CN-121981331-A - Short-term wind power prediction method and system based on dynamic graph neural network

CN121981331ACN 121981331 ACN121981331 ACN 121981331ACN-121981331-A

Abstract

The invention discloses a short-term wind power prediction method and a system based on a dynamic graph neural network, which belong to the field of wind power prediction, fill a wind power field missing power value through a space-time graph filling method, divide a power sequence into a high-frequency fluctuation set and a low-frequency stable set by utilizing VMD-CEEMDAN joint decomposition and SSA clustering optimization, dynamically construct a local subgraph to extract spatial characteristics by adopting a DGAT module to fuse geography, data and a leachable priori, simultaneously capture high-frequency mutation and low-frequency trend in parallel by utilizing a multi-scale convolution block through an MS-TCN module, generate multi-scale time characteristics, guide the fusion of the temporal characteristics through spatial and geographic joint characterization, realize collaborative modeling of multi-source heterogeneous data, finally input the fusion characteristics into a full-connection layer prediction fluctuation set prediction power, combine with a stable set prediction power and output a short-term wind power prediction value.

Inventors

TAN LING
MENG LINGJIE
XIA JINGMING

Assignees

南京信息工程大学

Dates

Publication Date: 20260505
Application Date: 20260131

Claims (7)

1. The short-term wind power prediction method based on the dynamic graph neural network is characterized by comprising the following steps of: s1, acquiring meteorological data, fan power and topography related data of a wind power plant, and constructing a wind power plant data set; s2, filling missing values in wind power data by using a space-time diagram filling method, carrying out data normalization processing, and dividing the wind power data into a training set and a test set; s3, performing VMD-CEEMDAN joint decomposition on the S2 stroke power data to obtain an IMF component, and automatically dividing the IMF component into a high-frequency fluctuation set and a low-frequency stationary set through an SSA-HSC mixed similarity clustering module; S4, inputting the fluctuation set in the step S3 into a DGAT spatial feature extraction module, and aggregating adjacent fan information through self-adaptive learning of topological relations among fans, so that spatial features of fans of a wind power plant are obtained; S5, inputting the fluctuation set in the step S3 into an MS-TCN time feature extraction module, and extracting time features of different scales from the fluctuation set data; S6, inputting the spatial characteristics of the wind power plant fans extracted in the step S4, the time characteristics of different scales generated in the step S5 and the topographic data in the step S1 into a multisource collaborative cross attention module, dynamically fusing the time characteristic representation through the spatial characteristics and the attention weight guided by the geographic information, and obtaining a fluctuation set prediction result through a full connection layer; And S7, inputting the stationary set in the step S3 to a LightGBM module to obtain a stationary set prediction result, and adding the stationary set prediction result and the fluctuation set prediction result obtained in the step S6 to obtain a final short-term wind power prediction result.
2. The method for predicting short-term wind power based on a dynamic graph neural network according to claim 1, wherein in step S2, the missing values in the wind power data are filled by a space-time graph filling method, specifically comprising the following steps: S21, a space-time diagram filling method comprises a first time convolution layer, a diagram convolution layer and a second time convolution layer; S22, regarding a wind farm data set Wherein the input of each time step Representing F-dimensional feature vectors of N fans at time t, constructing a space-time diagram structure G= (V, E, N), wherein N represents a fan node set, E represents an edge set, As an adjacency matrix, the elements in the adjacency matrix are dynamically calculated according to the relative positions of fans and the real-time wind directions, and the following formula is adopted: Wherein, the For the adjacency matrix between fans i and j at time t, the left part of the plus sign is the influence degree between fans i and j, d ij is the Euclidean distance between fan i and fan j, For the distance scaling parameter(s), For the direction angle from fan i to fan j at time t, For the dominant wind direction angle at time t, the right side of the plus sign represents the connection weight of the fan node itself, In order to be a self-connecting weight, Is an indication function; S23, inputting data Inputting a first time convolution layer of a space-time diagram convolution network, and extracting local time sequence characteristics of all fan nodes; S24, local time sequence characteristics are to be obtained Graph convolution layer of input space-time graph convolution network to obtain space characteristics among fans The spatial information of all fans is aggregated by using the A (t) calculated in the step S21, and the graph roll lamination adopts a first-order form of Chebyshev polynomial approximation, and the