A wind power generation prediction method based on a hybrid deep learning model

Abstract

This invention discloses a wind power prediction method based on a hybrid deep learning model, comprising: constructing a multi-source heterogeneous dataset of the target wind farm, including measured data, weather forecast data, and global forecast system data; constructing a wind speed correction model for weather forecast data and a wind speed downscaling and correction model for global forecast system data for short-term and medium-term wind speed correction tasks, respectively, and outputting wind speed correction products as input wind speeds; dividing the wind speed range into low-wind-speed segments and high-wind-speed segments; constructing prediction models for low-wind-speed segments and high-wind-speed segments, respectively, and outputting predicted wind turbine power for the wind farm; determining the wind speed segment of the input wind speed for the target wind farm, and calculating the predicted wind power at the target prediction time or multiple future times; this invention aims to solve the problems of long-term forecast accuracy decay and model switching jumps in coastal wind farms, and achieve refined prediction of power generation for the next 10-15 days.

Description

Technical Field

[0001] This invention relates to the field of wind power prediction technology driven by meteorological data, and specifically to a wind power prediction method based on a hybrid deep learning model. Background Technology

[0002] As wind power plays an increasingly important role in the power system, accurate forecasting of wind power generation over the next 10 to 15 days (i.e., medium to long term) has become a key technological support for ensuring the safe and stable operation of the power grid and optimizing energy market dispatch. However, due to the uncertainty of medium- and long-term meteorological evolution, existing forecasting technologies face severe challenges in practical applications, primarily due to knowledge gaps in cross-time-domain meteorological data and the physical discontinuity of segmented power characteristics.

[0003] First, within the medium-term forecast domain, the accuracy of meteorological products declines non-linearly with increasing time. Particularly when the forecast period exceeds 10 days, products such as the Global Forecast System (GFS) not only exhibit low temporal resolution (e.g., 6 hours) but also significant systematic drift. Existing time-series downscaling methods often process data in isolation, failing to consider the anchoring effect of short-term high-precision correction data on medium-term low-resolution data. In actual scheduling, the lack of an effective cross-time-domain prior guidance mechanism will lead to significant logical gaps in forecast results before and after 10 days at the handover point, severely impacting the continuity of power generation forecasts.

[0004] Secondly, existing multi-source meteorological data fusion technologies fail to fully utilize the complementarity of different forecast time domains. In the short-term time domain (0-10 days), although numerical weather prediction (NWP) can achieve high accuracy through meta-learning correction, this correction experience has not been translated into prior knowledge to guide the calibration of medium-term (10-15 days) GFS data. Current technology lacks an innovative constraint parameter that can quantify the weight of "short-term guiding medium-term," resulting in a lack of physical basis for the medium-term wind speed correction process and an inability to effectively correct the path bias of GFS under long-term forecasts.

[0005] Finally, at the level of wind power modeling, the output characteristics of wind turbines exhibit high heterogeneity across wind speed ranges. Wind speeds are widely distributed in coastal areas, and the power fluctuation mechanisms in low-speed and high-speed segments are completely different. Using a single model makes it difficult to balance stability at low wind speeds with sensitivity at high wind speeds, while simple piecewise modeling is prone to producing "step-like" jumps in predicted power near wind speed switching thresholds. Current technologies lack a loss function with a smooth transition penalty mechanism, making it impossible to mathematically guarantee the physical consistency of piecewise models at switching boundaries.

[0006] In summary, in order to solve the above-mentioned technical challenges, there is an urgent need for a wind power forecasting method that integrates a cross-time-domain prior guidance mechanism, a segmented hybrid architecture, and smooth constraint optimization, so as to meet the accuracy requirements for 10-15 day medium- and long-term power forecasting under complex meteorological conditions. Summary of the Invention

[0007] The purpose of this invention is to provide a wind power generation prediction method based on a hybrid deep learning model. This method aims to solve the problems of long-term forecast accuracy decay and model switching jumps in coastal wind farms, and to achieve refined prediction of wind power generation in the next 10-15 days.

[0008] To achieve the above functions, this invention designs a wind power prediction method based on a hybrid deep learning model. For a target wind farm, the following steps S1-S4 are executed to complete the 10-15 day wind power prediction:

[0009] Step S1: Construct a multi-source heterogeneous dataset of the target wind farm, including measured data, weather forecast data, and global forecast system data; after spatiotemporal alignment and noise removal, and preprocessing, form a standard input tensor;

[0010] Step S2: Based on the standard input tensor, the wind speed correction task is divided into a short-term stage and a medium-term stage. For the short-term stage of 0-10 days, a wind speed correction model for weather forecast data is built based on a meta-learner. For the medium-term stage of 10-15 days, a wind speed downscaling and correction model for global forecast system data is built based on a residual learning model. The wind speed correction products are output as input wind speeds.

[0011] Step S3: Based on the input wind speed, the wind speed range is divided into low wind speed segment and high wind speed segment according to the wind farm turbine power curve; for the low wind speed segment, a low wind speed segment prediction model is constructed based on the multilayer perceptron direct prediction model; for the high wind speed segment, a high wind speed segment prediction model is constructed based on the XGBoost model. The low wind speed segment prediction model and the high wind speed segment prediction model are trained and the parameters are optimized to output the predicted power of the wind farm turbine.

[0012] Step S4: For the target wind farm, determine the wind speed segment to which the input wind speed belongs, calculate the segmented fused power, and further calculate the wind power prediction value at the target prediction time or multiple future times to complete the 10-15 day wind power prediction.

[0013] Beneficial effects: Compared with the prior art, the advantages of the present invention include:

[0014] 1. A cascaded correction architecture based on cross-temporal prior guidance coefficients was constructed. The cross-temporal prior guidance coefficients are dynamically adjusted with the forecast step size, which quantifies the constraint strength in the process of extrapolating short-term knowledge to medium-term forecasts and solves the problem of temporal connection discontinuity in traditional downscaling methods.

