A wind farm power prediction method and device based on space-time cooperation and a storage medium

Abstract

The application discloses a kind of wind farm power prediction method, device and storage medium based on space-time cooperation, belong to deep learning technical field, the method includes: obtaining the historical wind field data and geographical space feature of wind farm, wherein wind field data includes the same time period wind turbine operation data and meteorological characteristic data;The historical wind field data and geographical feature are combined with the corresponding wind farm total power label as sample set, and the sample set is used to train wind farm power prediction model;Wherein, wind farm power prediction model extracts local and overall double-layer similarity based on historical wind field data, according to double-layer similarity, from historical wind turbine operation data, the time period clustering cluster with high similarity is obtained with meteorological characteristic data, and spatial clustering is carried out to time period clustering cluster based on historical geographical feature;The wind field data to be measured and the geographical feature to be measured are input into the trained wind farm power prediction model, and the wind farm total power prediction value is obtained, improve the complex terrain wind power prediction precision and efficiency.

Description

Technical Field

[0001] This invention belongs to the field of deep learning technology, and in particular relates to a method, device and storage medium for predicting wind farm power based on spatiotemporal coordination. Background Technology

[0002] Wind power forecasting has become a research area of ​​great interest in recent years, especially when dealing with wind power in complex terrain, where efficient and accurate forecasting methods are urgently needed. As the global energy structure accelerates its transition towards a low-carbon model, wind energy, as a clean, renewable energy source with large-scale development potential, is gradually becoming an important component of modern power systems. The integration of wind power enhances the sustainability of power systems, but its intermittency and volatility also pose significant challenges to system operation and dispatch, particularly in areas with abundant wind resources but complex terrain.

[0003] To date, numerous studies have attempted to address wind power forecasting from the perspectives of physical modeling, statistical modeling, and machine learning. First, physical modeling methods typically rely on numerical weather prediction (NWMR) and fluid dynamics equations to simulate the evolution of wind fields. While these methods can provide reasonable predictions on a large scale, their accuracy is significantly insufficient on small scales in complex terrain due to inadequate terrain characterization and difficulty in modeling local disturbances. Furthermore, NWMR has limited temporal resolution and high computational cost, making it unsuitable for direct use in ultra-short-term forecasting. Second, statistical modeling methods fit the time dependencies of time series data using historical data. These methods are computationally efficient but generally rely on the assumption of data stationarity, lacking the ability to model complex couplings of nonlinearity and multiple variables. Therefore, their effectiveness is limited when dealing with highly nonlinear time-series data such as wind power.

[0004] In recent years, with the development of data-driven methods, machine learning and deep learning methods have been widely used in wind power forecasting. Traditional machine learning methods, such as Support Vector Machines (SVM) and Random Forests (RF), can improve prediction accuracy to some extent, but they have limitations in handling long-sequence dependencies and high-dimensional spatiotemporal features. Deep learning methods, such as Long Short-Term Memory Networks (LSTM) and Gated Recurrent Units (GRU), have advantages in modeling temporal dependencies; the CNN-LSTM model, which combines Convolutional Neural Networks (CNN) and Recurrent Networks (RNN), can simultaneously extract local patterns and global dependencies; and the Transformer architecture, which has been widely used in recent years, significantly enhances the ability to model long dependencies through its self-attention mechanism. These methods can usually achieve good prediction accuracy in plains or low-to-medium complexity terrains. However, when applied to complex terrain scenarios, these methods still have significant shortcomings. On the one hand, there is a spatial scale mismatch between single-turbine observations and NWP data. The meteorological features provided by NWP represent regional averages, while single-turbine observations contain a large amount of local noise and non-stationarity. This scale difference often leads to prediction bias. On the other hand, wind fields in complex terrain exhibit significant spatial dependence and directionality (e.g., the wake effect of upwind turbines on downwind turbines). Traditional deep learning models mostly rely solely on time-series modeling, neglecting the spatial interaction between wind turbines. Furthermore, wind power sequences often simultaneously contain long-period patterns such as daily and semi-daily cycles, as well as rapid abrupt changes on the minute scale. Single-time-scale modeling methods struggle to accommodate both of these characteristics. Summary of the Invention

[0005] The purpose of this invention is to provide a method, device, and storage medium for wind farm power prediction based on spatiotemporal coordination. By performing local + global two-layer similarity modeling on wind turbine operation data and meteorological field data, the temporal characteristic data of wind farms is clustered in time series. Then, the spatial clustering effect is superimposed on the temporal clustering results according to the geographical characteristics of wind turbines, providing spatiotemporally linked cluster-level features and improving the accuracy of subsequent cluster-level temporal coding and inter-cluster relationship modeling.

