Wind speed inversion method and system based on active power of wind power plant

Abstract

The invention discloses a wind speed inversion method and system based on active power of a wind power plant, and belongs to the technical field of crossing of power systems and weather prediction.The method comprises the steps that obtained time series data of multiple stations of the wind power plant are input into a pre-trained space-time coupling wind speed inversion model for prediction, and a wind speed prediction result in the future step is output; the model comprises a Transform encoder and a full connection layer, a space-time dynamic fusion module is introduced between the Transform encoder and the full connection layer, and the module comprises a dynamic graph convolution layer and a feature aggregation layer; the encoder is used for extracting global context features of each time sequence data of each site; the dynamic graph convolution layer is used for determining a static space matrix and a dynamic time matrix; the feature aggregation layer is used for fusing the static space matrix and the dynamic time matrix to obtain a space-time adjacency matrix and taking the space-time adjacency matrix as a weight, and performing space-time fusion on global context features; and the full connection layer maps the space-time fusion feature and outputs a wind speed prediction result in the future step. The method can improve the precision of wind speed prediction.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of power systems and meteorological forecasting, and more specifically to a method and system for wind speed inversion based on the active power of a wind farm. Background Technology

[0002] In modern power systems, the volatility and intermittency of wind energy resources lead to significant uncertainties in power supply, seriously threatening the safe and stable operation of the power system. Therefore, achieving high-precision power weather forecasting has become a key link in ensuring the reliability of the power system. Constructing a correlation inversion model between the active power of a wind farm and the wind speed at its center point, based on historical power and wind speed data of wind farms, is an effective technical approach to improve the accuracy of power weather forecasting.

[0003] Existing inversion models often focus on static mapping of single sites, making it difficult to balance temporal continuity and physical rationality. This leads to distortion or loss of information in the data obtained from studies that infer meteorological data from active power. Summary of the Invention

[0004] To address the problems existing in the above-mentioned fields, this invention proposes a wind speed inversion method and system based on the active power of wind farms. The constructed spatiotemporal coupled wind speed inversion model can fuse the static spatial matrix and the dynamic time matrix to obtain a spatiotemporal adjacency matrix, which is used as a weight. The global context features of each time series data are spatiotemporally fused. Through multi-dimensional feature decoupling and cross-scale mapping mechanism, the problem of feature distortion and information loss caused by inconsistent sampling frequencies can be effectively solved.

[0005] To address the aforementioned technical problems, this invention discloses a wind speed inversion method based on the active power of a wind farm, comprising the following steps: Acquire historical wind speed time-series data, historical wind direction time-series data, and historical wind power time-series data from multiple sites in the wind farm; Historical wind speed time series data, historical wind direction time series data, and historical wind power time series data from multiple sites of the wind farm are input into a pre-trained spatiotemporal coupled wind speed inversion model for prediction, and the future step wind speed prediction results are output. The spatiotemporal coupled wind speed inversion model includes a Transformer encoder and a fully connected layer. A spatiotemporal dynamic fusion module is introduced between the Transformer encoder and the fully connected layer. The spatiotemporal dynamic fusion module includes a dynamic graph convolutional layer and a feature aggregation layer. The Transformer encoder is used to extract global context features of each time series data of each station; the dynamic graph convolutional layer obtains a static spatial matrix by determining the static spatial correlation strength between stations, and obtains a dynamic temporal matrix by determining the vector dot product of the time dimension between stations; the feature aggregation layer is used to fuse the static spatial matrix and the dynamic temporal matrix to obtain a spatiotemporal adjacency matrix, which is used as a weight to perform spatiotemporal fusion on the global context features and output spatiotemporal fusion features; the fully connected layer maps the spatiotemporal fusion features and outputs the future step wind speed prediction result.

[0006] Preferably, the spatiotemporal coupled wind speed inversion model further includes an embedding layer, a location encoding layer, and a CNN feature extraction layer: The embedding layer is used to map the acquired time-series data to the high-dimensional hidden dimension of the Transformer encoder to extract the time-series features of wind speed and wind direction. The location encoding layer is used to encode the temporal features of the extracted wind speed and wind direction, label the location information for each time step, and generate a spatiotemporal joint feature representation of each temporal data that integrates time, location and physical context. The spatiotemporal joint feature representation is input into the 1D convolutional layer of the CNN feature extraction layer. By capturing the local temporal patterns of wind speed and wind direction, the local spatiotemporal feature sequence of each temporal data is output, which is aligned with the hidden dimension of the Transformer encoder. The local spatiotemporal feature sequence is used as the input to the Transformer encoder.

[0007] Preferably, the embedding layer is used to map the acquired time-series data to the high-dimensional hidden dimension of the Transformer encoder to extract the time-series features of wind speed and wind direction, and further includes: Wind direction features were extracted from the time-series data of multiple sites in the wind farm. The wind direction characteristics are processed using a cyclic feature, converting the wind direction angle into sine and cosine values. The embedding layer maps the acquired time-series data, sine values, and cosine values ​​to the high-dimensional hidden dimension of the Transformer encoder to extract the time-series features of wind speed and wind direction.

[0008] Preferably, the step of acquiring historical wind speed time-series data, historical wind direction time-series data, and historical wind power time-series data from multiple sites of the wind farm further includes: The acquired historical wind speed time series data, historical wind direction time series data, and historical wind power time series data are time-stamped and synchronized. Based on the statistical distribution characteristics of wind turbine power curves, parameter estimation is performed on calibrated historical wind speed time series data, historical wind direction time series data, and historical wind power time series data. By constructing dynamic confidence boundaries, the preset confidence boundary range is used as the core interval of effective data, and outliers exceeding this range are screened out. After removing outliers, historical wind speed time series data, historical wind direction time series data, and historical wind power time series data from multiple sites are obtained, and interpolation processing is performed to obtain continuous historical wind speed time series data, historical wind direction time series data, and historical wind power time series data.

