A wind power prediction method and system

Abstract

This invention relates to the field of wind power prediction technology. It provides a wind power prediction method and system. The steps are as follows: acquire multi-dimensional meteorological time-series data, three-dimensional elevation data, and turbine operation data of the target wind farm; construct a spatiotemporal feature fusion network, extract temporal dynamic features, and weightedly fuse spatial correlation features and temporal dynamic features to obtain a fused feature vector; establish a hybrid prediction model and use the fused feature vector as input to obtain the initial wind power prediction result; introduce a terrain correction factor and construct a turbulence intensity compensation function to correct the initial wind power prediction result for micro-topographic disturbances; output the final power prediction curve and confidence interval. This invention addresses the problems in existing wind power prediction methods, such as insufficient accuracy of physical models for complex terrain and microclimate modeling, high computational complexity, and difficulty in meeting real-time requirements; and the limited ability of statistical learning methods to express high-dimensional nonlinear time-series features, leading to significantly increased prediction errors under abnormal operating conditions.

Description

Technical Field

[0001] This invention relates to the field of wind power prediction technology, and more specifically, to a wind power prediction method and system. Background Technology

[0002] As the global energy structure accelerates its transition towards cleaner and lower-carbon energy, wind power, as a core form of renewable energy, continues to expand its installed capacity. However, wind power output is highly volatile and random due to the influence of multiple meteorological factors such as wind speed, wind direction, and temperature. This poses a severe challenge to the frequency regulation, peak shaving capabilities, and safe and stable operation of the power system when large-scale wind power grid integration is implemented. Against this backdrop, high-precision wind power forecasting technology has become a key means to improve grid absorption capacity, optimize unit combination decisions, and reduce reserve capacity allocation. Its forecasting performance directly affects the economic efficiency and reliability of the power system.

[0003] Traditional wind power forecasting methods are mainly divided into two categories: physical model-driven and statistical learning. Physical models are based on numerical weather prediction (NWP) data, simulating the meteorological conditions of wind farms through fluid dynamics equations and calculating output power by combining the wind turbine power curves. However, their modeling accuracy for complex terrain and microclimates is insufficient, and their computational complexity is high, making it difficult to meet the real-time requirements of minute-level short-term forecasts. Statistical learning methods (such as ARIMA and support vector machines) rely on historical data to mine the mapping relationship between input features and power output. Although they have strong adaptability, their ability to express high-dimensional nonlinear time-series features is limited, especially under abnormal operating conditions such as extreme weather or turbine failures, where the prediction error increases significantly. Summary of the Invention

[0004] The purpose of this invention is to provide a wind power prediction method and system, which aims to solve the problems in existing wind power prediction methods, such as insufficient accuracy of physical models in modeling complex terrain and microclimate, high computational complexity and difficulty in meeting real-time requirements, limited ability of statistical learning methods to express high-dimensional nonlinear time series features, and significant increase in prediction error under abnormal operating conditions.

[0005] The present invention is achieved through the following technical solution:

[0006] A method for predicting wind power includes the following steps:

[0007] Acquire multi-dimensional meteorological time-series data, three-dimensional elevation data, and unit operation data of the target wind farm;

[0008] Based on multi-dimensional meteorological time-series data and unit operation data, a spatiotemporal feature fusion network is constructed to extract the spatial correlation features of multi-dimensional meteorological time-series data and the temporal dynamic features of unit operation data. The spatial correlation features and temporal dynamic features are then fused by weighted fusion through an attention mechanism to obtain a fused feature vector.

[0009] A hybrid prediction model is established based on a neural network, and the fused feature vector is used as input. An adaptive weighting algorithm is used to dynamically adjust the output weights of the data-driven and physical models to obtain the initial prediction results of wind power.

[0010] A terrain correction factor is introduced, and a turbulence intensity compensation function is constructed based on the three-dimensional elevation data of the target wind farm to correct the initial wind power prediction results for micro-terrain disturbance.

[0011] The final power prediction curve and confidence interval are output, and the parameters of the hybrid prediction model are optimized through a sliding time window feedback mechanism to form a closed-loop iterative update.

[0012] Optionally, the specific process of acquiring multi-dimensional meteorological time-series data, three-dimensional elevation data, and unit operation data of the target wind farm is as follows:

[0013] By deploying meteorological observation equipment at the target wind farm and connecting it to the numerical weather prediction system, multi-dimensional meteorological time-series data including wind speed, wind direction, temperature, air pressure, and air humidity are obtained.

[0014] By using digital elevation models and on-site topographic surveys, three-dimensional elevation data of the target wind farm, including longitude, latitude, altitude, and terrain roughness, are obtained.

[0015] The wind farm monitoring and data acquisition system acquires unit operation data including active power output, rotor speed, pitch angle, generator temperature, and gearbox oil temperature. It also performs time alignment and outlier cleaning on multi-dimensional meteorological time series data, three-dimensional elevation data, and unit operation data.

