Wind power cluster space-time fusion output prediction method based on branch network and dynamic source selection

Abstract

The invention discloses a wind power cluster space-time fusion output prediction method based on branch network and dynamic source selection. In view of significant differences of sounding data, reanalysis data and data acquisition and monitoring system data in temporal-spatial resolution and information content, the method designs a special branch network structure, performs independent feature extraction on each data source, and performs fusion to obtain a wind power output predicted value; the method comprises the following steps: firstly, extracting time sequence features from historical power data of each wind power plant, and introducing a multi-head attention mechanism to generate wind power cluster spatial features; secondly, performing convolution operation on the reanalysis data to obtain regional meteorological features; secondly, a dynamic source selection mechanism is introduced, contribution weights of data sources are adjusted in a self-adaptive mode on different prediction step lengths, and a preliminary prediction value is obtained; and finally, obtaining a corrected prediction value based on the sounding data, and realizing wind power cluster output prediction. The method supports synchronous prediction of multiple step lengths and multiple wind power plants, and is suitable for high-precision power prediction application of a large-scale wind power cluster.

Description

TECHNICAL FIELD

[0001] The present application relates to a wind power cluster spatiotemporal fusion output prediction method based on a branch network and dynamic source selection, which handles the modeling challenges brought by the heterogeneity of multi-source data, improves the accuracy of wind power cluster output prediction, and belongs to the technical field of data processing and wind power prediction. BACKGROUND

[0002] Currently, wind power is being widely deployed and rapidly developed as one of the core means to achieve low-carbon transformation of energy systems. However, due to the intermittent and volatile nature of wind energy resources, it is easy to cause delay or insufficient response of power grid dispatching, thereby affecting the stable operation of the power system and increasing the uncertainty of power generation and the cost of power grid operation and maintenance. Therefore, accurate wind power prediction has become a key technical means to alleviate the above problems.

[0003] The scheduling of wind power needs to rely on different prediction time scales. For example, day-ahead scheduling usually relies on hourly prediction results of wind power output for the next 24 hours; while day-ahead scheduling is usually based on 15-minute power prediction results within the next 6 hours. The former is called short-term prediction, and the latter is called ultra-short-term prediction. Although these two types of prediction models have certain commonalities in structure, due to the differences in time resolution and prediction time range, their prediction performance often differs significantly. Generally speaking, as the prediction step increases, the prediction error also increases, which is due to the increasing uncertainty of future states and the decreasing correlation with input features.

[0004] In order to cope with this challenge, researchers have tried to introduce external factors to extract more predictive feature information. However, despite the many advances in data-driven wind power prediction in recent years, there are still several key research gaps that need to be addressed. First, existing research generally ignores the importance of multi-source heterogeneous data. Previous methods have focused on using homogeneous data for modeling, i.e., only using data from the same source as wind power, which naturally has consistent time and spatial resolution, thereby simplifying the modeling process. However, models that rely solely on homogeneous data are difficult to fully utilize the rich information contained in multi-source heterogeneous data. Multi-source heterogeneous data not only provides complementary perspectives, but also may contain important forward-looking information, which is of great significance to improving prediction accuracy. Second, the current fusion strategy for heterogeneous data lacks detailed design. Although some research has attempted to integrate multi-source data, most methods still use a unified data processing flow, failing to fully consider the differences in characteristics between different data sources, limiting further improvement in model performance. Finally, data from different sources has inconsistent time and spatial resolution, and the impact of this mismatch varies at different prediction steps. However, previous modeling methods generally do not address this issue, limiting the model's ability to generalize in multiple scenarios. SUMMARY

[0005] The application aims at the problems and deficiencies of the prior art, and provides a wind power cluster spatio-temporal fusion output prediction method based on branch network and dynamic source selection, which overcomes the limitation of single data source information, effectively handles the modeling challenges brought by multi-source data heterogeneity, and significantly improves the wind power cluster output prediction accuracy.

[0006] The technical scheme is a wind power cluster spatio-temporal fusion output prediction method based on branch network and dynamic source selection, comprising the following steps: Step 1: Collect data from different sources, including wind power output data from a data acquisition and monitoring system, sounding data, and wind speed data from reanalysis data.

[0007] Step 2: Build a special branch network for wind power data, i.e. a wind power special branch network, for wind power data; the branch network extracts the historical power time sequence features of each wind farm through an independent fully connected neural network, and generates spatial features representing the wind power output in the wind power cluster through a multi-head attention mechanism according to the historical power time sequence features of multiple wind farms with different geographical positions; the time sequence features and the spatial features are spliced along the feature dimension, and are fused through a fully connected neural network to obtain the spatio-temporal features output by the wind power special branch network, which are used for subsequent fusion and wind power prediction.