[0009] Preferably, the step of acquiring historical wind speed time-series data, historical wind direction time-series data, and historical wind power time-series data from multiple sites after removing outliers, and performing interpolation processing to obtain continuous historical wind speed time-series data, historical wind direction time-series data, and historical wind power time-series data, specifically includes: The local outlier factor algorithm is used to identify outliers in historical wind speed time series data, historical wind direction time series data, and historical wind power time series data. The weighted distance outlier algorithm was used to clean up outliers, and the historical wind speed time series data, historical wind direction time series data and historical wind power time series data of multiple sites were obtained after removing outliers. Based on the Akaike Information Criterion (AIC), the AIC value of the combination of historical wind speed time series data and historical wind power time series data of multiple sites after removing outliers is determined. The combination of historical wind speed time series data and historical wind power time series data of multiple sites after removing outliers with the smallest AIC value is selected as the global optimal order of the autoregressive moving average (ARMA) data interpolation algorithm. By using the ARMA model corresponding to the globally optimal order, interpolation results with the same style as the original historical wind speed time series data, historical wind direction time series data, and historical wind power time series data are generated, that is, continuous historical wind speed time series data, historical wind direction time series data, and historical wind power time series data.

[0010] Preferably, the static spatial matrix obtained by determining the static spatial association strength between sites is generated by encoding site embedding representations through a Transformer encoder, calculating the static correlation between sites, and determining the static spatial association strength between sites.

[0011] Preferably, it further includes a wind speed inversion system based on the active power of a wind farm, characterized in that it comprises: The data acquisition module is used to acquire historical wind speed time-series data, historical wind direction time-series data, and historical wind power time-series data from multiple sites in the wind farm. The wind speed inversion module is used to input historical wind speed time series data, historical wind direction time series data, and historical wind power time series data from multiple sites of the wind farm into a pre-trained spatiotemporal coupled wind speed inversion model for prediction, and output the future step wind speed prediction results. A spatiotemporal coupled wind speed inversion model construction module is included. The spatiotemporal coupled wind speed inversion model comprises a Transformer encoder and a fully connected layer. A spatiotemporal dynamic fusion module is introduced between the Transformer encoder and the fully connected layer. This module includes a dynamic graph convolutional layer and a feature aggregation layer. The Transformer encoder extracts global context features from each time series data of each station. The dynamic graph convolutional layer obtains a static spatial matrix by determining the static spatial correlation strength between stations and a dynamic temporal matrix by determining the vector dot product of the time dimensions between stations. The feature aggregation layer fuses the static spatial matrix and the dynamic temporal matrix to obtain a spatiotemporal adjacency matrix, which is then used as weights to perform spatiotemporal fusion on the global context features, outputting spatiotemporal fused features. The fully connected layer maps the spatiotemporal fused features and outputs the future step wind speed prediction result.

[0012] Compared with the prior art, the present invention has the following beneficial effects: This invention proposes a wind speed inversion method based on the active power of wind farms. The constructed spatiotemporally coupled wind speed inversion model, through the self-attention mechanism of the Transformer encoder, captures long-distance temporal dependencies and implicit cross-site correlations in historical data of each site. Compared to traditional RNNs or CNNs, the Transformer avoids the gradient vanishing / exploding problem and is more suitable for handling nonlinear fluctuations in long-term wind farm data. By obtaining the static spatial matrix and dynamic temporal matrix between sites through dynamic graph convolutional layers, the spatiotemporal evolution of wind speed propagation can be simulated more realistically. The spatiotemporal adjacency matrix is ​​obtained and used as weights to spatiotemporally fuse the global contextual features of each time series data of each site extracted by the Transformer encoder, outputting a spatiotemporally fused feature. This spatiotemporal fused feature shares multi-site information, improving the robustness of local site predictions. This spatiotemporal fused feature can dynamically adjust the spatiotemporal focus according to real-time data, avoiding the limitations of a fixed graph structure. A fully connected layer performs the final mapping of the spatiotemporal fused feature, fitting the complex relationship between wind speed and power through multi-layer nonlinear transformations. The constructed spatiotemporal coupled wind speed inversion model fuses static and dynamic features between sites. Through multi-dimensional feature decoupling and cross-scale mapping mechanisms, it effectively solves the problems of feature distortion and information loss caused by inconsistent sampling frequencies. For single-site scenarios, the spatiotemporal coupled wind speed inversion model constructed in this invention can extend the single-site time series model to multi-site scenarios. The established multi-site spatiotemporal coupled wind speed prediction model has higher prediction accuracy than the single-site time series model. This method achieves high-precision, dynamic, and multimodal collaborative prediction of wind speed at multiple wind farm sites through the synergistic effect of Transformer global feature extraction, dynamic graph convolution spatiotemporal modeling, and feature aggregation and fusion. Attached Figure Description

[0013] Figure 1 This is a flowchart of the wind speed inversion method based on the active power of a wind farm proposed in this invention. Figure 2 A power probability distribution curve of multiple identical wind turbines at a site level provided for embodiments of the present invention; Figure 3 The network architecture of the traditional wind speed regression model provided in this embodiment of the invention; Figure 4 The network architecture of the traditional wind direction branching model provided in the embodiments of the present invention; Figure 5 The network architecture of the spatiotemporal coupled wind speed prediction model provided in this embodiment of the invention; Figure 6 A schematic diagram of the topology of a dynamic graph convolutional layer and its mathematical adjacency matrix provided in an embodiment of the present invention; Figure 7The diagram shows sample data and sliding window for different timescale transformation methods provided in the embodiments of the present invention. Detailed Implementation

[0014] The following will refer to the appendices in the embodiments of the present invention. Figures 1-7 The technical solutions in the embodiments of the present invention will be clearly and completely described. It should be understood that the terminology used in the present invention is only for describing particular implementation methods and is not intended to limit the present invention.

