Wind farm power generation prediction method based on improved numerical prediction and artificial intelligence algorithm

Abstract

This invention belongs to the field of power prediction technology, specifically relating to a wind farm power prediction method based on improved numerical weather prediction and artificial intelligence algorithms. The steps include: acquiring multi-source data from the wind farm, performing time alignment, anomaly detection, and standardization, and then inputting the data into a generative adversarial network (GAN) with fused physical constraints for data completion; extracting physical features, temporal features, and spatial features through a multi-branch feature extraction network and fusing them to obtain a comprehensive fused feature; identifying error-sensitive areas in numerical weather prediction using a multi-scale attention module and correcting the numerical weather prediction wind speed; based on the corrected wind speed data, outputting ultra-short-term power prediction results and short-term power prediction results through a multi-task learning framework including trend consistency constraints; and performing uncertainty quantification analysis on the ultra-short-term and short-term power prediction results. This invention achieves high-precision, highly robust, interpretable, and easily deployable wind power prediction.

Description

Technical Field

[0001] This invention belongs to the field of power prediction technology, specifically relating to a wind farm power generation prediction method based on improved numerical forecasting and artificial intelligence algorithms. Background Technology

[0002] As the global energy structure continues to shift towards cleaner and lower-carbon energy, wind power, as a renewable energy source with abundant reserves and huge development potential, has seen rapid growth in its grid-connected capacity and has become an important component of the power system. However, wind power is characterized by significant volatility, randomness, and intermittency. The output power of wind farms is severely affected by weather conditions, easily leading to problems such as grid power fluctuations, deviations in power generation plans, difficulties in dispatching and operation, and inadequate absorption of new energy sources. Therefore, high-precision and robust power generation forecasting has become a core technological support for ensuring the safe and stable operation of the power system, optimizing dispatching decisions, and improving the level of wind power absorption. In recent years, the rapid development of numerical weather prediction and artificial intelligence algorithms has provided new pathways for wind power forecasting. However, existing technologies still have significant shortcomings in complex scenarios and cannot meet the practical application requirements of engineering, high precision, and high reliability.

[0003] Current numerical weather prediction products typically have a spatial resolution of only kilometers, making it difficult to accurately depict the micro-meteorological distribution within wind farms under complex terrain, complex underlying surfaces, and extreme weather conditions. This results in significant local deviations in key meteorological elements such as wind speed and direction, particularly in complex environments like mountains, coastal areas, and high altitudes. Furthermore, conventional data error correction methods often rely on statistical fitting, lacking physical consistency constraints and prone to producing correction results that contradict meteorological evolution patterns. Multi-source data generated from wind farm operations, including meteorological monitoring, turbine operation, and power output, frequently suffers from missing data, anomalies, noise, and packet loss. Traditional data cleaning and completion methods struggle to simultaneously guarantee temporal continuity, spatial correlation, and physical plausibility. Moreover, the spatiotemporal coupling characteristics of airflow propagation and power correlation among multiple wind farms within a region are not fully explored, further limiting the predictive performance of poor-quality, sparsely sampled, and newly commissioned wind farms.

[0004] Furthermore, most existing wind power forecasting methods model independently at a single time scale, failing to balance the rapid response of ultra-short-term forecasts to high-frequency fluctuations with the overall grasp of trend evolution in short-term forecasts. Inconsistencies in trends are easily observed in forecast results across different scales. While mainstream deep learning models possess strong fitting capabilities, they generally suffer from insufficient interpretability, limited uncertainty quantification capabilities, large parameter scales, high training costs, and strong dependence on computing power. This makes them difficult to deploy rapidly at the edge of wind farms and unsuitable for the rapid modeling needs of newly built wind farms with insufficient historical data. Overall, existing technologies have failed to systematically address issues such as poor adaptability to complex terrain, low quality of multi-source data, insufficient utilization of regional spatiotemporal correlations, weak synergy in multi-time-scale forecasts, and low applicability to model engineering. Significant room for improvement remains in forecast accuracy, robustness, and practical application effectiveness. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the purpose of this invention is to provide a wind farm power prediction method based on improved numerical forecasting and artificial intelligence algorithms. By integrating physical constraints and multi-branch feature extraction, multi-scale spatiotemporal attention correction, dynamic weight multi-task learning, probability quantization and lightweight transfer learning, it achieves high-precision, high-robustness, interpretability and easy engineering deployment of wind power prediction.