[0008] Step 3: Organize the wind speed data in a regular latitude-longitude grid form, which has a two-dimensional spatial structure like an image, and use a convolutional neural network to build a reanalysis data special branch network to extract regional meteorological features that are helpful for wind power output prediction from the wind speed data of the reanalysis data.

[0009] The reanalysis data provides a near-surface wind speed field covering the target area, which is continuous in space and time, and has a spatial resolution significantly higher than the density of wind farm layout. Wind speed is a direct physical driving factor for wind power output, and using wind speed information as a prediction input has a clear physical basis; at the same time, compared with historical wind power output data that only reflect the power generation state of discrete wind farm locations, the reanalysis wind speed field can comprehensively depict the local wind resource heterogeneity and regional scale wind system structure of the continuous spatial distribution in the target area.

[0010] The convolution operation can effectively identify local high-value areas, gradient zones and other meteorological structures in the wind speed field that are closely related to wind power output through a local receptive field, and its translational invariance enables the model to generalize the recognition of similar wind systems appearing at different geographical locations, thereby learning spatial representations with physical interpretability, rather than relying only on statistical correlations of geographical positions.

[0011] Step4: adaptively fuse the spatio-temporal features output by the wind power dedicated branch network and the regional meteorological features from the reanalysis data dedicated branch network according to the prediction time step.

[0012] Considering that the reanalysis data has high spatial resolution but low temporal resolution, and the wind power data has high temporal resolution but lacks spatial context, the effectiveness of the two differs in different prediction time domains: in ultra-short-term prediction, historical wind power is more valuable for reference; while in longer-term prediction, regional meteorological structures provide key prior information. Therefore, the invention dynamically calculates the fusion weights of the two types of features, historical power spatio-temporal features and regional meteorological features, for each prediction time step, to adaptively balance the contributions of historical power spatio-temporal features and regional meteorological background.

[0013] Step5: adjust the prediction results using sounding data; use the pressure change feature to adjust the amplitude of the predicted wind power fluctuation. Considering that although the sounding data has low spatial and temporal resolution, its seasonal variation feature of pressure sequence is similar to the seasonal fluctuation of wind power output, the invention extracts the statistical features of pressure change from the sounding data and generates a modulation factor through a lightweight linear layer and an activation function; the factor acts on the difference (i.e. prediction increment) between adjacent time points of the preliminary prediction result, adaptively scales the change amplitude of the wind power, and finally obtains the corrected prediction value. This design increases the ability to characterize the fluctuation characteristics of wind power output in different seasons through sounding data, avoiding excessive smoothing of the prediction results in active weather periods or false oscillation in stable periods.

[0014] A branch network and dynamic source selection wind power cluster spatio-temporal fusion output prediction system, comprising the following modules: Data collection module: used to collect data from different sources, including wind power output data from the data acquisition and monitoring system, sounding data, and wind speed data from reanalysis data; Wind power dedicated branch network construction module: a dedicated branch network, wind power dedicated branch network, is built for wind power data of wind power output data; the branch network extracts historical power time series features of each wind farm through an independent fully connected neural network, and generates spatial features representing wind power output in the wind power cluster through a multi-head attention mechanism according to historical power time series features of multiple wind farms with different geographical locations; the time series features and the spatial features are spliced along the feature dimension, and fused through a fully connected neural network to obtain spatio-temporal features output by the wind power dedicated branch network, which are used for subsequent fusion and wind power prediction; Reanalysis data dedicated branch network construction module: a convolutional neural network is used to construct a reanalysis data dedicated branch network to extract regional meteorological features from wind speed data of reanalysis data that are helpful for wind power output prediction; Fusion module: adaptively fuse the spatio-temporal features output by the wind power dedicated branch network at the predicted time step with the regional meteorological features from the reanalysis data dedicated branch network; Prediction dynamic adjustment module: adjust the prediction results using sounding data; adjust the amplitude of the predicted wind power fluctuation using pressure change characteristics; extract the statistical characteristics of pressure change from the sounding data, and generate a modulation factor through a linear layer and an activation function; the modulation factor acts on the difference of the adjacent time of the preliminary prediction result, and adaptively scales the change amplitude of the wind power, to finally obtain the corrected prediction value.

