CN-122000884-A - Wind farm group power space-time prediction method based on dynamic graph convolution and frequency domain learning

CN122000884ACN 122000884 ACN122000884 ACN 122000884ACN-122000884-A

Abstract

The invention discloses a wind power plant group power space-time prediction method based on dynamic graph convolution and frequency domain learning, which is characterized by comprising the steps of S1, obtaining wind power plant group historical power data and meteorological data, constructing a multi-dimensional space-time embedding layer, carrying out joint coding on time, position and meteorological features to generate initial space-time features, S2, constructing a dynamic self-adaptive graph network, combining a learnable global adjacent matrix and local time sequence features to extract dynamic space features, S3, designing a frequency domain multi-layer perceptron, extracting power sequence frequency domain features through Fourier transformation, S4, fusing the dynamic space features and the frequency domain features, and generating future time power prediction values through an output layer. The method can effectively capture complex space-time coupling characteristics among fans, and remarkably improve wind power prediction precision and robustness.

Inventors

ZHANG GANG
YANG HONGYU
TIAN XIANGGUO
LIU HONGDA
ZHANG FENG

Assignees

烟台哈尔滨工程大学研究院
山东大学

Dates

Publication Date: 20260508
Application Date: 20260128

Claims (4)

1. A wind power plant group power space-time prediction method based on dynamic graph convolution and frequency domain learning is characterized by comprising the following steps of: s1, acquiring historical power data and meteorological data of a wind power plant group, constructing a multi-dimensional space-time embedded layer, and performing joint coding on a time period, node positions and meteorological features to generate an initial space-time feature vector containing multi-dimensional information; s2, constructing a dynamic self-adaptive graph Network (DYNAMIC ADAPTIVE GRAPH Network), capturing an inherent topological structure among fans by utilizing a learnable global adjacency matrix, generating a time-varying spatial dependency relationship by combining local time sequence characteristics, and extracting dynamic spatial characteristics of a wind power plant group; S3, designing a Frequency domain multi-layer perceptron (Frequency-domain MLP), projecting node characteristics to a Frequency domain through discrete Fourier transform, decoupling periodic components and trend components of a power sequence in a node dimension, and extracting Frequency domain evolution characteristics of the power sequence through inverse transformation reduction after complex operation processing; and S4, fusing the extracted dynamic space features with the frequency domain evolution features, and generating a power predicted value of the wind power plant group at a future moment through a predicted output layer.
2. The wind farm group power space-time prediction method based on dynamic graph convolution and frequency domain learning according to claim 1, wherein the space-time correlation analysis module is: (1) Time correlation Mutual information (Mutual Information, MI) is used to describe the degree of dependence between two random variables, the larger the value is, the stronger the association is, and when the mutual information is close to 0, the two are approximately independent. For time series, the time correlation can be quantified by calculating the mutual information between the current value of the series and its hysteresis value, as follows: in the formula, the joint probability distribution and the edge probability distribution are adopted. In time series, the time correlation is evaluated by analyzing the mutual information of the sequence and its lag sequence. (2) Lag time correlation Wherein the value of the time series at the current time point represents the joint probability distribution of sum of the values at the lag time point and the edge probability distribution of sum respectively. (3) Spatial correlation The spatial correlation is quantified, and the pearson correlation coefficient is used herein to measure the linear correlation between power sequences of different wind farms, as follows: Wherein: 、 respectively the power values of different fans or wind farms, 、 Is the average value thereof. 0.8 Shows strong correlation, 0.5< Less than or equal to 0.8 indicates a moderate correlation, Less than or equal to 0.5 indicates weak correlation.
3. The method for wind farm group power spatiotemporal prediction based on dynamic graph convolution and frequency domain learning of claim 1, wherein the dynamic adaptive graph network module of the spatiotemporal frequency network (STF-Net) is: (1) Graph convolution definition Extracting spatial structural features in a wind farm, herein in the form of a graph convolution based on first-order Chebyshev polynomial approximation The following is disclosed: Wherein: Representing a diagram structure matrix; Is that A degree matrix of (2); Is an input feature matrix; is an output feature matrix; Representing the weight; Representing the bias. Spatial feature extraction using a graph rolling network (GCN) for the first Layer input features Using the generated adaptive adjacency matrix And carrying out map propagation. (2) Adaptive dynamic graph network A) Global adaptive adjacency matrix An adaptive adjacency matrix is employed herein to learn global spatial dependencies from data. The specific method is realized by randomly initializing two parameters with learning ability. Wherein the adaptive adjacency matrix is: Wherein: Embedding for the source node; embedded for the target node. B) Local dynamic graph filtering A dynamic graph convolution method is designed for learning local graph information. At each time step, a local adjacency matrix is generated from the historical window data to reflect the local spatial dependence of the node time over time. And when in time step, learning short-term spatial dependence at the current moment according to the historical window data, and transforming the input signal by using the MLP to obtain a dynamically filtered signal. Wherein: Is a filtered dynamic signal. C) Dynamic graph information fusion And in the time sequence range, fusing the dynamic graph characteristics of each time step to generate dynamic characteristic representation of the node in the whole time window. Generating dynamic graph information by performing element-by-element multiplication operation on the sum: Wherein: Representation of Dynamic map information at the time step. D) Dynamic graph convolution update Wherein: a dynamic graph adjacency matrix for the current time step; A degree matrix; For a time of Input of time; Is a weight matrix which can be learned; To activate the function.
4. The method for wind farm group power space-time prediction based on dynamic graph convolution and frequency domain learning according to claim 1, wherein the frequency domain feature extraction module of the space-time frequency network (STF-Net) is: (1) Sequence frequency domain transformation To extract the periodicity and trend pattern of the time series, a fourier transform is used to project the original input into the frequency domain. The formula is as follows: Wherein: Is a frequency variable; Is an integral variable; is defined as the square root of-1 in imaginary units. The first half is Is abbreviated as the real part of (2) The latter half is Is abbreviated as Then in the equation Can be expressed as 。 (2) Complex feature decoupling learning The frequency domain MLP is redesigned for complex frequency components to efficiently capture the time series key pattern and energy compression through global views. Formally, for complex inputs Given a complex weight matrix And complex deviations The frequency domain MLP can be customized as: Wherein: is the final output; Represent the first The layer of the material is formed from a layer, Is an activation function. Due to And Are complex numbers, and according to complex multiplication rules, the further generalized equations are: Wherein: , . The MLP in the frequency domain can be implemented by calculating the real and imaginary parts of the frequency components, respectively, according to the above equation, and then superimposing them into one complex number to obtain the final result. (3) Frequency domain reconstruction After the frequency domain signal is processed, the frequency domain signal is restored to a time domain by utilizing inverse Fourier transform so as to keep the processing result of the period information and the global trend.

