Method for predicting generating capacity of wind power plant

Abstract

The invention discloses a method for predicting the generating capacity of a wind power plant, and the method specifically comprises the following steps: 1, carrying out the collection and preprocessing of multi-modal data, and synchronously obtaining the numerical weather forecast data of the wind power plant. NWP weather, SCADA operation state and geographic information multi-modal data are synchronously collected, feature weights are dynamically distributed by adopting a modal self-adaptive attention mechanism after preprocessing, the problems of single-modal information short plates and low efficiency of simple splicing are avoided, and the problem of low efficiency of the single-modal information short plates is solved; the dependence on accurate physical parameters of the wind field is reduced; meanwhile, a dynamic graph structure is constructed based on fan coordinates and real-time wind directions, dynamic space-time coupling characteristics among fan clusters are mined in combination with a space-time graph neural network, the dynamic environment of a wind field is effectively adapted, and the defect that a traditional method ignores space correlation or adopts a fixed space structure is overcome.

Description

Technical Field

[0001] This invention relates to the field of wind power generation prediction technology, specifically a method for predicting wind power generation. Background Technology

[0002] As an important component of clean and renewable energy, wind power has seen continuous growth in installed capacity globally. However, the inherent intermittent, volatile, and random characteristics of wind energy make its output power highly unstable. This poses significant challenges to the stable operation of the power grid, power dispatch, and electricity market transactions. Therefore, achieving high-precision, early-lead-time wind farm power generation forecasting is crucial for ensuring grid security, improving wind energy absorption capacity, and reducing operating costs.

[0003] Existing methods for predicting wind power generation can be mainly categorized into physical modeling methods, traditional statistical and time series methods, and shallow machine learning methods. However, in existing technologies, physical modeling methods rely on precise physical parameters of the wind farm, such as terrain and wind turbine aerodynamic characteristics, making modeling complex and sensitive to parameter errors. Traditional statistical and time series methods, such as ARIMA, struggle to handle the coupling relationships of multimodal data. Shallow machine learning methods, such as traditional neural networks and support vector machines, lack the ability to deeply mine spatiotemporal correlation features. A common drawback is that they either only utilize single-modal data, or simply splice multimodal data without achieving deep fusion, or ignore spatial correlation, or adopt a fixed spatial structure that cannot adapt to the dynamic environment of the wind farm, resulting in limited prediction accuracy, especially with large errors in long lead-time prediction scenarios. Therefore, this invention provides a method for predicting wind power generation. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for predicting the power generation of wind farms. It solves the problems of existing technologies, such as physical modeling methods relying on precise physical parameters of the wind farm, such as terrain and wind turbine aerodynamic characteristics, which are complex and sensitive to parameter errors; traditional statistical and time series methods, such as ARIMA, struggling to handle the coupling relationships of multimodal data; and shallow machine learning methods, such as traditional neural networks and support vector machines, lacking the ability to deeply mine spatiotemporal correlation features. These methods share common drawbacks: they either utilize only single-modal data, simply splice multimodal data without achieving deep fusion, ignore spatial correlations, or adopt fixed spatial structures that cannot adapt to the dynamic environment of the wind farm, resulting in limited prediction accuracy, especially with large errors in long-lead-time prediction scenarios.

[0005] To achieve the above objectives, the present invention provides a method for predicting the power generation of a wind power plant, characterized by the following steps:

[0006] Step 1: Multimodal data acquisition and preprocessing. Simultaneously acquire numerical weather prediction (NWP) data, wind turbine operation status monitoring (SCADA) data, and geographic information data of the wind farm. Perform outlier removal, missing value completion, and standardization on various types of data to obtain a standardized multimodal dataset.

[0007] Step 2: Multimodal data feature hierarchical fusion. Construct a feature fusion module. First, extract meteorological features from the standardized NWP data, extract operational status features from the SCADA data, and extract spatial topology features from the geographic information data. Then, use an attention mechanism to assign dynamic weights to different modal features to generate initial fused features.

[0008] Step 3: Enhance the spatiotemporal correlation features. Construct a dynamic graph structure based on the geographical coordinates of the wind turbines. Adjust the connection relationship and weight of the wind turbine nodes in the graph according to real-time wind direction data to form a spatiotemporal correlation graph that changes with wind direction. Input the initial fused features into the spatiotemporal graph neural network to mine the dynamic spatiotemporal coupling features between wind turbine clusters.

