before the Representative The index value of the most similar information, Represents the selection of the largest value Information about individual data; Update the timing hint pool information and index value to the timing hint pool: in, It is a splicing operation. The data in the prompt pool after the index values ​​of the top k most similar information after matching are updated to the prompt pool.

4. The multi-energy data long-term prediction method integrating prompt information according to claim 3 is characterized in that: The constructing of a multi-head prompt attention module, and modeling and analyzing the cross information of the seasonal input sequence and the encoder input through the multi-head prompt attention module, further includes: Building a Tip Pool , Represents the size of the prompt pool, represents the dimension of the prompt pool and is related to The dimensions are consistent; and construct the Prompt Key , Represents the size of PromptKey, which is the same as the size of the prompt pool. Indicates the dimension of the Prompt Key; Build Relationship with Prompt Key, Relationship with Prompt Key: in It is obtained after projection. is obtained after projection; O represents the length of the prediction sequence given each time in the model prediction task; By selecting the most similar A reminder message, as well as The sequence is supplemented: in Before the Representative The value of the most similar prompt information, Before the Representative The index value of the most similar information, Represents the selection of the largest value Information about individual data; Update the prompt pool with the most similar prompt information: Will and Splice to get Sequence, respectively for multi-head channels And data dimensions Perform dual-channel attention processing: in, For learnable parameters, the final weight sequence is obtained by concatenation: in, For projection operation.

5. A multi-energy data long-term prediction system integrating prompt information, characterized in that: include: The multi-energy data collection module is used to collect multi-energy data and perform missing filling and normalization processing on the collected data; A timing prompt module construction module is used to construct a timing prompt module in an encoder, model and analyze the original input sequence of the multivariate energy data, perform average pooling through a sequence decomposition module, decompose the original input sequence, and obtain a trend sequence and a seasonal input sequence; An input module, used for performing sequence update on the original input sequence through a feedforward module and a sequence decomposition module with a residual connection to obtain cross information, and inputting the cross information into a decoder; A fusion module, used for modeling and analyzing the seasonal input sequence in the decoder through the timing prompt module, and fusing the calculated seasonal data with the trend data through the sequence decomposition module; A multi-head prompt attention module construction module, used to construct a multi-head prompt attention module, and to model and analyze the cross information of the seasonal input sequence and the encoder input through the multi-head prompt attention module; The prediction result output module is used to update the sequence through the feedforward module and the sequence decomposition module with residual connection, and input the output result into the fully connected layer to obtain the prediction result.

6. The multi-energy data long-term prediction system integrating prompt information according to claim 5 is characterized in that: The multi-energy data acquisition module is further used for: The following formula is used for normalization: in, Represents each data, Represents the minimum value of the entire data, Indicates the maximum value of the entire data.

7. The multi-energy data long-term prediction system integrating prompt information according to claim 6 is characterized in that: The timing prompt module construction module is further used to: Build a Timing Hint Pool , Represents the size of the timing prompt pool, i.e., the prompt key; Constructing the Prompt Key , is an adjustable hyperparameter, representing the dimension of PromptKey; Build Relationship with Prompt Key: in It is through Obtained after projection; By selecting the most similar A reminder message, The sequence is supplemented: in Before the Representative The value of the most similar prompt information, Before the Representative The index value of the most similar information, Represents the selection of the largest value Information about individual data; Update the timing hint pool information and index value to the timing hint pool: in, It is a splicing operation. The data in the prompt pool after the index values ​​of the top k most similar information after matching are updated to the prompt pool.

8. The multi-energy data long-term prediction system integrating prompt information according to claim 7 is characterized in that: The multi-head prompt attention module building module is used to: Building a Tip Pool , Represents the size of the prompt pool, represents the dimension of the prompt pool and is related to The dimensions are consistent and the Prompt Key is constructed , Represents the size of PromptKey, which is the same as the size of the prompt pool. Indicates the dimension of the Prompt Key; Build Relationship with Prompt Key, Relationship with Prompt Key: in It is obtained after projection. is obtained after projection; O represents the length of the prediction sequence given each time in the model prediction task; By selecting the most similar A reminder message, as well as The sequence is supplemented: in Before the Representative The value of the most similar prompt information, Before the Representative The index value of the most similar information, Represents the selection of the largest value Information about individual data; Update the prompt pool with the most similar prompt information: Will and Splice to get Sequence, respectively for multi-head channels And data dimensions Perform dual-channel attention processing: in, For learnable parameters, the final weight sequence is obtained by concatenation: in, For projection operation.

9. A terminal comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is used to operate according to the instructions to execute the steps of the method for long-term prediction of multivariate energy data integrating prompt information according to any one of claims 1-4.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method for long-term prediction of multivariate energy data integrating prompt information described in any one of claims 1 to 4 are implemented.

Images