[0015] The implementation process of the system is the same as the above method, and will not be described again.

[0016] A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor executes the above-mentioned computer program to realize the steps of the branch network and dynamic source selection wind power cluster spatio-temporal fusion output prediction as described above.

[0017] A computer readable storage medium has a computer program / instruction stored thereon, and the computer program / instruction is executed by a processor to realize the steps of the branch network and dynamic source selection wind power cluster spatio-temporal fusion output prediction as described above.

[0018] Advantages: Compared with the prior art, the present application has the following advantages: (1) Enhanced multi-source heterogeneous data utilization capability: by integrating multiple sources of data, including historical wind power output data, sounding data, and wind speed data from reanalysis data, the present application can effectively mine and utilize the unique information contained in each data source, overcoming the information limitation problem caused by using only homogeneous data in traditional methods, thereby improving the prediction accuracy.

[0019] (2) Fine data fusion strategy: Unlike previous unified processing methods, the present application designs targeted data processing and fusion strategies according to the characteristics of different data sources, ensuring that the mismatch of various data in terms of spatial and temporal resolution is properly solved, thereby enhancing the generalization ability and adaptability of the model in complex application scenarios.

[0020] (3) Dynamic source selection mechanism: the present application innovatively introduces a dynamic source selection mechanism, which can adaptively adjust the contribution weight of each data source for each prediction step in the multi-step prediction process. This mechanism naturally adapts to the demand for one-time output of multi-time scale prediction results, effectively reflecting the differentiated dependence of different prediction time domains on information sources.

[0021] In summary, this invention effectively improves the accuracy of wind power forecasting by fusing multi-source heterogeneous data and dynamically adjusting the contribution weights of each data source based on the prediction step size. This method fully utilizes the advantages of data from different sources and adapts to the actual need for generating multi-step prediction values ​​at once through a dynamic source selection mechanism, thereby enhancing the reliability of day-ahead and intraday multi-timescale forecasting tasks. Attached Figure Description

[0022] Figure 1 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0023] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0024] like Figure 1 As shown, the spatiotemporal fusion power output prediction method for wind power clusters using branch networks and dynamic source selection includes the following steps: Step 1: Collect data from different sources, including wind power output data from the data acquisition and monitoring system, radiosonde data, and wind speed data from reanalysis data.

[0025] Specifically, this includes: collecting measured wind power data from wind power clusters within the target area, provided by the data acquisition and monitoring systems of each wind farm; acquiring radiosonde data (air pressure in the vertical atmospheric profile) from radiosonde stations within the target area over the same time period; and acquiring reanalysis data covering the target area within the same time period, including 10-meter and 50-meter wind speeds. The variables and spatiotemporal resolutions in each dataset are shown in Table 1. In the ultra-short-term forecasting task, predictions for the next 6 hours are required at 15-minute intervals. Therefore, both radiosonde and reanalysis data are padded forward with the most recent valid values ​​and upsampled to the required 15-minute time resolution. In the short-term forecasting task, the goal is to predict hourly wind power for the next 24 hours. In this case, radiosonde and reanalysis data are also upsampled to hourly resolution using the same method, while the wind power data acquired from the data acquisition and monitoring system is obtained by accumulating the values ​​within each hour to obtain hourly data.

[0026] Table 1: Comparison Table of Variables and Spatiotemporal Resolution in Multi-Source Datasets

[0027] Step 2: Construct a dedicated branch network for the wind power data acquired by the data acquisition and monitoring system. The branch network extracts the historical power time-series features of each wind farm using independent fully connected neural networks. Based on the historical power time-series features of multiple wind farms with different geographical locations, a multi-head attention mechanism is used to generate spatial features representing wind power output within the wind power cluster. The time-series features and spatial features are concatenated along the feature dimension and fused using a fully connected neural network to obtain the spatiotemporal features output by the dedicated wind power branch network, which are used for subsequent fusion and wind power prediction. The specific implementation process includes the following steps: Step 2.1: Capture the temporal characteristics of a single wind farm using a fully connected neural network. For each individual wind farm... , This represents the total number of wind farms; in this example, there are 10 wind farms. First, their historical wind power output is used... Data is used to model its time patterns, where This represents the length of the time series. Specifically, a fully connected neural network (FCN) consisting of two linear layers with a ReLU activation function in between is used for modeling, as shown in the formula below.

[0028]

[0029] in, and These are learnable parameters. This represents the ReLU activation function. Output Indicates the first The temporal characteristics of a wind farm are given, where D represents the prediction step size of the target variable in the prediction task. In this embodiment, D=24 is set.