[0009] Step 4: Power generation prediction model construction and training. Using spatiotemporal coupling features as input, a prediction model based on the Transformer-graph neural network hybrid architecture is constructed. The model is trained using historical power generation data, and the model parameters are adjusted through an adaptive learning rate optimization algorithm to obtain the trained prediction model.

[0010] Step 5: Power generation prediction output. The preprocessed multimodal data for the period to be predicted is input into the trained prediction model after feature fusion and spatiotemporal enhancement, and the power generation prediction results of the wind farm are output.

[0011] Preferably, the NWP data in step one includes spatiotemporal series data of wind speed, wind direction, air pressure, temperature and humidity, with a time resolution of not less than 15 minutes and a spatial resolution matching the range of the wind farm; the SCADA data includes wind turbine speed, pitch angle, nacelle azimuth angle, active power and gearbox temperature data; the geographic information data includes wind turbine coordinates, altitude and terrain slope data.

[0012] Preferably, the attention mechanism in step two is a modality adaptive attention mechanism, which initializes the weights by calculating the Pearson correlation coefficient between different modal features and historical power generation data, and then dynamically updates the weights by combining the prediction error during model training, so that features with high correlation get higher weights.

[0013] Preferably, when constructing the dynamic graph structure in step three, a single wind turbine is used as a node. When the angle between the line connecting two wind turbines and the real-time wind direction is less than 45° and the straight-line distance is less than a preset threshold, a connection is established between the two nodes. The connection weight is inversely proportional to the distance between the two wind turbines and directly proportional to the consistency of the wind direction.

[0014] Preferably, the Transformer-Graph Neural Network hybrid architecture described in step four includes parallel Transformer branches and graph neural network branches. The Transformer branch is used to extract long-term temporal dependencies of features, and the graph neural network branch is used to extract spatial correlations of wind turbine clusters. The outputs of the two branches are fused through a fully connected layer to output the prediction result.

[0015] Preferably, the loss function used in the model training process in step four is a hybrid loss function, which is a weighted sum of mean squared error loss and mean absolute percentage error loss. The weight coefficients are dynamically adjusted according to the prediction lead time. The longer the lead time, the greater the weight of the mean absolute percentage error loss.

[0016] Preferably, after step five, there is also a prediction result correction step, which combines the deviation between the real-time power output data of the wind farm and the prediction result to establish an error correction model. The error correction model adopts the gradient boosting tree algorithm with a sliding window to correct the prediction result in real time.

[0017] Preferably, the hybrid loss function equation is as follows: Let the number of predicted samples be n, and the nth... The predicted power generation value for each sample is Actual value The weight of the mean squared error loss (MSE) corresponding to the forecast lead time is: The weight of the mean absolute percentage error loss (MAPE) is: , , These are dynamic weighting coefficients that adaptively adjust with the forecast lead time. The final output value of the hybrid loss function reflects the overall error magnitude of the prediction model. The weighting coefficients of the mean squared error loss (MSE) are used to adjust the contribution of the MSE term to the total loss. The weighting coefficients of the mean absolute percentage error loss (MAPE) satisfy α + β = 1; the longer the lead time, the larger β becomes. n: the total number of samples participating in model training or prediction. The actual measured value of power generation for the i-th sample. : The model-predicted power generation value of the i-th sample.

[0018] Beneficial effects

[0019] This invention provides a method for predicting the power generation of a wind farm. Compared with existing technologies, it has the following advantages:

[0020] 1. This invention achieves deep fusion of multimodal data and accurate capture of dynamic spatiotemporal correlation, overcoming the limitations of traditional methods in data utilization and environmental adaptability: By simultaneously collecting multimodal data from NWP meteorology, SCADA operation status, and geographic information, and after preprocessing, a modality-adaptive attention mechanism is used to dynamically allocate feature weights. This avoids the shortcomings of single-modal information and the inefficiency of simple splicing, while also reducing the dependence on precise physical parameters of the wind field. At the same time, a dynamic graph structure is constructed based on wind turbine coordinates and real-time wind direction, and combined with a spatiotemporal graph neural network to mine the dynamic spatiotemporal coupling characteristics between wind turbine clusters, effectively adapting to the dynamic environment of the wind field and solving the defects of traditional methods that ignore spatial correlation or use fixed spatial structures.