[0015] Example like Figure 1 The figure shows a wind speed inversion method based on the active power of a wind farm proposed in this invention, which includes the following steps: S1: Obtain historical wind speed time-series data, historical wind direction time-series data, and historical wind power time-series data from multiple sites in the wind farm; S2: Input the historical wind speed time series data, historical wind direction time series data, and historical wind power time series data from multiple sites of the wind farm into the pre-trained spatiotemporal coupled wind speed inversion model for prediction, and output the future step wind speed prediction results. The spatiotemporal coupled wind speed inversion model includes a Transformer encoder and a fully connected layer. A spatiotemporal dynamic fusion module is introduced between the Transformer encoder and the fully connected layer. The spatiotemporal dynamic fusion module includes a dynamic graph convolutional layer and a feature aggregation layer. The Transformer encoder extracts global contextual features from each time series data of each station. The dynamic graph convolutional layer obtains a static spatial matrix by determining the static spatial correlation strength between stations and a dynamic temporal matrix by determining the vector dot product of the time dimension between stations. The feature aggregation layer fuses the static spatial matrix and the dynamic temporal matrix to obtain a spatiotemporal adjacency matrix, which is used as a weight to perform spatiotemporal fusion of global contextual features and outputs spatiotemporal fused features. The fully connected layer maps the spatiotemporal fused features and outputs the future step wind speed prediction result.

[0016] In step S1, by analyzing the inherent power curve characteristics of a single wind turbine, the research scope is then expanded to multiple wind turbines of the same model at the site level to systematically explore the aggregate distribution characteristics of their wind power and wind speed data.

[0017] By studying the probabilistic wind speed-power distribution characteristics in wind farms and the relationship between wind speed and power in actual wind power station datasets, this study investigates data cleaning and interpolation algorithms to process the original datasets and establish high-quality datasets.

[0018] By analyzing the wind speed-power curves of a single wind turbine, and considering the randomness and fluctuation of wind speed, spatial differences of wind turbines, and wake effect, we obtain a probability distribution-type wind speed-power curve for the overall input wind speed and output wind power of several identical wind turbines at the site.

[0019] like Figure 2 The figure shown is a power curve diagram of the probability distribution of multiple wind turbines of the same model at a site level provided by an embodiment of the present invention. By analyzing the probability distribution characteristics of the wind speed data and power data distribution of multiple wind turbines of the same model at a site level, the specific details are as follows: Wind speed-wind power data points exhibit a statistical distribution characteristic of "core clustering and peripheral diffusion" centered on the Wind Turbine Power Curve (WSPC). The probability density of data points decays exponentially with respect to the WSPC axis. Typically, a dynamic confidence boundary is constructed based on Weibull distribution parameter estimation, using the WSPC ±20% range as the core interval for valid data (represented as the upper and lower bounds of the quantiles in the figure). Outliers exceeding this range require physical verification in conjunction with equipment operating status and environmental parameters to ensure the reliability of data preprocessing.

[0020] The acquired historical wind speed time series data, historical wind direction time series data, and historical wind power time series data are time-stamped and synchronized. Based on the statistical distribution characteristics of wind turbine power curves, parameter estimation is performed on calibrated historical wind speed time series data, historical wind direction time series data, and historical wind power time series data. By constructing dynamic confidence boundaries, the preset confidence boundary range is used as the core interval of effective data, and outliers exceeding this range are screened out. After removing outliers, historical wind speed time series data, historical wind direction time series data, and historical wind power time series data from multiple sites are obtained, and interpolation processing is performed to obtain continuous historical wind speed time series data, historical wind direction time series data, and historical wind power time series data. Continuous historical wind speed time series data, historical wind direction time series data, and historical wind power time series data are used as inputs to the spatiotemporal coupled wind speed inversion model.

[0021] By analyzing the differences between the theoretical probability distribution wind speed-power curve and the actual wind speed-power scatter plot, the characteristics of outliers in the actual wind speed-power scatter plot are analyzed. Based on the characteristics of outliers, a two-step cleaning method combining the Local Outlier Factor (LOF) algorithm and the Weighted Distance Outlier Factor (WDOF) algorithm is adopted as the data cleaning algorithm.

[0022] The data cleaning algorithm is based on a two-step method using the Local Outlier Factor (LOF) algorithm. The first step is to identify outliers in the historical wind speed and historical wind power data using the LOF algorithm. The second step is to obtain the historical wind speed and historical wind power data of the sites after removing outliers using the Weighted Distance Outlier Factor (WDOF) algorithm.

[0023] 1) Core concepts of the local outlier factor algorithm: (1) k-distance and k-distance neighborhood For sample points Its k-distance Defined as the minimum distance that satisfies the following conditions : (1) In the formula: Given a dataset, p and q are different two-dimensional data points, with wind speed on the horizontal axis and power on the vertical axis. k-distance neighborhood Defined as all that satisfy The set of sample points contains at least Sample points (including) itself).

[0024] (2) Reachable distance Sample points arrive reachable distance Defined as: (2) This definition is achieved through fusion. k-distance and The actual distance between points avoids distance measurement errors caused by local density differences between neighboring points.

[0025] (3) Locally accessible density Sample points The local reachability density is defined as the distance from all points in its k-distance neighborhood to... The reciprocal of the average reachable distance: (3) in, The larger the value, the stronger the... The denser the distribution of surrounding sample points, the sparser the distribution.

[0026] (4) Local outlier LOF values ​​are compared The degree of outlier is quantified by the local reachability density of its neighboring points: (4) when hour, The density is consistent with that of the neighborhood, which is a normal point; when hour, The density is lower than that of the neighborhood. The larger the value, the higher the degree of outlier.

[0027] 2) Core concepts of the weighted distance outlier factor algorithm: Unlike the local outlier algorithm, the weighted distance outlier algorithm is specifically optimized by weighting the Euclidean distance metric calculation in the local outlier algorithm, while the rest is the same as the local outlier algorithm.

[0028] The weighted formula is as follows: (5) intermediate wind speed range ( ): (6) Low wind speed section and high wind speed section ): (7) In the formula: Weights; variables These are the cut-in wind speed, cut-out wind speed, and equivalent power output value, respectively.

[0029] Based on the above core concepts, the specific steps for data cleaning are as follows: The first step is to use the local outlier factor algorithm to clean up the large number of clustered outliers and the small number of discrete outliers in the historical wind speed and historical wind power data: (1) Neighborhood construction: For each sample point Calculate its and ; (2) Reachability distance calculation: Construct the reachability distance matrix of the dataset. ,in (8) (3) Local reachability density estimation: for each Based on neighborhood calculate ; (4) Outlier quantification: For each Through neighboring points With oneself The mean of the ratios is obtained. ; (5) Anomaly filtering: Preset The parameter (specifies the proportion of outliers in the data) is used to calculate the proportion of all samples. The corresponding percentile of the value is used as the threshold (e.g.) (Take the 95th percentile) Sample points exceeding this threshold are identified as outliers.

