CN-121983965-A - Multi-mode fusion photovoltaic cluster power prediction method

CN121983965ACN 121983965 ACN121983965 ACN 121983965ACN-121983965-A

Abstract

The invention relates to the technical field of intelligent power grid dispatching and artificial intelligence intersection, and discloses a multi-mode fusion photovoltaic cluster power prediction method, which comprises the following steps: the method comprises the steps of collecting historical time sequence data, numerical weather forecast data and satellite cloud picture data corresponding to a forecast time period of a photovoltaic cluster, wherein the historical time sequence data is a time sequence feature sequence comprising a plurality of time steps, a plurality of photovoltaic sites and a plurality of feature dimensions. Complex environmental factors such as cloud layer movement, shielding change and the like can be more fully perceived, so that accuracy of photovoltaic cluster power prediction under complex weather conditions is improved.

Inventors

FAN HANG
WANG SHUAIKANG
ZHANG ZIEN
JIA HEPING
XU XIAOFENG
LIU DUNNAN
WANG HANFU
Run Wencai
TAN XIAOWEI

Assignees

华北电力大学（保定）

Dates

Publication Date: 20260505
Application Date: 20260208

Claims (6)

1. The multi-mode fusion photovoltaic cluster power prediction method is characterized by comprising the following steps of: Collecting historical time sequence data, numerical weather forecast data and satellite cloud image data corresponding to a prediction time period of a photovoltaic cluster, wherein the historical time sequence data is a time sequence characteristic sequence comprising a plurality of time steps, a plurality of photovoltaic sites and a plurality of characteristic dimensions; Performing Patch division on the time sequence feature sequence, dividing the time sequence feature sequence into a plurality of time sequence patches according to preset Patch lengths and step sizes, and performing embedded representation on each time sequence Patch; Constructing a retrieval enhanced time sequence backbone network based on the time sequence Patch, and performing feature modeling on the time sequence Patch, wherein the retrieval enhanced time sequence backbone network comprises a local memory path and a global memory path, the local memory path comprises a Patch memory bank, the similarity between the current time sequence Patch and a historical Patch in the Patch memory bank is calculated, and a plurality of historical Patches similar to the current time sequence Patch are retrieved; fusing the characteristics output by the local memory path and the characteristics output by the global memory path through a gating mechanism to obtain time sequence prediction characteristics, and generating a first prediction result based on the time sequence prediction characteristics; Extracting visual features of continuous multi-frame images in the satellite cloud image data by using a pre-trained CLIP model to form a visual feature sequence, and performing time sequence modeling on the visual feature sequence by using a transducer encoder containing position codes to obtain visual time sequence features; Generating text information describing the running state of the photovoltaic cluster based on the historical time sequence data and the numerical weather forecast data, carrying out semantic coding on the text information by using a text encoder of a CLIP model, modeling a numerical value characteristic by using a trainable time sequence branch, and injecting the time sequence numerical value information into the text semantic characteristic by using a cross-attention mechanism to obtain a text enhancement characteristic; Constructing a multi-modal prediction branch based on the visual timing feature and the text enhancement feature, and generating a second prediction result based on the multi-modal prediction branch; Constructing a spatial relationship diagram among photovoltaic stations, wherein the spatial relationship diagram comprises an adjacency relationship constructed based on the geographic distance of the photovoltaic stations and a K neighbor adjacency relationship constructed based on the feature similarity, and transmitting information among the photovoltaic stations by using a diagram attention network; And calculating a modal gating weight according to the input characteristics, carrying out weighted fusion on the first prediction result and the second prediction result according to the modal gating weight, and outputting a power prediction result of the photovoltaic cluster.
2. The method of claim 1, wherein the Patch division is performed with a length of 16 and a step size of 8 to generate 35 sequential patches for a sequence of sequential features with a length of 288.
3. The method for predicting the power of the multi-modal fusion photovoltaic cluster according to claim 1, wherein the Patch memory bank has a preset maximum capacity, the similarity between the current time sequence Patch and the historical patches in the Patch memory bank is calculated in a dot product mode, and a plurality of historical patches with the highest similarity are selected as search results.
4. The method for predicting the power of the multi-modal fusion photovoltaic cluster according to claim 1, wherein in the local memory path, the plurality of historical patches obtained by searching are processed by a multi-layer perceptron and then are fused with the characteristics of the current time sequence Patch in a residual connection mode.
5. The multi-modal fusion photovoltaic cluster power prediction method according to claim 1, wherein the aggregation of the visual timing features adopts a hybrid pooling manner, and the hybrid pooling manner comprises mean pooling of the visual feature sequence and weighted fusion of the last frame visual features.
6. The method for predicting the power of a multi-modal fusion photovoltaic cluster according to claim 1, wherein the modal gating weights are adaptively generated according to input features through a gating network, and the gating network comprises a linear layer, a normalization layer, an activation layer and a Sigmoid output layer.

