Multi-modal electricity load forecasting method and system based on news and time series embedding fusion technology, and medium

Abstract

The application relates to a multi-modal power load prediction method and system based on news and time sequence embedding fusion technology and a medium, the method comprising the following steps: acquiring multi-source news texts and a multivariate load sequence of a current period; performing correlation screening on the multi-source news texts by adopting an RRFU framework based on a generative agent, and generating news abstracts; for the multivariate load sequence and the news abstracts, performing coding by adopting a dual-modal coding architecture based on knowledge enhancement, so as to respectively obtain time sequence embedding and text semantic embedding; based on the time sequence embedding and the text semantic embedding, performing cross-modal alignment and fusion to obtain fused features; and performing prediction on the fused features to output power load prediction values of future steps. Compared with the prior art, the application has the advantages that the prediction bottleneck of a traditional method in a complex scene can be broken through, and the accuracy, robustness and explainability of power load prediction are improved.

Description

Technical Field

[0001] This invention relates to the field of electricity load forecasting, and in particular to a multimodal electricity load forecasting method, system and medium based on news and time-series embedding fusion technology. Background Technology

[0002] Load forecasting is the cornerstone of ensuring the safe, stable, and economical operation of the power system. Its accuracy and reliability directly determine the rationality of power generation planning, the efficiency of reserve capacity allocation, and the scientific nature of grid dispatch decisions. High-quality load forecasting helps achieve coordination between power generation, grid, and load, thereby significantly reducing overall power generation and operation and maintenance costs, minimizing unnecessary energy consumption and carbon emissions, and ultimately ensuring the continuity of power supply and the stability of social electricity consumption. However, current mainstream forecasting methods still face significant technical bottlenecks when dealing with increasingly complex power environments.

[0003] Traditional methods based on statistical learning and shallow machine learning are limited primarily by their insufficient utilization of input information. These methods often rely excessively on the temporal patterns inherent in historical load data for modeling, failing to systematically and deeply integrate highly correlated dynamic external factors such as meteorological conditions, holiday types, changes in user behavior patterns, macroeconomic indicators, and sudden social events. This results in a significant decrease in the model's generalization ability and robustness when encountering complex scenarios such as extreme heat waves, cold waves, major public events, or sudden changes in local economic activity. The model struggles to accurately capture and respond to non-stationary and nonlinear characteristics such as rapid load fluctuations and sudden peaks.

[0004] On the other hand, while deep learning-based methods demonstrate advantages in automatically extracting deep temporal dependencies and complex patterns from load sequences, they also have significant shortcomings. First, these models still fall short in the deep fusion of multimodal data. Existing research and applications mostly focus on modeling structured time-series data (such as historical load and temperature), often neglecting the rich semantics contained in unstructured information sources such as news texts, social media, and industry reports. These texts actually contain a large number of event signals that may trigger load disturbances, such as announcements of large-scale project construction, adjustments to industrial policies, or reports of sudden public events. Existing models struggle to effectively analyze and fuse this crucial event-driven information. Second, the inherent "black box" nature of deep learning models leads to poor interpretability of their prediction process and opaque decision-making logic. This, to some extent, affects dispatchers' trust in the prediction results and hinders the deep application and iterative optimization of these models in safety-sensitive, high-reliability power dispatching scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a multimodal power load forecasting method, system, and medium based on news and time-series embedding fusion technology to achieve more reliable and accurate forecasting.

[0006] The objective of this invention can be achieved through the following technical solutions: A multimodal electricity load forecasting method based on news and time-series embedding fusion technology includes the following steps: Obtain multi-source news text and multivariate load series for the current period ; The RRFU framework based on generative agents is used to perform relevance filtering on the multi-source news texts and generate news summaries. ; For the multivariate load sequence and news summary A knowledge-enhanced bimodal coding architecture is used for encoding, resulting in temporal embeddings. and text semantic embedding ; Based on the timing embedding and text semantic embedding Cross-modal alignment and fusion are performed to obtain the fused features. ; The fused features Make predictions and output the future. The predicted electricity load for the step.