[0016] Optionally, the specific process of constructing a spatiotemporal feature fusion network based on multi-dimensional meteorological time-series data and unit operation data, extracting spatial correlation features from the multi-dimensional meteorological time-series data and temporal dynamic features from the unit operation data, and then weighting and fusing the spatial correlation features and temporal dynamic features through an attention mechanism to obtain a fused feature vector is as follows:

[0017] Multi-dimensional meteorological time-series data are input into a spatiotemporal feature fusion network composed of a multi-layer graph convolutional neural network. A spatial graph structure of meteorological parameters is constructed based on the location of meteorological stations in wind farms. The spatial gradient relationship of wind speed, air pressure and temperature between nodes is calculated through the adjacency matrix. The meteorological feature information of adjacent nodes is aggregated layer by layer, and the spatial correlation feature vector of multi-dimensional meteorological time-series data is output.

[0018] The unit operation data is input into a time encoder composed of a bidirectional long short-term memory network. The unit operation data sequence is segmented by a sliding time window. The temporal dependencies of historical and future moments are captured by forward and backward LSTM units, respectively, and the temporal dynamic feature vector of the unit operation data is output.

[0019] A dynamic feature fusion module based on a multi-head attention mechanism is constructed. The cosine similarity matrix between the spatially correlated feature vector and the temporally dynamic feature vector is calculated. The spatial-temporal attention weights are generated by normalization using the Softmax function. The two feature vectors are fused by weighted summation to obtain a fused feature vector with spatiotemporal correlation.

[0020] Optionally, the specific process of establishing a hybrid prediction model based on a neural network, using the fused feature vector as input, and dynamically adjusting the output weights of the data-driven and physical models using an adaptive weighting algorithm to obtain the initial prediction results of wind power is as follows:

[0021] A hybrid prediction model consisting of a data-driven sub-model and a physics-driven sub-model is constructed. The data-driven sub-model adopts a deep neural network architecture, receives the fused feature vector, and outputs data-driven prediction values. The physics-driven sub-model is based on numerical weather forecast data, simulates the boundary layer flow field of the wind farm by solving the Reynolds-averaged Navier-Stokes equation, and generates physical prediction values ​​by combining the wind turbine power-wind speed characteristic curve.

[0022] The historical prediction error sequences of the data-driven sub-model and the physics-driven sub-model are collected in real time, and the exponentially weighted moving average error of the two is calculated respectively. The historical prediction error sequence is used to extract the absolute error values ​​of the most recent N moments through a sliding time window.

[0023] Based on the ratio of the exponentially weighted moving average error of the data-driven sub-model and the physics-driven sub-model, dynamic weight allocation coefficients are generated by mapping through the sigmoid function, so that the sub-model with lower error receives higher weight.

[0024] The data-driven predictions and physical predictions are linearly combined according to the dynamic weighting coefficients to generate the initial prediction results of wind power. The weighting coefficients are updated at each prediction time through an online error feedback mechanism.

[0025] Optionally, the specific process of introducing a terrain correction factor and constructing a turbulence intensity compensation function based on the three-dimensional elevation data of the target wind farm to correct the initial wind power prediction results for micro-terrain disturbances is as follows:

[0026] Based on the altitude, terrain roughness, and height difference between adjacent grid points in the three-dimensional elevation data of the target wind farm, the equivalent terrain slope factor and surface roughness index at the hub height of each wind turbine are calculated.

[0027] Based on the equivalent terrain slope factor and surface roughness index, and combined with the Reynolds stress transfer model, a turbulent kinetic energy correction equation is established to solve the horizontal and vertical turbulence intensity components at the location of the wind turbine.

[0028] The turbulence intensity components at each wind turbine location are spatially interpolated using a three-dimensional convolution kernel to generate a turbulence intensity distribution matrix covering the target wind farm.

[0029] Based on the turbulence intensity distribution matrix and the wind turbine power-turbulence response characteristic curve, the turbulence attenuation coefficient at each time step is determined, and a compensation function with turbulence intensity as the independent variable is constructed.

[0030] The initial wind power prediction results are input into the compensation function, and the power correction corresponding to different turbulence intensity ranges is calculated by the piecewise linearization method. The correction is then added to the initial prediction results to compensate for the wake effect and turbulence dissipation loss caused by micro-topography.

[0031] Optionally, the specific process of optimizing the hybrid prediction model parameters through a sliding time window feedback mechanism to form a closed-loop iterative update of the output final power prediction curve and confidence interval is as follows:

[0032] The initial wind power prediction results are fitted with a probability distribution based on the quantile regression algorithm. The probability density function of the power prediction value is generated by Monte Carlo sampling. The 5% and 95% quantiles are selected as the upper and lower boundaries of the confidence interval, and the final power prediction curve with confidence interval is output.

[0033] A sliding time window with a length of T hours is set, and the actual measured value of wind power and the initial prediction result are collected in real time at each moment within the window. The mean square error, mean absolute percentage error and maximum deviation index are calculated as feedback signals.

[0034] A joint optimization objective function is constructed, which includes the weights of the spatiotemporal feature fusion network, the dynamic weight allocation coefficients of the hybrid prediction model, and the parameters of the compensation function. The model parameters are updated along the negative gradient direction of the error using a stochastic gradient descent algorithm with a driving term.