[0006] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0007] In a first aspect, the present invention provides a wind farm power prediction method based on spatiotemporal coordination, comprising:

[0008] Acquire wind farm data and geographical features to be measured for a wind farm, wherein the wind farm data includes wind turbine operation data and meteorological feature data for the corresponding time period;

[0009] The wind farm data and geographical features to be measured are input into a pre-trained wind farm power prediction model to predict the total power of the wind farm.

[0010] The training methods for wind farm power prediction models include:

[0011] Acquire historical wind field data and historical geospatial characteristics of wind farms;

[0012] By combining historical wind farm data and historical geographical features with the corresponding total power labels of wind farms, a sample set is obtained;

[0013] The sample set is input into the pre-built wind farm power prediction model to obtain the trained wind farm power prediction model. The wind farm power prediction model extracts two-layer similarity at the local change level and the overall trend level based on historical wind farm data. Based on the two-layer similarity, it obtains time-period clusters with high similarity to historical meteorological feature data from historical wind turbine operation data, and performs spatial clustering of time-period clusters based on historical geographical features.

[0014] Optionally, acquiring historical wind farm data and historical geospatial characteristics of the wind farm includes:

[0015] Historical operating data of each wind turbine is collected from the wind farm monitoring system, and meteorological field data at the same time step is exported from the meteorological system; among them, the wind turbine operating data includes wind turbine power, wind speed and wind direction;

[0016] Historical operational data and weather forecast wind speeds are synchronized and cleaned in time, and then normalized to construct a unified time-series input matrix.

[0017] Optionally, the structure of the wind farm power prediction model includes a subsequence hierarchical clustering module, a cluster-level time encoder module, an inter-cluster interaction graph network module, and a power prediction module connected in series.

[0018] The subsequence hierarchical clustering module includes a subsequence extraction unit, a two-layer similarity modeling unit, a subsequence hierarchical clustering unit, a geospatial clustering unit, and a cluster-level signal generation unit connected in series.

[0019] The cluster-level temporal encoder module includes an input embedding layer, a multi-scale downsampling layer, and a pyramid attention layer connected in series.

[0020] The inter-cluster interaction graph network module includes a series of dynamic graph construction units and a graph attention propagation unit;

[0021] The power prediction module includes a power regression unit and a power aggregation unit connected in series.

[0022] Optionally, the step of inputting the wind farm data to be measured and the geographical features to be measured into a pre-trained wind farm power prediction model to predict the total power of the wind farm includes:

[0023] The subsequence hierarchical clustering module performs temporal clustering on the input wind field data to be measured, performs spatial clustering on the temporal clustering results based on the geographic features to be measured, and obtains the cluster-level time series based on the spatiotemporal clustering results.

[0024] Cluster-level time features are obtained by performing time-series encoding based on cluster-level power features using a cluster-level time encoder module.

[0025] The cluster-level features, weighted by cross-cluster attention, are obtained through the inter-cluster interaction graph network module based on the cluster-level temporal features.

[0026] The power prediction module maps the power values ​​of the cluster-level features after cross-cluster attention weighting to obtain the predicted power values ​​of the wind farm.

[0027] Optionally, the step of performing temporal clustering on the input wind field data to be measured through the subsequence hierarchical clustering module, performing spatial clustering on the temporal clustering results based on the geographic features to be measured, and obtaining cluster-level time series based on the spatiotemporal clustering results includes:

[0028] The input wind field data to be measured is divided into time sequences by the subsequence extraction unit to obtain multiple overlapping subsequences;

[0029] The two-level similarity modeling unit performs local and global two-level similarity calculations based on overlapping subsequences;

[0030] Temporal clusters are obtained by performing temporal clustering on each overlapping subsequence using a subsequence hierarchical clustering unit based on two-level similarity.