[0015] 2. A precise segmented prediction mechanism based on the physical characteristics of wind turbine power curves has been realized. This invention breaks through the limitation of a single model processing the entire wind speed range. By deeply deconstructing the correlation characteristics between wind speed and wind turbine power, a high degree of coupling between the prediction model and physical characteristics has been achieved.

[0016] 3. A dual-objective joint optimization strategy is introduced. In the model training stage, this invention innovatively proposes a parameter optimization method based on a dual-objective loss function. By introducing a piecewise collaborative penalty term on the basis of mean absolute error (MAE), a unified optimization framework is constructed, which effectively eliminates the "step-like" numerical jumps common in piecewise prediction and significantly improves the robustness and smoothness of 10-15 day full wind speed time domain wind power prediction. Attached Figure Description

[0017] Figure 1 This is a flowchart of a wind power prediction method based on a hybrid deep learning model according to an embodiment of the present invention;

[0018] Figure 2 This is a schematic diagram of the overall architecture for wind speed correction and downscaling provided in an embodiment of the present invention;

[0019] Figure 3 This is a schematic diagram of wind speed segment division and hybrid modeling based on the wind turbine power curve of a wind farm, provided according to an embodiment of the present invention;

[0020] Figure 4 This is a schematic diagram of the optimization mechanism of the dual-objective loss function provided in an embodiment of the present invention;

[0021] Figure 5 This is an integrated stitched image of the full-time domain prediction results for April 10-15, 2023, provided according to an embodiment of the present invention;

[0022] Figure 6 This is an integrated stitched image of the full-time domain prediction results for May 10-15, 2023, provided according to an embodiment of the present invention;

[0023] Figure 7 This is an integrated stitched image of the full-time domain prediction results for August 10-15, 2023, provided according to an embodiment of the present invention. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0025] This invention provides a wind power prediction method based on a hybrid deep learning model, referring to... Figure 1 For the target wind farm, perform the following steps S1-S4 to complete the 10-15 day wind power forecast:

[0026] Step S1: Construct a multi-source heterogeneous dataset of the target wind farm, including measured data, weather forecast data, and global forecast system data; after spatiotemporal alignment and noise removal, and preprocessing, form a standard input tensor;

[0027] The specific steps of step S1 are as follows:

[0028] Step S1.1: Construct a multi-source heterogeneous dataset for the target wind farm, including historical SCADA system measured data of the target wind farm within a preset time period, as well as 0-10 day Numerical Weather Prediction (NWP) data and 0-15 day Global Forecast System (GFS) data for the region where the target wind farm is located; perform spatiotemporal alignment and noise cleaning.

[0029] In this embodiment, the target wind farm is the Yancheng coastal wind farm. To ensure that the present invention has excellent universality, adaptability and engineering implementation effect when facing complex marine climates, this embodiment specifically selects a long-term multi-source heterogeneous dataset located in the coastal wind farm for empirical research. The data collection period strictly covers the complete cycle of seven consecutive months from March 2023 to September 2023. This cycle accurately covers typical extreme conditions such as the spring windy season, summer monsoon season and typhoon season in coastal areas, ensuring the diversity and completeness of the samples in terms of meteorological characteristics.

[0030] The multi-source heterogeneous dataset consists of three main parts: The first part is the historical SCADA system measured data of the target wind farm, including but not limited to the actual output power of the generator, the measured wind speed at the hub height, and the unit operating status, with the time resolution uniformly aligned to 15 minutes; the second part is the weather forecast (NWP) data covering the short-term time domain (0 to 240h, i.e. 0 to 10 days), which has a high sampling frequency of 15 minutes and serves as the accuracy anchor point for the short-to-medium-term dimension; the third part is the global forecast system (GFS) data covering the medium-to-long-term full time domain (0 to 360h, i.e. 0 to 15 days), with an initial time resolution of 6 hours, which serves as the basic feature stream for medium-to-long-term wind power prediction and provides input for subsequent downscaling processing.

[0031] Step S1.2: For the multi-source heterogeneous dataset, the physical threshold elimination method based on wind speed is used to remove outlier values; for the missing points in the multi-source heterogeneous dataset, the inverse distance weighted method (IDW) is used for interpolation imputation to complete the preprocessing of the multi-source heterogeneous dataset and form the standard input tensor.

[0032] To address the sensor data anomalies caused by high salt spray corrosion, lightning interference, and unstable communication links at the Yancheng coastal wind farm, this embodiment constructs a high-quality dataset using physical threshold removal and temporal interpolation algorithms. First, an outlier removal operation is performed, setting a maximum physical wind speed threshold according to the wind turbine safety operation specifications. When the observed wind speed exceeds the maximum physical wind speed threshold, it is identified as a non-physical jump anomaly and marked as missing. In the inverse distance weighted method (IDW), the weight index p is taken as an empirical value of 2, and four nearby valid observation points before and after the missing time point are selected to participate in the interpolation calculation to ensure the physical smoothness of the 15-minute interval data.

[0033] To address missing points in the data stream, a high-precision interpolation imputation method based on inverse distance weighting in the time dimension is employed. This constructs a physically reasonable dataset with 15-minute intervals, temporal continuity, and rigorous logic, while ensuring the continuity of physical evolution. This provides a true benchmark for subsequent cascaded corrections and segmented predictions.

[0034] Step S2: Based on the standard input tensor, the wind speed correction task is divided into a short-term stage and a medium-term stage. For the short-term stage of 0-10 days, a wind speed correction model for weather forecast data is built based on a meta-learner. For the medium-term stage of 10-15 days, a wind speed downscaling and correction model for global forecast system data is built based on a residual learning model. The wind speed correction products are output as input wind speeds.