[0030] By stacking the time-series characteristics of all wind farms, we obtain , representing the time-series characteristics of all wind farms.

[0031] Step 2.2: Utilize multi-head attention mechanisms to capture the spatial characteristics between multiple wind farms (wind power clusters). For each wind farm... Temporal characteristics As the basis for the query, the time-series characteristics of all wind farms (including themselves) This serves as the basis for both keys and values; based on this, each attention head generates its own unique query vector through independent learnable transformations. Key vector Sum value vector Each attention head is based on its and The similarity between them, for Weighted combination calculations are performed, allowing each wind farm to learn from information from all other wind farms in a learnable way. Single attention head (SHOOL) The formula for ) is as follows.

[0032]

[0033] Where Attention() is the attention head function; softmax() is the normalized exponential function; and T represents the transpose. Represents the dimension of the key vector. Representing the The outputs from each attention head result in a total of M attention output heads. After concatenating all attention output heads and performing a linear transformation, the spatial characteristics of wind power output within the wind power cluster are obtained. The specific formula is as follows:

[0034] in, It is a learnable parameter that serves to weight and fuse information from multiple attention heads.

[0035] Step 2.3: Integrate the temporal characteristics of a single wind farm with the spatial characteristics of wind power output within a wind power cluster. Temporal characteristics and spatial features The data is then spliced ​​together and fused using a fully connected neural network to obtain the spatiotemporal characteristics of the wind power-specific branch network output. .

[0036]

[0037] in, This indicates a splicing operation along the first dimension. Spatiotemporal features. It integrates temporal and spatial features and will be further used for feature fusion and final prediction in subsequent modules.

[0038] Step 3: Construct a dedicated branch network for reanalysis data to extract regional meteorological characteristics that are helpful for wind power output prediction; The reanalysis data provides a near-surface wind speed field that covers the target area, is spatiotemporally continuous, and has a spatial resolution significantly higher than that of wind farms. Given that wind speed is the direct physical driver of wind power output, using this wind speed information as prediction input has a clear physical basis. Furthermore, compared to historical wind power output data that only reflects the power generation status of discrete wind farm locations, the reanalysis wind speed field can comprehensively characterize the heterogeneity of local wind resources and the regional-scale wind system structure within the target area. Since the wind speed data is organized in a regular latitude and longitude grid, possessing an image-like two-dimensional spatial structure, this invention employs a convolutional neural network as the core architecture of this branch.

[0039] Convolution operations, through local receptive fields, can effectively identify meteorological structures closely related to wind power output, such as local high-value areas and gradient zones in the wind speed field. Their translation invariance allows the dedicated branch network for reanalysis data to generalize and identify similar wind systems occurring at different geographical locations, thereby learning physically interpretable spatial representations rather than relying solely on statistical correlations based on geographical location. The tensor shape of the reanalysis data is (N, 2, K, J), where N represents the number of samples, 2 represents two climate factors (wind speeds at two different altitudes), K represents the number of sampling points obtained after equally dividing the target area along the longitude direction (in this embodiment, K=9), and J represents the number of sampling points obtained after equally dividing the target area along the latitude direction (in this embodiment, J=15), where 9×15 is the dimension of the corresponding latitude and longitude grid points.

[0040] First, let's analyze the wind speed data. Through a convolution kernel size of The convolutional layer is followed by an activation function. This process can be represented as:

[0041] in Represents a convolution kernel, This represents the ReLU activation function.

[0042] Subsequently, another convolutional layer with a kernel size of 1×1 is applied to reduce the feature dimensionality while preserving feature information, thereby obtaining shared features. .

[0043]

[0044] Among them shared features It contains D channels, each channel corresponding to a shared feature map of size 9×15; This represents the ReLU activation function.

[0045] To adapt to changes in the region of interest under different prediction step sizes, a spatial attention mechanism is introduced. This is achieved through a convolutional layer... Generate a spatial attention map with D channels. :

[0046] in, The function ensures that the sum of attention weights at each spatial location is 1.

[0047] Next, shared features will be used. Spatial attention map Element-wise multiplication yields the enhanced features. This operation dynamically adjusts the weights of each spatial location through an attention mechanism, thereby highlighting areas that are more critical to the wind power forecasting task. This allows the reanalysis data branch network to adaptively focus on the most informative spatial locations in the input meteorological field data, thus improving the accuracy of subsequent wind power forecasting.