[0035] Every K prediction cycles, the updated model parameters are synchronized to the online prediction system, and 10% of the samples within the current sliding time window are retained as a validation set. When the validation set error exceeds a preset threshold, the model parameter rollback mechanism is triggered.

[0036] The updated hybrid prediction model is reconnected to the real-time data stream for the next round of prediction, forming a closed-loop iterative update process of prediction-feedback-optimization.

[0037] Optionally, real-time monitoring of sudden wind speed changes and abnormal unit operating conditions is performed. When the predicted residual is detected to exceed the threshold, local model retraining is triggered and the output weight coefficients of the data-driven and physical models are updated.

[0038] Based on the same inventive concept, the present invention also provides a wind power prediction system for implementing the aforementioned wind power prediction method, comprising:

[0039] The data acquisition and processing module is used to acquire multi-dimensional meteorological time-series data, three-dimensional elevation data and unit operation data of the target wind farm, and to perform time alignment and outlier cleaning on the multi-dimensional meteorological time-series data, the three-dimensional elevation data and the unit operation data.

[0040] The spatiotemporal feature fusion module includes a graph convolutional neural network and a bidirectional long short-term memory network, which is used to extract the spatial correlation features of multi-dimensional meteorological time series data and the temporal dynamic features of unit operation data, and generate a fused feature vector through a multi-head attention mechanism.

[0041] The hybrid prediction model module includes a data-driven sub-model and a physics-driven sub-model, which are used to dynamically adjust the output weights of the two sub-models using an adaptive weighting algorithm to generate initial wind power prediction results.

[0042] The terrain correction module is used to calculate the turbulence intensity compensation function based on three-dimensional elevation data and to correct the initial prediction results for micro-terrain disturbances.

[0043] The prediction output and optimization module is used to output the final power prediction curve with confidence interval, and optimize the model parameters through a sliding time window feedback mechanism and a stochastic gradient descent algorithm.

[0044] The anomaly monitoring and retraining module is used to detect sudden changes in wind speed and abnormal operating conditions of the unit in real time. When the prediction residual exceeds the threshold, it triggers local model retraining and updates the weight coefficients.

[0045] Based on the same inventive concept, the present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the above-described wind power prediction method.

[0046] Based on the same inventive concept, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described wind power prediction method.

[0047] The technical solution of the present invention has at least the following advantages and beneficial effects:

[0048] By constructing a spatiotemporal feature fusion network, it is possible to extract the spatial correlation features of multi-dimensional meteorological time series data and the time series dynamic features of unit operation data respectively. The targeted extraction of features of different types of data, compared with traditional methods, more comprehensively and deeply explores the effective information in the data, laying a solid foundation for subsequent accurate prediction.

[0049] By using an attention mechanism to weightedly fuse spatial correlation features and temporal dynamic features, a fused feature vector is obtained. The attention mechanism can automatically focus on important features, making the fused feature vector more accurately reflect the complex relationship between wind power and various factors. This overcomes the shortcomings of traditional methods in handling multi-feature fusion and improves feature utilization efficiency and prediction accuracy.

[0050] A hybrid prediction model based on neural networks is established, and an adaptive weighting algorithm is used to dynamically adjust the output weights of the data-driven and physical models. This fully leverages the data-driven model's ability to fit complex nonlinear relationships and the physical model's ability to grasp the fundamental laws of wind power. By dynamically adjusting the weights, the model can flexibly adapt to different operating conditions, improving the accuracy and robustness of predictions and overcoming the limitations of traditional single models in dealing with complex wind power scenarios.

[0051] By introducing a terrain correction factor and constructing a turbulence intensity compensation function based on the three-dimensional elevation data of the target wind farm, the initial wind power prediction results are corrected for micro-topographic disturbances. Since wind power output is significantly affected by terrain factors, traditional methods do not adequately consider this. However, this invention, through this correction mechanism, can more accurately reflect the impact of micro-topography on wind power, further improving prediction accuracy. It is especially suitable for power prediction of wind farms with complex terrain.

[0052] The system outputs the final power prediction curve and confidence interval, and optimizes the parameters of the hybrid prediction model through a sliding time window feedback mechanism to form a closed-loop iterative update. This allows the model parameters to be continuously adjusted based on the actual prediction results, enabling the model to continuously learn and adapt to the dynamic changes of the wind power system. As time goes by and data accumulates, the model's prediction performance will continue to improve, ensuring the long-term effectiveness and accuracy of the prediction method. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the wind power prediction method according to an embodiment of the present invention.

[0054] Figure 2 This is a schematic diagram of the wind power prediction system according to an embodiment of the present invention. Detailed Implementation

[0055] The following is a detailed description of the embodiments, in conjunction with the accompanying drawings.

[0056] Reference Figure 1 A wind power prediction method includes the following steps:

[0057] Step 1: Obtain multi-dimensional meteorological time-series data, three-dimensional elevation data, and unit operation data of the target wind farm.

[0058] In some embodiments, the specific process of acquiring multi-dimensional meteorological time-series data, three-dimensional elevation data, and unit operation data of a target wind farm is as follows:

[0059] By deploying meteorological observation equipment at the target wind farm and connecting it to the numerical weather prediction system, multi-dimensional meteorological time-series data including wind speed, wind direction, temperature, air pressure, and air humidity are obtained.