[0035] Reference Figure 2 The specific steps of step S2 are as follows:

[0036] Step S2.1: For the wind speed correction task, the wind speed correction task is divided into a short-term phase of 0-10 days and a medium-term phase of 10-15 days;

[0037] Step S2.2: In the short-term phase, based on the weather forecast data, a wind speed correction model is constructed to remove biases from the weather forecast data, generating a wind speed correction product as an anchor reference. The meta-learner adopts a stacked generalization structure, which is an application based on existing ensemble learning frameworks and is trained based on the data of this invention.

[0038] The specific steps of step S2.2 are as follows:

[0039] Step S2.2.1: Construct a weather forecast data wind speed correction model based on a meta-learner. The meta-learner consists of different base learners. In this embodiment, it includes Random Forest (an existing open-source implementation model) and Extreme Gradient Boosting (XGBoost, an existing open-source implementation model). After each base learner is trained independently, its output is used as the input to the meta-learner. The number of trees in the Random Forest is set to 200, and the maximum depth is 8. The number of trees in the XGBoost model is set to 300, the maximum depth is 6, and the learning rate is 0.1. The initial weather forecast wind speed values ​​to be corrected are used by the base learners. Preliminary mapping is performed; initial weather forecast wind speed values ​​to be corrected. It refers to the raw wind speed sequence data output by the numerical weather prediction system, specifically the background wind speed forecast field at the latitude and longitude coordinates and hub height of the corresponding wind farm after time bilinear interpolation.

[0040] The meta-learner adopts a stacked generalization structure. Each base learner selects heterogeneous models such as random forest and gradient boosting regression, forming a decision tree. An explicit residual connection mechanism is introduced into each base learner as follows:

[0041] ;

[0042] In the formula, This represents the preprocessed weather forecast data. This represents a nonlinear feature transformation used to capture higher-order fluctuation features in wind speed sequences. This represents the weight matrix, used to learn the complex mapping relationship between input features and wind speed residuals. For activation functions; The residual weight matrix is ​​used to achieve linear projection of the input features, ensuring that the model can still retain effective information of the original wind speed signal as the depth increases; These are learnable bias term vectors used to enhance the model's ability to adapt to translational shifts in the predicted target distribution; This represents the feature map output by the base learner, which has residual compensation properties and is transmitted through a linear path. and nonlinear pathways The superposition of these values ​​resulted in the initial weather forecast wind speed values ​​to be corrected. Refined characterization of forecast bias;

[0043] The joint output matrix of all basic learners is expressed as follows:

[0044] ;

[0045] In the formula, These represent the wind speed prediction values ​​output by each basic learner. Defined as the first type of basic learner receiving preprocessed weather forecast data as input. Then, the wind speed prediction is output through a weighted voting process using its internal decision tree cluster; similarly, to Representing others Parallel prediction results generated by heterogeneous models. Each element All these correspond to wind speed predictions under the same spatiotemporal coordinates. By vectorizing and stacking these wind speed predictions with different bias characteristics, a multi-dimensional "expert suggestion" feature space is provided for the subsequent meta-learner. The index k is the number of basic learners, and the index k takes a value of 4. The joint output matrix represents, in a physical sense, the set of preliminary estimates of wind speed at the same forecast time by basic learners with different mechanisms. The robustness of the system is improved by the error complementarity of the heterogeneous models.

[0046] In this embodiment, the XGBoost model is used as the master learner (a direct call to an existing open-source model without structural modification), and the model parameters are set as follows: maximum depth. Learning rate Subsampling rate Column sampling rate The number of decision trees is 200, and the regularization parameter is 1; the input features are processed before entering the model. Normalization method to process interval;

[0047] Step S2.2.2: Construct the input feature vector for the wind speed correction model based on weather forecast data. as follows:

[0048] ;

[0049] This includes periodic time characteristics and sliding statistics of measured data; where, The initial weather forecast wind speed value is yet to be corrected. , These are the arithmetic mean and standard deviation of the measured wind speed of the target wind farm's SCADA system within the rolling window before the forecast time, respectively. Physically, they are used to characterize the systematic deviation benchmark of the local meteorological environment and the intensity of turbulent fluctuations in the airflow. The hour number (0-23) corresponding to the forecast time; It is a sine function. For cosine functions, through and The function will Mapped onto the unit circle, it is used to quantitatively characterize the diurnal periodic evolution of wind speed caused by the influence of thermal circulation;

[0050] Step S2.2.3: Construct the optimization objective function for the wind speed correction model based on weather forecast data. As shown in the following formula:

[0051] ;

[0052] in, This represents the actual wind speed value obtained by the monitoring equipment at the target wind farm at time t. For meta-learners in The fitted wind speed values ​​output at each time step are used to optimize the objective function. The first term in the equation is the mean squared error loss function that minimizes the prediction error, where N is the total number of time points. In order to target the Internal decision tree of the individual learner The regularization penalty term is set to a coefficient of 0.1; the model training uses a 7:2:1 ratio of training set, validation set, and test set, and adopts an early stopping strategy, stopping training when the validation set error does not decrease within 10 consecutive rounds; M is the total number of decision trees.

[0053] Penalty terms are applied via regularization. Wind speed correction products from weather forecast data wind speed correction models :

[0054]

[0055] In the formula, This represents the fitted wind speed value output by the meta-learner; This indicates a product with wind speed correction.

[0056] This formula represents the fitted wind speed value calculated by the meta-learner. As the final short-term wind speed correction product This value not only incorporates the characteristics of multi-source numerical forecasts, but also effectively eliminates the spatiotemporal representativeness error of the original weather forecast data in local geographical environments by introducing measured statistics and time characteristics.