[0021] 2. This invention optimizes the model architecture and error control system, significantly improving prediction accuracy, especially the reliability of long lead time predictions: It adopts a Transformer-graph neural network hybrid architecture to extract long temporal dependencies and spatial correlations in parallel, making up for the weak coupling processing capability of traditional statistical methods and the insufficient spatiotemporal feature mining of shallow machine learning; By dynamically adjusting the weights of the hybrid loss function to adapt to different lead time error characteristics, and with the sliding window gradient boosting tree error correction step, it specifically reduces the prediction error of long lead time, comprehensively enhances the model's ability to express complex features and the accuracy of prediction results, and provides strong support for power grid dispatching and power trading. Attached Figure Description

[0022] Figure 1 This is a flowchart of the wind power generation prediction method of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please see Figure 1 The present invention provides a technical solution:

[0025] A method for predicting the power generation of a wind power plant, characterized by the following steps:

[0026] Step 1: Multimodal Data Acquisition and Preprocessing. Simultaneously acquire numerical weather prediction (NWP) data, wind turbine monitoring and control (SCADA) data, and geographic information data from the wind farm. Perform outlier removal, missing value completion, and standardization on various types of data to obtain a standardized multimodal dataset. Specifically, the data is cleaned, including outlier removal (e.g., using box plots or the 3σ principle), missing value completion (e.g., using interpolation or machine learning methods), and standardization (e.g., Z-score standardization) to form a standardized multimodal dataset.

[0027] Step 2: Multimodal data feature hierarchical fusion. A feature fusion module is constructed. First, meteorological features are extracted from the standardized NWP data, operational status features from the SCADA data, and spatial topological features from the geographic information data. Then, an attention mechanism is used to assign dynamic weights to different modal features to generate initial fusion features. Feature engineering methods are then used to extract features from different modalities: meteorological features such as wind speed and direction are extracted from the NWP data; operational status features such as rotational speed and power are extracted from the SCADA data; and spatial topological features such as coordinates and altitude are extracted from the geographic information data. These features are then dynamically weighted and fused using an attention mechanism (such as modality adaptive attention) to generate initial fusion features.

[0028] Step 3: Enhance the spatiotemporal correlation features. Construct a dynamic graph structure based on the geographical coordinates of the wind turbines. Adjust the connection relationship and weight of the wind turbine nodes in the graph according to real-time wind direction data to form a spatiotemporal correlation graph that changes with wind direction. Input the initial fused features into the spatiotemporal graph neural network to mine the dynamic spatiotemporal coupling features between wind turbine clusters.

[0029] The spatiotemporal correlation features are enhanced by constructing a dynamic graph structure based on the geographical coordinates of the wind turbines. The connection relationships and weights of the wind turbine nodes in the graph are adjusted according to real-time wind direction data to form a spatiotemporal correlation graph that changes with wind direction. The initial fusion features are input into the spatiotemporal graph neural network to mine the dynamic spatiotemporal coupling features between wind turbine clusters.

[0030] Step 4: Power generation prediction model construction and training. Using spatiotemporal coupling features as input, a prediction model based on the Transformer-graph neural network hybrid architecture is constructed. The model is trained using historical power generation data, and the model parameters are adjusted through an adaptive learning rate optimization algorithm to obtain the trained prediction model.

[0031] Power generation prediction model construction and training: Using spatiotemporal coupling features as input, a prediction model based on the Transformer-graph neural network hybrid architecture is constructed. The model is trained using historical power generation data, and the model parameters are adjusted through an adaptive learning rate optimization algorithm to obtain the trained prediction model.

[0032] Step 5: Power generation prediction output. The preprocessed multimodal data for the period to be predicted is input into the trained prediction model after feature fusion and spatiotemporal enhancement, and the power generation prediction results of the wind farm are output.

[0033] Power generation prediction output: The preprocessed multimodal data of the period to be predicted is input into the trained prediction model after feature fusion and spatiotemporal enhancement, and the power generation prediction result is output. Combining the deviation between the real-time power output data of the wind farm and the prediction result, an error correction model is constructed through the gradient boosting tree algorithm of the sliding window to correct the prediction result in real time.