[0007] Furthermore, the aforementioned generative agent-based RRFU framework is used to generate news digests. The steps include: Based on the set prediction window Candidate sets are retrieved from the multi-source news texts. ,in For the first News text; Inference agent based on current load baseline For candidate set Each news article text Perform saliency reasoning to obtain The potential factors affecting electricity demand are expressed as follows: , In the formula, for Potential factors influencing electricity demand For saliency inference function, For regional context information; The filtering agent filters the candidate set. Filter to retain only the potential influencing factors. The final news collection is obtained from news texts that are significant and highly time-relevant. For each news text in the final news collection, a corresponding news summary is generated. .

[0008] Furthermore, the generative agent-based RRFU framework also introduces an evaluation agent to update the filtering logic of the filtering agent based on the electricity load forecast value predicted in the previous cycle. Specific steps include: Obtain the predicted electricity load value from the previous period. With the corresponding true value Calculate the prediction error ; Determine if it satisfies If so, a reflection operation is triggered to update the filtering logic of the filtering proxy: , In the formula, For the preset threshold, For the filtering logic in the next cycle, For reflection function, For the current period's filtering logic, The selected news text. These are the key news texts that were missed, leading to prediction bias. If not, no update will be performed.

[0009] Furthermore, the knowledge-enhanced bimodal coding architecture includes parallel temporal coding branches and semantic coding branches, wherein the temporal coding branches are used to process the multivariate load sequence. Encode to obtain temporal embedding The semantic encoding branch is used for the news summary Encode to obtain text semantic embedding .

[0010] Furthermore, the execution steps of the timing coding branch include: For the multivariate load sequence arranging them according to each channel Divide into multiple block sequences; For each block sequence, the corresponding temporal embedding is obtained through linear projection. Finally, a multivariate load sequence was obtained. Corresponding timing embedding The timing embedding Represented as: , In the formula, For feature embedding operations, For channel The Middle Features of each block sequence, subscript The number of block sequences, For position encoding.

[0011] Furthermore, the execution steps of the semantic encoding branch include: Based on the multivariate load sequence Calculate its statistics, including one or more of the mean, variance, and peak value; Obtain supplementary variables for the current period, including at least one of holidays and weather warnings; The news summary Statistics and supplementary variables are processed using natural language to construct prompt word templates. , is represented as: , In the formula, For statistical descriptors, For splicing operations, To summarize the operation, For contextual information; The prompt word template Input pre-trained large language model Processing will be performed in the process, The output vector of the last token in the hidden layer is used as the global semantic representation; The global semantic representation is then subjected to linear transformation and layer normalization sequentially to obtain the final text semantic embedding. , is represented as: , In the formula, For layer normalization operations, This is a linear transformation operation for fully connected layers.

[0012] Furthermore, the step of performing cross-modal alignment includes: For the timing embedding Perform global average pooling to obtain the results for each channel. Channel characteristics ; Calculate the characteristics of each channel With the text semantic embedding Correlation weight matrix between The calculation expression is: , In the formula, and It is a linear transformation matrix. For normalization operations, This is the scaling factor; Based on the correlation weight matrix embedding text semantics Mapping to the corresponding channels generates aligned feature vectors. , is represented as: , In the formula, For multiplication operations, This is a mapping function.

[0013] Furthermore, the obtained fused features The steps include: Introducing learnable adaptive coefficients to align the cross-modal feature vectors and the timing embedding Weighted integration is performed to obtain the fused features. , is represented as: , In the formula, These are learnable adaptive coefficients. It is a non-linear activation function.

[0014] This invention also provides a multimodal electricity load forecasting system based on news and time-series embedding fusion technology, comprising: Data acquisition module: used to acquire multi-source news text and multivariate load series for the current period. ; Summary generation module: Used to perform relevance filtering on the multi-source news text using a generative agent-based RRFU framework, and generate news summaries. ; Encoding module: used for encoding the multivariate load sequence and news summary A knowledge-enhanced bimodal coding architecture is used for encoding, resulting in temporal embeddings. and text semantic embedding ; Alignment and fusion module: used for alignment and fusion based on the temporal embedding and text semantic embedding Cross-modal alignment and fusion are performed to obtain the fused features. ; Prediction module: used for the fused features Make predictions and output the future. The predicted electricity load for the step.