[0006] To achieve the above objectives, this invention provides a wind farm power generation prediction method based on improved numerical forecasting and artificial intelligence algorithms, comprising the following steps:

[0007] S1. Acquire multi-source data from the wind farm, including numerical weather prediction data, wind farm measured data, and geographic information data;

[0008] S2. After performing time alignment, anomaly detection, and standardization on the multi-source data, the data is input into a generative adversarial network that integrates physical constraints for data completion. The loss function of the generative adversarial network includes generative adversarial loss and physical bias loss based on atmospheric motion equations.

[0009] S3. Based on the completed data, physical features, temporal features and spatial features are extracted through a multi-branch feature extraction network, and then fused through an attention mechanism to obtain comprehensive fused features;

[0010] S4. Based on the integrated features and historical wind speed error sequence, the numerical weather forecast error sensitive area is identified through the multi-scale attention module, and the wind speed error correction amount is generated by combining the physical constraint loss function to correct the wind speed in the numerical weather forecast.

[0011] S5. Based on the corrected wind speed data, the ultra-short-term power prediction results and short-term power prediction results are output through a multi-task learning framework that includes trend consistency constraints.

[0012] S6. Based on the probabilistic prediction model, perform uncertainty quantification analysis on the ultra-short-term power prediction results and the short-term power prediction results.

[0013] As a preferred embodiment of the present invention, in S1, the numerical weather forecast data includes wind speed, wind direction and air pressure of the wind farm and surrounding area; the measured data of the wind farm includes the total power of the wind farm, the wind speed of the wind turbine, the active power and the ambient temperature of the wind turbine; the geographic information data includes the digital elevation model of the wind farm, the coordinates of the wind farm, the coordinates of the wind turbine, the distance between the stations and the average altitude of the wind farm.

[0014] As a preferred embodiment of the present invention, in S3, the multi-branch feature extraction includes:

[0015] The physical features branch extracts physical features using a convolutional neural network, based on downscaled numerical weather prediction data and geographic information data.

[0016] The temporal feature branch extracts temporal features based on the modal components obtained by dynamic mode decomposition of the total power sequence of the wind farm through a long short-term memory network.

[0017] The spatial feature branch extracts spatial features based on graph structure data that characterizes the spatial relationships between wind farms through graph neural networks. The graph structure data is constructed based on wind farm coordinates, distances between wind farms, and average elevation of wind farms.

[0018] As a preferred embodiment of the present invention, in the time series feature branch, after dynamic mode decomposition, the modes are classified into periodic operation modes, slowly changing trend modes and random disturbance modes according to the mode frequency and decay characteristics, and time series features are extracted based on the classification results.

[0019] As a preferred embodiment of the present invention, in S4, the time scale of the multi-scale attention module includes hourly and daily scales, and the spatial scale includes neighborhood scales of different ranges around the wind turbine.

[0020] By utilizing the multi-scale convolution and pooling structures in the multi-scale attention module, features are dynamically extracted at different temporal and spatial scales, and spatiotemporal attention weights are adaptively calculated. By analyzing the spatiotemporal attention weight distribution, sensitive areas for numerical weather prediction errors are identified.

[0021] As a preferred embodiment of the present invention, in S4, the physical constraint loss function includes: the error term between the corrected wind speed and the measured wind speed, and the constraint term for the difference in wind speed gradient at adjacent wind turbine locations.

[0022] As a preferred embodiment of the present invention, in S5, the multi-task learning framework uses dynamic weights to weight the ultra-short-term prediction loss and the short-term prediction loss, and introduces trend consistency constraints to ensure that the prediction trends at different time scales remain consistent.

[0023] As a preferred embodiment of the present invention, in S6, the probability prediction model adopts a Bayesian neural network model, and performs multiple random samplings using the Monte Carlo Dropout method to calculate the mean and standard deviation of the prediction results.

[0024] As a preferred embodiment of the present invention, it further includes:

[0025] S7. Model Lightweighting and Transfer Learning: The prediction model is lightweighted through knowledge distillation and pruning; for newly commissioned wind farms, transfer learning is performed based on the pre-trained model parameters, where the domain adaptive loss is calculated based on the maximum mean difference.

[0026] The beneficial effects of this invention are:

[0027] This invention addresses the multi-source heterogeneous data characteristics of wind power forecasting by constructing a collaborative feature extraction architecture involving three branches: physics, time series, and space. It achieves numerical weather prediction downscaling and data completion through a generative adversarial network integrating atmospheric motion equations, and enhances the rationality of physical features by incorporating terrain slope constraints. Dynamic mode decomposition (DMD) is employed to decompose the power sequence into three modes: periodic, trend, and stochastic, selectively extracting multi-scale temporal evolution features. Graph neural networks are used to model the spatial correlation topology of wind farm clusters, accurately capturing the spatial coupling patterns between farms. This comprehensively mines data value from three dimensions: physical mechanism, temporal evolution, and spatial correlation, providing high-quality feature input for subsequent forecasts and fundamentally solving the problems of insufficient feature extraction and inadequate physical consistency in traditional methods.