[0034] Based on the above, deep fusion of multimodal data is achieved, breaking through the limitations of insufficient data utilization in traditional methods. It integrates three key types of information: meteorological, operational status, and geospatial data, forming complementary information enhancement. Through dynamic graph structures and spatiotemporal graph neural networks, it accurately captures the dynamic spatiotemporal coupling mechanism of wind turbine clusters changing with wind direction, solving the problem of traditional models ignoring spatial correlations or simply aggregating spatial information. Based on a Transformer-graph neural network hybrid architecture, it simultaneously considers the extraction of long-term temporal dependencies and spatial correlations, improving the model's ability to express complex spatiotemporal features. Optional error correction steps further reduce prediction bias, significantly improving the accuracy and reliability of power generation prediction, providing strong support for stable grid operation, power dispatch, and power market transactions.

[0035] In this embodiment of the invention, the NWP data in step one includes spatiotemporal series data of wind speed, wind direction, air pressure, temperature, and humidity, with a time resolution of not less than 15 minutes and a spatial resolution matching the wind farm range. The SCADA data includes wind turbine speed, pitch angle, nacelle azimuth angle, active power, and gearbox temperature data. The geographic information data includes wind turbine coordinates, altitude, and terrain slope data. Collecting various types of data according to the aforementioned defined range and indicators provides standardized data input for subsequent preprocessing and feature extraction, defines the specific content and technical parameters of multimodal data, avoids blind data collection, and ensures the completeness, relevance, and effectiveness of the input data. The requirement of NWP data with a time resolution of not less than 15 minutes ensures the timeliness and precision of meteorological data. The setting of spatial resolution matching the wind farm range ensures the adaptability of geographic spatial data to the actual scene, laying a data foundation for subsequent high-precision feature extraction and prediction.

[0036] In this embodiment of the invention, the attention mechanism described in step two is a modal adaptive attention mechanism. Weights are initialized by calculating the Pearson correlation coefficients between different modal features and historical power generation data. Then, the weights are dynamically updated based on prediction errors during model training, giving higher weights to highly correlated features. The modal adaptive attention mechanism is used for feature weight allocation. First, the Pearson correlation coefficients between different modal features (meteorological features, operational status features, and spatial topology features) and historical power generation data are calculated. Based on these coefficients, the weights of each modal feature are initialized. During model training, prediction errors are monitored in real time, and the weights of each modal feature are dynamically updated based on error feedback. This gives higher weights to features highly correlated with power generation and lower weights to features with low correlation, achieving dynamic optimization of feature weights. This avoids the problems of weakened effective features and interference with prediction caused by traditional fixed weight allocation. Initializing weights using the Pearson correlation coefficient ensures the scientific nature of weight allocation. Combined with dynamic updates based on prediction errors, the weights adapt to changes in data patterns during model training, significantly improving the quality of feature fusion and providing more valuable initial fused features for subsequent predictions.

[0037] In this embodiment of the invention, when constructing the dynamic graph structure in step three, a single wind turbine is used as a node. When the angle between the line connecting two wind turbines and the real-time wind direction is less than 45° and the straight-line distance is less than a preset threshold, a connection between the two nodes is established. The connection weight is inversely proportional to the distance between the two wind turbines and directly proportional to the consistency of the wind direction.

[0038] Specifically, when constructing the dynamic graph structure, a single wind turbine is used as an independent node in the graph. Connection criteria are set: when the angle between the line connecting any two wind turbines and the real-time wind direction is less than 45°, and the straight-line distance between the two wind turbines is less than a preset threshold (e.g., 500 meters, which can be adjusted according to the actual layout of the wind farm), a connection is established between these two nodes. Connection weight rules are set: the connection weight is inversely proportional to the straight-line distance between the two wind turbines (the closer the distance, the greater the weight), and directly proportional to the consistency of the connection line and wind direction between the two wind turbines (the smaller the angle, the higher the consistency, and the greater the weight). Based on real-time wind direction data changes, the node connection relationships and weights are dynamically adjusted to form a spatiotemporal correlation graph that changes dynamically with the wind direction. This breaks through the limitations of traditional fixed graph structures that cannot adapt to the dynamic environment of wind fields. Through wind direction-driven node connections and weight adjustments, the dynamic spatial relationships between wind turbine clusters are accurately depicted. The setting of connection conditions and weight rules conforms to the actual physical laws of wind fields, enabling the spatiotemporal correlation graph to truly reflect the mutual influence of wind turbines during wind propagation. This provides a structural support that fits the actual scenario for mining spatiotemporal coupling features and improves the accuracy of feature extraction.