[0048] Subsequently, Apply adaptive pooling to reduce its spatial dimension from the original Uniform downsampling to Where B and F are configurable positive integer parameters that satisfy the condition , as well as For the target number of wind farms, parameters B and F are set according to actual application requirements to control the spatial range of the local receptive field or feature extraction region. In this embodiment, B=2 and F=5 are set to obtain a tensor of shape (N,D,B,F). , where N represents the number of samples and D corresponds to the prediction step size in the prediction task.

[0049] Finally, Flattened along the spatial dimensions, reshaped into a shape of (N,D, Regional meteorological characteristics output from the dedicated branch network of reanalysis data. Thus achieving from Spatial grid to A one-to-one mapping of each target wind farm. This transformation aligns the feature structure with the output dimension of the wind farm, ensuring that predictions at each time step are based on a spatial context that matches the task requirements.

[0050] Step 4: Fusion module, adaptively fuses the spatiotemporal characteristics output by the wind power dedicated branch network with the regional meteorological characteristics from the reanalysis data dedicated branch network according to the prediction time step; Given that reanalysis data has high spatial resolution but low temporal resolution, while wind power data has high temporal resolution but lacks spatial context, their effectiveness differs across prediction time domains: historical power is more valuable for ultra-short-term predictions, while regional meteorological structure provides crucial prior information for longer-term predictions. Therefore, this invention dynamically calculates the fusion weights of the two types of features for each prediction time step, adaptively balancing the contributions of historical power spatiotemporal features and regional meteorological background. Regional meteorological features from a dedicated branch network of reanalysis data are fused. Spatiotemporal characteristics of wind power dedicated branch network output This allows wind power cluster output prediction to simultaneously utilize climate background information and the spatiotemporal characteristics of wind power generation. To align the feature dimensions, firstly... Perform the transpose operation to obtain .

[0051] The outputs of the two branches are concatenated along the feature dimension and then input into a fully connected neural network to obtain the fused prediction output. .

[0052]

[0053] in, Each element in Indicates the first In the nth sample The wind farm in the first Predicted power output for one future time step.

[0054] To enable the wind power cluster output prediction method proposed in this invention to adaptively adjust the importance of reanalysis data according to different prediction step sizes, we introduce a dynamic source selection mechanism. Specifically, a set of learnable parameters is generated by using a lightweight fully connected network connected to a sigmoid activation function. This is used to dynamically fuse the contributions of the reanalysis data branch and the wind power branch, thereby obtaining the preliminary predicted value output by the fusion module. :

[0055]

[0056] This represents a fully connected neural network with an activation function added between two linear layers. This dynamic source selection mechanism enables the wind power cluster output prediction method to adaptively adjust the emphasis on each input feature according to different prediction step sizes, thereby highlighting the feature information most relevant to the current prediction task.

[0057] Step 5: Adjusting Prediction Results Using Radiosonde Data. Considering that although radiosonde data has low spatiotemporal resolution, its seasonal pressure sequence exhibits similarities to the seasonal fluctuations in wind power output. This invention extracts statistical features of pressure changes from radiosonde data and generates a modulation factor using a lightweight linear layer and activation function. This factor acts on the time-series difference (i.e., prediction increment) of the initial prediction results, adaptively scaling the amplitude of wind power variation to obtain the corrected prediction value. This design enhances the characterization of wind power output fluctuations in different seasons using radiosonde data, avoiding over-smoothing of prediction results during periods of active weather or producing spurious oscillations during stable periods. Specifically, both wind power output and surface pressure changes exhibit greater volatility in spring and winter, and less volatility in summer and autumn. Based on this phenomenon, the amplitude of predicted wind power fluctuations is adjusted using the characteristics of pressure changes.

[0058] set up The air pressure represented by the sounding data, where The sample size is 12, representing the historical observation data from the previous 6 days, twice a day, and 2 representing two radiosonde stations. Based on this radiosonde data sequence, several statistical characteristics were calculated to characterize the pressure changes, including: the average magnitude of change (…). Volatility (standard deviation) ), maximum change ( ), and trend consistency (the proportion of continuous changes in the same direction). These features are pieced together into a compact representation. Then, a fully connected network is input, and a modulation factor is generated using the Tanh activation function. :

[0059]

[0060] The modulation factor It is applied uniformly across all forecast time steps and wind farms. The motivation for this design is mainly twofold: (1) the seasonal fluctuation patterns have similar effects on all wind farms; and (2) atmospheric conditions are relatively stable within the short-term forecast timeframe.