formula is as follows: Wherein, the Is that Is used for the degree matrix of the (c), To add the dynamic adjacency matrix of the self-connection, Is a matrix of units which is a matrix of units, And A learnable parameter that is a layer of a drawing volume; S25, will Inputting a second time convolution layer of the space-time diagram convolution network to further extract global time sequence characteristics Introducing residual connection to prevent gradient disappearance; S26, generating missing wind power data through an output layer, and adding a physical constraint item to ensure that a filling value accords with the characteristic of a fan power curve, wherein the following formula is as follows: Wherein, the For generating the missing wind power data, the left part of the plus sign is an effective power data processing part, And In order to output the layer parameters, The function is activated for Sigmoid, For a valid data mask, the right part of the plus sign is the missing power data processing part, In order to miss the data mask, And Theoretical minimum and maximum power values based on wind speed and power curves, respectively; And S27, carrying out normalization processing on the data filled with the missing values to obtain a complete wind power sequence, and dividing the data set into a training set and a testing set.
3. The method for predicting short-term wind power based on a dynamic graph neural network according to claim 1, wherein the step S3 is based on an SSA-HSC hybrid similarity clustering module to divide IMFs into a high-frequency fluctuation set and a low-frequency stationary set, and specifically comprises the following steps: s31, performing VMD decomposition on the power sequence in the step S2 to obtain an initial IMF set, and calculating the sample entropy of each IMF; S32, selecting an IMF component with the maximum sample entropy, performing secondary decomposition by CEEMDAN to obtain a plurality of new sub-IMF components, and replacing the IMF component with the maximum sample entropy by the components to form an enhanced IMF set; S33, constructing a mixed similarity matrix of fused time-frequency characteristic similarity and time sequence structure similarity based on the enhanced IMF set, wherein the mixed similarity matrix is represented by the following formula: Wherein, the For the four-dimensional time-frequency characteristic vector corresponding to the ith IMF, A dominant frequency energy ratio of the ith IMF, Sample entropy for the ith IMF, Is the center frequency of the ith IMF, For the effective bandwidth of the ith IMF, The similarity matrix of the fan i and the fan j is the time-frequency characteristic similarity matrix at the left part of the plus sign, The bandwidth parameters representing the RBF core, the right part of the plus sign is a time sequence structure similarity matrix, Representing the dynamic time warping distance between the ith and jth IMF sequences, E [0,1] is a learnable fusion weight coefficient; s34, introducing a sparrow search algorithm SSA to automatically optimize the weight and the spectral clustering contour coefficient, and accordingly automatically dividing all IMFs into two types, namely a high-frequency fluctuation set and a low-frequency stable set.
4. The method for predicting short-term wind power based on the dynamic graph neural network according to claim 1, wherein in the step S4, the spatial features between wind turbines of the wind farm are extracted by a DGAT spatial feature extraction module, and specifically comprises the following steps: s41, constructing three types of space embedding matrixes, namely a geographic embedding matrix, a data embedding matrix and a learnable embedding matrix, which respectively reflect the physical distance between fans and the space-time correlation and nonlinear interaction relation under data driving; S42, taking the fluctuation set obtained in the step S3 as input, and calculating physical distances among fans according to longitude and latitude coordinates of the fans to obtain a geographic embedding matrix; the method comprises the steps of flattening a fluctuation set of each fan, mapping the fluctuation set to a uniform low-dimensional feature space through a shared linear projection layer to obtain a data embedding matrix, initializing a group of learnable vectors corresponding to the number of the fans, and adaptively optimizing in a model training process to obtain the learnable embedding matrix; S43, taking the target fan as a central node, taking three corresponding embedded matrixes as a query vector Q, a key vector K and a value vector V respectively, and generating a comprehensive score matrix after fusion through an attention mechanism; s44, selecting K-1 adjacent fans with highest score according to the comprehensive score matrix, forming a local subgraph together with the target fans, aggregating the spatial feature information of neighbor nodes on the subgraph through a graph attention network, and outputting the spatial features among the fans, wherein the calculation mode is as follows: Wherein, the As a spatial feature between the target fans, Representing the stitched output of the H attention headers, The neighbor set for fan i in the local subgraph contains the fans themselves, The normalized weight for the kth attention header, For the k-th attention head's learnable weight matrix, Is an input feature of the fan j.