[0028] This invention constructs a dynamic weight-driven multi-task learning framework. By combining ultra-short-term MSE loss, short-term MAE loss, and trend consistency constraints, it adaptively balances the weights of ultra-short-term and short-term prediction tasks, avoiding the problems of accuracy imbalance and trend deviation in traditional multi-task learning. At the same time, it designs a multi-scale spatiotemporal attention correction module, which uses historical wind speed error sequences as the driving force to generate time and space dual-dimensional attention weights. Combined with Transformer modeling, it generates wind speed error correction amounts. Through physical constraint loss, it ensures the spatial consistency of wind speed gradients between adjacent wind turbines, which greatly improves the correction accuracy of numerical weather prediction and achieves dual optimization of prediction accuracy and physical laws.

[0029] The probabilistic prediction model of this invention employs a Monte Carlo Dropout Bayesian neural network. By approximating the posterior normal distribution of predicted power through multiple forward sampling, it outputs the predicted mean and confidence interval, thereby quantifying the uncertainty of power prediction and providing probabilistic prediction support for wind power dispatch. Simultaneously, it adopts a model compression scheme of knowledge distillation and pruning. Using a complex multi-task large model as the teacher model, a lightweight student model is trained through KL divergence knowledge transfer. Combined with gradient pruning, the model parameters are compressed by more than 70% with an accuracy loss of less than 2%, solving the problem of complex models being difficult to deploy in engineering. It balances prediction accuracy and deployment efficiency and can be adapted to edge computing scenarios in wind farms.

[0030] This invention designs a rapid adaptation mechanism that combines pre-training with domain adaptation. By loading the pre-trained model parameters of already operational wind farms, freezing the bottom feature layer and only fine-tuning the top mapping layer, it achieves rapid initialization of the new wind farm model. It uses the maximum mean difference as the domain adaptation loss and periodically reduces the feature distribution difference between new and old wind farms based on newly accumulated data, achieving dynamic model adaptation and solving the problems of cold start and data scarcity for new wind farms. The entire solution is reproducible end-to-end, and the hyperparameters of each module can be obtained through engineering experience or adaptive learning, adapting to wind farm scenarios of different terrains and scales. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating the principle of this invention;

[0032] Figure 2 This is the power sequence mode component diagram of dynamic mode decomposition in Embodiment 1 of the present invention;

[0033] Figure 3 This is a structural diagram of a wind farm cluster in Embodiment 1 of the present invention;

[0034] Figure 4 This is a comparison chart of the wind speed correction effect in numerical weather forecasting in Embodiment 1 of the present invention;

[0035] Figure 5 This is an output diagram of the multi-task prediction model in Embodiment 1 of the present invention;

[0036] Figure 6 This is a graph showing the prediction uncertainty quantification results in Embodiment 2 of the present invention. Detailed Implementation

[0037] The embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0038] Example 1: The purpose of this example is to address the following problems in traditional forecasting methods: poor local adaptability of numerical weather prediction, large prediction bias due to data quality issues, underutilization of wind farm cluster association, weak generalization ability for extreme and small sample scenarios, insufficient synergy between long-term and short-term predictions, lack of physical interpretability, contradiction between model deployment and real-time performance, difficulty in modeling newly commissioned wind farms, poor capture of multimodal features, and insufficient quantification of prediction uncertainty.

[0039] like Figure 1 As shown, the wind farm power generation prediction method based on improved numerical weather prediction and artificial intelligence algorithms includes:

[0040] S1. Acquire multi-source data from the wind farm, including numerical weather prediction data, wind farm measured data, and geographic information data;

[0041] S2. After performing time alignment, anomaly detection, and standardization on the multi-source data, the data is input into a generative adversarial network that integrates physical constraints for data completion. The loss function of the generative adversarial network includes generative adversarial loss and physical bias loss based on atmospheric motion equations.

[0042] S3. Based on the completed data, physical features, temporal features and spatial features are extracted through a multi-branch feature extraction network, and then fused through an attention mechanism to obtain comprehensive fused features;

[0043] S4. Based on the integrated features and historical wind speed error sequence, the numerical weather forecast error sensitive area is identified through the multi-scale attention module, and the wind speed error correction amount is generated by combining the physical constraint loss function to correct the wind speed in the numerical weather forecast.