[0015] The present invention also provides a storage medium storing a program thereon, which, when executed, implements a multimodal power load prediction method based on news and time-series embedding fusion technology as described above.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention integrates structured multivariate load sequences with unstructured news text information, and utilizes the information filtering mechanism of the RRFU framework and the prediction framework of bimodal coding-cross-modal alignment to achieve collaborative modeling of multi-source heterogeneous data. Simultaneously, in the task of integrating time-series data and news text information, multivariate load sequences and news summaries are embedded and generated, and then the two are fused. In the bimodal coding and cross-modal alignment tasks, the fused embedding is aligned with the original time-series data in a multimodal manner before prediction, thereby obtaining more reliable and accurate prediction results.

[0017] (2) This invention uses a large model-driven “retrieval-reasoning-filtering-update” closed-loop framework to automatically identify, extract and evaluate key event summaries that are strongly related to power load from news texts, realize early perception and dynamic response to load changes, fundamentally solve the problem of delayed response and large prediction deviation of traditional methods to sudden events, and significantly improve the response speed and prediction accuracy of sudden social / policy events.

[0018] (3) The present invention introduces a dual-modal coding architecture. In this architecture, the temporal coding branch is used to capture the multivariate coupling dependence of the load, and the semantic coding branch enhances the semantic representation by generating knowledge through cue engineering and large language model. It also introduces external common sense and event causal knowledge, which significantly improves the interpretability and prediction robustness of complex fluctuations. This enables the model to learn long-term trends, short-term fluctuations and the influence of external events at the same time, and has a stronger fitting and generalization ability for nonlinear, non-stationary and strongly disturbed loads, thereby enhancing the ability to capture complex load fluctuations and the prediction robustness.

[0019] (4) The present invention adopts a channel-level similarity retrieval mechanism to achieve precise alignment of temporal embedding and text semantic embedding in the semantic space, so that text event information can be effectively injected into temporal features, solve the technical defects of insufficient multimodal data fusion and information misalignment, thereby achieving deep alignment of multimodal data and eliminating modal misalignment.

[0020] (5) This invention generates news summaries and knowledge-enhanced embeddings through a large language model, enabling the prediction process to have traceable and explainable event evidence, so that load fluctuations can not only be predicted but also explained, thereby improving the credibility and security of power grid dispatching decisions and providing reliable technical support for the intelligent dispatching and safe operation of the power system. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0023] Example 1 This embodiment provides a multimodal electricity load forecasting method based on news and time-series embedding fusion technology. It employs a strategy based on news and time-series embedding fusion to address the robustness, interpretability, and accuracy issues in traditional electricity load forecasting tasks. This method integrates structured time-series electricity data with unstructured news text information, leveraging the contextual modeling and knowledge reasoning capabilities of a large language model to construct an information filtering mechanism of "retrieval-reasoning-filtering-updating" and a prediction framework of bimodal coding and cross-modal alignment, achieving collaborative modeling of multi-source heterogeneous data. In the task of integrating time-series data and news text information, the large language model is used to embed and generate time-series data and news data respectively, and then the two are fused. In the bimodal coding and cross-modal alignment tasks, the obtained fused embedding is multimodally aligned with the original time-series data before prediction, resulting in more reliable and accurate prediction results. Specifically, as shown... Figure 1 As shown, the method includes the following steps: S1. Unstructured news text filtering and extraction based on the RRFU framework.

[0024] This step addresses the characteristics of power load being highly susceptible to interference from social events, policy changes, and extreme weather. It designs a "Retrieval-Inference-Filtering-Update" (RRFU) framework based on a generative agent. Utilizing a large language model-driven inference agent and evaluation agent, it performs deep filtering and feedback optimization on unstructured news text to extract highly relevant event summaries. Specifically: In the targeted retrieval and reasoning analysis stage, a prediction window is given. Candidate sets are retrieved from unstructured multi-source news texts acquired in the current cycle. , For the first News text. The inference agent is based on the current load baseline. The following hints and engineering logic are used to perform saliency inference on each news item: , in The potential impact factors (such as increase, decrease, or no change) of this news on electricity demand are used to infer its effects. For saliency inference function, This is regional contextual information, and its purpose is to provide news text. The regional context (such as the region's industrial structure, electricity consumption habits, climate characteristics, policy environment, etc.) serves as a reference for the regional dimension of saliency reasoning.