5. The method for predicting short-term wind power based on a dynamic graph neural network according to claim 1, wherein in step S5, time features of different scales in the fluctuation set data are extracted by an MS-TCN time feature extraction module, specifically comprising the following steps: s51, inputting historical power and meteorological data of a target fan in a fluctuation set to an MS-TCN module formed by stacking a plurality of multi-scale gating convolution blocks MSGCB, wherein each MSGCB comprises two parallel cavity causal convolution branches, and different expansion rates are adopted to capture high-frequency fluctuation and low-frequency trend respectively; S52, extracting time features of different scales through two parallel hole causal convolution branches in each MSGCB And And splicing the two along the channel dimension to obtain the primary fusion characteristic The calculation formula is as follows: Wherein W f and W g are respectively filter convolution kernel parameters and gating convolution kernel parameters, X is an input sequence, Representing element-by-element multiplication; S53, introducing a self-adaptive fusion mechanism between the two MSGCB, calculating channel weights through global average pooling and a multi-layer perceptron, and performing the method And (3) carrying out channel recalibration, enhancing a characteristic channel with larger contribution to prediction, adding the characteristic channel with an original input through residual connection to form the next MSGCB input, stacking layer by layer in the mode, and expanding receptive fields gradually, so that time characteristics of different scales are obtained, and the calculation formula is as follows: Where a is the channel weight, W 1 and W 2 are the learnable parameters of the multi-layer perceptron, Z is the current MSGCB output, and W skip is the jump connection projection matrix.
6. The method for predicting short-term wind power based on a dynamic graph neural network according to claim 1, wherein in step S6, collaborative modeling of space, time and geography is realized through a multi-source collaborative cross attention fusion module, specifically comprising the following steps: S61, carrying out dimension normalization processing on the spatial feature vectors between wind power plant fans output in the step S4 and the time feature vectors with different scales output in the step S5, and mapping the spatial feature vectors and the time feature vectors to a uniform spatial dimension d through an independent linear projection layer to obtain an aligned spatial query vector With time key value pairs The inconsistency of the space static characteristics and the time dynamic sequences in the dimension and the semantic scale is solved; S62, encoding the topographic data in the step S1 into geographic embedded vectors And injecting the data into the space query vector Q through a gating fusion mechanism to form a geographic enhancement type query, wherein the geographic enhancement type query is represented by the following formula: Wherein, the For geographic enhancement queries, W g ,W f is a learnable parameter; S63 based on And a time key value pair K, V, calculating the cross-modal attention weight and aggregating key time step information to generate a time context vector, splicing the time context vector with a space feature vector between wind power plant fans, and then sending the spliced time context vector into a full-connection layer to output a short-term wind power predicted value.
7. A short-term wind power prediction system based on a dynamic graph neural network, which is applicable to the short-term wind power prediction method based on the dynamic graph neural network as claimed in any one of claims 1 to 6, and is characterized in that the prediction system comprises: the preprocessing module is used for acquiring meteorological data, wind power data and topography related data and filling the missing wind power data through a space-time diagram filling method; The SSA-HSC mixed similarity clustering module is composed of a VMD-CEEMDAN decomposition algorithm and a mixed similarity clustering unit based on SSA optimization, and is used for constructing a mixed matrix with similar time-frequency and time sequence structures and dividing an IMF into a high-frequency fluctuation set and a low-frequency stationary set in a self-adaptive manner; the DGAT spatial feature extraction module is composed of three types of spatial embedding matrixes and dynamic GAT units and is used for fusing geography, data and learnable priori information, dynamically constructing local subgraphs, aggregating information between adjacent fans and outputting spatial features between the fans; the MS-TCN time feature extraction module is used for capturing dependency relations of different time scales in parallel through stacked MSGCB units, and generating time features of different scales by combining self-adaptive channel recalibration and residual connection; The multisource collaborative cross attention module is composed of asymmetric cross attention units, and the collaborative modeling of fan space association, power time sequence change and geographic information is realized by dynamically focusing information of key historical moments through an asymmetric cross attention mechanism.