[0044] S5. Based on the corrected wind speed data, the ultra-short-term power prediction results and short-term power prediction results are output through a multi-task learning framework that includes trend consistency constraints.

[0045] S6. Based on the probabilistic prediction model, perform uncertainty quantification analysis on the ultra-short-term power prediction results and the short-term power prediction results.

[0046] In S1, numerical weather prediction data includes:

[0047] Wind speed in and around the wind farm ,wind direction and air pressure ,in Represents spatial coordinates, where t is the timestamp (moment).

[0048] The measured data from the wind farm include:

[0049] Total power of wind farm, wind speed of individual wind turbines Active power and ambient temperature Where i is the index of the wind turbine, and the number of wind turbines is N, for example... Let represent the single-unit wind speed of fan i at time t;

[0050] Geographic information data includes:

[0051] Digital Elevation Model of Wind Farm Wind farm coordinates, wind turbine coordinates Distance between wind farms and average altitude of wind farms.

[0052] In S2, linear interpolation is used to map numerical weather prediction data, geographic information data, and other data to the same time granularity, eliminating time discrepancies between different data sources and achieving precise alignment of the time dimension. Subsequently, for the aligned time-series data such as wind speed, power, and temperature, [further details are needed]. The criteria are used to detect anomalies, clean the data by removing outliers and replacing them with interpolated valid values ​​from adjacent time points, and finally perform Z-score standardization to eliminate the influence of units and provide unified and reliable input data for subsequent processing.

[0053] In a Generative Adversarial Network (GAN) that incorporates physical constraints, the generator G takes random noise z as input and a known data fragment as input, and outputs completed data. The discriminator D distinguishes between real and generated data. In the loss function L of the GAN, in addition to the generative adversarial loss, a physical bias loss based on atmospheric motion equations is also introduced.

[0054] ;

[0055] In the formula, To generate adversarial loss (cross-entropy loss); For physical constraint weights; Loss due to physical deviation:

[0056] ;

[0057] In the formula, This is the predicted wind speed vector output by the GAN; For Coriolis parameters; air density; For gradient operators; The predicted air pressure output by the GAN;

[0058] By introducing constraints from atmospheric motion equations, physical distortion of the generated data can be avoided.

[0059] When constructing a physically constrained generative adversarial network, the basic GAN loss function (generative adversarial loss) is first constructed. Cross-entropy loss is used to measure the difference in distribution between generated data and real data, and then physical bias loss is defined. Based on the atmospheric momentum equation, the generation is calculated. , The sum of the absolute values ​​of the deviations between the physical quantities and the theoretical values; setting the weights of the physical constraints. (For example, 0.3), to obtain the total loss function L, which is used for model training.

[0060] In S3, multi-branch feature extraction includes:

[0061] The physical features branch extracts physical features using a convolutional neural network, based on downscaled numerical weather prediction data and geographic information data.

[0062] The downscaling process specifically involves using bilinear interpolation to downscale the kilometer-level numerical weather forecast data to the wind turbine coordinate scale, thereby obtaining preliminary refined data. , which is the numerical weather forecast wind speed at the i-th wind turbine.

[0063] Input in a convolutional neural network and terrain slope obtained based on wind farm digital elevation model The physical characteristics of local weather were obtained. .

[0064] Time-series feature branching, based on the total power sequence of wind farms The modal components obtained by multi-scale dynamic mode decomposition (DMD) are used to construct a time-frequency characteristic representation model of the power evolution of a wind farm cluster:

[0065] ;

[0066] In the formula, This represents the single-unit active power of wind turbine i at time t; is the spatial distribution feature vector of the k-th power dynamic mode, used to characterize the degree of collaborative participation of each wind farm in this mode, and K is the number of effective dynamic modes obtained by decomposition; Let be the complex characteristic frequency, where Characterizing the modal decay or growth rate, Characterizing the angular frequency of the dominant mode oscillation; The initial amplitude of the mode; J represents the imaginary unit;

[0067] A modal physical property constraint screening mechanism is introduced. After dynamic mode decomposition, the modes are classified into periodic operation modes, slowly changing trend modes, and random perturbation modes based on their frequency and decay characteristics. Temporal features are then extracted using a Long Short-Term Memory (LSTM) network based on the classification results. .