[0031] By spatially clustering temporal clusters based on the geographic features to be measured using geospatial clustering units, a spatiotemporal joint clustering sequence is obtained;

[0032] Cluster-level power characteristics are calculated based on spatiotemporal joint clustering sequences by a cluster-level signal generation unit, and cluster-level power characteristics are combined with corresponding cluster-level wind speed and direction characteristics to obtain cluster-level time series.

[0033] Optionally, the step of performing time-series encoding based on cluster-level power features using a cluster-level time-series encoder module to obtain cluster-level time-series features includes:

[0034] The cluster-level time series is mapped to a high-dimensional space by an input embedding layer, and sinusoidal positional encoding is added to obtain the feature tensor.

[0035] A temporal pyramid structure is generated by downsampling the feature tensor layer by layer through a multi-scale downsampling layer.

[0036] Cluster-level temporal features are obtained by applying adjacent-scale attention weights to the temporal pyramid results through a pyramid attention layer.

[0037] Optionally, the step of obtaining cross-cluster attention-weighted cluster features based on cluster-level temporal features through the inter-cluster interaction graph network module includes:

[0038] A graph structure is constructed based on cluster-level time series using dynamic graph construction units;

[0039] The graph attention propagation unit calculates cross-cluster attention weights based on the graph structure and cluster-level temporal features, and generates cluster-level features after cross-cluster attention weighting.

[0040] Optionally, the step of mapping the power values ​​of the cluster-level features after cross-cluster attention weighting through the power prediction module to obtain the wind farm power prediction value includes:

[0041] The power prediction matrix of a single cluster feature is obtained by performing power prediction on the cluster-level features after cross-cluster attention weighting through the power regression unit.

[0042] The total power prediction value of the wind farm is output by weighted averaging of the prediction matrices of all cluster features by the power aggregation unit.

[0043] Secondly, the present invention provides a wind farm power prediction device based on spatiotemporal coordination, comprising:

[0044] The wind farm data acquisition module is used to acquire wind farm data and geographical features to be measured, wherein the wind farm data includes wind turbine operation data and meteorological feature data for the corresponding time period.

[0045] Wind farm total power prediction module: This module is used to input the wind farm data and geographical features to be measured into a pre-trained wind farm power prediction model to predict the total power of the wind farm.

[0046] The training methods for wind farm power prediction models include:

[0047] Acquire historical wind field data and historical geospatial characteristics of wind farms;

[0048] By combining historical wind farm data and historical geographical features with the corresponding total power labels of wind farms, a sample set is obtained;

[0049] The sample set is input into the pre-built wind farm power prediction model to obtain the trained wind farm power prediction model. The wind farm power prediction model extracts two-layer similarity at the local change level and the overall trend level based on historical wind farm data. Based on the two-layer similarity, it obtains time-period clusters with high similarity to historical meteorological feature data from historical wind turbine operation data, and performs spatial clustering of time-period clusters based on historical geographical features.

[0050] Thirdly, the present invention provides a computer storage medium having a computer program stored thereon, which, when executed by a processor, implements the spatiotemporal collaborative wind farm power prediction method as described in any of the first aspects.

[0051] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: By simultaneously performing similarity measurement on wind farm characteristic data at both the local dynamic change level and the overall trend level, it can capture the overall evolution trend of wind power sequences and identify key dynamic features such as short-term fluctuations and local disturbances. This enables refined matching in complex non-stationary time series. Based on the two-layer similarity, the wind farm characteristic data is clustered in time series. At the same time, the spatial distribution characteristics of the wind farm under complex terrain are combined to perform spatial clustering on the time series clustering results, resulting in spatiotemporal joint clustering features. This provides basic data that can truly reflect the flow field structure, meteorological distribution law, and dynamic wind field propagation law in the region for subsequent cluster-level time series coding and inter-cluster interaction modeling, greatly improving the prediction accuracy of the model under complex wind direction changes and local disturbance conditions. Attached Figure Description

[0052] Figure 1 The diagram shown is an architecture diagram of a wind farm power prediction model in one embodiment of the present invention;

[0053] Figure 2 The diagram shown is a flowchart of subsequence hierarchical clustering in one embodiment of the present invention;

[0054] Figure 3 The diagram shown is an architecture diagram of a cluster-level timing encoder module in one embodiment of the present invention;

[0055] Figure 4 The diagram shown is an inter-cluster interaction graph network module architecture diagram in one embodiment of the present invention. Detailed Implementation