Description

Short-term wind power prediction method and system based on dynamic graph neural network Technical Field The invention belongs to the field of wind power prediction, and particularly relates to a short-term wind power prediction method and system based on a dynamic graph neural network. Background Wind energy plays a key role in global energy conversion as a renewable energy source. However, wind power output has strong randomness, intermittence and volatility, and particularly in a short period of an hour scale, wind power often presents high-frequency abrupt change and local abrupt change characteristics, which brings serious challenges to power grid dispatching. Therefore, the development of a short-term wind power prediction technology with high precision and high robustness has important significance in improving new energy consumption capability and guaranteeing safe and stable operation of a power grid. The current short-term wind power prediction method mainly comprises a physical method, a statistical method and a deep learning method. The physical method predicts by simulating the response relation between the atmospheric motion and the fan based on the numerical weather forecast, but the physical method depends on high-resolution meteorological data and has high calculation cost, so that the real-time requirement of ultra-short-term prediction is difficult to meet. The statistical method such as autoregressive integral moving average, support vector machine and the like models through mining statistical rules between historical power and meteorological variables, and although the calculation efficiency is high, the complex nonlinear space-time coupling relationship inside the wind field is difficult to effectively describe, and particularly, the prediction performance is remarkably reduced when the wind field is used for coping with sudden wind speed changes. In recent years, deep learning methods have shown great potential in short-term wind power prediction. Convolutional and recurrent neural networks are used to extract local patterns in time series and long-term dependencies, and graph neural networks attempt to model spatial correlations between fans. However, the existing method still has obvious limitations that most models adopt static or predefined fan adjacency relations, dynamic interaction driven by geographical priors and data cannot be fused, so that insufficient space heterogeneity response to wake effects, terrain shielding and the like is caused, time modeling generally depends on a single receptive field structure, high-frequency mutation and low-frequency periodic components are difficult to capture simultaneously, sensitivity to severe fluctuation is insufficient, in addition, multisource heterogeneous data such as weather, power and terrain are obvious in difference in spatial resolution and semantic level, the existing fusion strategy mostly adopts shallow splicing or symmetrical attention mechanisms, directional interaction of physical guidance is lacked, redundant noise is easy to introduce, and effective modeling of space-time and geographical features is difficult to realize. Therefore, a short-term wind power prediction method with dynamic collaborative modeling capability is needed to break through the bottleneck of the prior art. Disclosure of Invention The invention aims to overcome the defects of the prior art and provide a short-term wind power prediction method based on a dynamic graph neural network. The invention provides a short-term wind power prediction method based on a dynamic graph neural network, which comprises the following steps: S1, acquiring meteorological data, fan power, terrain and other related data of a wind power plant, and constructing a wind power plant data set; s2, filling missing values in wind power data by using a space-time diagram filling method, carrying out data normalization processing, and dividing the wind power data into a training set and a test set; s3, performing VMD-CEEMDAN joint decomposition on the S2 stroke power data to obtain an IMF component, and automatically dividing the IMF component into a high-frequency fluctuation set and a low-frequency stationary set through an SSA-HSC mixed similarity clustering module; S4, inputting the fluctuation set in the step S3 into a DGAT spatial feature extraction module, and aggregating adjacent fan information through self-adaptive learning of topological relations among fans, so that spatial features of fans of a wind power plant are obtained; S5, inputting the fluctuation set in the step S3 into an MS-TCN time feature extraction module, and extracting time features of different scales from the fluctuation set data; S6, inputting the spatial characteristics of the wind power plant fans extracted in the step S4, the time characteristics of different scales generated in the step S5 and the topographic data in the step S1 into a multisource collaborative cross attention module, dynamic