[0039] In this embodiment of the invention, the Transformer-Graph Neural Network hybrid architecture described in step four includes parallel Transformer branches and graph neural network branches. The Transformer branch is used to extract long-term temporal dependencies of features, and the graph neural network branch is used to extract spatial correlations of wind turbine clusters. The outputs of the two branches are fused through a fully connected layer to output the prediction result.

[0040] Specifically, a Transformer-Graph Neural Network hybrid architecture is constructed. This architecture includes two parallel branches: a Transformer branch and a Graph Neural Network branch. The Transformer branch receives spatiotemporally coupled feature inputs and extracts long-term temporal dependencies (such as the correlation between changes in wind speed and power at different times) from the features through a self-attention mechanism. The Graph Neural Network branch extracts spatial correlations between wind turbine clusters (such as the wind power transmission influence between upstream and downstream turbines) based on a dynamic spatiotemporal correlation graph. The fusion output is obtained by inputting the outputs of the two branches into a fully connected layer. Through weight allocation and feature integration, the final power generation prediction result is output. This achieves parallel extraction and complementary fusion of features in the temporal and spatial dimensions, solving the deficiency of a single architecture that can only focus on the extraction of features in one dimension (e.g., traditional Transformers are good at temporal but weak in spatial, and simple Graph Neural Networks are good at spatial but weak in long-term temporal). The hybrid architecture design fully leverages the advantages of both networks, enabling it to capture long-term temporal dependencies and complex spatial correlations simultaneously, significantly improving the model's ability to express multi-dimensional coupled features, thereby improving prediction accuracy.

[0041] In this embodiment of the invention, the loss function used in the model training process in step four is a hybrid loss function, which is a weighted sum of mean squared error loss and mean absolute percentage error loss. The weight coefficients are dynamically adjusted according to the prediction lead time. The longer the lead time, the greater the weight of the mean absolute percentage error loss.

[0042] In this embodiment of the invention, after step five, there is also a prediction result correction step. An error correction model is established by combining the deviation between the real-time power output data of the wind power plant and the prediction result. The error correction model adopts the gradient boosting tree algorithm with a sliding window to correct the prediction result in real time.

[0043] In this embodiment of the invention, the hybrid loss function equation is: Let the number of predicted samples be n, and the nth... The predicted power generation value for each sample is Actual value The weight of the mean squared error loss (MSE) corresponding to the forecast lead time is: The weight of the mean absolute percentage error loss (MAPE) is: , , These are dynamic weighting coefficients that adaptively adjust with the forecast lead time. The final output value of the hybrid loss function reflects the overall error magnitude of the prediction model. The weighting coefficient of the mean squared error loss (MSE) is used to adjust the contribution of this loss term to the total loss. The weighting coefficient of the mean absolute percentage error loss (MAPE) satisfies α + β = 1, with β increasing as the lead time increases. n: the total number of samples participating in model training or prediction. The actual measured value of power generation for the i-th sample. The model prediction value of the power generation of the i-th sample provides a clear and quantifiable mathematical basis for the hybrid loss function, avoiding the subjectivity of weight adjustment and error calculation, and ensuring the scientific and standardized nature of model training.

[0044] Specifically, the model training employs a hybrid loss function, which is a weighted sum of mean squared error loss (MSE) and mean absolute percentage error loss (MAPE). A dynamic adjustment rule for the weight coefficients is set: the weight coefficients α (MSE weight) and β (MAPE weight) satisfy α + β = 1, and are dynamically adjusted according to the prediction lead time (e.g., 1-3 hours for short-term predictions, 4-24 hours for medium-term predictions). The adjustment logic is as follows: the longer the prediction lead time, the greater the uncertainty of meteorological and operational data, and the more significant the impact of relative error on the prediction results; in this case, the weight of MAPE β is increased. Conversely, the shorter the lead time, the higher the reference value of absolute error; in this case, the weight of MSE α is increased. This achieves adaptive optimization of the loss function for the error characteristics of different prediction lead times, avoiding the limitations of a single loss function in different lead time scenarios (e.g., MSE is sensitive to outliers, while MAPE focuses more on relative error). Increasing the MAPE weight for long lead times can effectively reduce relative error and improve the stability of prediction results; increasing the MSE weight for short lead times can precisely control absolute error and improve prediction accuracy, enabling model training to adapt to the prediction needs of different lead times and expanding the model's applicability.