[0068] Periodic operation mode satisfies , This refers to the oscillation period of the operating mode in the time domain, used to characterize the regular changes in daily wind resource cycles, load dispatch cycles, etc.; the gradually changing trend mode satisfies and It is used to characterize the long-term power evolution trend of wind farm clusters; among which, the threshold , The modal energy ratios can be set separately; random perturbation modes satisfy... and It exhibits broadband distribution characteristics and is used to reflect random behaviors such as local meteorological disturbances and transient fluctuations of generating units;

[0069] Through the above modal classification mechanism, only the effective modal components that are strongly correlated with the stable operation and power scheduling decisions of wind farms are retained, thereby realizing the adaptive dimensionality reduction expression of the power dynamic characteristics of wind farm groups and avoiding the problems of severe noise coupling and insufficient physical interpretability in traditional time series analysis methods.

[0070] Figure 2 The diagram presents the modal component plot of the power sequence from dynamic mode decomposition, which is the result of dynamic mode decomposition. It decomposes the total power sequence of the wind farm into periodic modes (daily period), trend modes, and random modes. Figure 2 The horizontal axis represents the normalized time scale, and the vertical axis represents the power amplitude (kW). Periodic modes exhibit clear sinusoidal fluctuations, reflecting the diurnal variation of wind speed; trend modes show linear growth, reflecting long-term power change trends; and stochastic modes exhibit irregular fluctuations, corresponding to random factors such as turbulence. The decomposed data can be input into the temporal feature branch (LSTM) to extract targeted features, addressing the problem of single models struggling to capture multimodal characteristics. Dynamic mode decomposition provides the foundation for subsequent feature fusion, enabling the model to specifically handle power characteristics at different time scales and improving prediction robustness in complex scenarios.

[0071] The spatial feature branch extracts spatial features based on graph-structured data representing the spatial relationships between wind farms, using the message-passing mechanism of a graph neural network (GNN). The graph structure data is constructed based on wind farm coordinates, distances between wind farms, and average wind farm elevation.

[0072] The graph structure is constructed by treating the M wind farms within the selected area as graph nodes and building an adjacency matrix. , Let M×M be the M×M dimensional real space, and let the elements be... Defined as:

[0073] ;

[0074] In the formula, Let m be the straight-line distance between wind farms m and n; For the power sequence of the wind farm m and wind farm n power sequence The Pearson correlation coefficient; , are the average elevations of wind farms m and n, respectively; , These are the scale parameters for distance and altitude, respectively.

[0075] Figure 3 The structure of the wind farm cluster graph is presented, and the corresponding adjacency matrix is ​​constructed. With 5 wind farms as nodes, the weight of the edges is determined by the straight-line distance, power correlation and altitude difference. Figure 3 The nodes in the graph correspond to the geographical coordinates of the wind farm, and the thickness of the edges indicates the strength of the association (e.g., nodes that are close together and have high power correlation are connected with thicker edges). The horizontal axis represents the east-west geographical coordinates of the wind farm, and the vertical axis represents the north-south geographical coordinates. This graph structure allows for the discovery of spatial relationships, such as the airflow propagation delay from upstream wind farms to downstream wind farms, addressing the problem of neglecting cluster collaboration in traditional single-wind-farm modeling. This graph visually presents the spatial dependencies between wind farms, providing a topological foundation for GNNs to extract spatial features and improving prediction accuracy in sparse data scenarios.

[0076] The fusion is achieved through an attention mechanism to obtain the comprehensive fusion feature F:

[0077] ;

[0078] During fusion, the attention weights for the three types of features are:

[0079] ;

[0080] In the formula, Represents the r-th class of features The attention weights, r takes values ​​of 1, 2, or 3, respectively representing... , or ; Let be the scoring function; a higher score indicates a greater potential contribution of the feature flow to the prediction result. , W and b are the weight matrix and bias term obtained during model training. Let l represent the l-th feature, which is used to sum the scores of all feature categories.

[0081] In S4, the time scale of the multi-scale attention module includes hourly and daily scales, and the spatial scale includes different neighborhood scales around the wind turbine. The multi-scale attention module is constructed by inputting the fused feature F and the historical wind speed error sequence. Output time attention weights (e.g., hourly level) Japanese level ) and spatial attention weights (e.g., within 500m of the wind turbine) 1km range ):

[0082] ;

[0083] ;

[0084] In the formula, Softmax is the activation function; Linear represents a linear layer (fully connected layer); LayerNnorm represents layer normalization;

[0085] Design a correction function (dynamic correction model):

[0086] ;

[0087] In the formula, This represents the corrected wind speed value of fan i at time t; The error correction amount is obtained using the Transformer model:

[0088] ;

[0089] In the formula, Represents tensor product;

[0090] During training, a physical constraint loss function is introduced. (The total loss function of the calibration model, used to measure the calibration effect), includes the error term between the calibrated wind speed and the measured wind speed, as well as the constraint term for the difference in wind speed gradient at adjacent wind turbine locations:

[0091] ;

[0092] In the formula, The weighting coefficients for the physical constraint term (i.e., the constraint term for the difference in wind speed gradient at adjacent wind turbine locations) are used to adjust the degree of influence of physical constraints on the total loss. , Let be the corrected wind speed gradients in the x-direction at the i-th and j-th wind turbines (adjacent to the i-th wind turbine), respectively. The wind speed gradient between adjacent wind turbines can be calculated using the central difference method based on the wind turbine coordinate grid.