[0057] Step S2.3: In the medium-term phase, wind speed correction products are generated based on global forecast system data and within a preset historical time period. As prior guidance knowledge, a prior guidance coefficient with cross-temporal guidance coefficients is constructed based on the Cascaded Residual Learning model. A global forecast system (GRS) data wind speed downscaling and correction model was developed. The residual learning model employs an improved ResNet structure, a shallow modification of the existing ResNet network. Instead of pre-trained weights, the network is retrained. The ResNet network has six fully connected layers with 128, 128, 64, 64, 32, and 16 neurons per layer, respectively. The ReLU activation function is used, the Adam optimizer is employed, the learning rate is set to 0.001, the batch size is 64, and the training iterations are 150. This model performs temporal downscaling and numerical calibration on GRS data within a preset time range, outputting a corrected wind speed product. ;

[0058] The specific steps of step S2.3 are as follows:

[0059] Step S2.3.1: To address the sparsity of the 6-hour timescale data in the medium-term (10-15 days) global forecast system data, perform time-series downscaling based on piecewise cubic Hermitian interpolation (PCHIP) to maintain the monotonicity and physical continuity of wind speed evolution, obtaining the interpolated 15-minute resolution baseline background wind speed values. ;

[0060] Step S2.3.2: Construct the residual to be predicted at time t As shown in the following formula:

[0061] ;

[0062] In the formula, The endogenous weighting coefficient. This indicates the predicted wind speed increment. The background wind speed baseline value after interpolation at time t; Indicates the step size; Represents the prior guidance coefficients across the time domain; Indicates wind speed correction product The anchor point value at the end, Indicates the end time of the weather forecast data; The decay characteristic time constant is used to characterize the exponential decay rate of cross-temporal prior guidance information with increasing forecast step size. The cross-temporal prior guidance coefficients are... Set as: Endogenous weighting coefficient Set to 0.3; decay characteristic time constant Set to 24; predict wind speed increment The output of the above 6-layer ResNet network; This represents an exponential function with the natural constant e as its base.

[0063] Defined as the wind speed correction product output in step S2.2 The final wind speed correction value at the 240th hour (i.e. the end time of the short-term weather forecast data) serves as the "time domain anchor data" for the entire forecast cycle, providing a physical starting reference for the medium-term forecast based on high-precision measured corrections, and solving the problem of numerical discontinuity between the start time of the medium-term forecast and the end time of the short-term forecast.

[0064] Endogenous weighting coefficient At the data level, it is used to adjust the predicted wind speed increment learned by the residual neural network (ResNet) based on mid-term GFS features. In terms of its proportion in the overall residual composition, it physically acts as a regulating valve to balance the weights of "short-term experience-guided" and "medium-term feature-driven" approaches, ensuring that as the step size increases, the model can smoothly evolve from relying on short-term anchor points to relying on the deep correction of its own mid-term features.

[0065] Step S2.3.2: Construct the input feature vector for the global forecast system data wind speed downscaling and correction model. As shown in the following formula:

[0066] ;

[0067] In the formula, This is the interpolated baseline value for the background field wind speed. To reflect the historical lag characteristic, the first three moments were selected. All input features are standardized (mean 0, standard deviation 1) to compensate for the time lag effect of atmospheric circulation. Indicates latitude, Longitude is represented, which constitutes the model's spatial location perception capability, enabling the model to identify the impact of the specific geographical environment of coastal wind farms on wind speed; Indicates the wind direction angle; Indicates the step size, used to distinguish the prediction bias characteristics under different timeframes; This represents the predicted wind speed increment, which is the first-order difference feature of wind speed, reflecting the slope and evolution of wind speed over time. and This refers to the wind direction angle. By performing trigonometric function mapping, the wind direction is transformed into a continuous vector representation, thus accurately describing the vector characteristics of wind direction in a two-dimensional plane.

[0068] Step S2.3.3: Output the wind speed downscaling and correction product of the Global Forecasting System data wind speed correction model. As shown in the following formula:

[0069] ;

[0070] In the formula, The wind speed correction product at time t , Represents the residual at time t. Represents the ReLU activation function, used as a physical non-negativity constraint term to ensure the result To ensure the rigor of training on coastal wind farm data, a rolling cross-validation strategy was adopted, with a time window length of 30 days and a step size of 5 days. The training set... With test set Strictly adhere to physical isolation according to the forecast time (forecast_time): This effectively avoids information leakage and ensures the numerical stability and generalization accuracy of the model under ultra-long forecast lead times of 10-15 days.

[0071] The present invention defines the key time parameters as follows: It refers to the forecast start time recorded in the database of sample data, that is, the physical point in time when the meteorological characteristics are observed or generated. This is defined as the absolute start timestamp of the test set within the current validation period. Physically, it means that during model training, only historical samples whose predicted times are earlier than the test start point are allowed to be retrieved. The model is strictly prohibited from accessing any "future information" occurring after the test node. This is based on... The physical isolation measures effectively avoid information leakage caused by time series correlation, ensuring the numerical stability and generalization accuracy of the model under ultra-long forecast lead times of 10-15 days.

[0072] Wind speed correction product , As input wind speed.

[0073] In step S2, time-series downscaling and numerical calibration are performed on the low-resolution global forecast system data series within the range of 240h-360h. The "anchor effect" of the short-term correction results is used to constrain the drift direction of the medium-term forecast, thereby achieving the inherent unity of cross-time-dependent wind speed prediction accuracy and physical logic at the mathematical level.

[0074] Step S3: Based on the input wind speed, the wind speed range is divided into low-wind-speed and high-wind-speed segments according to the wind farm turbine power curve. For the low-wind-speed segment, a low-wind-speed segment prediction model is constructed based on the multilayer perceptron direct prediction model. For the high-wind-speed segment, a high-wind-speed segment prediction model is constructed based on the XGBoost model. The low-wind-speed segment prediction model and the high-wind-speed segment prediction model are trained and their parameters are optimized to output the predicted power of the wind farm turbines. Among them, the multilayer perceptron direct prediction model is a direct regression improvement model based on the standard feedforward neural network. It does not use pre-trained weights and is trained entirely based on the data of this invention. The XGBoost model is an existing open-source implementation model. Its basic structure is not changed, only the parameters are set and trained.

