Multi-agent stacked integration offshore wind power prediction method and system based on dual learner

Abstract

The application discloses a kind of multi-agent stacked integrated offshore wind power prediction method and system based on dual learning ware.The method obtains offshore wind farm historical data to carry out outlier processing and feature reconstruction to construct training set;Build double-layer stacked integrated learning model, its basic learning ware layer contains multiple heterogeneous basic learning ware, and the meta-learner layer is the dual parallel structure containing nonlinear and linear meta-learner;Training set is input into basic learning ware layer and is predicted and spliced to obtain meta-feature;Meta-feature is input into meta-learner layer, and nonlinear and linear predicted values are obtained respectively;Finally, the final prediction result is obtained by integrating the two predicted values using a weighted fusion strategy.The application considers the nonlinear fitting capability and the robustness of model structure, improves the generalization ability and effectively reduces the prediction error.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation and artificial intelligence, and in particular to a multi-agent stacked integrated offshore wind power prediction method and system based on a binary learner. Background Technology

[0002] As the global carbon neutrality process continues to advance, renewable energy plays a vital role in modern power systems. Compared to onshore wind power, offshore wind power has advantages such as high wind speed, high wind energy density, and no land occupation, making it an important direction for future wind power development. However, due to the complex and variable offshore environment, wind energy resources exhibit strong randomness and intermittency, making it difficult to accurately predict wind power output. This not only hinders the optimization of power system dispatch but may also threaten the safe and stable operation of the power grid.

[0003] Existing wind power prediction methods are mainly divided into physical methods, statistical methods, and artificial intelligence methods: 1. Physical methods rely on the topography and weather information of wind farms to build physical models, but the modeling process is complex, highly dependent on the accuracy of input data, and has a large computational load, making it difficult to guarantee prediction accuracy in real-world scenarios.

[0004] 2. Statistical methods (such as Markov chain models and autoregressive models) are highly dependent on the data distribution characteristics and linear relationships. When encountering extreme scenarios such as complex weather or sudden changes in wind speed, they are difficult to accurately capture the power change patterns, leading to a significant decrease in prediction performance.

[0005] 3. Single artificial intelligence models (such as BP neural networks, CNNs, LSTMs, etc.) can automatically learn the mapping relationships between data. However, the changes in offshore wind power involve multi-dimensional meteorological coupling and deep spatiotemporal dependence. Traditional single prediction models can often only characterize data features from a single dimension, making it difficult to fully capture the complex and diverse fluctuation characteristics of offshore wind power generation.

[0006] To overcome the limitations of single models, the industry has gradually begun to adopt stacking ensemble learning frameworks, combining multiple base models. However, existing conventional stacking ensemble methods typically use only a single model for fitting and integrating results at the top level (i.e., the meta-learner layer). Since the feature matrices output by the underlying base learners contain highly complex spatial, temporal, and nonlinear information, a single meta-learner, when facing scenarios like offshore wind power with strong nonlinearity and high volatility, is ill-suited to the following scenarios: if a simple linear model is used, it cannot deeply explore the complex nonlinear relationships in the residuals of the base models; if a complex tree model or deep network is used, it is prone to overfitting, failing to consider the structural robustness of the model, resulting in insufficient generalization ability under extreme or boundary conditions.

[0007] Furthermore, in the conventional data preprocessing stage, existing technologies often directly use time and wind direction as linear numerical inputs, which can easily lead to distortion of periodic boundaries; and they fail to effectively identify invalid meteorological data with zero power when the wind turbine is in a real shutdown state, resulting in noise from abnormal operating conditions severely interfering with the stable training of the model.

[0008] In summary, how to effectively reconstruct and process the raw offshore wind power data in a targeted manner, and how to break through the limitations of single-element learners in the integrated learning framework to construct a new fusion prediction architecture that takes into account both nonlinear feature mining and robust linear combination, so as to further improve the accuracy and robustness of offshore wind power prediction, are the technical problems that urgently need to be solved in this field. Summary of the Invention

[0009] To address the shortcomings of the existing technologies, the technical problem to be solved by this invention is: how to provide a multi-agent stacked integrated offshore wind power prediction method and system based on a binary learner, so as to overcome the shortcomings of existing single prediction models in dealing with the strong volatility and complex spatiotemporal coupling characteristics of offshore wind power, and at the same time solve the problems of overfitting or underfitting and insufficient generalization ability of single learners in conventional stacked integrated methods.