[0093] Error term between corrected wind speed and measured wind speed It can reflect the deviation between the corrected value and the true value; it designs physical constraint terms and calculates the absolute value of the difference in wind speed gradient between adjacent wind turbines. and multiplied by (This can be set based on engineering experience) to ensure that spatial variations in wind speed conform to physical laws. It can be used to train dynamic calibration models, so that the calibration results take into account both data fitting and physical rationality.

[0094] By utilizing the multi-scale convolution and pooling structures in the multi-scale attention module, features are dynamically extracted at different temporal and spatial scales, and spatiotemporal attention weights are adaptively calculated. By analyzing the spatiotemporal attention weight distribution, sensitive areas for numerical weather prediction errors are identified.

[0095] Figure 4 The chart presents a comparison of the wind speed correction effects in numerical weather forecasts. The horizontal axis represents the normalized time scale, and the vertical axis represents the wind speed (m / s). Figure 4 Comparing the wind speeds before and after dynamic correction with measured values, the original forecast had a large error (deviating from the measured value by approximately ±100), while the error was reduced to within ±50 after correction. The correction process used a multi-scale attention module to identify error-sensitive areas (such as within 500m of the wind turbine) and combined this with physical constraint losses to ensure the results conformed to the atmospheric momentum equation. The corrected wind speeds are closer to the measured values, addressing the issue of poor local adaptability in numerical weather prediction. This step provides high-precision input for subsequent power prediction and is a crucial step in improving overall prediction accuracy.

[0096] Figure 5 The graph presents the output results of the multi-task prediction model. The horizontal axis represents time (min or h), and the vertical axis represents active power (kW). The solid line represents the actual active power of the wind turbines, and the dashed line represents the ultra-short-term / short-term power prediction values. The ultra-short-term (0-4 hours) prediction (10-minute interval) shows high agreement with the measured power high-frequency fluctuations (error less than 5%), while the short-term (1-2 days) prediction (1-hour interval) accurately captures trend changes. The dual outputs employ LSTM, attention, and Transformer respectively, and dynamic weight loss ensures consistency of results. This solves the problem of contradictory trends in long-term and short-term predictions in traditional single-scale models. Ultra-short-term predictions support real-time scheduling, while short-term predictions assist in planning, synergistically improving the consistency and reliability of power grid scheduling.

[0097] In S5, the multi-task learning framework uses dynamic weights to weight the ultra-short-term prediction loss and the short-term prediction loss, and introduces trend consistency constraints to ensure that the prediction trends at different time scales remain consistent.

[0098] by Using F as input, construct a dual-output model:

[0099] Ultra-short-term forecast (0-4 hours): Employing LSTM and attention mechanism, output power sequence ,in Indicates a time interval (10 minutes). ;

[0100] Short-term forecast (1-3 days): Using the Transformer model, the output power sequence is... ,at this time , 1 hour interval.

[0101] Total loss function of multi-task learning framework for:

[0102] ;

[0103] In the formula, , These are the ultra-short-term and short-term predicted losses, respectively. This represents the actual active power of the wind turbine. For the active power predicted in the very short term, The active power is the short-term forecast, MSE is the mean square error, and MAE is the mean absolute error. For trend consistency constraints, Trend represents the trend extraction function (e.g., sliding window averaging to extract the trend); dynamic weights , ; These are constraint coefficients;

[0104] In S6, the probabilistic prediction model uses a Bayesian neural network model. Multiple random samples are performed using the Monte Carlo Dropout method to calculate the mean and standard deviation of the prediction results, thus obtaining the probability distribution of the predicted power.

[0105] ;

[0106] In the formula, The number of samples (e.g., 50 times); Indicates predicted power; Indicates the first Dropout random parameters for each sample; Indicates the first The predicted output of the neural network when the input is F after the second sampling; Indicates that given F, The posterior probability distribution;

[0107] Based on the probability distribution, output the mean. and standard deviation ,satisfy , Let P represent a normal distribution, where P is a random variable representing the active power of the wind turbine, used to quantify the uncertainty of power prediction. Reflecting the central trend of the forecast, It reflects the predicted range of fluctuations.

