CN-121998694-A - Power supply and demand prediction method and system based on multi-model collaboration

CN121998694ACN 121998694 ACN121998694 ACN 121998694ACN-121998694-A

Abstract

The application relates to a power supply and demand prediction method and system based on multi-model cooperation. The method comprises the steps of obtaining multi-source data corresponding to a current prediction period and at least comprising structured data of an electric market and unstructured text data related to electric operation, processing the structured data to form basic time sequence features, calling a large language model customized in the electric field to analyze the unstructured text data and map the unstructured text data into structured event features, aligning and fusing the basic time sequence features and the structured event feature time sequence to obtain multi-mode fusion feature vectors, inputting the vectors into prediction models corresponding to at least two different preset scenes to obtain electric supply and demand predictor results, representing the current market scenes by the vectors, calculating scene similarity between the current market scenes and each preset scene in a historical scene feature library, giving normalization weights to each predictor result according to the scene similarity, carrying out weighted fusion, and outputting final electric supply and demand predictor results.

Inventors

DAI CHANGCHUN
LI YONGBO
Qian Hanhan
QI HUI
ZHANG WEISHI
JI CHAO
HAO YUXING
XIE DAOQING
FU JINGYU
ZHOU TAO
JIANG HAILONG
ZHAO XUETING
ZHANG WEI
Cheng Honggu
LIN ZHEMIN

Assignees

安徽电力交易中心有限公司

Dates

Publication Date: 20260508
Application Date: 20260129

Claims (10)