[0045] Furthermore, after outputting the prediction results in step five, a prediction result correction step is added: First, real-time power output data of the wind farm (i.e., actual power generation data) is collected, and the deviation between the real-time power output data and the prediction results is calculated. This deviation value is used as the input to the error correction model to construct the error correction model: A gradient boosting tree algorithm with a sliding window (such as XGBoost or LightGBM) is adopted. The window size can be set according to the real-time data update frequency (such as 30 minutes, with rolling updates as new data is generated). The error correction model is used to adjust the initial prediction results in real time, and the corrected final prediction results are output. This compensates for the real-time dynamic deviations that cannot be covered during model training (such as sudden weather changes, fluctuations in the temporary operating status of wind turbines, etc.), and realizes dynamic calibration of the prediction results. The application of the sliding window algorithm can capture the latest error patterns in a timely manner, and the gradient boosting tree algorithm has a strong nonlinear fitting capability, which can accurately fit the error change trend, further reducing the prediction error and improving the real-time performance and reliability of the prediction results.

[0046] Additional notes: Outlier removal: Outliers in NWP and SCADA data are processed using the 3σ principle (i.e., removing data that deviates from the mean by more than 3 times the standard deviation); for geographic information data, errors such as coordinates and altitude are removed by comparing with actual wind farm survey data.

[0047] Missing value completion: Missing values ​​in NWP and SCADA data are completed using linear interpolation (for short-term missing values, such as 1-2 time periods) or LSTM interpolation (for long-term missing values, such as 3 or more time periods); when geographic information data is missing, it is filled by field survey or by the average of nearby wind turbine data.

[0048] Standardization: Z-score standardization is used. (where μ is the mean and σ is the standard deviation) converts the three types of data to the same magnitude to avoid the interference of data scale differences on model training.

[0049] Example of preset threshold values: The preset threshold for the straight-line distance between two wind turbines in the dynamic graph structure is set according to the wind farm's wind turbine layout density, and is generally set to 300-800 meters. For example, it can be set to 500 meters for small and medium-sized wind farms (installed capacity below 50MW) and 800 meters for large wind farms (installed capacity above 100MW). Sliding window size: The sliding window size in the error correction model can be set to 20-50 data points. If the real-time data update frequency is 15 minutes, the corresponding window duration is 5-12.5 hours to ensure that the recent error change trend can be captured.

[0050] Selection of adaptive learning rate optimization algorithm: The AdamW algorithm is preferred for adaptive learning rate optimization in step four. This algorithm combines the adaptive learning rate characteristics of the Adam algorithm with the weight decay mechanism, which can effectively avoid model overfitting and accelerate training convergence. The initial learning rate can be set to 1e-4 and the decay coefficient can be set to 0.99.

[0051] Specific parameter settings for model training: Batch Size: Set according to hardware computing power, preferably 32 or 64; Epochs: 50-100 epochs, using an early stopping strategy, stopping training when the validation set loss does not decrease for 10 consecutive epochs to avoid overfitting; Transformer branch parameters: Encoder layers set to 3-5 layers, multi-head attention heads set to 4-8; Graph Neural Network branch parameters: Use GAT (Graph Attention Network), with 2-3 layers and 2-4 attention heads.

[0052] Furthermore, supplementary experimental data and performance verification are provided, along with a quantitative comparison table with existing technologies: Three typical existing methods (ARIMA, traditional CNN, and single graph neural network model) are selected, and the prediction errors (MAE, MAPE) are compared for different lead times (1 hour, 6 hours, and 24 hours) under the same wind farm dataset, as shown in the following example:

[0053]

[0054] Clarify the effect of the sliding window gradient boosting tree algorithm on the correction of prediction results.