[0075] The specific steps of step S3 are as follows:

[0076] Step S3.1: Refer to Figure 3 Based on the wind speed correction product, a wind speed sample set is constructed. Power output of wind turbine With wind speed There is a significant difference between them. The piecewise nonlinear relationship typically includes three key stages: cut-in, rated, and cut-out. Specifically, in the low wind speed range (near the cut-in wind speed), the wind turbine generator has not yet reached a stable operating state, and the power increases slowly with wind speed. However, it is significantly affected by wind speed measurement errors and turbulence noise, leading to high prediction uncertainty and modeling difficulty in the low wind speed range. In the high wind speed range near the rated wind speed, power changes are relatively drastic, with a high degree of nonlinearity, and it is easily affected by meteorological disturbances and turbine control strategies. Based on these physical characteristics, wind speed thresholds are set according to the key points of the wind farm turbine power curve. , wind speed sample set Divided into low wind speed zone High wind speed section as follows:

[0077] ;

[0078] ;

[0079] In the formula, This represents the input wind speed at time t, and the value is the wind speed correction product at time t. or ; The wind speed threshold is typically set as an empirically optimal value between the wind turbine's cut-in wind speed and its rated wind speed, or determined through data-driven analysis by minimizing the integrated mean square error (IMSE) of the piecewise prediction error. This piecewise approach decomposes the complex regression problem across the entire wind speed range into two sub-regression problems of varying difficulty. This allows subsequent models to differentiate their modeling of power variation characteristics within different wind speed ranges, thereby significantly improving the accuracy and robustness of the overall prediction performance.

[0080] Step S3.2: Considering the characteristics of wind farm turbine power output variation being gradual and relatively low in nonlinearity but with high noise levels in low wind speed sections, a low wind speed section prediction model is constructed based on the Multilayer Perceptron Direct Prediction (MLP-D) model. This model is used for direct regression prediction of wind farm turbine power in low wind speed sections. The MLP-D model has advantages such as simple structure, stable convergence, and suitability for fitting gradual nonlinear relationships. The input feature vector at time t... As shown in the following formula:

[0081] ;

[0082] In the formula, This represents the temperature at time t. This represents the humidity at time t. Indicates time periodicity characteristics; , These represent the input wind speeds at times t-1 and t-2, respectively.

[0083] in, Defined as the wind speed correction product output from the global forecast system data downscaling and correction model constructed in step S2.3; this data is introduced into the power prediction model as a "medium-term time-domain guiding feature," and its physical significance lies in providing the model with information on the wind energy evolution trend over a span of 10-15 days, along with the high-precision correction results representing the short-term (0-10 days). Together, they constitute a feature coverage across the entire time scale, ensuring that the low-wind-speed prediction model can perceive the long-range evolution constraints from the Global Forecast System (GFS) after physical correction when dealing with the power mapping of wind turbines in low-wind-speed wind farms under different forecast lead times.

[0084] The low-wind-speed prediction model consists of 4 hidden layers. The number of neurons in each layer is 128, 128, 64, and 32, respectively, and the ReLU function is used uniformly; the model parameters are initialized using the He initialization method; the optimizer is... The optimizer has a learning rate of 0.001, a batch size of 64, and a training epoch of 200. It employs an early stopping strategy (training stops if the validation set error does not decrease for 10 consecutive epochs). Its forward propagation process is expressed as follows:

[0085] ;

[0086] in, This represents the output vector of the l-th hidden layer at time t. denoted as the output vector of the (l-1)th hidden layer at time t, its physical meaning is the deep abstract expression of the input features after high-dimensional space mapping, used to extract the nonlinear relationship between wind speed and wind turbine power in wind farms layer by layer; and Here, are the weight matrix and bias term of the l-th hidden layer, respectively. This determines the strength and direction of signal transmission between neurons, and This is responsible for shifting and adjusting the input of the activation function to enhance the model's flexibility in fitting small power fluctuations in the low wind speed range of wind farms; L represents the total number of hidden layers, which determines the depth at which the network can mine complex nonlinear features.

[0087] When l=1, Represents the input feature vector The weight mapping matrix between the first hidden layer and the first hidden layer. for The corresponding bias term, This represents the high-dimensional latent space representation of the original physical input variables after the first linear transformation and ReLU nonlinear activation. Its function is to complete the initial mapping from the input space to the hidden feature space; when At that time, each hidden layer further performs deep feature abstraction and reconstruction based on the output of the previous layer, so as to realize the layer-by-layer enhanced expression of the subtle pattern of wind turbine power change in wind farms in low wind speed section.

[0088] The final output layer uses a linear transformation to output the predicted power of wind turbines in low-wind-speed wind farms:

[0089] ;

[0090] in, Let be the weight matrix of the output layer, and This is the bias term for the output layer, defined as the intercept correction parameter for the predicted power values ​​of wind turbines in the wind farm. It ensures that the hidden layer features after linear transformation can accurately regress to the physically meaningful power value space. This represents the output vector of the Lth hidden layer at time t. This represents the predicted power of wind turbines in the wind farm output by the prediction model for the low wind speed segment at time t.

[0091] To improve the model's robustness in low wind speed ranges and the smoothness of segment transitions, refer to Figure 4This invention introduces a dual-objective loss function to optimize model parameters during model training. Based on the mean absolute error (MAE) loss, the dual-objective loss function introduces a segment collaboration penalty term, specifically designed for prediction models in low wind speed ranges. As shown in the following formula:

[0092] ;

[0093] In the formula, This represents the core prediction error term. This indicates a segmented collaborative penalty term. This represents a hyperparameter, set to 0.5, determined through optimization using a grid search on the validation set.