[0010] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A multi-agent stacked ensemble offshore wind power prediction method based on a binary learner includes the following steps: S1. Obtain historical meteorological data and historical wind power data of offshore wind farms, perform outlier processing and feature reconstruction, and construct a training set; S2. Construct a two-layer stacked ensemble learning model, which includes a first-layer base learner layer and a second-layer meta-learner layer; wherein, the base learner layer includes multiple heterogeneous base learners, and the meta-learner layer is a dual-parallel network structure, including parallel nonlinear meta-learners and linear meta-learners. S3. Input the training set into each basic learner in the basic learner layer for prediction, obtain the preliminary prediction results of each basic learner, and concatenate the preliminary prediction results to construct the meta-features; S4. Input the meta-features into the meta-learner layer, and use the nonlinear meta-learner and the linear meta-learner to predict the meta-features respectively, thereby obtaining nonlinear prediction values ​​and linear prediction values. S5. The nonlinear prediction value and the linear prediction value are integrated using a weighted fusion strategy to obtain the final offshore wind power prediction result.

[0011] Further, in step S1, the outlier processing and feature reconstruction include: The DBSCAN algorithm was used to identify and remove abnormal samples from the historical meteorological data and historical wind power data. The time variables in the historical meteorological data are mapped to sine and cosine periodic phase features; The wind direction and wind speed in the historical meteorological data are orthogonally vectorized to obtain the eastward continuous component and the northward continuous component, respectively. Identify the zero-power interval in the historical wind power data and force the meteorological characteristics corresponding to the zero-power interval to zero.

[0012] Furthermore, in step S2, the multiple heterogeneous base learners include the CNN-LSTM-AM model, the CatBoost model, the ExtraTrees model, and the SVR model.

[0013] Furthermore, the CNN-LSTM-AM model is composed of a convolutional neural network layer, a long short-term memory network layer, and an attention mechanism layer cascaded together. After the training set is input into the CNN-LSTM-AM model, local feature patterns are extracted through the convolutional neural network layer, dynamic dependencies of the power sequence are extracted through the long short-term memory network layer, and preliminary prediction results are output after the time steps are weighted by the attention mechanism layer.

[0014] Furthermore, in step S2, the nonlinear meta-learner is an XGBoost model, and the linear meta-learner is a RidgeCV model.

[0015] Furthermore, in step S5, the nonlinear predicted value and the linear predicted value are weighted and fused using the following formula: In the formula, For the final offshore wind power forecast results, The nonlinear predicted value output by the XGBoost model is... The linear prediction value output by the RidgeCV model. For weight fusion.

[0016] Another aspect of the present invention provides a multi-agent stacked integrated offshore wind power prediction system based on a binary learner, comprising: The data acquisition and preprocessing module is used to acquire historical meteorological data and historical wind power data of offshore wind farms, perform outlier processing and feature reconstruction, and build a training set. A dual-layer stacked model building module is used to build a dual-layer stacked ensemble learning model, which includes a first-layer base learner layer and a second-layer meta-learner layer; wherein, the base learner layer includes multiple heterogeneous base learners, and the meta-learner layer is a dual-parallel network structure, including parallel nonlinear meta-learners and linear meta-learners. The basic learner layer prediction module is used to input the training set into each basic learner in the basic learner layer for prediction, obtain the preliminary prediction results of each basic learner, and concatenate the preliminary prediction results to construct the meta-features. The bi-element meta-learner layer prediction module is used to receive the meta-features and predict the meta-features by the nonlinear meta-learner and the linear meta-learner in the meta-learner layer, respectively, to obtain nonlinear prediction values ​​and linear prediction values. The weighted fusion output module is used to integrate the nonlinear prediction value and the linear prediction value using a weighted fusion strategy, and output the final offshore wind power prediction result.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention effectively reduces data noise introduced by equipment failures and meteorological changes through specific data preprocessing mechanisms such as outlier filtering based on local density, time phase mapping, and zero-power data correction, thereby improving the stability of the model when learning the mapping relationship between meteorological conditions and wind power output.

[0018] 2. In the first layer, this invention employs multiple heterogeneous basic learners with different mechanisms to process the complex and multidimensional features of wind power scenarios. By integrating different basic learners, the spatiotemporal coupling characteristics and nonlinear mapping relationships are fully explored, overcoming the overfitting problem and insufficient local prediction accuracy that are prone to occur in conventional single models.

[0019] 3. This invention introduces a dual-architecture system in the second layer, employing both a nonlinear meta-learner and a linear meta-learner in parallel. The nonlinear meta-learner further explores the nonlinear residual relationship between the prediction results of the base learner and the actual power, while the linear meta-learner uses regularization to mitigate the impact of feature correlation. This dual-architecture balances nonlinear fitting capability with model robustness, improving the generalization ability of the prediction model under complex sea conditions and effectively reducing power prediction errors. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the vectorized decomposition of wind speed and direction in an embodiment of the present invention.