[0061] Finally, the modulation factor is applied to the output of the fusion module. The corrected predicted value is obtained. This refers to the final predicted value of wind power, as detailed below:

[0062] in, The temporal differences of the predicted sequence between adjacent time steps were calculated; and Broadcasting over time to match The shape; This represents the standard element-wise addition operation. The formula dynamically adjusts the prediction results by scaling the temporal variation of the predicted wind power sequence based on atmospheric conditions extracted from radiosonde data.

[0063] A spatiotemporal fusion power output prediction system for wind power clusters based on branch networks and dynamic source selection includes the following modules: Data collection module: used to collect data from different sources, including wind power output data from the data acquisition and monitoring system, as well as sounding data and wind speed data from reanalysis data; Wind Power Dedicated Branch Network Construction Module: A dedicated branch network—the wind power dedicated branch network—is built for wind power output data. This branch network extracts the historical power time-series features of each wind farm through independent fully connected neural networks. Based on the historical power time-series features of multiple wind farms with different geographical locations, a multi-head attention mechanism is used to generate spatial features representing wind power output in the wind power cluster. The time-series features and the spatial features are concatenated along the feature dimension and fused through a fully connected neural network to obtain the spatiotemporal features output by the wind power dedicated branch network, which are used for subsequent fusion and wind power prediction. Reanalysis Data Dedicated Branch Network Construction Module: A convolutional neural network is used to construct a dedicated branch network for reanalysis data, which extracts regional meteorological features that are helpful for wind power output prediction from the meteorological data of the reanalysis data; Fusion module: Adaptively fuses the spatiotemporal characteristics output by the wind power dedicated branch network with the regional meteorological characteristics from the reanalysis data dedicated branch network according to the prediction time step; The prediction dynamic adjustment module adjusts the prediction results using radiosonde data; it uses air pressure change characteristics to adjust the amplitude of predicted wind power fluctuations; it extracts statistical features of air pressure changes from radiosonde data and generates a modulation factor through a linear layer and activation function; this modulation factor acts on the time difference between adjacent moments of the preliminary prediction results to adaptively scale the amplitude of wind power changes, and finally obtains the corrected prediction value.

[0064] Obviously, those skilled in the art should understand that the steps of the spatiotemporal fusion power output prediction method for wind power clusters with branch networks and dynamic source selection, or the modules of the spatiotemporal fusion power output prediction system for wind power clusters with branch networks and dynamic source selection described in the above embodiments of the present invention, can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by the computing device. In some cases, the steps shown or described can be performed in a different order than those described herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of the present invention are not limited to any particular hardware and software combination.

Claims

1. A spatiotemporal fusion power output prediction method for wind power clusters based on branch networks and dynamic source selection, characterized in that, Includes the following steps: Step 1: Collect data from different sources, including wind power output data from the data acquisition and monitoring system, as well as sounding data and wind speed data from reanalysis data; Step 2: Construct a dedicated branch network for wind power output data – the Wind Power Dedicated Branch Network. This branch network extracts the historical power time-series features of each wind farm through independent fully connected neural networks. Based on the historical power time-series features of multiple wind farms with different geographical locations, a multi-head attention mechanism is used to generate spatial features representing wind power output in the wind power cluster. The time-series features and spatial features are concatenated along the feature dimension and fused through a fully connected neural network to obtain the spatiotemporal features output by the Wind Power Dedicated Branch Network, which are used for subsequent fusion and wind power prediction. Step 3: Construct a dedicated branch network for reanalysis data using a convolutional neural network to extract regional meteorological features from the wind speed data of the reanalysis data that are helpful for wind power output prediction; Step 4: Adaptively fuse the spatiotemporal characteristics output by the wind power dedicated branch network with the regional meteorological characteristics from the reanalysis data dedicated branch network according to the prediction time step; Step 5: Adjust the prediction results using radiosonde data; adjust the amplitude of predicted wind power fluctuations using air pressure change characteristics; extract statistical features of air pressure changes from radiosonde data and generate a modulation factor through a linear layer and activation function; The modulation factor is applied to the time difference between adjacent moments of the preliminary prediction result, adaptively scaling the variation of wind power, and finally obtaining the corrected prediction value.

2. The spatiotemporal fusion power output prediction method for wind power clusters based on branch networks and dynamic source selection according to claim 1, characterized in that, In Step 1, the radiosonde data and reanalysis data are upsampled to hourly resolution, and the wind power data obtained from the data acquisition and monitoring system are obtained by accumulating the values ​​within each hour to obtain hourly data.