[0060] By using digital elevation models and on-site topographic surveys, three-dimensional elevation data of the target wind farm, including longitude, latitude, altitude, and terrain roughness, are obtained.

[0061] The wind farm monitoring and data acquisition system acquires unit operation data including active power output, rotor speed, pitch angle, generator temperature, and gearbox oil temperature. It also performs time alignment and outlier cleaning on multi-dimensional meteorological time series data, three-dimensional elevation data, and unit operation data.

[0062] Step 2: Based on multi-dimensional meteorological time-series data and unit operation data, construct a spatiotemporal feature fusion network, extract the spatial correlation features of multi-dimensional meteorological time-series data and the temporal dynamic features of unit operation data, and obtain the fused feature vector by weighted fusion of spatial correlation features and temporal dynamic features through an attention mechanism.

[0063] In some embodiments, a spatiotemporal feature fusion network is constructed based on multi-dimensional meteorological time-series data and unit operation data to extract spatial correlation features from the multi-dimensional meteorological time-series data and temporal dynamic features from the unit operation data. The spatial correlation features and temporal dynamic features are then weighted and fused through an attention mechanism to obtain the fused feature vector. The specific process is as follows:

[0064] Multi-dimensional meteorological time-series data are input into a spatiotemporal feature fusion network composed of a multi-layer graph convolutional neural network. A spatial graph structure of meteorological parameters is constructed based on the location of meteorological stations in wind farms. The spatial gradient relationship of wind speed, air pressure and temperature between nodes is calculated through the adjacency matrix. The meteorological feature information of adjacent nodes is aggregated layer by layer, and the spatial correlation feature vector of multi-dimensional meteorological time-series data is output.

[0065] The unit operation data is input into a time encoder composed of a bidirectional long short-term memory network. The unit operation data sequence is segmented by a sliding time window. The temporal dependencies of historical and future moments are captured by forward and backward LSTM units, respectively, and the temporal dynamic feature vector of the unit operation data is output.

[0066] A dynamic feature fusion module based on a multi-head attention mechanism is constructed. The cosine similarity matrix between the spatially correlated feature vector and the temporally dynamic feature vector is calculated. The spatial-temporal attention weights are generated by normalization using the Softmax function. The two feature vectors are fused by weighted summation to obtain a fused feature vector with spatiotemporal correlation.

[0067] Step 3: Establish a hybrid prediction model based on neural networks, and use the fused feature vector as input. Use an adaptive weighting algorithm to dynamically adjust the output weights of the data-driven and physical models to obtain the initial prediction results of wind power.

[0068] In some embodiments, a hybrid prediction model is established based on a neural network, and the fused feature vector is used as input. An adaptive weighting algorithm is then used to dynamically adjust the output weights of the data-driven and physical models to obtain the initial prediction results of wind power. The specific process is as follows:

[0069] Construct a hybrid prediction model consisting of a data-driven sub-model and a physics-driven sub-model.

[0070] The data-driven sub-model employs a deep neural network architecture, receiving fused feature vectors and outputting data-driven predictions; the deep neural network can be LSTM, CNN, etc. The physics-driven sub-model, based on numerical weather prediction data, simulates the boundary layer flow field of the wind farm by solving the Reynolds-averaged Navier-Stokes equations, and generates physical predictions by combining the wind turbine power-wind speed characteristic curves.

[0071] The historical prediction error sequences of the data-driven sub-model and the physics-driven sub-model are collected in real time, and the exponentially weighted moving average error of the two is calculated. The absolute error values ​​of the most recent N moments are extracted from the historical prediction error sequences through a sliding time window.

[0072] At each time step t, the actual power is obtained, and the absolute error between the data-driven sub-model and the physical-driven sub-model is calculated, as shown in equation (1) below:

[0073]

[0074] Among them, e d (t) represents the absolute error of the data-driven sub-model; D(t) represents the predicted value of the data-driven sub-model; y(t) represents the actual wind power; e p P(t) represents the absolute error of the physical-driven sub-model; P(t) represents the predicted value of the material-driven sub-model.

[0075] Maintain a sliding window queue of length N to store the absolute errors of the data-driven sub-model and the physics-driven sub-model for the most recent N time steps.

[0076] Based on the ratio of the exponentially weighted moving average error of the data-driven sub-model to that of the physics-driven sub-model, dynamic weight allocation coefficients are generated through sigmoid function mapping, so that the sub-model with lower error receives higher weight; whereby the exponentially weighted moving average error of the data-driven sub-model is as shown in equation (2):

[0077]

[0078] Among them, E d (t) represents the exponentially weighted moving average error of the data-driven sub-model; t∈(0,1), α represents the decay factor; N represents the sliding window length;

[0079] The exponentially weighted moving average error of the physics-driven sub-model is shown in equation (3) below:

[0080]

[0081] Among them, E p (t) represents the exponentially weighted moving average error of the physics-driven sub-model; the error is shown in equation (4) below:

[0082]

[0083] Where r(t) represents the error ratio;

[0084] The weights of the data-driven sub-model and the physics-driven sub-model are mapped using the sigmoid function. The weights of the data-driven sub-model and the physics-driven sub-model are shown in Equations (5) and (6) respectively:

[0085]

[0086] w p (t)=1-w d (t)(6)

[0087] Among them, w d (t) represents the weights of the data-driven sub-model; w p (t) represents the weights of the physics-driven sub-model; k represents the sensitivity parameter, and k>0.