[0108] Example 2: Based on Example 1, it also includes:

[0109] S7. Model Lightweighting and Transfer Learning: The prediction model is lightweighted through knowledge distillation and pruning; for newly commissioned wind farms, transfer learning is performed based on the pre-trained model parameters, where the domain adaptive loss is calculated based on the maximum mean difference.

[0110] A complete multi-task learning model, consisting of a multi-branch feature extraction network, a multi-scale attention module, a multi-task learning framework, and a probabilistic prediction model, is used as the teacher model T to train the student model S.

[0111] ;

[0112] In the formula, For mission losses; KL divergence (measures the difference in the distribution of teacher and student model outputs); Distillation weight; This represents the predicted output of the student model (with F as the input). This represents the predicted output of the teacher model (with F as the input).

[0113] The pruning algorithm is used to remove students whose gradient contributions are less than a threshold. (For example The model parameters are compressed by more than 70% while the accuracy loss is less than 2%.

[0114] For newly commissioned wind farms, load the pre-trained model parameters, freeze the bottom feature layer, and only fine-tune the top mapping layer:

[0115] ;

[0116] In the formula, The parameters are updated for the new wind farm model; These are the pre-trained model parameters (teacher / student model weights). The learning rate; These are the model parameters (trainable parameters of the model). The comprehensive integration characteristics of newly commissioned wind farms; The measured active power of newly commissioned wind farms; The task loss function for the new wind farm; Represents the loss function right The gradient.

[0117] The model parameters are trained on mature wind farm datasets, including the bottom feature layers (such as CNN, LSTM, and GNN layers for extracting physical, temporal, and spatial features) and the top mapping layer (a fully connected layer that maps fused features to power prediction values), which are used as the initial parameters for the newly commissioned wind farm model.

[0118] The parameters of the bottom feature layer are fixed and not updated during training. This is because the bottom feature layer learns the basic features common to wind farms (such as the general relationship between wind speed and power, and the physical characteristics of meteorological elements), which have a certain degree of universality for new scenarios. Freezing the parameters can avoid overfitting caused by small sample data.

[0119] Based on limited data from new wind farms , Only update the parameters of the top-level mapping layer. (By...) Formula, with learning rate Along the loss function The gradient direction is iteratively optimized to adapt the top-level network to the specific characteristics of the new wind farm (such as unit model and terrain detail differences).

[0120] For every 100 hours of new data accumulated, domain-adaptive loss is applied. Adjust the model and use Gaussian kernel calculations to achieve dynamic adaptation. The MMD represents the maximum mean difference, which is a comprehensive integration characteristic of wind farms in operation.

[0121] Figure 6 This graph presents the quantification results of prediction uncertainty, with the horizontal axis representing time (0-24h) and the vertical axis representing active power (kW). Corresponding to the Bayesian neural network output, it compares the mean ± standard deviation range (shaded area) of the ultra-short-term prediction with the measured power. The mean line closely approximates the measured value, while the standard deviation reflects uncertainty (e.g., a larger standard deviation during periods of high volatility). By generating 100 samples using Monte Carlo dropout, the error probability distribution is quantified, addressing the issue that deterministic predictions cannot support risk-based decision-making. This result provides a quantitative basis for grid reserve capacity configuration; for example, more reserves should be reserved during periods of high standard deviation to reduce wind curtailment by 8%-10%.

[0122] Example 3: Based on Example 1, an adaptive triggering module for extreme weather events is added between wind speed correction and multi-task prediction. When the multi-scale attention module identifies error-sensitive areas corresponding to sudden wind speed changes, extreme pressure gradients, or topographic funneling effects, the extreme weather-specific sub-model is automatically activated.

[0123] The extreme weather-specific sub-model employs a gated recurrent unit with enhanced physical constraints. It embeds the gradient constraints of atmospheric motion equations into the gating mechanism and freezes the weights of random perturbation modes in the time-series feature branches, retaining only periodic modes and slowly changing trend modes for prediction. This reduces the wind speed correction error under extreme weather conditions by more than 30%, and the predicted trend remains strictly consistent with physical laws.

[0124] Example 4: A wind farm power generation prediction device based on improved numerical forecasting and artificial intelligence algorithms, comprising:

[0125] One or more processors;

[0126] Memory, used to store one or more computer programs;

[0127] When one or more programs are executed by one or more processors, the one or more processors perform the methods in Embodiment 1, Embodiment 2, or Embodiment 3.

[0128] Example 5: A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to perform the method of Example 1, Example 2 or Example 3.