[0056] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0057] Example 1

[0058] This embodiment provides a wind farm power prediction method based on spatiotemporal coordination, which specifically includes the following steps:

[0059] Step S1: Data Input and Standardization Preprocessing. First, historical operating data for each wind turbine is collected from the wind farm monitoring system, including the turbine number. timestamp Measured power Measured wind speed and wind direction Simultaneously, the NWP numerical weather forecast wind speed at the same time step was exported from the meteorological system. Subsequently, time synchronization and cleaning were performed on the multi-source data, including outlier detection, linear interpolation to repair missing values, and normalization processing to construct a unified time-series input matrix. :

[0060] ,

[0061] The standardized time-series data stream provides a unified input for subsequent spatial clustering and cluster-level modeling.

[0062] Step S2: Construct a wind farm power prediction model, such as Figure 1 As shown, it includes a subsequence hierarchical clustering module, a cluster-level time-series encoder module, an inter-cluster interaction graph network module, and a power prediction module connected in sequence.

[0063] 1. The subsequence hierarchical clustering module includes the following structure and functions:

[0064] like Figure 2 As shown, the subsequence hierarchical clustering module first collects multi-source data from the wind farm, including wind turbine power, wind speed, wind direction, and geographical information (latitude, longitude, altitude, slope, etc.). It then performs sliding window segmentation on the time series, extracting several overlapping subsequence segments and simultaneously measuring similarity at both the local dynamic change and overall trend levels. By introducing a hierarchical clustering method, this module achieves dual-layer similarity modeling of global trends and local patterns. This captures both the overall evolution trend of the wind power sequence and identifies key dynamic features such as short-term fluctuations and local disturbances, thus achieving refined matching in complex non-stationary time series. Simultaneously, considering the spatial distribution characteristics of the wind farm under complex terrain, the module further utilizes multi-source features such as latitude, longitude, altitude, terrain slope, wind speed, and historical power to perform geospatial clustering, dividing wind turbines with similar spatial locations and wind conditions into several geographical clusters. The wind turbines within each cluster are influenced by the same terrain flow field, exhibiting similar turbulent structures and aerodynamic response characteristics. This process effectively reduces individual noise disturbances between wind turbines in space, enhances the statistical stability of regional modeling, and provides physically consistent spatial structural constraints for subsequent cluster-level temporal coding and inter-cluster relationship modeling.

[0065] (1) Sliding window construction and time series subsequence extraction. In order to capture the local dynamic features and short-term disturbance patterns in the wind power series, the sliding window length is set to be... The sliding step size is This divides a continuous time series into multiple overlapping subsequences. Then, for each wind turbine... This yields a set of time segments:

[0066] ,

[0067] in, For wind turbine The A time subsequence, For wind turbine In time Power value at time, This represents the total number of subsequences obtained by the sliding window. , This represents the total length of the time series. This subsequence extraction mechanism maintains temporal continuity while covering the feature space across the entire time period, enabling the model to identify both global trends and extract local variation patterns during subsequent clustering.

[0068] (2) Two-layer similarity modeling and feature encoding. To address the non-stationary characteristics and multi-timescale fluctuations of wind power sequences, a two-layer similarity modeling strategy combining global trend and local patterns is adopted. First, at the global level, cross-correlation coefficients are used to measure the similarity of the overall trend, reflecting the macroscopic consistency of wind turbine operating power evolution over time. Next, at the local level, local dynamic feature similarity is calculated based on the time-frequency domain features of subsequences to identify short-term abrupt changes, local turbulence effects, or wind speed jumps. Finally, the two-layer similarity is fused to obtain a comprehensive distance matrix. :

[0069] ,

[0070] in, These are weighting coefficients that adjust the balance between global and local similarity. Similarity at the global trend level. This represents the similarity at the local dynamic level.

[0071] (3) Hierarchical clustering of subsequences and selection of similar time periods. After obtaining the comprehensive distance matrix, a hierarchical clustering algorithm is used to recursively aggregate the subsequences. First, based on the geographic coordinates, the most similar subsequences are gradually merged according to the minimum distance criterion to form a tree-like hierarchical structure. Then, the target prediction time period is selected according to the meteorological feature data. Based on the statistical characteristics and power distribution characteristics of the target prediction time period, a similarity threshold is selected on the clustering tree. The corresponding hierarchical structure is extracted to obtain several clusters that are closest to the target prediction period. The sample sequences within each cluster are highly similar in terms of temporal pattern and meteorological characteristics, thus forming the optimal set of similar time periods for the target prediction period. This set will serve as the input of training samples for subsequent time series modeling.