[0021] Figure 2This is an overall structural diagram of the two-layer stacked ensemble learning prediction model in an embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram of the structure of the pre-stage CNN of the basic learner in an embodiment of the present invention.

[0023] Figure 4 This is a schematic diagram of the intermediate LSTM structure in the basic learner in an embodiment of the present invention.

[0024] Figure 5 This is a schematic diagram of the structure of the ExtraTrees model, the basic learner, in an embodiment of the present invention.

[0025] Figure 6 This is a schematic diagram of the structure of the basic learner SVR model in an embodiment of the present invention.

[0026] Figure 7 This is a schematic diagram of the weighted fusion structure of the meta-learning model in an embodiment of the present invention.

[0027] Figure 8 This is a first-stage comparison diagram of the predicted and actual values ​​of different meta-learners and the final weighted ensemble model on the test set in an embodiment of the present invention.

[0028] Figure 9 This is a second-stage comparison diagram of the predicted and actual values ​​of different meta-learners and the final weighted ensemble model on the test set in an embodiment of the present invention.

[0029] Figure 10 This is a third-stage comparison chart of the predicted and actual values ​​of different meta-learners and the final weighted ensemble model on the test set in this embodiment of the invention.

[0030] Figure 11 This is a fourth-stage comparison chart of the predicted and actual values ​​of different meta-learners and the final weighted ensemble model on the test set in this embodiment of the invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0032] This embodiment provides a multi-agent stacked integrated method for predicting offshore wind power based on a dual learner, including the following steps: Step S1: Obtain historical meteorological data and historical wind power data of offshore wind farms, perform outlier processing and feature reconstruction, and construct a training set.

[0033] To ensure the reliability and effectiveness of offshore wind power prediction models, data preprocessing techniques are needed to improve the quality of wind power data, standardize feature scales, and transform raw data into a format conducive to model learning. Data preprocessing includes the following four parts: (1) Extraction of time periodic features: Wind power typically exhibits certain intraday and annual periodic characteristics over time. Directly inputting time variables such as "hour / day sequence" as linear values ​​can easily lead to distance distortion at the periodic boundaries. For example, 11 PM and midnight are adjacent in real time, but numerically they differ significantly. The model will treat them as vastly different states, thus affecting the learning effect. To characterize the periodic changes in wind power and avoid discontinuities in time features at periodic boundaries, the sampling time is mapped onto a unit circle, constructing periodic time features in sine and cosine form. The constructed sine and cosine functions are as follows: (1) (2) In equations (1)-(2): D sin D cos Representing the daily cycle phase characteristics, Y sin Y cos Indicates the intra-annual periodic phase characteristic; t n T is the timestamp corresponding to the nth sample (time resolution on the order of seconds); day With T year These represent the lengths of the daily and annual cycles, respectively. The values ​​of these features range from [-1, 1], which can characterize time information in a phase manner, making the samples at the cycle boundaries close to each other in the feature space, thus making it easier for the model to learn intraday and seasonal variation patterns.

[0034] (2) Vectorization decomposition of wind speed and direction: Wind direction is an angle variable, 0 ° With 360 ° This indicates the same direction. If the wind direction angle is directly input as a regular numerical value, the model is prone to discontinuities in the vicinity. To avoid this problem, and to represent wind direction and wind speed as continuous features, wind speed and direction are decomposed into two orthogonal components: east (x-axis) and north (y-axis), as illustrated in the diagram below. Figure 1 As shown.

[0035] according to Figure 1 In the coordinate system shown, let the wind vector magnitude be V and the wind direction angle be θ (with north as 0° and clockwise as positive). From the projection relationship of the vector onto the orthogonal coordinate axes, the components of the wind vector on the x-axis and y-axis are as follows: (3) (4) Compared to directly inputting the wind direction angle, (v x v y At the same time, it retains information on wind speed magnitude and direction, which is more conducive to the model's stable learning of the wind speed-power mapping relationship.

[0036] (3) Correction of invalid data for zero-power samples: In historical wind power output data, periods with zero power usually correspond to units not generating electricity. If the original meteorological observations are retained as input in these samples, the training data may show situations where meteorological conditions fluctuate but the output is zero, which can weaken the model's learning of the wind speed-power relationship under normal power generation conditions. To unify the representation of zero-power samples and reduce the impact of noise, all meteorological features corresponding to zero-power records are set to zero, forming a consistent "non-power generation state" encoding. After this processing, meteorological fluctuations in the zero-power interval no longer participate in the fitting process, which helps to reduce the interference of factors such as abnormal shutdowns, power rationing, or acquisition errors on the training distribution, enabling the model to learn the effective mapping relationship between meteorological conditions and wind power output more stably, thereby improving the model training stability.