[0094] Core prediction error term As shown in the following formula:

[0095] ;

[0096] In the formula, This represents the predicted power of wind turbines in the wind farm output by the prediction model for the low wind speed segment at time t. Let t be the measured power of the wind turbines in the wind farm. Indicates a low wind speed range;

[0097] Segmented Collaborative Penalty Items This is used to constrain the behavior of the prediction model in low-wind-speed segments when dealing with boundary and high-wind-speed samples (even if these samples are not used for its primary training objective), enhance the model's perception of the overall wind farm turbine power curve, and avoid exceeding wind speed thresholds. Abrupt prediction jumps occur nearby. Segmented collaborative penalty term. As shown in the following formula:

[0098] ;

[0099] In the formula, This indicates the predicted power of the wind turbines in the i-th high-wind-speed sample using the low-wind-speed prediction model. Let be the measured wind turbine power of the wind farm for the i-th high wind speed sample. For a learnable weight coefficient (such as a scalar or function optimized via gradient descent), it can be derived from the i-th high wind speed sample to the wind speed threshold. The distance is dynamically adjusted to ensure that the segmented collaborative penalty is mainly concentrated on the wind speed threshold. On nearby samples; Indicates a high wind speed range;

[0100] Using a dual-objective loss function The design constrains the prediction model for low-wind-speed segments to avoid excessive prediction bias for high-wind-speed segments during training, thereby achieving soft coordination between the prediction results of the two segmented models and enhancing the continuity and smoothness of the overall power prediction.

[0101] Step S3.3: Addressing the characteristics of drastic power variations, complex nonlinear relationships, and susceptibility to strong disturbances within high-wind-speed ranges, a prediction model for these ranges is constructed based on the XGBoost model using Gradient Boosting Decision Tree. The XGBoost model possesses strong nonlinear fitting capabilities, robustness to feature scales, and a built-in regularization mechanism, making it highly suitable for accurately fitting the complex and variable mapping relationship between wind speed and wind farm turbine power within high-wind-speed ranges. The input feature vector of the high-wind-speed prediction model is shown below. as follows:

[0102] ;

[0103] exist Building upon this foundation, the study further emphasizes the multi-step lag information and nonlinear transformation of wind speed to capture the temporal dependence of rapidly changing high-wind-speed segments.

[0104] The XGBoost model, based on an additive model structure, outputs the predicted power of wind turbines in high-wind-speed wind farms, expressed in the form of: The accumulation of decision trees:

[0105] ;

[0106] in, Indicates the first The prediction function of a decision tree. The number of decision trees; The predicted power of wind turbines in the high-wind-speed segment at time t represents the expected active power output of the wind turbines under high-intensity wind energy input at that time. The predicted power of wind turbines in the high-wind-speed segment is determined by... Regression trees are used for input feature vectors The resulting residual increments are obtained by summing them up tree by tree, for each decision tree. Based on the prediction error of the previous decision tree, targeted fitting is performed, thereby achieving a high-precision approximation of the complex and violently fluctuating power characteristics in the high wind speed range through additive model logic;

[0107] Constructing a regularized objective function for the prediction model in the high wind speed range As shown in the following formula:

[0108] ;

[0109] in, This represents the total number of time points used for training samples in the high-wind-speed segment, and is used to characterize the sample size participating in the training of the high-wind-speed power prediction model. The loss function; For regularization penalty terms, The number of decision trees in the high-wind-speed prediction model; Let t be the measured power of the wind turbines in the wind farm at time t, which serves as the ground truth label for model training;

[0110] loss function The mean squared error (MSE) is calculated using the following formula:

[0111] ;

[0112] The physical meaning of this loss function lies in quantitatively calculating the squared deviation between the model output and the actual output, thereby providing an accurate gradient direction for optimizing model parameters under high wind speed conditions.

[0113] Regularization penalty term To suppress model complexity and avoid overfitting, it is calculated as follows:

[0114] ;

[0115] in, The number of leaf nodes in the decision tree; The weighted penalty coefficient for the number of leaf nodes in the decision tree. The score for the leaf nodes. The weighted penalty coefficient for penalizing the scores of leaf nodes; It is an L2 norm.

[0116] By minimizing the regularization penalty term, the XGBoost model can maintain prediction accuracy while... Restrictions on decision tree structure and By constraining the numerical smoothness and maintaining the optimal complexity of the model structure, accurate and stable predictions of wind turbine power in wind farms are achieved in the high wind speed range.

[0117] Step S4: For the target wind farm, determine the wind speed segment to which the input wind speed belongs, calculate the segmented fused power, and further calculate the wind power prediction value at the target prediction time or multiple future times to complete the 10-15 day wind power prediction.

[0118] The specific steps of step S4 are as follows:

[0119] Step S4.1: Determine the input wind speed The wind speed range to which it belongs, if Then the low-wind-speed prediction model is activated, and the predicted power of the wind turbines in the wind farm at time t is output. ;like Then the high-wind-speed prediction model is activated, and the predicted power of the wind turbines in the wind farm at time t is output. Final segmented fusion power Represented as:

[0120] ;

[0121] In the formula, The indicator function is used; the segmented fusion power effectively eliminates the "step" power fluctuations generated at the switching moment of the traditional segmented model, ensuring the physical consistency between the wind farm turbine power prediction curve and the actual power characteristic curve of the wind farm turbine.

[0122] Step S4.2: For a single prediction step size In this example, the time frame is 0.25 hours, and the instantaneous predicted wind power output is calculated. For medium-term dispatch demand of 10-15 days (240h-360h), the cumulative predicted wind power output is output. Complete 10-15 day wind power generation forecast.

[0123] This embodiment uses measured data from the target coastal wind farm from March to September 2023 as the verification basis. Through deep collaboration between the cascaded correction architecture and the piecewise hybrid model, the accuracy of medium- and long-term predictions is significantly improved. (Refer to...) Figure 5 , Figure 6 , Figure 7 These are integrated stitched images of the full-time domain prediction results for April, May, and August 2023, spanning 10-15 days respectively. Experimental results show that the introduction of cross-time domain prior guidance coefficients... Subsequently, the root mean square error (RMSE) of wind speed forecast in the 10-15 day time domain was reduced by 12.4%, effectively overcoming the problem of divergence of native GFS data over time; while in the power modeling stage, the prediction bias of the model near the rated wind speed was reduced by 8.7%, and the average rate of change of the wind turbine power prediction curve of the wind farm was more in line with the actual operating trajectory of the unit.