[0072] (4) Geospatial Clustering and Physical Structure Integration. Based on temporal hierarchical clustering, a geospatial joint clustering mechanism is introduced to further consider the spatial dependence characteristics of wind turbines under complex terrain. First, based on the longitude of the wind turbine... ,latitude ,altitude ,slope Slope aspect With terrain roughness Constructing spatial feature vectors from geographical features:

[0073] ,

[0074] Then, the DBSCAN algorithm is used to cluster wind turbines based on spatial proximity and wind condition similarity, grouping those located in the same flow field region and with similar terrain conditions into the same cluster. This spatial aggregation not only reduces random disturbances at the single-unit level but also ensures the consistency of temporal clustering results across geospatial dimensions. Finally, the subsequence temporal clustering results are combined with the geospatial clustering results to form a spatiotemporal joint two-layer clustering structure, providing a basic data framework for subsequent cluster-level temporal coding and inter-cluster interaction modeling.

[0075] (5) Generate cluster-level signals. The cluster-level power is obtained by weighted averaging of the power of the fans within the cluster. : ,in, For the first The set of wind turbines contained in each cluster. For wind turbine In time The actual power value, Weights are based on wind turbine capacity or credibility.

[0076] (6) Data transmission. Cluster-level power signal The corresponding cluster-level wind speed and direction features are packaged into a feature vector stream. The signal is then transmitted to the input of the cluster-level timing encoder module. This achieves the conversion of the spatial scale from single-machine to cluster-level, improves signal stability, reduces noise, and provides a stable aggregated input for timing modeling.

[0077] 2. The Cluster-Level Temporal Encoder (CLTE) module includes the following structure and functions:

[0078] After completing spatial clustering and time period selection, the cluster-level time-series encoder module uses the geographical clusters obtained in step 1 as input units to jointly model the time series of wind turbines within each cluster. For example... Figure 3As shown, the cluster-level temporal encoder module first performs statistical aggregation or weighted fusion of the power sequences of wind turbines within each cluster to form a representative cluster-level temporal input, thereby reducing local noise and enhancing the expression of regional features. Subsequently, the encoder introduces a multi-scale downsampling and pyramid attention mechanism module to extract information from different time scales, preserving both long-term evolutionary trends and capturing short-term abrupt dynamic features, achieving multi-scale dependent joint modeling. This increases the model's multi-scale modeling capability and significantly enhances the stability and noise resistance of temporal domain modeling.

[0079] (1) Input embedding layer. Receives cluster-level time series. It maps to a high-dimensional space through linear embedding and adds sinusoidal positional encoding to output a feature tensor. : ,

[0080] in, , Represents a linear mapping. This represents the sine wave position code.

[0081] (2) Multi-scale downsampling. For feature tensors Layer-by-layer downsampling is used to form sequence representations with different time resolutions: ,

[0082] Forming a time pyramid structure.

[0083] (3) Pyramid attention mechanism. Sparse connections are established between adjacent scales of the time pyramid structure to compute the attention between local neighborhoods and parent-child layers: in, As a mask matrix, only local and cross-layer interactions are allowed. For query vector, For key vectors, For value vectors, for Dimensions for The transpose of .

[0084] (4) Feature Transmission. Output cluster-level temporal feature representations and transmit these representations to the inter-cluster interaction graph network module. The expression is as follows: ,

[0085] in, For clusters Time series encoding input, This represents a cluster-level temporal encoder used to extract intra-cluster temporal dependent features.

[0086] This module extracts time-dependent features through multi-scale modeling, enabling the output features to have both short-term abrupt responses and long-term trends, thereby improving the accuracy of time prediction.