[0055] Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0056] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0057] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method of wind farm power production prediction, characterized by: Specifically comprising the following steps: Step one, multi-modal data acquisition and preprocessing, synchronously acquiring numerical weather prediction data, wind turbine operation state monitoring data and geographic information data of the wind farm, and performing outlier elimination, missing value completion and standardization processing on various data to obtain a standardized multi-modal data set; Step two, multi-modal data feature hierarchical fusion, constructing a feature fusion module, extracting meteorological features from the standardized NWP data, extracting operation state features from the SCADA data, and extracting spatial topology features from the geographic information data, and then generating initial fusion features by assigning dynamic weights to different modal features through an attention mechanism; Step three, time-space correlation feature enhancement, constructing a dynamic graph structure based on the geographic coordinates of the wind turbine, adjusting the connection relationship and weight of the wind turbine nodes in the graph according to the real-time wind direction data, forming a time-space correlation graph that changes with the wind direction, and inputting the initial fusion features into a time-space graph neural network to mine the dynamic time-space coupling features between the wind turbine clusters; Step four, power generation prediction model construction and training, taking the time-space coupling features as input, constructing a prediction model based on a hybrid architecture of Transformer-graph neural network, training the model using historical power generation data, adjusting the model parameters through a self-adaptive learning rate optimization algorithm, and obtaining the trained prediction model; Step five, power generation prediction output, inputting the preprocessed multi-modal data of the to-be-predicted period into the trained prediction model after feature fusion and time-space enhancement, and outputting the power generation prediction result of the wind farm.

2. The wind farm power production forecasting method of claim 1, wherein: The NWP data in step one includes time-space sequence data of wind speed, wind direction, air pressure, temperature and humidity, with a time resolution of not less than 15 minutes and a spatial resolution matching the range of the wind farm; the SCADA data includes wind turbine speed, pitch angle, nacelle azimuth angle, active power and gearbox temperature data; and the geographic information data includes wind turbine coordinates, elevation and terrain slope data.

3. The wind farm power production forecasting method of claim 1, wherein: The attention mechanism in step two is a modal adaptive attention mechanism, which initializes the weights by calculating the Pearson correlation coefficients of different modal features and historical power generation data, and dynamically updates the weights based on the prediction error during model training, so that features with high correlation obtain higher weights.

4. The wind farm power production forecasting method of claim 1, wherein: In step three, when the included angle between the connection line of two wind turbines and the real-time wind direction is less than 45° and the straight-line distance is less than a preset threshold, the connection between the two nodes is established, and the connection weight is inversely proportional to the distance between the two wind turbines and proportional to the wind direction consistency.

5. The wind farm power production forecasting method of claim 1, wherein: The hybrid architecture of Transformer-graph neural network in step four includes parallel Transformer branches and graph neural network branches, the Transformer branches are used to extract long-time sequence dependencies of features, and the graph neural network branches are used to extract spatial correlation relationships of wind turbine clusters, and the outputs of the two branches are fused through a fully connected layer to output the prediction result.

6. The wind farm power production forecasting method of claim 1, wherein: The loss function used in the model training process in step four is a hybrid loss function, which is a weighted sum of mean square error loss and mean absolute percentage error loss, and the weight coefficient is dynamically adjusted according to the prediction lead time, and the longer the lead time, the greater the weight of the mean absolute percentage error loss.

7. The wind farm power production forecasting method of claim 1, wherein: Step five is followed by a prediction result correction step six, combining the deviation of the real-time output data of the wind power plant and the prediction result, an error correction model is established, and the error correction model uses a sliding window gradient boosting tree algorithm to correct the prediction result in real time.

8. The wind farm power production forecasting method of claim 6, wherein: The mixed loss function equation is wherein n is the number of predicted samples, the power generation prediction value of the i-th sample is , the actual value is , the mean square error loss weight corresponding to the prediction lead time is , the average absolute percentage error loss weight is , , , is a dynamic weight coefficient, which is adaptively adjusted with the prediction lead time, The final output value of the mixed loss function reflects the overall error size of the prediction model. The weight coefficient of the mean square error loss is used to adjust the contribution proportion of the mean square error loss term in the total loss. The weight coefficient of the average absolute percentage error loss satisfies α+β=1. The longer the lead time is, the larger β is. n: the total number of samples participating in model training or prediction, : the actual measured value of the power generation of the i-th sample, : the model predicted value of the power generation of the i-th sample.

Images