Description

Wind farm group power space-time prediction method based on dynamic graph convolution and frequency domain learning Technical Field The invention relates to the technical field of operation and control of an electric power system and artificial intelligence application, in particular to a method for predicting new energy generated power by using a deep learning technology. Background With the advancement of global energy conversion, wind power generation is an important component of clean energy, and its installed capacity continues to increase. However, wind power has extremely strong volatility, randomness and intermittence, and large-scale wind power grid connection brings great challenges to safe and stable operation and scheduling of a power system. The high-precision wind power prediction is a key technology for relieving peak shaving pressure and improving the power grid digestion capacity. The current wind power prediction method is mainly divided into a physical model method, a statistical method and an artificial intelligence method. Early studies mostly employed autoregressive integrated moving average (ARIMA), support Vector Machine (SVM), or long term memory network (LSTM) time series models. These methods, while capable of capturing the historical time dependence of a single site, typically ignore the linking effects between geographically adjacent wind farms. In practice, there is a significant spatial correlation between groups of wind farms, and wind speed variations in upstream wind farms will typically affect downstream wind farms after a certain time delay. To take advantage of spatial features, researchers have introduced Convolutional Neural Networks (CNNs) to model wind farm data into grid images for processing. Wind farms, however, often exhibit Non-Euclidean (Non-Euclidean) irregular distribution geographically, and CNNs have difficulty in efficiently handling such graph topologies. In recent years, graph roll-up networks (GCNs) have been widely used for their advantage of processing unstructured data. Existing GCN-based prediction methods generally construct a static adjacency matrix based on geographical distances between wind farms or Pearson correlation coefficients. However, in actual operation, the correlation between wind farms is not constant, but rather exhibits significant dynamic time-varying characteristics (e.g., the direction of wake effects changes with wind direction, resulting in reversal of upstream and downstream dependencies) as meteorological conditions such as wind direction, wind speed, etc. The fixed static diagram structure is difficult to capture the complex dynamic space-time coupling mechanism, and further improvement of prediction precision is limited. In addition, the existing deep learning prediction model is mostly modeled end-to-end only in the time domain, and directly outputs the power predicted value, but often ignores valuable information contained in the model predicted residual (Residuals). The prediction error usually contains high-frequency random components or systematic deviations which are not captured by the main model, and the characteristics are difficult to identify in the time domain, but often have a learnable rule in the frequency domain. The existing scheme lacks an effective mechanism for carrying out frequency domain depth excavation and correction on the prediction error, so that the robustness of the model is insufficient when the model faces to severe fluctuation of wind speed. In summary, how to construct a graph structure capable of adaptively capturing the dynamic spatial dependency relationship of the wind power plant group and make full use of the frequency domain information to refine and correct the prediction result is a technical problem to be solved in the current wind power prediction field. Aiming at the problems, the invention provides a wind farm group power space-time prediction method based on dynamic graph convolution and frequency domain learning. The method realizes the accurate depiction of the time-varying topological relation of the wind power plant group by constructing the dynamic graph structure comprising the node embedding and the gating mechanism, simultaneously creatively introduces the frequency domain learning module, maps the time domain residual error to the frequency domain to perform characteristic decoupling and correction, thereby effectively overcoming the defect of poor robustness of the traditional model under complex meteorological conditions and realizing the high-precision time-space collaborative prediction of the power of the wind power plant group. Disclosure of Invention Symbol and definition: A. index, set, and matrix: B. variable, constant C. abbreviations (abbreviations) The invention aims to provide a wind power plant group power space-time prediction method based on dynamic graph convolution and frequency domain learning, which can accurately capture space-time dynamic coupli