[0030] The second step involves cleaning up the remaining discrete outliers using a weighted distance outlier algorithm. (1) Calculate the quantiles of effective power (i.e., wind power time series data) within an independent wind farm: ; (2) Statistical calculations: low power range The minimum wind speed in this interval is taken as the candidate cut-in wind speed and corresponding wind power for the day. (9) High power range Take the minimum wind speed in this interval as the candidate cut-out wind speed and corresponding wind power for the day: (10) Final cut-in wind speed: for satisfying : (11) Final wind speed cut-off: of Calculate the mean: (12) (3) Solving piecewise curves Based on the operating characteristics of wind turbine generators, the piecewise function of the equivalent power curve is as follows: (13) Among them, the coefficients of the cubic polynomial Solving using the square method, the constraints are: (14) (4) Substitute the parameters obtained in (2) and (3) into the weight formula and weighting formula in the core concept of the weighted distance outlier factor algorithm to obtain the reachable distance between every two sample points, and substitute them into formula (8) to calculate the reachable distance matrix. .

[0031] (5) Based on the calculated reachability matrix The system sequentially performs three steps: (3) local reachability density estimation, (4) outlier factor quantification, and (5) outlier screening, ultimately obtaining historical wind speed data and historical wind power data of the stations with outliers removed.

[0032] The effect diagram of the two-step cleaning method is compared and evaluated with the power curve of the wind turbine.

[0033] Since data cleaning disrupts the temporal sequence of the original dataset, data interpolation is needed to restore the temporal sequence of the cleaned data. Modeling based on the traditional Autoregressive Moving Average (ARMA) data interpolation algorithm requires manually setting the order, which can easily lead to overfitting (too high an order) or underfitting (too low an order).

[0034] Therefore, this invention employs intelligent order selection based on the Akaike Information Criterion (AIC) to perform data interpolation on historical wind speed data and historical wind power data, obtaining interpolated historical wind speed time-series data, historical wind direction time-series data, and historical wind power time-series data, specifically including: Based on the historical wind speed time series data, historical wind direction time series data, and historical wind power time series data of the sites where outliers have been removed, a combination of historical wind speed time series data and historical wind power time series data is determined. The AIC value of the combination of historical wind speed data and historical wind power data is determined based on the Akaike Information Criterion (AIC). 1) Traverse candidate orders: Set the maximum order , (Covering common timing patterns); 2) Model Fitting and Evaluation: For Combine the data, fit the data using the ARMA model, and calculate the AIC value. The smaller the AIC value, the better the fit of the ARMA model. 3) Stationarity filtering: Only models with stationary AR components (roots outside the unit circle) are retained; 4) Determine the optimal order: Select the combination of historical wind speed data and historical wind power data with the lowest AIC value. As the globally optimal order of the autoregressive moving average (ARMA) data interpolation algorithm, it balances the complexity of the ARMA model with the fitting accuracy.

[0035] Bidirectional interpolation is performed for missing points, including forward prediction and backward prediction; Forward prediction: using known data to the left of the missing point Input the ARMA model corresponding to the globally optimal order to obtain the prediction data for missing points. ; Backward prediction: using known data to the right of the missing points The sequence is inverted and treated as a "forward" sequence. It is then input into the ARMA model corresponding to the globally optimal order, predicted, and then inverted back to the original time series to obtain the prediction data for the missing points. ; The final data interpolation result is a weighted average of the forward and backward predictions: (15) in, and These represent the weights for forward and backward prediction, respectively, and the weight values ​​are exponentially negatively correlated with the AIC value. If both bidirectional predictions fail, the algorithm reverts to linear interpolation to ensure the robustness of the autoregressive moving average (ARMA) data interpolation algorithm.

[0036] The interpolated historical wind speed time series data and historical wind power time series data are verified to conform to the original time series by using autocorrelation function graphs and cross-correlation function graphs; and the interpolated historical wind speed time series data and historical wind power time series data are verified to conform to the original distribution pattern by using data distribution comparison graphs and quantile comparison graphs.

[0037] The time-series data of aligned historical wind direction, historical wind speed, and historical wind power were synchronized and calibrated (including data cleaning and data interpolation) to facilitate subsequent model research. Wind rose diagrams and wind autocorrelation function diagrams were used to verify that the distribution and time-series characteristics of the wind direction data conformed to the original pattern.

[0038] This invention presents a data processing method based on wind turbine power curves, comprising two parts: data cleaning and data interpolation. The data cleaning part studies and evaluates a two-step cleaning method combining Local Outlier Factor (LOF) with a weighted improved LOF algorithm (WDOF). The data interpolation part studies and evaluates a bidirectional interpolation method using an autoregressive moving average (ARMA) model to restore the continuity of the original time series. Finally, the method for processing wind direction data is investigated, comprehensively improving the data preprocessing of the three-dimensional dataset and providing theoretical support for subsequent steps.

[0039] This invention acquires historical wind speed time-series data, historical wind direction time-series data, and historical wind power time-series data from multiple sites in a wind farm, and can integrate complementary information from multiple data sources (wind speed, wind direction, and power).

[0040] This invention also includes the study of power system regression models and meteorological prediction models, comparing static regression models and dynamic time series models, and combining the requirements and characteristics of wind speed inversion models to determine the Transformer neural network architecture to establish the inversion model.

[0041] Traditional methods utilize the Transformer neural network architecture to build time-series data on wind speed, wind direction, and wind power for a single wind farm. The wind speed regression model and wind direction branch model are as follows: Figure 3 and Figure 4 As shown, flowcharts for the wind speed regression model and the wind direction branch model for a single site are presented respectively, intuitively demonstrating the specific architecture of the two models.