[0088] The data-driven prediction values ​​and physical prediction values ​​are linearly combined according to the dynamic weight allocation coefficients to generate the initial prediction result of wind power, as shown in the following equation (7):

[0089]

[0090] in, This represents the initial wind power forecast result; the weighting coefficients are updated at each forecast time through an online error feedback mechanism. The error sequence within the sliding window can be updated in each forecast period, and E can be recalculated. d (t) and E p (t), and adjust the weights w d (t) and w p (t).

[0091] Step 4: Introduce a terrain correction factor and construct a turbulence intensity compensation function based on the three-dimensional elevation data of the target wind farm to correct the initial wind power prediction results for micro-terrain disturbances.

[0092] In some embodiments, a terrain correction factor is introduced, and a turbulence intensity compensation function is constructed based on the three-dimensional elevation data of the target wind farm. The specific process of correcting the initial wind power prediction results for micro-terrain disturbance is as follows:

[0093] Based on the elevation, terrain roughness, and height difference between adjacent grid points in the three-dimensional elevation data of the target wind farm, the equivalent terrain slope factor and surface roughness index at the hub height of each wind turbine are calculated; whereby the equivalent terrain slope factor is shown in the following formula (8):

[0094]

[0095] Among them, S i Represents the equivalent terrain slope factor; h represents the set of adjacent grid points of fan i; i and h j Let x represent the elevation of wind turbine i and the adjacent point j, respectively; i ,y o (x) represents the latitude and longitude coordinates of wind turbine i; (x) represents the latitude and longitude coordinates of wind turbine i. j ,y j ) represents the latitude and longitude coordinates of the adjacent point j.

[0096] Surface roughness index z 0,i =Lookup(c i ), where represents the terrain roughness classification parameter, which is mapped to a typical roughness value (e.g., grassland z) through a lookup table. 0,i =0.03m, forest z 0,i =1.0m).

[0097] Based on the equivalent terrain slope factor and surface roughness index, and combined with the Reynolds stress transfer model, a turbulent kinetic energy correction equation is established to solve for the horizontal and vertical turbulence intensity components at the location of the wind turbine. The turbulent kinetic energy correction equation is shown in equation (9) below:

[0098]

[0099] Where k represents turbulent kinetic energy; U i and U j τ represents the time-averaged velocity component. ij The Reynolds stress tensor represents the momentum transport caused by turbulent fluctuations and is defined by the fluctuation velocity covariance; ∈ represents the dissipation rate; v t denoted by tidal viscosity, used to simulate turbulent diffusion effects; β represents the terrain coupling coefficient; the modified Reynolds-averaged equation is discretized at the wind turbine location, and k and ∈ are solved using the finite volume method. The inlet boundary uses the wind speed profile provided by numerical weather prediction, and the ground boundary applies a wall function combined with the surface roughness index. Using the solved k value, combined with the hub height wind speed U... hub The turbulence intensity in the horizontal direction and the turbulence intensity in the vertical direction are calculated as shown in equation (10):

[0100]

[0101] Among them, I h,i Indicates the intensity of turbulence in the horizontal direction; I v,i γ represents the vertical turbulence intensity; z represents the vertical attenuation coefficient; z represents the height.

[0102] The turbulence intensity components at each wind turbine location are spatially interpolated using a three-dimensional convolution kernel to generate a turbulence intensity distribution matrix covering the target wind farm. The turbulence intensity distribution matrix can be generated using a Gaussian kernel, as shown in Equation (11) below:

[0103]

[0104] in, Represents the turbulence intensity distribution matrix; I i This represents the turbulence intensity component at the location of fan i, such as the horizontal turbulence intensity I. h,i Or vertical turbulence intensity I v,i (x,y,z) represents the three-dimensional coordinates of the target grid point; (x i ,y i ,z i (x, y, z) represents the three-dimensional coordinates of wind turbine i; σ represents the Gaussian kernel bandwidth, used to control the spatial correlation range; N represents the total number of wind turbines. The wind farm area is divided into a 100m × 100m grid. For each target grid point (x, y, z), its spatial distance to all wind turbines is calculated. Turbulence intensity values ​​are then superimposed with Gaussian weights, and the result is output. The three-dimensional matrix has the same resolution as the wind farm grid.