3. The spatiotemporal fusion power output prediction method for wind power clusters based on branch networks and dynamic source selection according to claim 1, characterized in that, The implementation of Step 2 includes the following steps: Step 2.1: Capture the temporal characteristics of a single wind farm using a fully connected neural network; For each individual wind farm First, utilize its historical wind power data. To the wind farm Modeling is performed using time patterns, where The length of the time series is represented by a fully connected neural network (FCN) consisting of two linear layers with a ReLU activation function in between, as shown in the following formula: in, and These are learnable parameters. Represents the ReLU activation function; output Indicates the first Temporal characteristics of individual wind farms; By stacking the time-series characteristics of all wind farms, we obtain , The time-series characteristics of all wind farms; Step 2.2: Utilize multi-head attention mechanisms to capture spatial characteristics between multiple wind farms; Each wind farm Temporal characteristics As a query, the time-series characteristics of all wind farms Then it serves as both key and value; each attention head generates a query vector through an independent learnable transformation. Key vector Sum value vector Each attention head is based on the query vector. With key vector The similarity between them, for value vectors The formula for a single attention head is as follows, calculated using a weighted combination. Where Attention() is the attention head function; softmax() is the normalized exponential function; and T represents the transpose. Represents the dimension of the key vector. Representing the The outputs from each attention head are obtained, resulting in a total of M attention output heads. After concatenating all attention output heads and performing a linear transformation, the spatial characteristics of wind power output within the wind power cluster are obtained. The specific formula is as follows: in, It is a learnable parameter that serves to weight and fuse information from multiple attention heads. Step 2.3: Integrate the temporal characteristics of a single wind farm with the spatial characteristics of wind power output within a wind power cluster. ; Time series characteristics and spatial features The data is then spliced ​​together and fused using a fully connected neural network to obtain the spatiotemporal characteristics of the wind power-specific branch network output. ; in, This indicates a splicing operation along the first dimension; spatiotemporal features. It integrates temporal and spatial features.

4. The spatiotemporal fusion power output prediction method for wind power clusters based on branch network and dynamic source selection according to claim 1, characterized in that, The implementation of Step 3 includes the following steps: First, let's analyze the wind speed data. Through a convolution kernel size of The convolutional layer is followed by an activation function; this process is represented as: in Represents a convolution kernel, Represents the ReLU activation function; Subsequently, another convolutional layer with a kernel size of 1×1 is applied to reduce the feature dimensionality while preserving feature information, thereby obtaining shared features. ; in It contains D channels, where D represents the prediction step size of the target variable in the prediction task. Each channel corresponds to a shared feature map of size K×J, where K represents the number of sampling points obtained after dividing the target area into equal intervals in the longitude direction, and J represents the number of sampling points obtained after dividing the target area into equal intervals in the latitude direction. To adapt to changes in the region of interest under different prediction step sizes, a spatial attention mechanism is introduced; this is achieved through a convolutional layer... Generate a spatial attention map with D channels. : in, The function ensures that the sum of attention weights at each spatial location is 1; Next, shared features will be used. Spatial attention map Element-wise multiplication yields the enhanced features. The weights of each spatial location are dynamically adjusted through an attention mechanism. Subsequently, Apply adaptive pooling to reduce its spatial dimension from the original Uniform downsampling to Where B and F are configurable positive integer parameters that satisfy the condition , as well as Let K represent the number of target wind farms, J represent the number of sampling points obtained after equally dividing the target area along the longitude direction, and B and F represent the number of sampling points obtained after equally dividing the target area along the latitude direction. The parameters B and F are set according to actual application requirements to control the spatial range of the local receptive field or feature extraction region. After adaptive pooling, a tensor of shape (N, D, B, F) is obtained. , where N represents the number of samples and D corresponds to the prediction step size of the target variable in the prediction task; Finally, Flattened along the spatial dimension, the regional meteorological characteristics are reshaped into a reanalysis data-specific branch network output with a shape of (N,D,B×F). Thus achieving from A one-to-one mapping from the spatial grid to B×F target wind farms ensures that the feature structure is aligned with the output dimension of the wind farms, guaranteeing that the prediction at each time step can be based on a spatial context that matches the task requirements.