[0042] Wind speed regression model exist Figure 3 In this study, the wind speed regression model employs a Transformer encoder architecture to capture the nonlinear temporal dependency between wind speed and wind power. The input to the wind speed regression model is historical data. The multivariate sequence of steps (historical wind speed, historical wind power, and future power fill value) outputs the future... Wind speed regression model. The core components of the wind speed regression model include a temporal embedding layer, a location encoding layer, a Transformer encoder, and a fully connected layer. The wind speed regression model is optimized using a loss function.

[0043] The core structure is described in detail below: The historical wind speed and historical wind power of the acquired single wind power station are preprocessed. After decomposing the wind speed features, the wind speed features are processed to convert the wind speed angle into sine and cosine values.

[0044] (1) The acquired time series data, sine and cosine values ​​are mapped to the hidden dimensions of the Transformer through the linear layer of the time embedding layer. The formula is: (16) By extracting the temporal features of wind speed, the transformation from low-dimensional input to high-dimensional feature space was achieved.

[0045] (2) The location encoding uses sine and cosine pairs to encode the temporal features of wind speed, and marks the location information for each time step to generate a spatiotemporal joint feature representation that integrates time, location and physical context.

[0046] The location coding formula is: (17) In the formula: For location index; This serves as a dimensional index. Two indexes ensure that signals of different frequencies are injected with timing information, resolving the Transformer's insensitivity to order.

[0047] (3) The Transformer encoder consists of three stacked Transformer encoder layers. Each layer includes an attention layer, a feedforward neural network, and a normalization layer. The attention layer utilizes a multi-head attention mechanism. The Transformer encoder captures the temporal dependencies of spatiotemporal joint feature representations from different subspaces in parallel; a feedforward neural network performs nonlinear transformations of these spatiotemporal joint feature representations; and a normalization layer stabilizes the training process. The Transformer encoder outputs a dimension... The characteristic sequence.

[0048] (4) Obtain the last result of the Transformer encoder After flattening the output feature sequence (corresponding to the prediction step number), it is mapped to a 1D wind speed prediction value through a fully connected layer: (18) That is, the wind speed prediction result output by the wind speed regression model is ultimately .

[0049] The wind speed regression model is trained and its parameters are tuned using a loss function and optimization strategy until the output wind speed prediction result is optimal.

[0050] The loss function is the mean squared error (MSE), and the formula is: (19) In the formula: For actual wind speed, This is the predicted wind speed output by the wind speed regression model.

[0051] Specific optimization strategies include: using AdamW as the optimizer and setting the learning rate to... The weight decay is set to Learning rate scheduling: linear warm-up for the first 10 epochs, followed by cosine annealing scheduler; early stopping mechanism: training is terminated if the loss does not decrease for 8 consecutive epochs to prevent overfitting.

[0052] Wind direction branch model The primary task of the wind direction bifurcation model is to build a wind speed regression model. Simultaneously, inputting wind direction data allows the generation of a wind direction prediction model based on the wind speed model's architecture. This model can provide spatial characteristics for multi-site inversion. However, because the wind direction bifurcation model is a generative model, and wind direction data is more frequently changing than wind power and wind speed data, establishing a wind direction bifurcation model requires adjusting the structure of the wind speed regression model according to the characteristics of the wind direction data.

[0053] exist Figure 4In this model, the wind direction branching approach employs a CNN-Transformer encoder-decoder architecture, aiming to fuse local spatial features with long temporal dependencies. The input to the wind direction branching approach is historical data. The sine and cosine values ​​of the wind speed and direction are taken as the output for the future. The wind direction branch model adds core components to the wind speed regression model, including a CNN feature extraction layer and a Transformer decoder.

[0054] The core differences between the wind direction bifurcation model and the wind speed regression model are as follows: The historical wind speed and direction data from the acquired historical steps are preprocessed. After splitting the wind direction features, the wind direction features are processed to convert the cyclic characteristics, converting the wind direction angle into sine and cosine values ​​to avoid angle cyclic errors (e.g., 350° and 10° are considered close). The input dimension is (3-dimensional historical steps), where 3 includes 1-dimensional wind speed and 2-dimensional wind direction. .

[0055] (1) The CNN feature extraction layer uses a 1D convolutional layer. The input is a spatiotemporal joint feature representation with 3 channels, and the output is a local spatiotemporal feature sequence with 64 channels. Convolutional operations capture local temporal patterns, and the output dimension is... The local spatiotemporal feature sequence is adapted to the input of the Transformer encoder after dimensionality adjustment.

[0056] (2) The Transformer encoder is similar to the wind speed regression model, but has hidden dimensions. (Wind direction changes more frequently, and wind direction prediction relies more on time series generation without mapping features, requiring a more complex model structure).

[0057] The Transformer decoder employs a two-layer decoder layer. The input is a future sequence initialized with all zeros. It generates future step wind direction features through self-attention and encoder-decoder attention, with an output dimension of [dimensionality missing]. The fully connected layer maps the generated future wind direction features to 2D (sin / cos values) through two fully connected network layers, and introduces the ReLU activation function and Dropout regularization. ReLU, through non-linear activation, enables the Transformer decoder to model the complex temporal dependencies in wind direction prediction, while Dropout, through random deactivation, prevents the model from overfitting the training data and improves the generalization ability to real wind direction changes. The combination of the two enables the Transformer decoder to maintain strong expressive power while possessing robustness and stability, making it suitable for long-term, multi-step wind direction prediction tasks.

[0058] The wind direction prediction result is output using the ReLU activation function and Dropout regularization: (20) Similar to the wind speed regression model, the wind direction branch model is also trained using a loss function (mean squared error, MSE) until the output wind direction prediction result is optimal.

[0059] The specific optimization strategy for the wind direction branch model is the same as that for the wind speed regression model.

[0060] In step S2 of this invention, a spatiotemporal coupled wind speed inversion model is constructed. Historical wind speed time-series data, historical wind direction time-series data, and historical wind power time-series data from multiple wind farm sites are input into the pre-trained spatiotemporal coupled wind speed inversion model for prediction, outputting the future wind speed prediction result. Specifically, this includes: For time-series data of wind speed, wind direction and wind power from multiple wind power stations, a spatiotemporal coupled wind speed prediction model is established to invert wind speed at multiple time scales, obtain inversion results, break the assumption of fixed spatiotemporal correlation, and dynamically adapt to complex meteorological conditions.