Claims

1. A method for predicting wind farm power generation based on improved numerical weather prediction and artificial intelligence algorithms, characterized in that, include: S1. Acquire multi-source data from the wind farm, including numerical weather prediction data, wind farm measured data, and geographic information data; S2. After performing time alignment, anomaly detection, and standardization on the multi-source data, the data is input into a generative adversarial network that integrates physical constraints for data completion. The loss function of the generative adversarial network includes generative adversarial loss and physical bias loss based on atmospheric motion equations. S3. Based on the completed data, physical features, temporal features and spatial features are extracted through a multi-branch feature extraction network, and then fused through an attention mechanism to obtain comprehensive fused features; S4. Based on the integrated features and historical wind speed error sequence, the numerical weather forecast error sensitive area is identified through the multi-scale attention module, and the wind speed error correction amount is generated by combining the physical constraint loss function to correct the wind speed in the numerical weather forecast. S5. Based on the corrected wind speed data, the ultra-short-term power prediction results and short-term power prediction results are output through a multi-task learning framework that includes trend consistency constraints. S6. Based on the probabilistic prediction model, perform uncertainty quantification analysis on the ultra-short-term power prediction results and the short-term power prediction results.

2. The wind farm power generation prediction method based on improved numerical forecasting and artificial intelligence algorithms according to claim 1, characterized in that, In S1, the numerical weather forecast data includes wind speed, wind direction and air pressure in and around the wind farm; the measured data of the wind farm includes the total power of the wind farm, the wind speed of the wind turbine, the active power and the ambient temperature; the geographic information data includes the digital elevation model of the wind farm, the coordinates of the wind farm, the coordinates of the wind turbine, the distance between the stations and the average altitude of the wind farm.

3. The wind farm power generation prediction method based on improved numerical forecasting and artificial intelligence algorithms according to claim 2, characterized in that, In S3, multi-branch feature extraction includes: The physical features branch extracts physical features using a convolutional neural network, based on downscaled numerical weather prediction data and geographic information data. The temporal feature branch extracts temporal features based on the modal components obtained by dynamic mode decomposition of the total power sequence of the wind farm through a long short-term memory network. The spatial feature branch extracts spatial features based on graph structure data that characterizes the spatial relationships between wind farms through graph neural networks. The graph structure data is constructed based on wind farm coordinates, distances between wind farms, and average elevation of wind farms.

4. The wind farm power generation prediction method based on improved numerical forecasting and artificial intelligence algorithms according to claim 3, characterized in that, In the time series feature branch, after dynamic mode decomposition, the modes are classified into periodic operation modes, slowly changing trend modes and random disturbance modes according to the mode frequency and decay characteristics, and time series features are extracted based on the classification results.

5. The wind farm power generation prediction method based on improved numerical forecasting and artificial intelligence algorithms according to claim 1, characterized in that, In S4, the time scale in the multi-scale attention module includes hourly and daily scales, and the spatial scale includes neighborhood scales of different ranges around the wind turbine. By utilizing the multi-scale convolution and pooling structures in the multi-scale attention module, features are dynamically extracted at different temporal and spatial scales, and spatiotemporal attention weights are adaptively calculated. By analyzing the spatiotemporal attention weight distribution, sensitive areas for numerical weather forecast errors are identified.

6. The wind farm power generation prediction method based on improved numerical forecasting and artificial intelligence algorithms according to claim 1, characterized in that, In S4, the physical constraint loss function includes: the error term between the corrected wind speed and the measured wind speed, and the constraint term for the difference in wind speed gradient at adjacent wind turbine locations.

7. The wind farm power generation prediction method based on improved numerical forecasting and artificial intelligence algorithms according to claim 1, characterized in that, In the S5 framework, the ultra-short-term prediction loss and the short-term prediction loss are weighted by dynamic weights, and a trend consistency constraint is introduced to ensure that the prediction trend remains consistent across different time scales.

8. The wind farm power generation prediction method based on improved numerical forecasting and artificial intelligence algorithms according to claim 1, characterized in that, In S6, the probability prediction model uses a Bayesian neural network model and performs multiple random samplings using the Monte Carlo Dropout method to calculate the mean and standard deviation of the prediction results.

9. The wind farm power generation prediction method based on improved numerical forecasting and artificial intelligence algorithms according to claim 1, characterized in that, Also includes: S7. Model Lightweighting and Transfer Learning: Lightweighting the prediction model through knowledge distillation and pruning; For newly commissioned wind farms, transfer learning is performed based on pre-trained model parameters, where the domain adaptive loss is calculated based on the maximum mean difference.

Images