[0087] 3. The Cluster Interaction Graph Network (CIGN) module includes the following structure and functions:

[0088] like Figure 4 As shown, based on cluster-level temporal coding results, the inter-cluster interaction graph network module achieves spatial dynamic dependence learning of wind power under complex terrain by explicitly modeling the spatial relationships between clusters. First, a dynamic adjacency matrix with directionality and time-varying weights is constructed based on the geographical location, prevailing wind direction, spatial distance, and terrain distribution characteristics of the wind turbine clusters to characterize the interaction intensity between different regions in the wind field. Then, a graph attention network is introduced to achieve adaptive propagation and weighted fusion of inter-cluster information, enabling the model to dynamically correct wind energy transfer relationships between different directions and regions, compensating for the systematic bias of local numerical weather prediction under complex terrain. Through graph structure propagation in this module, the model ultimately obtains a cluster-level feature representation that simultaneously integrates temporal dependence and spatial coupling characteristics. This representation is then input into the prediction layer to output short-term wind power prediction results, achieving global spatial consistency and high-precision dynamic wind power prediction. This module adds fully connected layers and pooling layers to the graph neural network structure to calculate the gross power of each cluster, while simultaneously calculating the error between the gross power and the predicted power, and then aggregating and predicting the total power of the wind farm.

[0089] (1) Dynamic graph construction. A time-varying adjacency matrix is ​​generated based on geographical location and wind direction: , in, , respectively wind turbine With wind turbine , For a moment Downdraft fan and Adjacency weights between them For wind turbine Geographical coordinates For wind turbine Geographical coordinates This is the distance decay coefficient, which controls the impact of geographical distance on the weights. This is a wind direction gating function.

[0090] (2) Graph attention propagation. Cluster-level timing features received from the CLTE module. The graph attention layer calculates cross-cluster information propagation and outputs a cluster-level representation dynamically corrected by graph attention, resulting in graph attention-weighted cluster-level features. The expression is as follows: ,in, For graph attention weights, The characteristic transformation matrix, For nodes Input features, For attention weight vectors, The input features are any node in the set of neighboring nodes.

[0091] (3) Data transmission. The corrected cluster-level features will be transmitted... Transmitted to the power prediction module.

[0092] 4. The structure and function of the power prediction module are as follows:

[0093] (1) The power regression unit receives the cluster-level feature vector output by the CIGN module. Output the future through a fully connected layer Step power prediction sequence: ,

[0094] in, Indicates the future The power prediction sequence of the step.

[0095] (2) The power aggregation unit aggregates the total power of the wind farm. It receives the predicted outputs from all clusters of the power regression unit. The total power output of the wind farm is predicted by calculating a weighted average based on the cluster capacity or the number of wind turbines. : ,

[0096] in, The index is the cluster number. For cluster capacity, For the first The weights of each cluster.

[0097] Step S3: Train the wind farm power prediction model.

[0098] The preprocessed historical data obtained in step S1 is combined with the known total power labels of wind farms to obtain sample sets. The sample sets are then divided into a 7:2:1 dataset, which are used as training, validation, and test sets, respectively, and training parameters are set.

[0099] The training set data is input into the wind farm power prediction model. The model is trained using cross-entropy loss and Adam optimizer. After multiple rounds of training, the model performance is calculated using the validation set data, and the model hyperparameters are adjusted. If the performance does not improve after multiple consecutive validations, the iteration is stopped. Finally, the final performance index of the model is calculated using the test set data.

[0100] Step S4: Application of wind farm power prediction model.

[0101] Based on the historical data collection steps, the operating data of each wind turbine in the wind farm under test is collected, and the meteorological field data of the corresponding time period is exported. The geospatial characteristics of each wind turbine in the wind farm under test are obtained. After processing according to the historical data preprocessing steps, the data is input into the trained wind farm power prediction model to obtain the predicted total power of the wind farm, thus achieving an accurate estimation of the wind farm power value.

[0102] Example 2

[0103] This embodiment provides a wind farm power prediction device based on spatiotemporal coordination, including:

[0104] The wind farm data acquisition module is used to acquire wind farm data and geographical features to be measured, wherein the wind farm data includes wind turbine operation data and meteorological feature data for the corresponding time period.

[0105] Wind farm total power prediction module: This module is used to input the wind farm data and geographical features to be measured into a pre-trained wind farm power prediction model to predict the total power of the wind farm.