[0124] This embodiment yielded the following research results: Traditional medium- and long-term wind power forecasting methods typically exhibit limitations when processing long-term meteorological data. In multi-day forecasts, traditional methods often employ simple linear interpolation or direct mapping using a single model, employing coarse ensemble forecast data to extrapolate future power trends. However, this method has a key limitation: as the forecast lead time extends (especially beyond 10 days), the error in numerical weather prediction exhibits non-linear divergence, and the coarse-grained temporal resolution masks the details of wind speed fluctuations, significantly impacting the final wind power forecast results and making it difficult to meet the needs of refined grid dispatching.

[0125] In summary, this invention not only achieves deep fusion of multi-source heterogeneous data in mathematical form, but also ensures the continuity and stability of wind power prediction in physical mechanism, providing a high-confidence decision-making basis for medium- and long-term energy trading and power balance scheduling in the power grid.

[0126] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A wind power generation prediction method based on a hybrid deep learning model, characterized in that, For the target wind farm, perform the following steps S1-S4 to complete the 10-15 day wind power forecast: Step S1: Construct a multi-source heterogeneous dataset of the target wind farm, including measured data, weather forecast data, and global forecast system data; after spatiotemporal alignment and noise removal, and preprocessing, form a standard input tensor; Step S2: Based on the standard input tensor, the wind speed correction task is divided into a short-term stage and a medium-term stage. For the short-term stage of 0-10 days, a wind speed correction model for weather forecast data is built based on a meta-learner. For the medium-term stage of 10-15 days, a wind speed downscaling and correction model for global forecast system data is built based on a residual learning model. The wind speed correction products are output as input wind speeds. Step S3: Based on the input wind speed, the wind speed range is divided into low wind speed segment and high wind speed segment according to the wind farm turbine power curve; for the low wind speed segment, a low wind speed segment prediction model is constructed based on the multilayer perceptron direct prediction model; for the high wind speed segment, a high wind speed segment prediction model is constructed based on the XGBoost model. The low wind speed segment prediction model and the high wind speed segment prediction model are trained and the parameters are optimized to output the predicted power of the wind farm turbine. Step S4: For the target wind farm, determine the wind speed segment to which the input wind speed belongs, calculate the segmented fused power, and further calculate the wind power prediction value at the target prediction time or multiple future times to complete the 10-15 day wind power prediction.

2. The wind power prediction method based on a hybrid deep learning model according to claim 1, characterized in that, The specific steps of step S1 are as follows: Step S1.1: Construct a multi-source heterogeneous dataset for the target wind farm, including measured data from the historical SCADA system of the target wind farm within a preset time period, as well as weather forecast data covering 0-10 days for the region where the target wind farm is located and global forecast system data covering 0-15 days; perform spatiotemporal alignment and noise cleaning. Step S1.2: For the multi-source heterogeneous dataset, the physical threshold elimination method based on wind speed is used to remove outlier values; for the missing points in the multi-source heterogeneous dataset, the inverse distance weighting method is used for interpolation imputation, thus completing the preprocessing of the multi-source heterogeneous dataset and forming the standard input tensor.

3. The wind power prediction method based on a hybrid deep learning model according to claim 2, characterized in that, The specific steps of step S2 are as follows: Step S2.1: For the wind speed correction task, the wind speed correction task is divided into a short-term phase of 0-10 days and a medium-term phase of 10-15 days; Step S2.2: In the short-term phase, based on the weather forecast data, a wind speed correction model for the weather forecast data is constructed using a meta-learner. Bias is removed from the weather forecast data, and a wind speed correction product is generated as an anchor reference. ; Step S2.3: In the medium-term phase, wind speed correction products are generated based on global forecast system data and within a preset historical time period. As prior guidance knowledge, a prior guidance coefficient with cross-temporal guidance is constructed based on the residual learning model. A global forecast system (GCS) data wind speed downscaling and correction model is used to perform temporal downscaling and numerical calibration on GCS data within a preset time range, outputting wind speed correction products. ; Wind speed correction product , As input wind speed.

4. The wind power prediction method based on a hybrid deep learning model according to claim 3, characterized in that, The specific steps of step S2.2 are as follows: Step S2.2.1: Construct a weather forecast data wind speed correction model based on a meta-learner. The meta-learner consists of multiple different base learners, forming a decision tree. An explicit residual connection mechanism is introduced into each base learner as follows: ； In the formula, This represents the preprocessed weather forecast data. Represents nonlinear characteristic transformation. Represents the weight matrix. For activation functions; This is the residual weight matrix; These are learnable bias term vectors; This represents the feature map that is the final output of the base learner; The joint output matrix of all basic learners is expressed as follows: ； In the formula, These represent the wind speed prediction values ​​output by each basic learner, with the subscript k indicating the number of basic learners. Represents the joint output matrix; Step S2.2.2: Construct the input feature vector for the wind speed correction model based on weather forecast data. as follows: ； In the formula, The initial weather forecast wind speed value is yet to be corrected. , These represent the arithmetic mean and standard deviation of the measured wind speeds by the SCADA system of the target wind farm within the rolling window preceding the forecast time, respectively. The hour number corresponding to the forecast time on that day; It is a sine function. It is a cosine function; Step S2.2.3: Construct the optimization objective function for the wind speed correction model based on weather forecast data. As shown in the following formula: ； in, express The real wind speed value obtained at all times by the monitoring equipment of the target wind farm. For meta-learners in The fitted wind speed value output at each time point, where N is the total number of time points; In order to target the Internal decision tree of the individual learner The regularization penalty term, M is the total number of decision trees; Penalty terms are applied via regularization. Wind speed correction products from weather forecast data wind speed correction models : ； In the formula, This represents the fitted wind speed value output by the meta-learner; This indicates a product with wind speed correction.