[0037] (4) Z-score normalization of input features: Input features used for wind power prediction usually have different dimensions and orders of magnitude (such as wind speed m / s, air pressure hPa, temperature ℃, wind power kW, etc.). If directly input into the model, features with larger values ​​tend to have a larger weight in training, affecting the stability of the optimization process. In order to unify the feature scale, the Z-score normalization method is used to normalize all input features.

[0038] For any feature x, the standardization is as follows: (5) In equation (5), μ is the mean of the feature in the dataset, and σ is the standard deviation. After standardization, the mean of each feature is close to 0, and the variance is close to 1, which effectively speeds up the model convergence and improves the prediction performance.

[0039] In addition to the four preprocessing steps mentioned above, special handling is required for outliers in the data to enhance the stability of model training and improve prediction accuracy.

[0040] Outlier Handling: During the operation of offshore wind farms, factors such as equipment failures, turbine shutdowns, communication anomalies, and sudden changes in weather conditions often result in outlier samples deviating from the normal distribution in the raw operational data. Directly using such data for wind power prediction model training can weaken the true mapping relationship between input features and output power, thus affecting the model's prediction accuracy and stability. Therefore, outlier identification and removal are performed on the raw samples before model construction to improve the robustness and generalization ability of the offshore wind power prediction model. Considering that outlier samples usually exhibit significant deviations in local density, this embodiment uses the DBSCAN method to filter outlier data. Compared to methods such as K-means, DBSCAN does not require a pre-defined number of clusters and can identify outlier samples based on local density, making it more suitable for anomaly screening in wind power operation data. The general steps are as follows: ① Construct a sample set based on environmental factors and power data, and set the neighborhood radius ε and the minimum number of samples MinPts; ② Identify clustering regions based on sample density distribution, and remove samples from low-density areas or those that cannot be classified into any cluster as outliers. The specific processing flow is shown in Table 1.

[0041] Table 1. Abnormal Sample Removal Process for Offshore Wind Power Prediction Based on DBSCAN Step S2: Construct a two-layer stacked ensemble learning model, which includes a first-layer base learner layer and a second-layer meta-learner layer; wherein, the base learner layer includes multiple heterogeneous base learners, and the meta-learner layer is a dual-parallel network structure, including parallel nonlinear meta-learners and linear meta-learners.

[0042] To improve the accuracy and stability of offshore wind power forecasting, it is necessary to first analyze the structure of the forecasting model. Offshore wind power data is characterized by strong fluctuations and significant nonlinearity. Different learners have their own advantages in feature extraction and pattern fitting, but a single learner can usually only characterize the data from one aspect and cannot fully reflect its complex changes. Therefore, a two-layer learning structure consisting of a base learner and a meta-learner can be adopted. The base learner first obtains diverse forecast information, and then the meta-learner integrates and optimizes it, thereby achieving complementary advantages between different models.

[0043] In contrast, while a single-layer structure or a single learner per layer results in a simpler model, it limits information utilization and lacks further fusion and correction of prediction results. Model performance is easily constrained by the modeling capabilities and applicability of a single learner. In the complex dynamics of offshore wind power, such structures often struggle to simultaneously account for local fluctuations and overall trends. Therefore, a two-layer stacked ensemble learning model is selected, with ensemble strategies introduced in each layer to enhance the model's expressive power, stability, and generalization performance. The overall model structure is as follows: Figure 2 As shown.

[0044] The first layer consists of four base learners to capture different features of the data: CNN-LSTM-AM focuses on capturing spatiotemporal features, the CatBoost model reduces prediction bias by optimizing the mapping relationship between meteorological factors and electricity data, SVR enhances the model's generalization ability in scenarios with limited data samples, and ExtraTrees effectively reduce the variance of the prediction results. The integration of different base learners effectively alleviates the model overfitting problem by fully exploring and utilizing their complementary advantages, thereby improving the overall prediction accuracy.

[0045] The second layer consists of two meta-learners: XGBoost (a non-linear structure) and RidgeCV (a linear structure). First, each meta-learner is trained based on the output of the first layer. Then, the two trained meta-learners are used to predict the data to be predicted, yielding their respective predictions. Finally, a weighted fusion strategy is used to integrate the two predictions to obtain the final prediction. This approach allows for the complementarity of different model characteristics, effectively balancing biases in the prediction process, enhancing adaptability to diverse data distributions, and thus obtaining more robust prediction results.

[0046] The learner model in the stacked ensemble learning framework proposed in this embodiment includes four base learners and two meta learners.