[0061] like Figure 5 As shown, a flowchart of the multi-site spatiotemporal coupled wind speed prediction model established in this invention is presented, demonstrating the specific architecture of the model.

[0062] The spatiotemporal coupled wind speed prediction model includes a time embedding layer, a location encoding layer, a 1D CNN feature extraction layer, a Transformer encoder, a spatiotemporal dynamic fusion module, and a fully connected layer. This invention extends the single wind power station wind speed regression model to a multi-wind power station scenario through the dynamic graph convolutional layer in the spatiotemporal dynamic fusion module.

[0063] Short-term abrupt changes in wind speed and wind power (such as minute-level wind speed spikes or power drops) have a significant impact on prediction accuracy. Therefore, the spatiotemporal coupled wind speed prediction model constructed in this invention includes a 1D CNN feature extraction layer comprising a 1D convolutional layer, a max-pooling layer, and a linear projection layer.

[0064] The input features of a 1D convolutional layer are composite temporal tensors ,in, B For batch size, T h The historical time step for each historical time series data. N For the number of stations, D =3 represents the feature dimension, including wind speed, wind direction, wind power, and future wind power feature filler values; the input composite time series tensor X undergoes a dimensionality transformation, merging the site dimension and the batch dimension to obtain... ; The kernel size of a 1D convolutional layer is Step length ,filling The activation function used is ReLU; local temporal patterns of wind speed, wind direction, and wind power characteristics are captured through 1D convolutional layers. ; (twenty one) In the formula, The convolution kernel parameters are a three-dimensional tensor. The number of output channels is D, where D represents the feature dimension of the input data. Indicates input data, This represents the convolution operation. The convolution bias term is a term of length . C out The vector.

[0065] The convolution bias term is a trainable parameter. By shifting the result of the convolution kernel operation, it enhances the non-linear expressive ability of the model and helps the model better fit the data.

[0066] The kernel size of the max pooling layer is Step size 2; use maximum pooling layer for... Downsampling is performed, and the feature dimensions of the output wind speed, wind direction, and wind power features are: ; Feature dimensions are reduced through a linear projection layer. Increased from 32 dimensions to an input dimension aligned with the Transformer encoder. The feature tensor of the local spatiotemporal feature sequence after processing by the linear projection layer is obtained. This provides suitable input features for constructing a spatiotemporally coupled wind speed prediction model.

[0067] The Transformer encoder is used to extract global contextual features from each time series data of each site. Through the self-attention mechanism of the Transformer encoder, it captures long-distance temporal dependencies and implicit cross-site associations in the historical data of each site. Compared with traditional RNN or CNN, Transformer avoids the gradient vanishing / exploding problem and is more suitable for processing nonlinear fluctuations in long-term wind farm sequence data.

[0068] Spatial coordination among multiple stations (such as synchronized wind speed fluctuations at adjacent stations) is key to improving prediction accuracy. To this end, the spatiotemporal coupled wind speed prediction model uses a designed dynamic graph convolutional layer to adaptively generate a spatiotemporal adjacency matrix to capture the dynamic correlation between stations. This breaks the assumption of fixed spatiotemporal correlation, dynamically adapts to complex meteorological conditions, and can more realistically simulate the spatiotemporal evolution of wind speed propagation.

[0069] like Figure 6 The diagram shown illustrates the topological structure of a dynamic graph convolutional layer and its mathematical adjacency matrix. Indicates wind power station Indicates wind power station With wind power station The strength of the correlation between them, and the range is . The value is determined by both wind speed and wind direction characteristics. Its practical significance lies in the wind power station... The higher the wind speed, the closer the wind direction is. direction, The larger the value (for wind power plants) The same applies to traditional graph convolution. These can be fixed parameters, such as the correlation strength between wind farms at a given moment (spatial adjacency matrix); alternatively, historical time points of the same node can be considered as "virtual nodes," constructing the dependency of this node in the time dimension (temporal adjacency matrix). The adjacency matrix of dynamic graph convolution is obtained by linearly adding the spatial and temporal adjacency matrices, with parameters... This will change as the time window moves. Its significance lies in the fact that the relationships between different stations change over time with variations in wind speed and direction, which aligns with practical considerations.

[0070] Dynamic graph convolutional layers use learnable node embeddings for representation. Determine the site and The static spatial correlation strength is used to determine the static spatial matrix between stations; ; (twenty two) In the formula: ,element Indicates site and The strength of static spatial associations (such as geographical location or physical connectivity). F long It is for F local Further utilization of long-term time-series information to obtain global contextual features of each time-series data; based on F long The system contains a sequence of local spatiotemporal features, which is used to calculate a dynamic time matrix and capture the dynamic temporal correlations between stations.

[0071] That is, based on the extracted global context features The dynamic time matrix is ​​calculated through the vector dot product in the time dimension; ; (twenty three) In the formula, This reflects the time-varying collaborative patterns between stations (such as the synchronization of wind speeds in the same direction). d model=64 is the input dimension of the Transformer.

[0072] The feature aggregation layer is used to add and normalize the static spatial matrix and the dynamic temporal matrix to obtain the final spatiotemporal adjacency matrix: ; (twenty four) in, The learnable fusion coefficient is taken in this invention. The initial value is 0.5, and it is continuously adjusted adaptively during training.

[0073] Using the spatiotemporal adjacency matrix as weights, a graph convolution operation is performed on the global context features to output spatiotemporal fusion features: (25) In the formula, The graph convolution weight matrix, For the graph convolution bias term, It is the ReLU activation function. The output is a spatiotemporal fusion feature, which integrates spatiotemporal correlation information from multiple sites.

[0074] By sharing multi-site information through spatiotemporal fusion features, the robustness of local site predictions is improved. This spatiotemporal fusion feature can dynamically adjust the spatiotemporal focus based on real-time data, avoiding the limitations of a fixed graph structure.

[0075] The fully connected layer consists of two fully connected layers; by extracting the spatiotemporal fusion features of the last future time step (output time step), the shape is... After being flattened, it is mapped sequentially through two fully connected layers to obtain: (26) in, and These represent two fully connected layers, describing the mapping relationship between the input dimensions and output dimensions of the two fully connected layers. Mapping the dimensions from 64 to 32. Mapping the dimension from 32 to 1 dimension, the final output is... The output dimension is the predicted wind speed for the future step. .