[0106] The training methods for wind farm power prediction models include:

[0107] Acquire historical wind field data and historical geospatial characteristics of wind farms;

[0108] By combining historical wind farm data and historical geographical features with the corresponding total power labels of wind farms, a sample set is obtained;

[0109] The sample set is input into the pre-built wind farm power prediction model to obtain the trained wind farm power prediction model. The wind farm power prediction model extracts two-layer similarity at the local change level and the overall trend level based on historical wind farm data. Based on the two-layer similarity, it obtains time-period clusters with high similarity to historical meteorological feature data from historical wind turbine operation data, and performs spatial clustering of time-period clusters based on historical geographical features.

[0110] Example 3

[0111] This embodiment provides a computer storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements the spatiotemporal collaborative wind farm power prediction method as described in any step of Embodiment 1.

[0112] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0113] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0114] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0115] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0116] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A wind farm power prediction method based on spatiotemporal coordination, characterized in that, include: Acquire wind farm data and geographical features to be measured for a wind farm, wherein the wind farm data includes wind turbine operation data and meteorological feature data for the corresponding time period; The wind farm data and geographical features to be measured are input into a pre-trained wind farm power prediction model to predict the total power of the wind farm. The training methods for wind farm power prediction models include: Obtain historical wind field data and historical geographical characteristics of wind farms; By combining historical wind farm data and historical geographical features with the corresponding total power labels of wind farms, a sample set is obtained; The sample set is input into the pre-built wind farm power prediction model to obtain the trained wind farm power prediction model. The wind farm power prediction model extracts two-layer similarity at the local change level and the overall trend level based on historical wind farm data. It performs time-series clustering based on the two-layer similarity to obtain time-period clusters, and performs spatial clustering of the time-period clusters based on historical geographical features. The structure of the wind farm power prediction model includes a subsequence hierarchical clustering module, a cluster-level time encoder module, an inter-cluster interaction graph network module, and a power prediction module connected in series. The subsequence hierarchical clustering module includes a subsequence extraction unit, a two-layer similarity modeling unit, a subsequence hierarchical clustering unit, a geospatial clustering unit, and a cluster-level signal generation unit connected in series. The cluster-level temporal encoder module includes an input embedding layer, a multi-scale downsampling layer, and a pyramid attention layer connected in series. The inter-cluster interaction graph network module includes a series of dynamic graph construction units and a graph attention propagation unit; The power prediction module includes a power regression unit and a power aggregation unit connected in series. The step of inputting the wind farm data to be measured and the geographical features to be measured into a pre-trained wind farm power prediction model to predict the total power of the wind farm includes: The subsequence hierarchical clustering module performs temporal clustering on the input wind field data to be measured, performs spatial clustering on the temporal clustering results based on the geographic features to be measured, and obtains the cluster-level time series based on the spatial clustering results. Cluster-level time features are obtained by performing time-series encoding based on cluster-level power features using a cluster-level time encoder module. The cluster-level features, weighted by cross-cluster attention, are obtained through the inter-cluster interaction graph network module based on the cluster-level temporal features. The power prediction module maps the power values ​​of the cluster-level features after cross-cluster attention weighting to obtain the predicted power values ​​of the wind farm.

2. The wind farm power prediction method based on spatiotemporal coordination according to claim 1, characterized in that, The acquisition of historical wind farm data and historical geographical features of the wind farm includes: Historical operating data of each wind turbine is collected from the wind farm monitoring system, and meteorological field data at the same time step is exported from the meteorological system; among them, the wind turbine operating data includes wind turbine power, wind speed and wind direction; Historical operational data and weather forecast wind speeds are synchronized and cleaned in time, and then normalized to construct a unified time-series input matrix.

3. The wind farm power prediction method based on spatiotemporal coordination according to claim 2, characterized in that, The process involves performing temporal clustering on the input wind field data using a sub-sequence hierarchical clustering module, spatial clustering of the temporal clustering results based on the geographic features to be measured, and obtaining cluster-level time series data based on the spatial clustering results. This includes: The input wind field data to be measured is divided into time sequences by the subsequence extraction unit to obtain multiple overlapping subsequences; The two-level similarity modeling unit performs local and global two-level similarity calculations based on overlapping subsequences; Temporal clusters are obtained by performing temporal clustering on each overlapping subsequence using a subsequence hierarchical clustering unit based on two-level similarity. By spatially clustering temporal clusters based on the geographic features to be measured using geospatial clustering units, a spatiotemporal joint clustering sequence is obtained; Cluster-level power characteristics are calculated based on spatiotemporal joint clustering sequences by a cluster-level signal generation unit, and cluster-level power characteristics are combined with corresponding cluster-level wind speed and direction characteristics to obtain cluster-level time series.