[0047] (1) Base Learner Model – CNN-LTSM-AM: CNN-LSTM-AM serves as the first base learner, used to characterize the spatiotemporal coupling features driven by multidimensional meteorological variables in the power sequence of offshore wind farms. Offshore wind power is affected by factors such as wind speed, wind direction, temperature, air pressure, and humidity. There are local correlations between different meteorological variables, and the power evolution process has obvious time dependence and stage fluctuation characteristics. Among them, the CNN stage is used to extract local feature patterns between multidimensional meteorological variables, the LSTM stage is used to capture the dynamic dependence of the power sequence over time, and the AM stage further strengthens the ability to represent information in key periods, thereby improving the model's ability to characterize the fluctuation process of offshore wind power. CNN-LSTM-AM is composed of CNN, LSTM, and AM cascaded in sequence. The input data first passes through the CNN module to extract features, then the obtained feature sequence is input into the LSTM module for time series modeling, and then the AM layer performs weighted fusion of the LSTM output. Finally, the wind power prediction result is obtained through a fully connected layer.

[0048] The specific structure and implementation principles of each part of CNN-LSTM-AM are described below: ① Pre-stage CNN stage: In offshore wind farm prediction scenarios, different meteorological variables do not act independently on the output power, but often jointly affect the unit's power output in a coupled manner. The CNN module extracts local pattern features from the input meteorological sequence through convolutional operations, thereby mining the combined information between variables such as wind speed, wind direction, and temperature that is indicative of changes in wind power. Its structure diagram is shown below. Figure 3 As shown.

[0049] The specific feature extraction process is as follows: Raw meteorological data generally consists of the following sequences: (6) In equation (6): T s Indicates the size of the time step in the original meteorological sequence; f met The dimension of meteorological features contained in each time step; Xmet t represents the meteorological feature vector of the t-th time step.

[0050] Taking one-dimensional convolution as an example, we first define a one-dimensional convolution kernel of size k, denoted as... Use this convolution kernel to perform a convolution operation on sequence X: (7) In equation (7): hCNN t represents the output of the convolutional layer at time step t; ReLU is the activation function; b CNN It is the offset; xmet t+i represents the offset at time step x. t+i Meteorological feature vector at the location.

[0051] The new sequence obtained after a series of convolution operations is represented as follows: (8) ② Intermediate LSTM stage: The LSTM module is used for temporal modeling of the feature sequences extracted by convolution. A gating mechanism is used to retain historical information that contributes to the current prediction time and suppress redundant or interfering information. Its structure diagram is shown below. Figure 4 As shown in the diagram. The forget gate determines which irrelevant information to discard, the input gate controls which new information needs to be stored and used to update the cell state, and the output gate extracts relevant information from the updated cell state as output.

[0052] LSTM on convolutional feature sequences H CNN The processing is carried out as follows: (9) (10) (11) (12) (13) (14) In equations (9)-(14): hCNN t represents the feature vector output by the CNN at time t; hLSTM t is the output; WLSTM(×) and bLSTM(×) are the weights and biases of the corresponding neurons, respectively; σ is the sigmoid activation function; tanh is the tanh activation function; fLSTM t is the forgetting gate; iLSTM t is the input gate; CLSTM t is the candidate vector value; and OLSTM t is the output gate.

[0053] ③ Subsequent AM stage: To further enhance the model's ability to perceive key time periods, this embodiment introduces an attention mechanism (AM) after the LSTM layer. AM assigns different weights to the output states at different time steps, thereby highlighting the time step information that is more critical to the current power prediction and enhancing the model's ability to identify key dynamic features under complex sea wind conditions. First, for the hidden state output by the LSTM layer at time step t, its attention score is calculated: (15) In equation (15): W AM It is a trainable weight matrix that maps the hidden states to the attention space; b AM It is a paranoid vector; v TIt is the transpose of the attention score vector. v itself is not directly obtained from the input data. Instead, it is used as a trainable attention parameter and gradually obtained through model training.

[0054] Then, for all time steps {eAM 1, eAM 2, ... … The scoring results of ,eAM T} are normalized to obtain the attention weight αAM t, which is expressed as: (16) Finally, the context vector c is obtained by weighted summation. AM : (17) AM can dynamically focus on the most important time period in the current time period, thereby improving the model's ability to predict wind fluctuation trends. It mainly allocates weights to the LSTM output at different time steps, rather than the contribution of a single meteorological feature at that time step. Therefore, it is difficult to accurately describe the degree of influence of a single meteorological factor (such as wind speed) on the model output.