[0025] Subsequently, the filtering agent retains the candidate set based on the reasoning results. News texts with significant potential impact factors and high time relevance are selected to obtain the final news collection; then, a high-quality news summary is generated for each news text in the final news collection. .

[0026] The RRFU framework also features a cyclical mechanism. To correct biases or omissions in the screening process, an evaluation agent is introduced. This evaluation agent compares the electricity load forecast from the previous cycle. With the corresponding true value Calculate the prediction error .like This triggers a reflection operation to update the filtering logic of the filtering proxy: , In the formula, For the preset threshold, For the filtering logic in the next cycle, For reflection function, For the current period's filtering logic, The selected news text. These are the key news texts that were missed, leading to prediction bias. Otherwise, the reflection operation will not be triggered, and no update will be performed.

[0027] Through this self-correcting cycle, the method in this embodiment can learn from past mistakes, thereby establishing a refined news filtering strategy for specific grid load characteristics.

[0028] S2. Knowledge-enhanced bimodal coding modeling.

[0029] This embodiment employs a bimodal coding architecture, which includes parallel temporal coding and semantic coding branches. The temporal coding branch extracts multivariate dependency features from the structured payload data, while the semantic coding branch constructs natural language prompts and inputs them into a large language model to obtain knowledge-enhanced representations. Specifically, this step is as follows: In the time-series coding branch, the acquired structured multivariate load sequence will be... ( For window length, The number of channels (a variable) is input into this branch for block processing. Each channel... The sequence is transformed into a block sequence, and then a temporal embedding is obtained through linear projection. : , in The number of block sequences, For feature dimension, For channel The Middle Features of each block sequence This is for location encoding. This branch aims to capture the inherent temporal dynamics of electrical load, such as periodicity and trends.

[0030] In the semantic coding branch, this embodiment first obtains the multivariate load sequence. The statistics (including mean, variance, and peak value) are compared with the news summaries filtered by RRFU. In addition, supplementary variables (such as holidays and weather warnings) are processed using natural language to construct prompt word templates. : , In the formula, For statistical descriptors, For splicing operations, To summarize the operation, from Extract key information from it. This provides contextual information, that is, supplementary background information of the current scene, to provide a more comprehensive reasoning context.

[0031] Will Input pre-trained large language model The deep hidden layer is extracted as a knowledge-enhanced representation. To obtain alignment-friendly features, this embodiment extracts the output vector of the last token in the hidden layer as a global semantic representation, and performs linear transformation and layer normalization to obtain the final knowledge-enhanced embedding, which serves as the text semantic embedding. : , In the formula, For layer normalization operations, This is a linear transformation operation for fully connected layers.

[0032] Text semantic embedding It integrates common-sense background knowledge about electricity acquired during the pre-training stage of a large language model with semantic information about breaking news events contained in real-time news.

[0033] S3, cross-modal alignment and feature fusion.

[0034] This step utilizes a channel-level similarity retrieval mechanism to achieve precise alignment and injection of temporal embeddings and textual semantic embeddings in the semantic space. The specific steps are as follows: To solve the problem of timing embedding With text semantic embedding To address the spatial heterogeneity between features, this invention designs a cross-modal alignment mechanism that achieves precise feature alignment by calculating channel-level semantic relevance.

[0035] First, the temporal embeddings extracted from the temporal coding branch Global average pooling is used to obtain channel features. Calculate the features of each channel and embed them into the text semantics. Correlation weight matrix between : , in and It is a linear transformation matrix. For normalization operations, This is the scaling factor. The relevance weight matrix characterizes the influence weight of specific news information on different electricity consumption components. Based on this relevance weight matrix, this embodiment accurately maps semantic knowledge to the corresponding time-series channels, generating aligned feature vectors. ,in For multiplication operations, This is a mapping function.