5. The wind power prediction method based on a hybrid deep learning model according to claim 4, characterized in that, The specific steps of step S2.3 are as follows: Step S2.3.1: For the global forecast system data, perform time-series downscaling based on piecewise cubic Hermitian interpolation to obtain the interpolated background wind speed baseline values. ; Step S2.3.2: Construct the residual to be predicted at time t As shown in the following formula: ； In the formula, The endogenous weighting coefficient. This indicates the predicted wind speed increment. The background wind speed baseline value after interpolation at time t; Indicates the step size; Represents the prior guidance coefficients across the time domain; Indicates wind speed correction product The anchor point value at the end; Indicates the end time of the weather forecast data; The decay characteristic time constant; This represents an exponential function with the natural constant e as its base. Step S2.3.2: Construct the input feature vector for the global forecast system data wind speed downscaling and correction model. As shown in the following formula: ； In the formula, This is the interpolated baseline value for the background field wind speed. Characterized by historical lag, Indicates latitude, Indicates longitude. Indicates the wind direction angle; Step S2.3.3: Output the wind speed downscaling and correction product of the Global Forecasting System data wind speed correction model. As shown in the following formula: ； In the formula, The wind speed correction product at time t , Represents the residual at time t. This represents the ReLU activation function.

6. The wind power generation prediction method based on a hybrid deep learning model according to claim 5, characterized in that, The specific steps of step S3 are as follows: Step S3.1: Construct a wind speed sample set based on the wind speed correction product. Wind speed thresholds are set based on key points in the wind turbine power curves of the wind farm. , wind speed sample set Divided into low wind speed zone High wind speed section as follows: ； ； In the formula, This represents the input wind speed at time t, and the value is the wind speed correction product at time t. or ; It is the wind speed threshold; Step S3.2: For the low wind speed segment, construct a low wind speed segment prediction model based on the multilayer perceptron direct prediction model, where the input feature vector at time t is... As shown in the following formula: ； In the formula, This represents the temperature at time t. This represents the humidity at time t. Indicates time periodicity characteristics; , These represent the input wind speeds at times t-1 and t-2, respectively. The low-wind-speed prediction model is based on The system consists of stacked hidden layers, and its forward propagation process is expressed as follows: ； in, This represents the output vector of the l-th hidden layer at time t. This represents the output vector of the (l-1)th hidden layer at time t; and Here, represents the weight matrix and bias term of the l-th hidden layer, respectively; L represents the total number of hidden layers; when l=1, ... Represents the input feature vector The weight mapping matrix between the first hidden layer and the first hidden layer. for The corresponding bias term, This represents the original physical input variables after the first linear transformation and... High-dimensional latent space representation after nonlinear activation; The final output layer uses a linear transformation to output the predicted power of wind turbines in low-wind-speed wind farms: ； in, Let be the weight matrix of the output layer, and For the bias term of the output layer, This represents the output vector of the Lth hidden layer at time t. This represents the predicted power of wind turbines in the wind farm output by the prediction model for the low wind speed segment at time t. Construct a bi-objective loss function for the prediction model in the low wind speed range. As shown in the following formula: ； In the formula, This represents the core prediction error term. This indicates a segmented collaborative penalty term. Indicates hyperparameters; Core prediction error term As shown in the following formula: ； In the formula, This represents the predicted power of wind turbines in the wind farm output by the prediction model for the low wind speed segment at time t. Let t be the measured power of the wind turbines in the wind farm. Indicates a low wind speed range; Segmented Collaborative Penalty Items As shown in the following formula: ； In the formula, This indicates the predicted power of the wind turbines in the i-th high-wind-speed sample using the low-wind-speed prediction model. Let represent the measured power of the wind turbines in the wind farm for the i-th high wind speed sample. It is a learnable weight coefficient; Indicates a high wind speed range; Step S3.3: For high wind speed sections, construct a high wind speed section prediction model based on the XGBoost model using gradient boosting decision trees. The input feature vector of the high wind speed section prediction model... as follows: ； The XGBoost model, based on an additive model structure, outputs the predicted power of wind turbines in high-wind-speed wind farms, expressed in the form of: The accumulation of decision trees: ； in, Indicates the first The prediction function of a decision tree. The number of decision trees; The predicted power of wind turbines in the wind farm is output by the prediction model for the high wind speed segment at time t. Constructing a regularized objective function for the prediction model in the high wind speed range As shown in the following formula: ； in, This represents the total number of time points in the training samples during the high-wind-speed segment. The loss function; For regularization penalty terms, The number of decision trees in the high-wind-speed prediction model; Let t be the measured power of the wind turbines in the wind farm. loss function The calculation is as follows: ； Regularization penalty term The calculation is as follows: ； in, The number of leaf nodes in the decision tree; The weighted penalty coefficient for the number of leaf nodes in the decision tree. The score for the leaf nodes. The weighted penalty coefficient for penalizing the scores of leaf nodes; It is an L2 norm.

7. The wind power prediction method based on a hybrid deep learning model according to claim 6, characterized in that, The specific steps of step S4 are as follows: Step S4.1: Determine the input wind speed The wind speed range to which it belongs, if Then the low-wind-speed prediction model is activated, and the predicted power of the wind turbines in the wind farm at time t is output. ;like Then the high-wind-speed prediction model is activated, and the predicted power of the wind turbines in the wind farm at time t is output. Final segmented fusion power Represented as: ； In the formula, For indicator functions; Step S4.2: For a single prediction step size Output instantaneous predicted wind power output For medium-term dispatch needs of 10-15 days, output the cumulative predicted wind power generation. Complete 10-15 day wind power generation forecast.

Images