[0076] The fully connected layer performs the final mapping of spatiotemporal fusion features and fits the complex relationship between wind speed and power through multi-layer nonlinear transformation.

[0077] The wind speed regression model was trained and its parameters tuned using a loss function and optimization strategy until the optimal wind speed prediction result was achieved. The loss function was set as the mean squared error of the future step wind speed prediction values ​​output by the fully connected layer. The accuracy of both single-wind-station and multi-wind-station wind speed prediction models was evaluated. The specific optimization strategy was the same as that for traditional wind speed regression models.

[0078] By utilizing the sliding time window feature in the spatiotemporal coupled wind speed prediction model, the output dimension is divided into output data and input data within a single time step window.

[0079] Based on the output and input data within a single time step window, the wind power data and wind speed data within the sliding time window are aligned. The wind power feature row vectors of multiple time steps are transformed into wind power differentiation feature column vectors of the same time step, thus preserving wind power data information to the greatest extent and ensuring the prediction accuracy of the spatiotemporal coupled wind speed prediction model.

[0080] like Figure 7 As shown, the following are given: (a) a schematic diagram of sample data and sliding window when wind speed and wind power data are balanced and aligned; (b) a time resolution diagram of wind speed data and wind power data. (a) is a schematic diagram of sample data and sliding window when the time resolution of wind power data is 15 min and the time resolution of wind speed data is 1 h. (b) is a schematic diagram of sample data and sliding window when the oversampling interpolation method is used for (b). (c) is a schematic diagram of sample data and sliding window when the undersampling fitting method is used for (b). (d) is a schematic diagram of sample data and sliding window when the feature differentiation and alignment method of wind power data is used in S302.

[0081] The prediction accuracy of the wind speed inversion method proposed in this invention is evaluated by comparing the wind power data feature differentiation and alignment method with undersampling and oversampling methods.

[0082] This invention addresses the power system's need for high-precision, high-timeliness, and operability of large-scale meteorological forecast models, as well as the shortcomings of current conventional meteorological services in power system applications. It constructs a spatiotemporally coupled wind speed prediction model to reflect the wind speed inversion process based on the active power of wind farms.

[0083] This invention is based on the entire process of wind speed inversion method for active power of wind farms. It studies the accurate conversion method of wind farm data at different time scales, and verifies and evaluates each link through real datasets, forming a complete technical research system.

[0084] This invention also proposes a wind speed inversion system based on the active power of a wind farm, characterized by comprising: The data acquisition module is used to acquire historical wind speed time-series data, historical wind direction time-series data, and historical wind power time-series data from multiple sites in the wind farm. The wind speed inversion module is used to input historical wind speed time series data, historical wind direction time series data, and historical wind power time series data from multiple sites of the wind farm into a pre-trained spatiotemporal coupled wind speed inversion model for prediction, and output the future step wind speed prediction results. A spatiotemporal coupled wind speed inversion model construction module is introduced. The spatiotemporal coupled wind speed inversion model includes a Transformer encoder and a fully connected layer. A spatiotemporal dynamic fusion module is introduced between the Transformer encoder and the fully connected layer. The spatiotemporal dynamic fusion module includes a dynamic graph convolutional layer and a feature aggregation layer. The Transformer encoder extracts global contextual features from each time series data point at each station. The dynamic graph convolutional layer obtains a static spatial matrix by determining the strength of the static spatial association between stations and a dynamic temporal matrix by determining the vector dot product of the temporal dimensions between stations. The feature aggregation layer fuses the static spatial matrix and the dynamic temporal matrix to obtain a spatiotemporal adjacency matrix, which is then used as a weight to perform spatiotemporal fusion on the global contextual features of each time series data point, outputting the spatiotemporal fusion features. The fully connected layer maps the spatiotemporal fusion features and outputs the future wind speed prediction results.

[0085] The established spatiotemporal coupled wind speed prediction model is used as an inversion model. First, the overall requirements of the inversion model of this invention are introduced in general, and the advantages and disadvantages of each model are studied according to the requirements. Then, for single-site scenarios, the wind speed inversion model framework applicable to this invention is introduced. Finally, it is extended to multi-site scenarios, and a multi-site spatiotemporal coupled wind speed prediction model with higher prediction accuracy than the single-site time series model is established.

[0086] This paper proposes an accurate transformation method for wind power data across different time scales. Building upon the inversion model constructed in the previous step, it addresses the feature misalignment problem caused by differences in temporal resolution between wind power and wind speed data in reality. By systematically comparing traditional methods for handling data time scale mismatch, a wind power data feature differentiation and alignment method is proposed, aiming to maximize the preservation of feature information. This method effectively solves the problem of feature distortion and information loss caused by inconsistent sampling frequencies through multi-dimensional feature decoupling and cross-scale mapping mechanisms.

[0087] As can be seen, this method establishes a complete process for inversion modeling, from data preprocessing to model modeling and finally to model post-processing.

[0088] The wind speed inversion method proposed in this invention achieves high-precision, dynamic, and multimodal collaborative prediction of wind speed at multiple sites in a wind farm through the synergistic effect of global feature extraction by Transformer, spatiotemporal modeling by dynamic graph convolution, and feature aggregation and fusion.

[0089] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0090] Furthermore, unless otherwise stated, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. All references to this specification are incorporated by way of citation to disclose and describe methods relating to those references. In the event of any conflict with any incorporated reference, the content of this specification shall prevail.