[0036] In the feature fusion stage, learnable adaptive coefficients are introduced. The original temporal embeddings and aligned knowledge features are weighted and integrated, and the calculation formula is as follows: , in This is a non-linear activation function. The fused features It not only preserves the historical evolution of electricity consumption, but also incorporates prior events that can explain future load fluctuations.

[0037] S4. Predict the fused features.

[0038] In this step, the fused features Input a multivariate linear prediction layer to make predictions, and output the future. Predicted value of step This enables robust capture of complex load fluctuations.

[0039] For the above method, during the training phase, the prediction error can be minimized. Perform end-to-end training.

[0040] Example 2 This embodiment provides a multimodal electricity load forecasting system based on news and time-series embedding fusion technology. The system includes: Data acquisition module: used to acquire multi-source news text and multivariate load series for the current period. ; Summary generation module: Used to perform relevance filtering on the multi-source news text using a generative agent-based RRFU framework, and generate news summaries. ; Encoding module: used for encoding the multivariate load sequence and news summary A knowledge-enhanced bimodal coding architecture is used for encoding, resulting in temporal embeddings. and text semantic embedding ; Alignment and fusion module: used for alignment and fusion based on the temporal embedding and text semantic embedding Cross-modal alignment and fusion are performed to obtain the fused features. ; Prediction module: used for the fused features Make predictions and output the future. The predicted electricity load for the step.

[0041] The present invention also provides a storage medium storing a program thereon, which, when executed, implements a multimodal power load prediction method based on news and time-series embedding fusion technology as described above.

[0042] The rest are as in Example 1.

[0043] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0044] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0045] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0046] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0047] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0048] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0049] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A multimodal electricity load forecasting method based on news and time-series embedding fusion technology, characterized in that, Includes the following steps: Obtain multi-source news text and multivariate load series for the current period ; The RRFU framework based on generative agents is used to perform relevance filtering on the multi-source news texts and generate news summaries. The news digest generation uses the aforementioned generative agent-based RRFU framework. The steps include: Based on the set prediction window Candidate sets are retrieved from the multi-source news texts. ,in For the first News text; Inference agent based on current load baseline For candidate set Each news article text By performing saliency reasoning, we obtain The potential factors affecting electricity demand are expressed as follows: ， In the formula, for Potential factors influencing electricity demand For saliency inference function, For regional context information; The filtering agent filters the candidate set. Filter to retain only the potential influencing factors. The final news collection is obtained from news texts that are significant and highly time-relevant. For each news text in the final news collection, a corresponding news summary is generated. ; For the multivariate load sequence and news summary A knowledge-enhanced bimodal coding architecture is used for encoding, resulting in temporal embeddings. and text semantic embedding The knowledge-enhanced bimodal coding architecture includes parallel temporal coding branches and semantic coding branches, wherein the temporal coding branches are used to process the multivariate load sequence. Encode to obtain temporal embedding The semantic encoding branch is used for the news summary Encode to obtain text semantic embedding ; Based on the timing embedding and text semantic embedding Cross-modal alignment and fusion are performed to obtain the fused features. ; The fused features Make predictions and output the future. The predicted electricity load for the step.

2. The multimodal electricity load forecasting method based on news and time-series embedding fusion technology according to claim 1, characterized in that, The generative agent-based RRFU framework also introduces an evaluation agent to update the filtering logic of the filtering agent based on the electricity load forecast value predicted in the previous cycle. Specific steps include: Obtain the predicted electricity load value from the previous period. With the corresponding true value Calculate the prediction error ; Determine if it satisfies If so, a reflection operation is triggered to update the filtering logic of the filtering proxy: ， In the formula, For the preset threshold, For the filtering logic in the next cycle, For reflection function, For the current period's filtering logic, The selected news text. These are the key news texts that were missed, leading to prediction bias. If not, no update will be performed.