1. The utility model provides a power supply and demand prediction method based on multi-model cooperation, which is characterized by comprising the following steps: The method comprises the steps of obtaining multi-source data corresponding to a current prediction period, wherein the multi-source data at least comprises structured data of an electric power market and unstructured text data related to electric power operation; performing time stamp alignment, deletion and normalization on the structured data to form a basic time sequence feature; invoking a large language model customized in the electric power field to execute semantic analysis on the unstructured text data, identifying event types, influence objects and influence intensities related to power supply and demand, and mapping analysis results into structured event features; Carrying out time sequence alignment and fusion on the basic time sequence characteristics and the structured event characteristics to obtain a multi-mode fusion characteristic vector of the current prediction period; Inputting the multi-mode fusion feature vector into at least two prediction models corresponding to different preset scenes to obtain power supply and demand predictor results output by each prediction model, wherein the prediction models at least comprise a reference scene prediction model, an extreme weather scene prediction model, a holiday scene prediction model, a maintenance scene prediction model and a new energy high-duty ratio scene prediction model; Representing a current market scene by the multi-modal fusion feature vector, and calculating scene similarity of the current market scene and each preset scene in a historical scene feature library, wherein the preset scene comprises multi-modal fusion feature vectors of a plurality of historical prediction periods; Giving weight to the power supply and demand predictor results output by each prediction model according to the scene similarity of each preset scene in the current market scene and the historical scene feature library, wherein the weight is obtained by normalizing the scene similarity and the sum of the weights is 1; And carrying out weighted fusion on the power supply and demand forecasting sub-results to obtain a final power supply and demand forecasting result, wherein the final power supply and demand forecasting result at least comprises a power total demand forecasting value and a power total supply forecasting value in the current forecasting period.
2. The method according to claim 1, wherein the method further comprises: acquiring an actual observed total power demand value and an actual total power supply value after at least one prediction period is finished, and generating a corresponding actual supply and demand result; Comparing the final power supply and demand prediction result of each prediction period with the actual supply and demand result period by period according to a pre-constructed comprehensive evaluation index system for adapting the power supply and demand prediction, and calculating index values, wherein the comprehensive evaluation index system comprises a conventional time sequence prediction index and a special index, and the conventional time sequence prediction index comprises an average absolute error index and a root mean square error index; judging whether the predicted result meets the standard according to the index value and the dynamic error threshold value of each predicted period; And outputting a prediction period with exceeding error and a market scene corresponding to the prediction period.
3. The method according to claim 2, wherein the method further comprises: performing error tracing analysis on a prediction period with an exceeding error standard, wherein the error tracing analysis comprises the following steps: Verifying feature dimension integrity, timestamp consistency and updating timeliness of the multi-mode fusion feature vector, and positioning a data layer error source; disassembling a multi-model collaborative prediction process, evaluating the recognition accuracy of a market scene, the suitability and parameter stability of each prediction model, and positioning model layer error sources; acquiring an electric power market condition, carrying out external disturbance analysis according to the electric power market condition, and positioning an external disturbance error source; and generating a traceability conclusion of the error exceeding prediction period according to the data layer error source, the model layer error source and the external disturbance error source, wherein the traceability conclusion at least comprises an error source type and an optimization target object.
4. A method according to claim 3, characterized in that the method further comprises: Performing iterative updating based on the traceability conclusion, wherein the iterative updating comprises the following steps: updating a conversion rule from unstructured text to structured event features and a time-space alignment threshold value of the basic time sequence features and the structured event features when the error source type is a data layer error source; When the error source type is a model layer error source, updating the mapping relation between a preset scene and a prediction model, optimizing model parameters of the prediction model, and executing incremental training on the prediction model when an incremental training triggering condition is met, wherein the model parameters at least comprise one or more of tree model depth and learning rate; when the error source type is an external disturbance error source, updating a historical scene feature library for identifying market scenes so as to increase scene features corresponding to the external disturbance and improve the identification capability of novel market disturbance; After the updating is completed, the updated conversion rule, the space-time alignment threshold, the mapping relation, the model parameters and the historical scene feature library are called in the subsequent prediction period to execute the prediction, and the differentiated optimization instruction, the updating content and the updating effect are recorded to form an optimization log for long-term iteration.
5. The method of claim 1, wherein invoking the power domain customized large language model to perform semantic parsing on the unstructured text data comprises: The unstructured text data is subjected to event extraction to obtain an event element set related to power supply and demand, wherein the event element set at least comprises event types, influence objects and influence intensity, and the event types comprise extreme weather events, holiday events, overhaul events and new energy output abnormal events; Mapping the event element set into a structured event feature based on a preset feature mapping rule, wherein the structured event feature at least comprises event occurrence time, duration and influence intensity quantification value associated with an influence object; And when the unstructured text data has multi-source conflict, updating the structured event features based on the data source credibility or consistency checking structure.
6. The method of claim 1, wherein the time-aligning and fusing the base time-sequence feature with the structured event feature comprises: Mapping the structured event features to a time axis consistent with the basic time sequence features according to a preset time granularity, and executing time window slicing on the structured event features of which the event start-stop time spans multiple prediction periods; And performing one or more of feature stitching, gating fusion or attention fusion on the basic time sequence features and the aligned structured event features to obtain the multi-mode fusion feature vector.
7. The method according to claim 1, wherein the method further comprises: Calculating the scene similarity by adopting one or more of cosine similarity, euclidean distance similarity or Markov distance similarity; Based on scene similarity of the current market scene and each preset scene in the historical scene feature library, selecting the preset scene with K top ranking, and only giving weight to power supply and demand predictor results output by prediction models corresponding to the K preset scenes, wherein K is an integer greater than or equal to 2.
8. The method according to claim 1, wherein the method further comprises: calculating a supply-demand gap predicted value on the basis of the power total demand predicted value and the power total supply predicted value; And outputting the supply and demand gap predicted value as part of the final power supply and demand predicted result.
9. The method of claim 2, wherein the constructing of the dynamic error threshold comprises: Constructing a rolling evaluation window based on index values of a historical prediction period, and calculating statistics of the index values in the rolling evaluation window, wherein the statistics at least comprise one or more of mean, standard deviation and quantiles; Respectively determining a dynamic error threshold value for the conventional time sequence prediction index and the special index according to the statistic, and carrying out self-adaptive adjustment on the dynamic error threshold value when the working condition of the electric power market changes; And judging whether the index value of the subsequent prediction period meets the standard or not based on the adjusted dynamic error threshold value.
10. A multi-model collaboration-based power supply and demand prediction system, comprising: a processor and a memory; The processor is connected with the memory through a communication bus: The processor is used for calling and executing the program stored in the memory; The memory is used for storing a program at least for executing a power supply and demand prediction method based on multimodal synergy according to any one of claims 1 to 9.