Claims

1. A wind speed inversion method based on the active power of a wind farm, characterized in that, Includes the following steps: Acquire historical wind speed time-series data, historical wind direction time-series data, and historical wind power time-series data from multiple sites in the wind farm; Historical wind speed time series data, historical wind direction time series data, and historical wind power time series data from multiple sites of the wind farm are input into a pre-trained spatiotemporal coupled wind speed inversion model for prediction, and the future step wind speed prediction results are output. The spatiotemporal coupled wind speed inversion model includes a Transformer encoder and a fully connected layer. A spatiotemporal dynamic fusion module is introduced between the Transformer encoder and the fully connected layer. The spatiotemporal dynamic fusion module includes a dynamic graph convolutional layer and a feature aggregation layer. The Transformer encoder is used to extract global contextual features of each time series data from each site. The dynamic graph convolutional layer obtains a static spatial matrix by determining the static spatial association strength between sites, and obtains a dynamic temporal matrix by determining the vector dot product of the temporal dimension between sites; the feature aggregation layer is used to fuse the static spatial matrix and the dynamic temporal matrix to obtain a spatiotemporal adjacency matrix, which is used as a weight to perform spatiotemporal fusion on the global context features and output spatiotemporal fusion features. The fully connected layer maps the spatiotemporal fusion features and outputs the future wind speed prediction results.

2. The wind speed inversion method based on the active power of a wind farm according to claim 1, characterized in that, The spatiotemporal coupled wind speed inversion model also includes an embedding layer, a location encoding layer, and a CNN feature extraction layer: The embedding layer is used to map the acquired time-series data to the high-dimensional hidden dimension of the Transformer encoder to extract the time-series features of wind speed and wind direction. The location encoding layer is used to encode the temporal features of the extracted wind speed and wind direction, label the location information for each time step, and generate a spatiotemporal joint feature representation of each temporal data that integrates time, location and physical context. The spatiotemporal joint feature representation is input into the 1D convolutional layer of the CNN feature extraction layer. By capturing the local temporal patterns of wind speed and wind direction, the local spatiotemporal feature sequence of each temporal data is output, which is aligned with the hidden dimension of the Transformer encoder. The local spatiotemporal feature sequence is used as the input to the Transformer encoder.

3. The wind speed inversion method based on the active power of a wind farm according to claim 2, characterized in that, The embedding layer is used to map the acquired time-series data to the high-dimensional hidden dimension of the Transformer encoder, extracting the time-series features of wind speed and wind direction, and also includes: Wind direction features were extracted from the time-series data of multiple sites in the wind farm. The wind direction characteristics are processed using a cyclic feature, converting the wind direction angle into sine and cosine values. The embedding layer maps the acquired time-series data, sine values, and cosine values ​​to the high-dimensional hidden dimension of the Transformer encoder to extract the time-series features of wind speed and wind direction.

4. The wind speed inversion method based on the active power of a wind farm according to claim 1, characterized in that, The acquisition of historical wind speed time-series data, historical wind direction time-series data, and historical wind power time-series data from multiple sites of the wind farm also includes: The acquired historical wind speed time series data, historical wind direction time series data, and historical wind power time series data are time-stamped and synchronized. Based on the statistical distribution characteristics of wind turbine power curves, parameter estimation is performed on calibrated historical wind speed time series data, historical wind direction time series data, and historical wind power time series data. By constructing dynamic confidence boundaries, the preset confidence boundary range is used as the core interval of effective data, and outliers exceeding this range are screened out. After removing outliers, historical wind speed time series data, historical wind direction time series data, and historical wind power time series data from multiple sites are obtained, and interpolation processing is performed to obtain continuous historical wind speed time series data, historical wind direction time series data, and historical wind power time series data.

5. The wind speed inversion method based on the active power of a wind farm according to claim 4, characterized in that, The process of acquiring historical wind speed time-series data, historical wind direction time-series data, and historical wind power time-series data from multiple sites after removing outliers, and performing interpolation processing to obtain continuous historical wind speed time-series data, historical wind direction time-series data, and historical wind power time-series data, specifically includes: The local outlier factor algorithm is used to identify outliers in historical wind speed time series data, historical wind direction time series data, and historical wind power time series data. The weighted distance outlier algorithm was used to clean up outliers, and the historical wind speed time series data, historical wind direction time series data and historical wind power time series data of multiple sites were obtained after removing outliers. Based on the Akaike Information Criterion (AIC), the AIC value of the combination of historical wind speed time series data and historical wind power time series data of multiple sites after removing outliers is determined. The combination of historical wind speed time series data and historical wind power time series data of multiple sites after removing outliers with the smallest AIC value is selected as the global optimal order of the autoregressive moving average (ARMA) data interpolation algorithm. By using the ARMA model corresponding to the globally optimal order, interpolation results with the same style as the original historical wind speed time series data, historical wind direction time series data, and historical wind power time series data are generated, that is, continuous historical wind speed time series data, historical wind direction time series data, and historical wind power time series data.

6. The wind speed inversion method based on the active power of a wind farm according to claim 1, characterized in that, The static spatial matrix obtained by determining the static spatial association strength between sites is generated by encoding site embedding representations through a Transformer encoder, calculating the static correlation between sites, and determining the static spatial association strength between sites.

7. A wind speed inversion system based on the active power of a wind farm, characterized in that, include: The data acquisition module is used to acquire historical wind speed time-series data, historical wind direction time-series data, and historical wind power time-series data from multiple sites in the wind farm. The wind speed inversion module is used to input historical wind speed time series data, historical wind direction time series data, and historical wind power time series data from multiple sites of the wind farm into a pre-trained spatiotemporal coupled wind speed inversion model for prediction, and output the future step wind speed prediction results. A spatiotemporal coupled wind speed inversion model construction module is provided. The spatiotemporal coupled wind speed inversion model includes a Transformer encoder and a fully connected layer. A spatiotemporal dynamic fusion module is introduced between the Transformer encoder and the fully connected layer. The spatiotemporal dynamic fusion module includes a dynamic graph convolutional layer and a feature aggregation layer. The Transformer encoder is used to extract the global context features of each time series data of each station. The dynamic graph convolutional layer obtains a static spatial matrix by determining the static spatial association strength between sites, and obtains a dynamic temporal matrix by determining the vector dot product of the temporal dimension between sites; the feature aggregation layer is used to fuse the static spatial matrix and the dynamic temporal matrix to obtain a spatiotemporal adjacency matrix, which is used as a weight to perform spatiotemporal fusion on the global context features and output spatiotemporal fusion features. The fully connected layer maps the spatiotemporal fusion features and outputs the future wind speed prediction results.

Images