5. The spatiotemporal fusion power output prediction method for wind power clusters based on branch network and dynamic source selection according to claim 1, characterized in that, Step 4 adaptively fuses the spatiotemporal characteristics output by the wind power dedicated branch network with the regional meteorological characteristics from the reanalysis data dedicated branch network according to the prediction time step. The fusion weights of the two types of features are dynamically calculated for each prediction time step, and the contributions of historical power spatiotemporal features and regional meteorological background are adaptively balanced. Regional meteorological characteristics derived from reanalysis data branch networks Spatiotemporal characteristics of wind power dedicated branch network output This allows for the simultaneous utilization of climate background information and the spatiotemporal dynamic characteristics of wind power generation during the wind power cluster output prediction process; to align the feature dimensions, firstly... Perform the transpose operation to obtain ; The regional meteorological characteristics output by the reanalysis data branch network will be further analyzed. Spatiotemporal characteristics of wind power dedicated branch network output The data is concatenated along the feature dimension, then input into a fully connected neural network to obtain the fused prediction output. ; in, Each element in Indicates the first In the nth sample The wind farm in the first Predicted power output for one future time step; A dynamic source selection mechanism is introduced; a set of learnable parameters is generated by connecting a fully connected network to a sigmoid activation function. This is used to dynamically fuse the contributions of the reanalysis data branch and the wind power branch, thereby obtaining the fused output value. : This represents a fully connected neural network with an activation function added between two linear layers.

6. The spatiotemporal fusion power output prediction method for wind power clusters based on branch network and dynamic source selection according to claim 1, characterized in that, In Step 5, the predicted wind power fluctuation is adjusted by utilizing air pressure change characteristics in adjusting the prediction results based on sounding data. set up The air pressure represented by the radiosonde data is calculated based on the radiosonde data sequence to characterize the air pressure changes, including: average variation amplitude, volatility, maximum variation, and trend consistency; these statistical features are concatenated into a single representation. Then, a fully connected network is input, and a modulation factor is generated using the Tanh activation function. , For the sample size: The average variation and standard deviation represent the average variation. Indicates volatility. This indicates the largest change, representing the proportion of continuous changes in the same direction. Indicates trend consistency; tanh() represents the Tanh activation function. Finally, the modulation factor Applied to the fused output value The corrected predicted value is obtained. The details are as follows: in, The temporal differences of the predicted sequence between adjacent time steps were calculated; and Broadcasting over time to match The shape; The above formula represents the standard element-wise addition operation; it scales the temporal variation of the predicted wind power sequence based on the atmospheric conditions extracted from the radiosonde data, thereby achieving dynamic adjustment of the prediction results.

7. A spatiotemporal fusion power output prediction system for wind power clusters with branch network and dynamic source selection, characterized in that, Includes the following modules: Data collection module: used to collect data from different sources, including wind power output data from the data acquisition and monitoring system, as well as sounding data and wind speed data from reanalysis data; Wind Power Dedicated Branch Network Construction Module: A dedicated branch network—the wind power dedicated branch network—is built for wind power output data. This branch network extracts the historical power time-series features of each wind farm through independent fully connected neural networks. Based on the historical power time-series features of multiple wind farms with different geographical locations, a multi-head attention mechanism is used to generate spatial features representing wind power output in the wind power cluster. The time-series features and the spatial features are concatenated along the feature dimension and fused through a fully connected neural network to obtain the spatiotemporal features output by the wind power dedicated branch network, which are used for subsequent fusion and wind power prediction. Reanalysis Data Dedicated Branch Network Construction Module: A convolutional neural network is used to construct a dedicated branch network for reanalysis data, which extracts regional meteorological features that are helpful for wind power output prediction from the wind speed data of the reanalysis data; Fusion module: Adaptively fuses the spatiotemporal characteristics output by the wind power dedicated branch network with the regional meteorological characteristics from the reanalysis data dedicated branch network according to the prediction time step; Prediction Dynamic Adjustment Module: Adjusts prediction results using radiosonde data; adjusts the amplitude of predicted wind power fluctuations using air pressure change characteristics; extracts statistical features of air pressure changes from radiosonde data and generates a modulation factor through a linear layer and activation function; The modulation factor is applied to the time difference between adjacent moments of the preliminary prediction result, adaptively scaling the variation of wind power, and finally obtaining the corrected prediction value.

8. A computer device, characterized in that: The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the step of spatiotemporal fusion power output prediction of wind power clusters with branch network and dynamic source selection as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that: When the computer program / instruction is executed by the processor, it implements the steps of spatiotemporal fusion power output prediction of wind power clusters with branch network and dynamic source selection as described in any one of claims 1-6.

Images