Description

Power supply and demand prediction method and system based on multi-model collaboration Technical Field The application relates to the technical field of power supply and demand prediction, in particular to a power supply and demand prediction method and system based on multi-model cooperation. Background With the continuous advancement of the evolution of the electric power marketing, the operation of the electric power system is gradually changed from planning to market trading and real-time scheduling. Especially in business scenes such as spot market and bidding space, the electric power supply and demand relation not only determines the safe and stable operation of the electric power system, but also directly influences market clearing results, transaction strategy formulation and risk management and control level. Therefore, the construction of high-precision and strong-robustness power supply and demand prediction capability for short-period, strong-fluctuation and strong-disturbance-oriented power market environments has become an important requirement for power transaction institutions, power selling companies and dispatching operation departments. The existing power supply and demand prediction technology generally takes structured data of a power market as main input, such as a load curve, unit output, new energy output, maintenance plan, meteorological data, transaction data and the like, and adopts a statistical model or a machine learning/deep learning model to perform time sequence prediction. However, the mechanism of power supply and demand is complex, and is significantly affected by multi-factor coupling, and there is a lot of key disturbance information generated and propagated in unstructured form in market operation, such as policy notification, market announcement, overhaul announcement, weather early warning, emergency notification, public opinion information, etc. Because unstructured information is difficult to directly enter a traditional prediction model, the sensitivity of a prediction system to typical scenes such as extreme weather, holiday effect, overhaul overlap, new energy output mutation and the like is insufficient, prediction deviation is more prominent in peak time periods and extreme scenes, and the requirements of bidding space and the like on peak value, extreme deviation and period matching degree are difficult to meet. Disclosure of Invention The application provides a power supply and demand prediction method and system based on multi-model cooperation, which aims to solve the problem that a power supply and demand prediction model in the related technology is poor in generalization of multiple scenes and limited in special scene prediction capability due to insufficient integration capability of multi-source heterogeneous data at least to a certain extent. The scheme of the application is as follows: According to a first aspect of an embodiment of the present application, there is provided a power supply and demand prediction method based on multi-model collaboration, including: The method comprises the steps of obtaining multi-source data corresponding to a current prediction period, wherein the multi-source data at least comprises structured data of an electric power market and unstructured text data related to electric power operation; performing time stamp alignment, deletion and normalization on the structured data to form a basic time sequence feature; invoking a large language model customized in the electric power field to execute semantic analysis on the unstructured text data, identifying event types, influence objects and influence intensities related to power supply and demand, and mapping analysis results into structured event features; Carrying out time sequence alignment and fusion on the basic time sequence characteristics and the structured event characteristics to obtain a multi-mode fusion characteristic vector of the current prediction period; Inputting the multi-mode fusion feature vector into at least two prediction models corresponding to different preset scenes to obtain power supply and demand predictor results output by each prediction model, wherein the prediction models at least comprise a reference scene prediction model, an extreme weather scene prediction model, a holiday scene prediction model, a maintenance scene prediction model and a new energy high-duty ratio scene prediction model; Representing a current market scene by the multi-modal fusion feature vector, and calculating scene similarity of the current market scene and each preset scene in a historical scene feature library, wherein the preset scene comprises multi-modal fusion feature vectors of a plurality of historical prediction periods; Giving weight to the power supply and demand predictor results output by each prediction model according to the scene similarity of each preset scene in the current market scene and the historical scene feature library, wherein the wei