4. The wind farm power prediction method based on spatiotemporal coordination according to claim 2, characterized in that, The process of performing time-series encoding based on cluster-level power characteristics using a cluster-level time-series encoder module to obtain cluster-level time-series characteristics includes: The cluster-level time series is mapped to a high-dimensional space by an input embedding layer, and sinusoidal positional encoding is added to obtain the feature tensor. A temporal pyramid structure is generated by downsampling the feature tensor layer by layer through a multi-scale downsampling layer. Cluster-level temporal features are obtained by applying adjacent-scale attention weights to the temporal pyramid structure through a pyramid attention layer.

5. The wind farm power prediction method based on spatiotemporal coordination according to claim 2, characterized in that, The process of obtaining cross-cluster attention-weighted cluster features based on cluster-level temporal features through the inter-cluster interaction graph network module includes: A graph structure is constructed based on cluster-level time series using dynamic graph construction units; The graph attention propagation unit calculates cross-cluster attention weights based on the graph structure and cluster-level temporal features, and generates cluster-level features after cross-cluster attention weighting.

6. The wind farm power prediction method based on spatiotemporal coordination according to claim 2, characterized in that, The step of mapping power values ​​of cluster-level features after cross-cluster attention weighting through the power prediction module to obtain wind farm power prediction values ​​includes: The power prediction matrix of a single cluster feature is obtained by performing power prediction on the cluster-level features after cross-cluster attention weighting through the power regression unit. The total power prediction value of the wind farm is output by weighted averaging of the prediction matrices of all cluster features by the power aggregation unit.

7. A wind farm power prediction device based on spatiotemporal coordination, characterized in that, include: The wind farm data acquisition module is used to acquire wind farm data and geographical features to be measured, wherein the wind farm data includes wind turbine operation data and meteorological feature data for the corresponding time period. Wind farm total power prediction module: This module is used to input the wind farm data and geographical features to be measured into a pre-trained wind farm power prediction model to predict the total power of the wind farm. The training methods for wind farm power prediction models include: Obtain historical wind field data and historical geographical characteristics of wind farms; By combining historical wind farm data and historical geographical features with the corresponding total power labels of wind farms, a sample set is obtained; The sample set is input into the pre-built wind farm power prediction model to obtain the trained wind farm power prediction model. The wind farm power prediction model extracts two-layer similarity at the local change level and the overall trend level based on historical wind farm data. It performs time-series clustering based on the two-layer similarity to obtain time-period clusters, and performs spatial clustering of the time-period clusters based on historical geographical features. The structure of the wind farm power prediction model includes a subsequence hierarchical clustering module, a cluster-level time encoder module, an inter-cluster interaction graph network module, and a power prediction module connected in series. The subsequence hierarchical clustering module includes a subsequence extraction unit, a two-layer similarity modeling unit, a subsequence hierarchical clustering unit, a geospatial clustering unit, and a cluster-level signal generation unit connected in series. The cluster-level temporal encoder module includes an input embedding layer, a multi-scale downsampling layer, and a pyramid attention layer connected in series. The inter-cluster interaction graph network module includes a series of dynamic graph construction units and a graph attention propagation unit; The power prediction module includes a power regression unit and a power aggregation unit connected in series. The step of inputting the wind farm data to be measured and the geographical features to be measured into a pre-trained wind farm power prediction model to predict the total power of the wind farm includes: The subsequence hierarchical clustering module performs temporal clustering on the input wind field data to be measured, performs spatial clustering on the temporal clustering results based on the geographic features to be measured, and obtains the cluster-level time series based on the spatial clustering results. Cluster-level time features are obtained by performing time-series encoding based on cluster-level power features using a cluster-level time encoder module. The cluster-level features, weighted by cross-cluster attention, are obtained through the inter-cluster interaction graph network module based on the cluster-level temporal features. The power prediction module maps the power values ​​of the cluster-level features after cross-cluster attention weighting to obtain the predicted power values ​​of the wind farm.

8. A computer storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the spatiotemporal collaborative wind farm power prediction method as described in any one of claims 1-6.

Images