[0055] (2) Base Learner Model – CatBoost Learner Model: CatBoost, as the second base learner, is mainly used to supplement the advantages of deep learning models in modeling nonlinear relationships. For offshore wind farm power prediction, there is a clear nonlinear mapping relationship between meteorological variables and wind power. CatBoost can effectively learn the complex relationship between meteorological characteristics and target power by iteratively constructing regression trees, thus characterizing the variation law of offshore wind power output from another modeling perspective and providing complementary prediction information for subsequent stacking and fusion. Its prediction process is achieved by stacking multiple regression trees: (18) In equation (18): h m (x i ) represents the output of the m-th tree; γ m ŷCatBoost i represents the weight of the corresponding m trees; ŷCatBoost i is the predicted value.

[0056] (3) Base Learner Model – Extratrees Learner Model: ExtraTrees, as the third base learner, is a machine learning algorithm that rapidly constructs tree models through high randomness, making it suitable for processing large amounts of wind power generation data with strong fluctuations. Its prediction result is the average of the outputs of multiple tree models. Its structural model diagram is shown below. Figure 5 As shown, the prediction function can be expressed as: (19) In equation (19): T represents the total number of trees; ht (x i ) represents the t-th random tree for input sample x. i The predicted output.

[0057] (4) Base Learner Model – SVR Learner Model: As the fourth base learner, SVR is mainly used to enhance the stable prediction capability of the ensemble framework under small sample areas and complex operating conditions. In offshore wind power prediction, the sample distribution under different wind conditions is usually uneven, and there are few samples under some extreme or transitional operating conditions, which can easily lead to insufficient generalization ability of complex models in local areas. Based on the idea of ​​kernel function mapping and structural risk minimization, SVR can establish a relatively robust nonlinear regression relationship under limited sample conditions. Therefore, it can be used as an effective supplement to other base learners to further improve the prediction stability of the ensemble model under different operating conditions. This model can robustly model wind power output under small sample size or strong nonlinear conditions and is applicable to boundary wind condition prediction. Its basic modeling process is as follows: Figure 6 As shown, the prediction function can be expressed as: (20) Its optimization objective is: (twenty one) In equations (20) to (21): Φ(x) i ) is the feature vector after kernel function mapping; C represents the penalty coefficient, used to balance model complexity and empirical risk; δ i and δ i * represents slack variables; ω and b SVR These are model parameters.

[0058] Meta-learner model: In ensemble learning, the prediction results of base learners often have certain biases and limitations. Therefore, it is necessary to further introduce a meta-learner to learn and integrate the outputs of multiple base learners. To balance linear and nonlinear modeling capabilities, this embodiment introduces two different types of models, RidgeCV and XGBoost, into the meta-learner part, and combines the prediction results of the two through weighted fusion, thereby improving the overall prediction accuracy and robustness.

[0059] (1) XGBoost Learning Model: In the meta-learning stage, XGBoost is mainly used to further learn the nonlinear relationship between the prediction results of each base learner and the actual power. Since different base learners have different focuses on modeling the complex wind conditions of offshore wind farms, their output results usually contain certain complementary information. By improving the modeling of these meta-features, XGBoost can further explore the information that has not been fully expressed in the combined results of the base learners, thereby improving the overall prediction accuracy of the stacked ensemble model. In the meta-learner, the input is the prediction result generated by the base learner during the training stage. There are 4 base learners in total. For sample i, the prediction given by the j-th base learner is denoted as ŷj i. Then the meta-feature of the sample can be expressed as: (twenty two) In equation (22): X i,meta Let represent the meta-feature vector corresponding to the i-th sample.

[0060] Therefore, the meta-feature matrix composed of all samples is X. meta =[x 1,meta ,x 2,meta ,…,x N,meta ] T Where N is the number of samples; meanwhile, the corresponding true label vector can be represented as: (twenty three) In equation (23): y meta The vector represents the true label vector corresponding to the training phase of meta-learning; N represents the total number of samples; y i This represents the actual wind power value of the i-th sample.

[0061] XGBoost iteratively constructs a regression tree, fitting the residuals between the base learner's predictions and the true values. Its training process can be described as optimizing the following regularized objective function: (twenty four) In equation (24): For the squared error loss, For the first The predicted value for each sample, Let k be the k-th regression tree, where k is the number of regression trees. The regularization term is used to limit the complexity of the tree; the final prediction result is a weighted sum of multiple trees. (25) (2) RidgeCV Learning Model: RidgeCV, as another meta-learner, is mainly used to robustly combine the outputs of each base learner from a linear fusion perspective. Compared to XGBoost's further modeling of nonlinear relationships, RidgeCV focuses more on reducing the impact of feature correlation through regularization, thereby improving the stability of the fusion result and enhancing the generalization ability of the ensemble model. RidgeCV directly combines the prediction results of the base learners using a linear weighting method, and its objective function is: (26) In equation (26): For the regression coefficient vector, is the regularization parameter automatically selected through cross-validation. Its prediction result is: (27) To fully leverage the complementary advantages of the two models, the outputs of XGBoost and RidgeCV are weighted and fused, such as... Figure 7 As shown, the final prediction result is obtained: (28) In equation (28): The fusion weights are calculated through the validation set tuning process.