[0105] Based on the turbulence intensity distribution matrix and the power-turbulence response characteristic curve of the wind turbine, the turbulence attenuation coefficient at each time step is determined, and a compensation function with turbulence intensity as the independent variable is constructed. The actual power and prediction deviation of the wind turbine are recorded under different turbulence intensities, and the piecewise linear relationship between the compensation function and the turbulence attenuation coefficient is obtained by fitting; wherein, the expression of the turbulence attenuation coefficient is shown in the following equation (12):

[0106]

[0107] Where β(t) represents the turbulence attenuation coefficient; w m,n Indicates the weighting coefficient; This represents the elements of the turbulence intensity distribution matrix, i.e., the turbulence intensity value at grid point (m,n) at time t;

[0108] The compensation function is shown in equation (13) below:

[0109]

[0110] The compensation function is a piecewise function, where ΔP(t) represents the compensation value; a k and b k The linear coefficients for each interval are calibrated using historical data; θ k This indicates a preset threshold; the threshold is set based on the calibration results.

[0111] The initial wind power prediction results are input into the compensation function. A piecewise linearization method is used to calculate the power correction for different turbulence intensity ranges, which is then superimposed onto the initial prediction results to compensate for the wake effect and turbulence dissipation losses caused by micro-topography. Input initial wind power prediction values. The compensation interval is selected based on the turbulence attenuation coefficient β(t) at the current moment, the compensation value ΔP(t) is calculated, the correction is superimposed, and the final predicted value is output.

[0112] Step 5: Output the final power prediction curve and confidence interval, and optimize the parameters of the hybrid prediction model through a sliding time window feedback mechanism to form a closed-loop iterative update.

[0113] In some embodiments, the specific process of outputting the final power prediction curve and confidence interval, and optimizing the parameters of the hybrid prediction model through a sliding time window feedback mechanism to form a closed-loop iterative update is as follows:

[0114] The probability distribution of the initial wind power prediction results is fitted based on the quantile regression algorithm. The probability density function of the power prediction value is generated by Monte Carlo sampling. The 5% and 95% quantiles are selected as the upper and lower boundaries of the confidence interval, and the final power prediction curve with confidence interval is output.

[0115] A sliding time window with a length of T hours is set, and the actual measured value of wind power and the initial prediction result are collected in real time at each moment within the window. The mean square error, mean absolute percentage error and maximum deviation index are calculated as feedback signals.

[0116] A joint optimization objective function is constructed, which includes the weights of the spatiotemporal feature fusion network, the dynamic weight allocation coefficients of the hybrid prediction model, and the parameters of the compensation function. The model parameters are updated along the negative gradient direction of the error using a stochastic gradient descent algorithm with a driving term.

[0117] Every K prediction cycles, the updated model parameters are synchronized to the online prediction system, and 10% of the samples within the current sliding time window are retained as a validation set. When the validation set error exceeds a preset threshold, the model parameter rollback mechanism is triggered.

[0118] The updated hybrid prediction model is reconnected to the real-time data stream for the next round of prediction, forming a closed-loop iterative update process of prediction-feedback-optimization.

[0119] In some embodiments, real-time monitoring of sudden wind speed changes and abnormal turbine operating conditions is implemented. When the predicted residual exceeds a threshold, local model retraining is triggered, and the output weight coefficients of the data-driven and physical models are updated. A dedicated monitoring module is established in the wind power prediction system. This module continuously receives real-time data from meteorological observation equipment and the Supervisory Control and Data Acquisition (SCADA) system for wind farms, including key parameters such as wind speed, wind direction, and active power output of the turbines. Data analysis algorithms are used to dynamically analyze the real-time data to identify sudden wind speed changes (such as a sudden increase or decrease in wind speed) and abnormal turbine operating conditions (such as abnormal fluctuations in power output and abnormal temperature increases). At each prediction time, the system generates a predicted wind power value based on the current hybrid prediction model; simultaneously, the system acquires the actual wind power measurement value at that time; the difference between the predicted value and the actual measurement value is calculated, i.e., the prediction residual. Based on historical data and experience, a reasonable prediction residual threshold is set. This threshold reflects the error range of the prediction model under normal conditions, and the real-time calculated prediction residual is compared with the set threshold. When the predicted residual exceeds a threshold, the system determines that the current prediction model cannot accurately reflect the actual conditions of the wind farm, thus triggering a local model retraining mechanism. Local model retraining can be performed on the data-driven sub-model, the physics-driven sub-model, or both simultaneously. During retraining, historical data from the most recent period (including multi-dimensional meteorological time-series data, turbine operation data, etc.) is used to retrain the model to adjust its parameters and structure, making it better suited to the current wind farm conditions. After local model retraining is complete, the system needs to re-evaluate the predictive performance of the data-driven and physics-driven sub-models; based on the retrained model performance, an adaptive weighting algorithm is used to recalculate the output weight coefficients of the data-driven and physics-driven models; the new weight coefficients are then applied to the hybrid prediction model to generate more accurate wind power prediction results.

[0120] Based on the same inventive concept, and corresponding to any of the above embodiments, refer to... Figure 2 This invention provides a wind power prediction system for implementing the aforementioned wind power prediction method, comprising:

[0121] The data acquisition and processing module is used to acquire multi-dimensional meteorological time-series data, three-dimensional elevation data and unit operation data of the target wind farm, and to perform time alignment and outlier cleaning on the multi-dimensional meteorological time-series data, three-dimensional elevation data and unit operation data.