3. The multimodal electricity load forecasting method based on news and time-series embedding fusion technology according to claim 1, characterized in that, The execution steps of the timing coding branch include: For the multivariate load sequence arranging them according to each channel Divide into multiple block sequences; For each block sequence, the corresponding temporal embedding is obtained through linear projection. Finally, a multivariate load sequence was obtained. Corresponding timing embedding The timing embedding Represented as: ， In the formula, For feature embedding operations, For channel The Middle Features of each block sequence, subscript The number of block sequences, For position encoding.

4. The multimodal electricity load forecasting method based on news and time-series embedding fusion technology according to claim 1, characterized in that, The execution steps of the semantic encoding branch include: Based on the multivariate load sequence Calculate its statistics, including one or more of the mean, variance, and peak value; Obtain supplementary variables for the current period, including at least one of holidays and weather warnings; The news summary Statistics and supplementary variables are processed using natural language to construct prompt word templates. , represented as: ， In the formula, For statistical descriptors, For splicing operations, To summarize the operation, For contextual information; The prompt word template Input pre-trained large language model Processing will be performed in the process, The output vector of the last token in the hidden layer is used as the global semantic representation; The global semantic representation is then subjected to linear transformation and layer normalization sequentially to obtain the final text semantic embedding. , represented as: ， In the formula, For layer normalization operations, This is a linear transformation operation for fully connected layers.

5. The multimodal electricity load forecasting method based on news and time-series embedding fusion technology according to claim 1, characterized in that, The steps for performing cross-modal alignment include: For the timing embedding Perform global average pooling to obtain the results for each channel. Channel characteristics ; Calculate the characteristics of each channel With the text semantic embedding Correlation weight matrix between The calculation expression is: ， In the formula, and It is a linear transformation matrix. For normalization operations, This is the scaling factor; Based on the correlation weight matrix embedding text semantics Mapping to the corresponding channels generates aligned feature vectors. , represented as: ， In the formula, For multiplication operations, This is a mapping function.

6. The multimodal electricity load forecasting method based on news and time-series embedding fusion technology according to claim 5, characterized in that, The obtained fused features The steps include: Introducing learnable adaptive coefficients to align the cross-modal feature vectors and the timing embedding Weighted integration is performed to obtain the fused features. , represented as: ， In the formula, These are learnable adaptive coefficients. It is a non-linear activation function.

7. A multimodal electricity load forecasting system based on news and time-series embedding fusion technology, characterized in that, include: Data acquisition module: used to acquire multi-source news text and multivariate load series for the current period. ; Summary generation module: Used to perform relevance filtering on the multi-source news text using a generative agent-based RRFU framework, and generate news summaries. The news digest generation uses the aforementioned generative agent-based RRFU framework. The steps include: Based on the set prediction window Candidate sets are retrieved from the multi-source news texts. ,in For the first News text; Inference agent based on current load baseline For candidate set Each news article text By performing saliency reasoning, we obtain The potential factors affecting electricity demand are expressed as follows: ， In the formula, for Potential factors influencing electricity demand For saliency inference function, For regional context information; The filtering agent filters the candidate set. Filter to retain only the potential influencing factors. The final news collection is obtained from news texts that are significant and highly time-relevant. For each news text in the final news collection, a corresponding news summary is generated. ; Encoding module: used for encoding the multivariate load sequence and news summary A knowledge-enhanced bimodal coding architecture is used for encoding, resulting in temporal embeddings. and text semantic embedding The knowledge-enhanced bimodal coding architecture includes parallel temporal coding branches and semantic coding branches, wherein the temporal coding branches are used to process the multivariate load sequence. Encode to obtain temporal embedding The semantic encoding branch is used for the news summary Encode to obtain text semantic embedding ; Alignment and fusion module: used for alignment and fusion based on the temporal embedding and text semantic embedding Cross-modal alignment and fusion are performed to obtain the fused features. ; Prediction module: used for the fused features Make predictions and output the future. The predicted electricity load for the step.

8. A storage medium having a program stored thereon, characterized in that, When the program is executed, it implements a multimodal electricity load forecasting method based on news and time-series embedding fusion technology as described in any one of claims 1-6.

Images