[0062] Experimental Results Verification: This embodiment uses all numerical weather forecast data for a specific offshore wind farm in a given year, including meteorological conditions such as temperature, humidity, and wind speed at different times, as well as actual power output. The ratio of the training set, validation set, and test set is set to 7:1.5:1.5. To comprehensively evaluate the model's performance on the test set, root mean square error (RMSE) and coefficient of determination (R²) are used. 2 The mean absolute error (MAE) and relative root mean square error (RRMSE) are used as performance indicators.

[0063] Table 2 shows the wind power prediction evaluation metrics under different models. The prediction results indicate that the final model significantly outperforms any single base learner. Compared to the best single model, the final model's prediction accuracy (R²) is significantly higher. 2 The prediction accuracy improved by approximately 6%, while the error metrics (MAE and RMSE) decreased by approximately 15% and 18%, respectively, demonstrating stronger generalization ability and stability. In the comparison of meta-learners, the nonlinear meta-learner (XGBoost) outperformed the linear meta-learner (RidgeCV) overall, further improving prediction accuracy by approximately 2% and reducing error by approximately 5%. Furthermore, through weighted fusion of the two meta-learners, the final model comprehensively leveraged the advantages of both linear and nonlinear methods, achieving optimal prediction performance and validating the effectiveness of the multi-level ensemble strategy in wind power prediction.

[0064] Table 2 Wind power prediction and evaluation indicators under different models Figures 8-11 The results demonstrate the comparison between predicted and actual values ​​on the test set using different meta-learners (XGBoost, RidgeCV) and the final weighted ensemble model, with the results presented across four time periods. Overall, the predicted values ​​of each model are highly consistent with the actual power trends in most time periods, especially in intervals with periodic fluctuations, demonstrating a good ability to learn and reproduce the main characteristics of wind power time series.

[0065] Specifically, the XGBoost meta-learner fits the periodic fluctuations well, but lags behind during sudden power changes (such as rapid increases or decreases), underestimating peak values ​​or overestimating trough values. The RidgeCV meta-learner has a smoother overall fit and captures periodic changes well, but its bias is more pronounced during abrupt changes, often overestimating peak values. In contrast, the final weighted ensemble model combines the advantages of different meta-learners, predicting curves that are closer to the true curves for most periods. It maintains good learning ability for periodic fluctuations and alleviates the bias problem of single models during peak-trough abrupt changes to some extent.

[0066] However, in a few intervals with dramatic power shifts, the ensemble model still exhibits some lag and local bias. This indicates that the model has limitations in capturing sudden changes, possibly due to insufficient extreme sample sizes or weak sensitivity of features to abrupt changes. Future research could improve its predictive ability for extreme fluctuations by incorporating more comprehensive meteorological data or further optimizing the model structure.

[0067] In summary, the method of this embodiment has the following advantages: 1) This embodiment employs a two-layer stacked integrated structure. The first layer consists of four base learners: CNN-LSTM-AM, CatBoost, ExtraTrees, and SVR. These learners predict offshore wind farm power from different perspectives and output their respective prediction results. These results further constitute meta-features, serving as input to the second-layer meta-learner. In the second layer, XGBoost and RidgeCV relearn the relationship between the base learner outputs and the actual power, obtaining a higher-level fusion result. In other words, the base learners are responsible for providing differentiated initial predictions, while the meta-learners are responsible for further combining and refining these prediction results, thereby improving the overall prediction accuracy and stability.

[0068] 2) In the meta-learning stage, this embodiment selects XGBoost and RidgeCV primarily to balance nonlinear modeling capabilities and fusion stability. XGBoost can perform nonlinear enhancement modeling on the meta-features formed by the outputs of each base learner, further uncovering the complex relationships between the predicted results and the actual power that are not fully expressed. RidgeCV, on the other hand, employs a regularized linear fusion method, which can alleviate the correlation problem between meta-features, making the fusion result more stable and reliable. For offshore wind farm power prediction, the outputs of different base learners often contain complementary information as well as some correlation. Therefore, combining XGBoost and RidgeCV helps to leverage the advantages of both nonlinear fitting and robust fusion.