[0122] The spatiotemporal feature fusion module includes a graph convolutional neural network and a bidirectional long short-term memory network, which is used to extract the spatial correlation features of multi-dimensional meteorological time series data and the temporal dynamic features of unit operation data, and generate a fused feature vector through a multi-head attention mechanism.

[0123] The hybrid prediction model module includes a data-driven sub-model and a physics-driven sub-model, which are used to dynamically adjust the output weights of the two sub-models using an adaptive weighting algorithm to generate initial wind power prediction results.

[0124] The terrain correction module is used to calculate the turbulence intensity compensation function based on three-dimensional elevation data and to correct the initial prediction results for micro-topographic disturbances.

[0125] The prediction output and optimization module is used to output the final power prediction curve with confidence interval, and optimize the model parameters through a sliding time window feedback mechanism and a stochastic gradient descent algorithm.

[0126] The anomaly monitoring and retraining module is used to detect sudden changes in wind speed and abnormal operating conditions of the unit in real time. When the prediction residual exceeds the threshold, it triggers local model retraining and updates the weight coefficients.

[0127] Based on the same inventive concept, corresponding to any of the above embodiments, the present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the wind power prediction method of the embodiment.

[0128] Alternatively, the aforementioned electronic device may be a server.

[0129] In addition, this embodiment also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the wind power prediction method of the embodiment.

[0130] It is understood that the processor in the embodiments of the present invention can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. The general-purpose processor can be a microprocessor or any conventional processor.

[0131] The method steps in the embodiments of the present invention can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0132] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a storage medium or transmitted through a storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

Claims

1. A method for predicting wind power output, characterized in that, Includes the following steps: Acquire multi-dimensional meteorological time-series data, three-dimensional elevation data, and unit operation data of the target wind farm; Based on multi-dimensional meteorological time-series data and unit operation data, a spatiotemporal feature fusion network is constructed to extract the spatial correlation features of multi-dimensional meteorological time-series data and the temporal dynamic features of unit operation data. The spatial correlation features and temporal dynamic features are then fused by weighted fusion through an attention mechanism to obtain a fused feature vector. A hybrid prediction model based on neural networks is established, and the fused feature vector is used as input. An adaptive weighting algorithm is employed to dynamically adjust the output weights of the data-driven and physical models to obtain initial wind power prediction results. The specific process is as follows: a hybrid prediction model is constructed, consisting of a data-driven sub-model and a physical-driven sub-model. The data-driven sub-model uses a deep neural network architecture, receives the fused feature vector, and outputs data-driven prediction values. The physical-driven sub-model, based on numerical weather prediction data, simulates the boundary layer flow field of the wind farm by solving the Reynolds-averaged Navier-Stokes equations, and generates physical prediction values ​​by combining the wind turbine power-wind speed characteristic curve. Historical prediction error sequences of the data-driven sub-model and the physics-driven sub-model are collected in real time, and their exponentially weighted moving average errors are calculated. The historical prediction error sequences are truncated to the absolute error values ​​of the most recent N moments through a sliding time window. Based on the ratio of the exponentially weighted moving average errors of the data-driven sub-model and the physics-driven sub-model, dynamic weight allocation coefficients are generated through sigmoid function mapping, so that the sub-model with lower error receives higher weight. The data-driven prediction values ​​and the physics prediction values ​​are linearly combined according to the dynamic weight allocation coefficients to generate the initial prediction results of wind power, wherein the weight allocation coefficients are updated at each prediction moment through an online error feedback mechanism. A terrain correction factor is introduced, and a turbulence intensity compensation function is constructed based on the three-dimensional elevation data of the target wind farm to correct the initial wind power prediction results for micro-topographic disturbances. Specifically, based on the altitude, terrain roughness, and height difference between adjacent grid points in the three-dimensional elevation data of the target wind farm, the equivalent terrain slope factor and surface roughness index at the hub height of each wind turbine are calculated. Based on the equivalent terrain slope factor and surface roughness index, and combined with the Reynolds stress transfer model, a turbulence kinetic energy correction equation is established to solve for the turbulence intensity in the horizontal and vertical directions at the wind turbine location. The turbulence intensity components at each wind turbine location are spatially interpolated using a 3D convolution kernel to generate a turbulence intensity distribution matrix covering the target wind farm. Based on the turbulence intensity distribution matrix and the wind turbine power-turbulence response characteristic curve, the turbulence attenuation coefficient at each time step is determined, and a compensation function with turbulence intensity as the independent variable is constructed. The initial wind power prediction results are input into the compensation function, and the power correction corresponding to different turbulence intensity intervals is calculated using a piecewise linearization method. This correction is then superimposed on the initial prediction results to compensate for the wake effect and turbulence dissipation loss caused by micro-topography. The final power prediction curve and confidence interval are output. The parameters of the hybrid prediction model are optimized through a sliding time window feedback mechanism to form a closed-loop iterative update. Specifically, the initial wind power prediction results are fitted with a probability distribution based on a quantile regression algorithm. A probability density function for generating the power prediction value is generated using Monte Carlo sampling. The 5% and 95% quantiles are selected as the upper and lower boundaries of the confidence interval, and the final power prediction curve with the confidence interval is output. A sliding time window of length T hours is set, and the actual measured wind power values ​​and initial prediction results are collected in real time at each moment within the window. The mean square error, mean absolute percentage error, and maximum deviation are calculated. The difference index is used as a feedback signal; a joint optimization objective function is constructed, which includes the weights of the spatiotemporal feature fusion network, the dynamic weight allocation coefficients of the hybrid prediction model, and the parameters of the compensation function. The stochastic gradient descent algorithm with a driving term is used to update the model parameters along the negative gradient direction of the error. Every K prediction cycles, the updated model parameters are synchronized to the online prediction system, and 10% of the samples in the current sliding time window are retained as a validation set. When the error of the validation set exceeds a preset threshold, the model parameter rollback mechanism is triggered. The hybrid prediction model with updated parameters is reconnected to the real-time data stream for the next round of prediction, forming a closed-loop iterative update process of prediction-feedback-optimization.