Description

Multi-mode fusion photovoltaic cluster power prediction method Technical Field The invention relates to the technical field of intelligent power grid dispatching and artificial intelligence intersection, in particular to a multi-mode fusion photovoltaic cluster power prediction method. Background Along with the continuous expansion of the new energy grid-connected scale, the photovoltaic power generation is obviously influenced by weather conditions, has obvious intermittent and fluctuation characteristics, and brings great challenges to the real-time load balance of a power grid, the overload prevention and control of equipment and the safe and stable operation. In order to improve the dispatching and running reliability of the power system, accurate prediction of photovoltaic power generation power has become an important technical problem in new energy grid-connected operation. Existing photovoltaic power prediction techniques still have shortcomings in terms of environmental awareness. The traditional prediction method mainly relies on historical power data and numerical weather forecast data for modeling, and can reflect the change trend of the photovoltaic output to a certain extent, but is difficult to fully describe complex dynamic environment factors which have significant influence on the photovoltaic output, such as cloud cover movement, cloud cover change, shielding degree and the like. Because of the lack of effective utilization of visual-level environment information, under complex working conditions such as cloudiness, rapid shadow change and the like, the prediction result is often difficult to accurately capture the output fluctuation rule, so that the accuracy of the prediction result is limited. Meanwhile, when long time sequence data is processed, the problem that the existing prediction model forgets long sequence information is common. When predicting future power by using historical data of past days, the traditional recurrent neural network or the common transducer model has limited capability in capturing long-term dependency and periodic patterns, and early key information is easy to lose under the condition of long sequence length, so that the prediction performance is influenced. In addition, the existing multi-mode photovoltaic power prediction method is stiff in mode fusion mode. Although the partial method introduces multisource data, the multisource data is generally fused by adopting a simple characteristic splicing or weighting mode, an effective alignment mechanism for visual characteristics, text characteristics and numerical characteristics in semantic space is lacked, and complementary advantages among the modal information are difficult to fully develop. Meanwhile, the existing method generally adopts a fixed fusion strategy, and cannot adaptively adjust the influence weights of different mode information according to the change of weather conditions or running states, so that the adaptability and the robustness of the model in a complex environment are insufficient. Disclosure of Invention The invention aims to solve the defects in the prior art and provides a multi-mode fusion photovoltaic cluster power prediction method. In order to achieve the above purpose, the present invention adopts the following technical scheme: A multi-mode fusion photovoltaic cluster power prediction method comprises the following steps: Collecting historical time sequence data, numerical weather forecast data and satellite cloud image data corresponding to a prediction time period of a photovoltaic cluster, wherein the historical time sequence data is a time sequence characteristic sequence comprising a plurality of time steps, a plurality of photovoltaic sites and a plurality of characteristic dimensions; Performing Patch division on the time sequence feature sequence, dividing the time sequence feature sequence into a plurality of time sequence patches according to preset Patch lengths and step sizes, and performing embedded representation on each time sequence Patch; Constructing a retrieval enhanced time sequence backbone network based on the time sequence Patch, and performing feature modeling on the time sequence Patch, wherein the retrieval enhanced time sequence backbone network comprises a local memory path and a global memory path, the local memory path comprises a Patch memory bank, the similarity between the current time sequence Patch and a historical Patch in the Patch memory bank is calculated, and a plurality of historical Patches similar to the current time sequence Patch are retrieved; fusing the characteristics output by the local memory path and the characteristics output by the global memory path through a gating mechanism to obtain time sequence prediction characteristics, and generating a first prediction result based on the time sequence prediction characteristics; Extracting visual features of continuous multi-frame images in the satellite cloud image data by usin