[0069] 3) This embodiment selects these four base learners primarily because offshore wind farm power prediction is characterized by multi-dimensional meteorological factor coupling, strong temporal correlation, significant nonlinearity, and drastic fluctuations, making it difficult for a single model to fully characterize its changing patterns. Among them, CNN-LSTM-AM is suitable for extracting local features, time dependence, and key period information from meteorological inputs; CatBoost excels at establishing complex nonlinear mapping relationships between meteorological variables and power; ExtraTrees improves the model's adaptability to fluctuating samples and abnormal wind conditions through highly randomized tree ensembles; and SVR exhibits good stability in small sample areas and complex boundary conditions. Using these four models with different mechanisms as base learners facilitates the formation of more differentiated and complementary prediction results, providing more effective input for subsequent meta-learner fusion.

[0070] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-agent stacked integrated method for offshore wind power prediction based on a binary learner, characterized in that, Includes the following steps: S1. Obtain historical meteorological data and historical wind power data of offshore wind farms, perform outlier processing and feature reconstruction, and construct a training set; S2. Construct a two-layer stacked ensemble learning model, which includes a first-layer base learner layer and a second-layer meta-learner layer; wherein, the base learner layer includes multiple heterogeneous base learners, and the meta-learner layer is a dual-parallel network structure, including parallel nonlinear meta-learners and linear meta-learners. S3. Input the training set into each basic learner in the basic learner layer for prediction, obtain the preliminary prediction results of each basic learner, and concatenate the preliminary prediction results to construct the meta-features; S4. Input the meta-features into the meta-learner layer, and use the nonlinear meta-learner and the linear meta-learner to predict the meta-features respectively, thereby obtaining nonlinear prediction values ​​and linear prediction values. S5. The nonlinear prediction value and the linear prediction value are integrated using a weighted fusion strategy to obtain the final offshore wind power prediction result.

2. The multi-agent stacked integrated offshore wind power prediction method based on a binary learner according to claim 1, characterized in that, In step S1, the outlier processing and feature reconstruction include: The DBSCAN algorithm was used to identify and remove abnormal samples from the historical meteorological data and historical wind power data. The time variables in the historical meteorological data are mapped to sine and cosine periodic phase features; The wind direction and wind speed in the historical meteorological data are orthogonally vectorized to obtain the eastward continuous component and the northward continuous component, respectively. Identify the zero-power interval in the historical wind power data and force the meteorological characteristics corresponding to the zero-power interval to zero.

3. The multi-agent stacked integrated offshore wind power prediction method based on a binary learner according to claim 1, characterized in that, In step S2, the multiple heterogeneous base learners include the CNN-LSTM-AM model, the CatBoost model, the ExtraTrees model, and the SVR model.

4. The multi-agent stacked integrated offshore wind power prediction method based on a binary learner according to claim 3, characterized in that, The CNN-LSTM-AM model is composed of a convolutional neural network layer, a long short-term memory network layer, and an attention mechanism layer in sequence. After the training set is input into the CNN-LSTM-AM model, local feature patterns are extracted through the convolutional neural network layer, dynamic dependencies of the power sequence are extracted through the long short-term memory network layer, and preliminary prediction results are output after the time steps are weighted by the attention mechanism layer.

5. The multi-agent stacked integrated offshore wind power prediction method based on a binary learner according to claim 1, characterized in that, In step S2, the nonlinear meta-learner is an XGBoost model, and the linear meta-learner is a RidgeCV model.

6. The multi-agent stacked integrated offshore wind power prediction method based on a binary learner according to claim 5, characterized in that, In step S5, the nonlinear predicted value and the linear predicted value are weighted and fused using the following formula: In the formula, For the final offshore wind power forecast results, The nonlinear predicted value output by the XGBoost model is... The linear prediction value output by the RidgeCV model. For weight fusion.

7. A multi-agent stacked integrated offshore wind power prediction system based on a binary learner, characterized in that, include: The data acquisition and preprocessing module is used to acquire historical meteorological data and historical wind power data of offshore wind farms, perform outlier processing and feature reconstruction, and build a training set. A dual-layer stacked model building module is used to build a dual-layer stacked ensemble learning model, which includes a first-layer base learner layer and a second-layer meta-learner layer; wherein, the base learner layer includes multiple heterogeneous base learners, and the meta-learner layer is a dual-parallel network structure, including parallel nonlinear meta-learners and linear meta-learners. The basic learner layer prediction module is used to input the training set into each basic learner in the basic learner layer for prediction, obtain the preliminary prediction results of each basic learner, and concatenate the preliminary prediction results to construct the meta-features. The bi-element meta-learner layer prediction module is used to receive the meta-features and predict the meta-features by the nonlinear meta-learner and the linear meta-learner in the meta-learner layer, respectively, to obtain nonlinear prediction values ​​and linear prediction values. The weighted fusion output module is used to integrate the nonlinear prediction value and the linear prediction value using a weighted fusion strategy, and output the final offshore wind power prediction result.

Images