2. The wind power prediction method as described in claim 1, characterized in that, The specific process for acquiring multi-dimensional meteorological time-series data, three-dimensional elevation data, and unit operation data of the target wind farm is as follows: By deploying meteorological observation equipment at the target wind farm and connecting it to the numerical weather prediction system, multi-dimensional meteorological time-series data including wind speed, wind direction, temperature, air pressure, and air humidity are obtained. By using digital elevation models and on-site topographic surveys, three-dimensional elevation data of the target wind farm, including longitude, latitude, altitude, and terrain roughness, are obtained. The wind farm monitoring and data acquisition system acquires unit operation data including active power output, rotor speed, pitch angle, generator temperature, and gearbox oil temperature. It also performs time alignment and outlier cleaning on multi-dimensional meteorological time series data, three-dimensional elevation data, and unit operation data.

3. The wind power prediction method as described in claim 1, characterized in that, The specific process of constructing a spatiotemporal feature fusion network based on multi-dimensional meteorological time-series data and unit operation data, extracting spatial correlation features from the multi-dimensional meteorological time-series data and temporal dynamic features from the unit operation data, and then weighting and fusing the spatial correlation features and temporal dynamic features through an attention mechanism to obtain the fused feature vector is as follows: Multi-dimensional meteorological time-series data are input into a spatiotemporal feature fusion network composed of a multi-layer graph convolutional neural network. A spatial graph structure of meteorological parameters is constructed based on the location of meteorological stations in wind farms. The spatial gradient relationship of wind speed, air pressure and temperature between nodes is calculated through the adjacency matrix. The meteorological feature information of adjacent nodes is aggregated layer by layer, and the spatial correlation feature vector of multi-dimensional meteorological time-series data is output. The unit operation data is input into a time encoder composed of a bidirectional long short-term memory network. The unit operation data sequence is segmented by a sliding time window. The temporal dependencies of historical and future moments are captured by forward and backward LSTM units, respectively, and the temporal dynamic feature vector of the unit operation data is output. A dynamic feature fusion module based on a multi-head attention mechanism is constructed. The cosine similarity matrix between the spatially correlated feature vector and the temporally dynamic feature vector is calculated. The spatial-temporal attention weights are generated by normalization using the Softmax function. The two feature vectors are fused by weighted summation to obtain a fused feature vector with spatiotemporal correlation.

4. The wind power prediction method as described in claim 1, characterized in that, Real-time monitoring of sudden wind speed changes and abnormal unit operating conditions; when the prediction residual is detected to exceed the threshold, local model retraining is triggered and the output weight coefficients of the data-driven and physical models are updated.

5. A wind power prediction system, used to implement the wind power prediction method according to any one of claims 1-4, characterized in that, include: The data acquisition and processing module is used to acquire multi-dimensional meteorological time-series data, three-dimensional elevation data and unit operation data of the target wind farm, and to perform time alignment and outlier cleaning on the multi-dimensional meteorological time-series data, the three-dimensional elevation data and the unit operation data. The spatiotemporal feature fusion module includes a graph convolutional neural network and a bidirectional long short-term memory network, which is used to extract the spatial correlation features of multi-dimensional meteorological time series data and the temporal dynamic features of unit operation data, and generate a fused feature vector through a multi-head attention mechanism. The hybrid prediction model module includes a data-driven sub-model and a physics-driven sub-model, which are used to dynamically adjust the output weights of the two sub-models using an adaptive weighting algorithm to generate initial wind power prediction results. The terrain correction module is used to calculate the turbulence intensity compensation function based on three-dimensional elevation data and to correct the initial prediction results for micro-terrain disturbances. The prediction output and optimization module is used to output the final power prediction curve with confidence interval, and optimize the model parameters through a sliding time window feedback mechanism and a stochastic gradient descent algorithm. The anomaly monitoring and retraining module is used to detect sudden changes in wind speed and abnormal operating conditions of the unit in real time. When the prediction residual exceeds the threshold, it triggers local model retraining and updates the weight coefficients.

6. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform the wind power prediction method according to any one of claims 1-4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the wind power prediction method according to any one of claims 1-4.

Images